diff --git a/llvm/include/llvm/CodeGen/MachineScheduler.h b/llvm/include/llvm/CodeGen/MachineScheduler.h
index fe4a3d7c551bd..486508e760dc5 100644
--- a/llvm/include/llvm/CodeGen/MachineScheduler.h
+++ b/llvm/include/llvm/CodeGen/MachineScheduler.h
@@ -304,10 +304,6 @@ class LLVM_ABI ScheduleDAGMI : public ScheduleDAGInstrs {
   /// The bottom of the unscheduled zone.
   MachineBasicBlock::iterator CurrentBottom;
 
-  /// Record the next node in a scheduled cluster.
-  const SUnit *NextClusterPred = nullptr;
-  const SUnit *NextClusterSucc = nullptr;
-
 #if LLVM_ENABLE_ABI_BREAKING_CHECKS
   /// The number of instructions scheduled so far. Used to cut off the
   /// scheduler at the point determined by misched-cutoff.
@@ -368,10 +364,6 @@ class LLVM_ABI ScheduleDAGMI : public ScheduleDAGInstrs {
   /// live ranges and region boundary iterators.
   void moveInstruction(MachineInstr *MI, MachineBasicBlock::iterator InsertPos);
 
-  const SUnit *getNextClusterPred() const { return NextClusterPred; }
-
-  const SUnit *getNextClusterSucc() const { return NextClusterSucc; }
-
   void viewGraph(const Twine &Name, const Twine &Title) override;
   void viewGraph() override;
 
@@ -1295,6 +1287,9 @@ class LLVM_ABI GenericScheduler : public GenericSchedulerBase {
   SchedBoundary Top;
   SchedBoundary Bot;
 
+  ClusterInfo *TopCluster;
+  ClusterInfo *BotCluster;
+
   /// Candidate last picked from Top boundary.
   SchedCandidate TopCand;
   /// Candidate last picked from Bot boundary.
@@ -1335,6 +1330,9 @@ class LLVM_ABI PostGenericScheduler : public GenericSchedulerBase {
   /// Candidate last picked from Bot boundary.
   SchedCandidate BotCand;
 
+  ClusterInfo *TopCluster;
+  ClusterInfo *BotCluster;
+
 public:
   PostGenericScheduler(const MachineSchedContext *C)
       : GenericSchedulerBase(C), Top(SchedBoundary::TopQID, "TopQ"),
diff --git a/llvm/include/llvm/CodeGen/ScheduleDAG.h b/llvm/include/llvm/CodeGen/ScheduleDAG.h
index 7a49e9ffedf8a..3a0a31b1930b0 100644
--- a/llvm/include/llvm/CodeGen/ScheduleDAG.h
+++ b/llvm/include/llvm/CodeGen/ScheduleDAG.h
@@ -17,6 +17,7 @@
 
 #include "llvm/ADT/BitVector.h"
 #include "llvm/ADT/PointerIntPair.h"
+#include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/iterator.h"
 #include "llvm/CodeGen/MachineInstr.h"
@@ -235,6 +236,10 @@ class TargetRegisterInfo;
     LLVM_ABI void dump(const TargetRegisterInfo *TRI = nullptr) const;
   };
 
+  /// Keep record of which SUnit are in the same cluster group.
+  typedef SmallSet<SUnit *, 8> ClusterInfo;
+  constexpr unsigned InvalidClusterId = ~0u;
+
   /// Scheduling unit. This is a node in the scheduling DAG.
   class SUnit {
   private:
@@ -275,6 +280,8 @@ class TargetRegisterInfo;
     unsigned TopReadyCycle = 0; ///< Cycle relative to start when node is ready.
     unsigned BotReadyCycle = 0; ///< Cycle relative to end when node is ready.
 
+    unsigned ParentClusterIdx = InvalidClusterId; ///< The parent cluster id.
+
   private:
     unsigned Depth = 0;  ///< Node depth.
     unsigned Height = 0; ///< Node height.
diff --git a/llvm/include/llvm/CodeGen/ScheduleDAGInstrs.h b/llvm/include/llvm/CodeGen/ScheduleDAGInstrs.h
index 7f7b3036af803..ba5b2da64fd80 100644
--- a/llvm/include/llvm/CodeGen/ScheduleDAGInstrs.h
+++ b/llvm/include/llvm/CodeGen/ScheduleDAGInstrs.h
@@ -181,6 +181,8 @@ namespace llvm {
     /// case of a huge region that gets reduced).
     SUnit *BarrierChain = nullptr;
 
+    SmallVector<ClusterInfo> Clusters;
+
   public:
     /// A list of SUnits, used in Value2SUsMap, during DAG construction.
     /// Note: to gain speed it might be worth investigating an optimized
@@ -384,6 +386,14 @@ namespace llvm {
     /// equivalent edge already existed (false indicates failure).
     bool addEdge(SUnit *SuccSU, const SDep &PredDep);
 
+    /// Returns the array of the clusters.
+    SmallVector<ClusterInfo> &getClusters() { return Clusters; }
+
+    /// Get the specific cluster, return nullptr for InvalidClusterId.
+    ClusterInfo *getCluster(unsigned Idx) {
+      return Idx != InvalidClusterId ? &Clusters[Idx] : nullptr;
+    }
+
   protected:
     void initSUnits();
     void addPhysRegDataDeps(SUnit *SU, unsigned OperIdx);
diff --git a/llvm/lib/CodeGen/MachineScheduler.cpp b/llvm/lib/CodeGen/MachineScheduler.cpp
index 2cc7f549a14d1..7b4f29e82fe6f 100644
--- a/llvm/lib/CodeGen/MachineScheduler.cpp
+++ b/llvm/lib/CodeGen/MachineScheduler.cpp
@@ -15,6 +15,7 @@
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/BitVector.h"
 #include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/EquivalenceClasses.h"
 #include "llvm/ADT/PriorityQueue.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallVector.h"
@@ -943,8 +944,6 @@ void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
 
   if (SuccEdge->isWeak()) {
     --SuccSU->WeakPredsLeft;
-    if (SuccEdge->isCluster())
-      NextClusterSucc = SuccSU;
     return;
   }
 #ifndef NDEBUG
@@ -967,12 +966,6 @@ void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
 
 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
-  // Reset the next successor, For example, we want to cluster A B C.
-  // After A is picked, we will set B as next cluster succ, but if we pick
-  // D instead of B after A, then we need to reset the next cluster succ because
-  // we have decided to not pick the cluster candidate B during pickNode().
-  // Leaving B as the NextClusterSucc just make things messy.
-  NextClusterSucc = nullptr;
   for (SDep &Succ : SU->Succs)
     releaseSucc(SU, &Succ);
 }
@@ -986,8 +979,6 @@ void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
 
   if (PredEdge->isWeak()) {
     --PredSU->WeakSuccsLeft;
-    if (PredEdge->isCluster())
-      NextClusterPred = PredSU;
     return;
   }
 #ifndef NDEBUG
@@ -1010,7 +1001,6 @@ void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
 
 /// releasePredecessors - Call releasePred on each of SU's predecessors.
 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
-  NextClusterPred = nullptr;
   for (SDep &Pred : SU->Preds)
     releasePred(SU, &Pred);
 }
@@ -1184,11 +1174,8 @@ findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
 }
 
 /// Identify DAG roots and setup scheduler queues.
-void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
-                               ArrayRef<SUnit*> BotRoots) {
-  NextClusterSucc = nullptr;
-  NextClusterPred = nullptr;
-
+void ScheduleDAGMI::initQueues(ArrayRef<SUnit *> TopRoots,
+                               ArrayRef<SUnit *> BotRoots) {
   // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
   //
   // Nodes with unreleased weak edges can still be roots.
@@ -2116,6 +2103,7 @@ void BaseMemOpClusterMutation::clusterNeighboringMemOps(
     ScheduleDAGInstrs *DAG) {
   // Keep track of the current cluster length and bytes for each SUnit.
   DenseMap<unsigned, std::pair<unsigned, unsigned>> SUnit2ClusterInfo;
+  EquivalenceClasses<SUnit *> Clusters;
 
   // At this point, `MemOpRecords` array must hold atleast two mem ops. Try to
   // cluster mem ops collected within `MemOpRecords` array.
@@ -2155,6 +2143,7 @@ void BaseMemOpClusterMutation::clusterNeighboringMemOps(
 
     SUnit *SUa = MemOpa.SU;
     SUnit *SUb = MemOpb.SU;
+
     if (!ReorderWhileClustering && SUa->NodeNum > SUb->NodeNum)
       std::swap(SUa, SUb);
 
@@ -2162,6 +2151,7 @@ void BaseMemOpClusterMutation::clusterNeighboringMemOps(
     if (!DAG->addEdge(SUb, SDep(SUa, SDep::Cluster)))
       continue;
 
+    Clusters.unionSets(SUa, SUb);
     LLVM_DEBUG(dbgs() << "Cluster ld/st SU(" << SUa->NodeNum << ") - SU("
                       << SUb->NodeNum << ")\n");
     ++NumClustered;
@@ -2201,6 +2191,21 @@ void BaseMemOpClusterMutation::clusterNeighboringMemOps(
                       << ", Curr cluster bytes: " << CurrentClusterBytes
                       << "\n");
   }
+
+  // Add cluster group information.
+  // Iterate over all of the equivalence sets.
+  auto &AllClusters = DAG->getClusters();
+  for (const EquivalenceClasses<SUnit *>::ECValue *I : Clusters) {
+    if (!I->isLeader())
+      continue;
+    ClusterInfo Group;
+    unsigned ClusterIdx = AllClusters.size();
+    for (SUnit *MemberI : Clusters.members(*I)) {
+      MemberI->ParentClusterIdx = ClusterIdx;
+      Group.insert(MemberI);
+    }
+    AllClusters.push_back(Group);
+  }
 }
 
 void BaseMemOpClusterMutation::collectMemOpRecords(
@@ -3688,6 +3693,9 @@ void GenericScheduler::initialize(ScheduleDAGMI *dag) {
   }
   TopCand.SU = nullptr;
   BotCand.SU = nullptr;
+
+  TopCluster = nullptr;
+  BotCluster = nullptr;
 }
 
 /// Initialize the per-region scheduling policy.
@@ -3997,13 +4005,11 @@ bool GenericScheduler::tryCandidate(SchedCandidate &Cand,
   // This is a best effort to set things up for a post-RA pass. Optimizations
   // like generating loads of multiple registers should ideally be done within
   // the scheduler pass by combining the loads during DAG postprocessing.
-  const SUnit *CandNextClusterSU =
-    Cand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  const SUnit *TryCandNextClusterSU =
-    TryCand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  if (tryGreater(TryCand.SU == TryCandNextClusterSU,
-                 Cand.SU == CandNextClusterSU,
-                 TryCand, Cand, Cluster))
+  const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster;
+  const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster;
+  if (tryGreater(TryCandCluster && TryCandCluster->contains(TryCand.SU),
+                 CandCluster && CandCluster->contains(Cand.SU), TryCand, Cand,
+                 Cluster))
     return TryCand.Reason != NoCand;
 
   if (SameBoundary) {
@@ -4262,11 +4268,25 @@ void GenericScheduler::reschedulePhysReg(SUnit *SU, bool isTop) {
 void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
   if (IsTopNode) {
     SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
+    TopCluster = DAG->getCluster(SU->ParentClusterIdx);
+    LLVM_DEBUG(if (TopCluster) {
+      dbgs() << "  Top Cluster: ";
+      for (auto *N : *TopCluster)
+        dbgs() << N->NodeNum << '\t';
+      dbgs() << '\n';
+    });
     Top.bumpNode(SU);
     if (SU->hasPhysRegUses)
       reschedulePhysReg(SU, true);
   } else {
     SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.getCurrCycle());
+    BotCluster = DAG->getCluster(SU->ParentClusterIdx);
+    LLVM_DEBUG(if (BotCluster) {
+      dbgs() << "  Bot Cluster: ";
+      for (auto *N : *BotCluster)
+        dbgs() << N->NodeNum << '\t';
+      dbgs() << '\n';
+    });
     Bot.bumpNode(SU);
     if (SU->hasPhysRegDefs)
       reschedulePhysReg(SU, false);
@@ -4303,6 +4323,8 @@ void PostGenericScheduler::initialize(ScheduleDAGMI *Dag) {
   if (!Bot.HazardRec) {
     Bot.HazardRec = DAG->TII->CreateTargetMIHazardRecognizer(Itin, DAG);
   }
+  TopCluster = nullptr;
+  BotCluster = nullptr;
 }
 
 void PostGenericScheduler::initPolicy(MachineBasicBlock::iterator Begin,
@@ -4367,14 +4389,12 @@ bool PostGenericScheduler::tryCandidate(SchedCandidate &Cand,
     return TryCand.Reason != NoCand;
 
   // Keep clustered nodes together.
-  const SUnit *CandNextClusterSU =
-      Cand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  const SUnit *TryCandNextClusterSU =
-      TryCand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  if (tryGreater(TryCand.SU == TryCandNextClusterSU,
-                 Cand.SU == CandNextClusterSU, TryCand, Cand, Cluster))
+  const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster;
+  const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster;
+  if (tryGreater(TryCandCluster && TryCandCluster->contains(TryCand.SU),
+                 CandCluster && CandCluster->contains(Cand.SU), TryCand, Cand,
+                 Cluster))
     return TryCand.Reason != NoCand;
-
   // Avoid critical resource consumption and balance the schedule.
   if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
               TryCand, Cand, ResourceReduce))
@@ -4571,9 +4591,11 @@ SUnit *PostGenericScheduler::pickNode(bool &IsTopNode) {
 void PostGenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
   if (IsTopNode) {
     SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
+    TopCluster = DAG->getCluster(SU->ParentClusterIdx);
     Top.bumpNode(SU);
   } else {
     SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.getCurrCycle());
+    BotCluster = DAG->getCluster(SU->ParentClusterIdx);
     Bot.bumpNode(SU);
   }
 }
diff --git a/llvm/lib/CodeGen/MacroFusion.cpp b/llvm/lib/CodeGen/MacroFusion.cpp
index 5bd6ca0978a4b..71053342f468e 100644
--- a/llvm/lib/CodeGen/MacroFusion.cpp
+++ b/llvm/lib/CodeGen/MacroFusion.cpp
@@ -61,6 +61,11 @@ bool llvm::fuseInstructionPair(ScheduleDAGInstrs &DAG, SUnit &FirstSU,
   for (SDep &SI : SecondSU.Preds)
     if (SI.isCluster())
       return false;
+
+  unsigned FirstCluster = FirstSU.ParentClusterIdx;
+  unsigned SecondCluster = SecondSU.ParentClusterIdx;
+  assert(FirstCluster == InvalidClusterId && SecondCluster == InvalidClusterId);
+
   // Though the reachability checks above could be made more generic,
   // perhaps as part of ScheduleDAGInstrs::addEdge(), since such edges are valid,
   // the extra computation cost makes it less interesting in general cases.
@@ -70,6 +75,14 @@ bool llvm::fuseInstructionPair(ScheduleDAGInstrs &DAG, SUnit &FirstSU,
   if (!DAG.addEdge(&SecondSU, SDep(&FirstSU, SDep::Cluster)))
     return false;
 
+  auto &Clusters = DAG.getClusters();
+
+  FirstSU.ParentClusterIdx = Clusters.size();
+  SecondSU.ParentClusterIdx = Clusters.size();
+
+  SmallSet<SUnit *, 8> Cluster{{&FirstSU, &SecondSU}};
+  Clusters.push_back(Cluster);
+
   // TODO - If we want to chain more than two instructions, we need to create
   // artifical edges to make dependencies from the FirstSU also dependent
   // on other chained instructions, and other chained instructions also
diff --git a/llvm/lib/CodeGen/ScheduleDAG.cpp b/llvm/lib/CodeGen/ScheduleDAG.cpp
index 26857edd871e2..e630b80e33ab4 100644
--- a/llvm/lib/CodeGen/ScheduleDAG.cpp
+++ b/llvm/lib/CodeGen/ScheduleDAG.cpp
@@ -365,6 +365,9 @@ LLVM_DUMP_METHOD void ScheduleDAG::dumpNodeName(const SUnit &SU) const {
 LLVM_DUMP_METHOD void ScheduleDAG::dumpNodeAll(const SUnit &SU) const {
   dumpNode(SU);
   SU.dumpAttributes();
+  if (SU.ParentClusterIdx != InvalidClusterId)
+    dbgs() << "  Parent Cluster Index: " << SU.ParentClusterIdx << '\n';
+
   if (SU.Preds.size() > 0) {
     dbgs() << "  Predecessors:\n";
     for (const SDep &Dep : SU.Preds) {
diff --git a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
index 706ae92c9e47c..0f80462050cda 100644
--- a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
@@ -589,12 +589,11 @@ bool GCNMaxILPSchedStrategy::tryCandidate(SchedCandidate &Cand,
   // This is a best effort to set things up for a post-RA pass. Optimizations
   // like generating loads of multiple registers should ideally be done within
   // the scheduler pass by combining the loads during DAG postprocessing.
-  const SUnit *CandNextClusterSU =
-      Cand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  const SUnit *TryCandNextClusterSU =
-      TryCand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  if (tryGreater(TryCand.SU == TryCandNextClusterSU,
-                 Cand.SU == CandNextClusterSU, TryCand, Cand, Cluster))
+  const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster;
+  const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster;
+  if (tryGreater(TryCandCluster && TryCandCluster->contains(TryCand.SU),
+                 CandCluster && CandCluster->contains(Cand.SU), TryCand, Cand,
+                 Cluster))
     return TryCand.Reason != NoCand;
 
   // Avoid increasing the max critical pressure in the scheduled region.
@@ -664,12 +663,11 @@ bool GCNMaxMemoryClauseSchedStrategy::tryCandidate(SchedCandidate &Cand,
 
   // MaxMemoryClause-specific: We prioritize clustered instructions as we would
   // get more benefit from clausing these memory instructions.
-  const SUnit *CandNextClusterSU =
-      Cand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  const SUnit *TryCandNextClusterSU =
-      TryCand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  if (tryGreater(TryCand.SU == TryCandNextClusterSU,
-                 Cand.SU == CandNextClusterSU, TryCand, Cand, Cluster))
+  const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster;
+  const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster;
+  if (tryGreater(TryCandCluster && TryCandCluster->contains(TryCand.SU),
+                 CandCluster && CandCluster->contains(Cand.SU), TryCand, Cand,
+                 Cluster))
     return TryCand.Reason != NoCand;
 
   // We only compare a subset of features when comparing nodes between
diff --git a/llvm/lib/Target/PowerPC/PPCMachineScheduler.cpp b/llvm/lib/Target/PowerPC/PPCMachineScheduler.cpp
index 03712879f7c49..5eb1f0128643d 100644
--- a/llvm/lib/Target/PowerPC/PPCMachineScheduler.cpp
+++ b/llvm/lib/Target/PowerPC/PPCMachineScheduler.cpp
@@ -100,12 +100,11 @@ bool PPCPreRASchedStrategy::tryCandidate(SchedCandidate &Cand,
   // This is a best effort to set things up for a post-RA pass. Optimizations
   // like generating loads of multiple registers should ideally be done within
   // the scheduler pass by combining the loads during DAG postprocessing.
-  const SUnit *CandNextClusterSU =
-      Cand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  const SUnit *TryCandNextClusterSU =
-      TryCand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
-  if (tryGreater(TryCand.SU == TryCandNextClusterSU,
-                 Cand.SU == CandNextClusterSU, TryCand, Cand, Cluster))
+  const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster;
+  const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster;
+  if (tryGreater(TryCandCluster && TryCandCluster->contains(TryCand.SU),
+                 CandCluster && CandCluster->contains(Cand.SU), TryCand, Cand,
+                 Cluster))
     return TryCand.Reason != NoCand;
 
   if (SameBoundary) {
@@ -190,8 +189,11 @@ bool PPCPostRASchedStrategy::tryCandidate(SchedCandidate &Cand,
     return TryCand.Reason != NoCand;
 
   // Keep clustered nodes together.
-  if (tryGreater(TryCand.SU == DAG->getNextClusterSucc(),
-                 Cand.SU == DAG->getNextClusterSucc(), TryCand, Cand, Cluster))
+  const ClusterInfo *CandCluster = Cand.AtTop ? TopCluster : BotCluster;
+  const ClusterInfo *TryCandCluster = TryCand.AtTop ? TopCluster : BotCluster;
+  if (tryGreater(TryCandCluster && TryCandCluster->contains(TryCand.SU),
+                 CandCluster && CandCluster->contains(Cand.SU), TryCand, Cand,
+                 Cluster))
     return TryCand.Reason != NoCand;
 
   // Avoid critical resource consumption and balance the schedule.
diff --git a/llvm/test/CodeGen/AArch64/argument-blocks-array-of-struct.ll b/llvm/test/CodeGen/AArch64/argument-blocks-array-of-struct.ll
index f364429b86c38..6019a62f4925e 100644
--- a/llvm/test/CodeGen/AArch64/argument-blocks-array-of-struct.ll
+++ b/llvm/test/CodeGen/AArch64/argument-blocks-array-of-struct.ll
@@ -581,9 +581,8 @@ define void @callee_in_memory(%T_IN_MEMORY %a) {
 ; CHECK-SD-NEXT:    add x8, x8, :lo12:in_memory_store
 ; CHECK-SD-NEXT:    ldr d0, [sp, #64]
 ; CHECK-SD-NEXT:    str d0, [x8, #64]
-; CHECK-SD-NEXT:    ldr q0, [sp, #16]
 ; CHECK-SD-NEXT:    str q2, [x8, #48]
-; CHECK-SD-NEXT:    ldr q2, [sp]
+; CHECK-SD-NEXT:    ldp q2, q0, [sp]
 ; CHECK-SD-NEXT:    stp q0, q1, [x8, #16]
 ; CHECK-SD-NEXT:    str q2, [x8]
 ; CHECK-SD-NEXT:    ret
diff --git a/llvm/test/CodeGen/AArch64/arm64-dagcombiner-load-slicing.ll b/llvm/test/CodeGen/AArch64/arm64-dagcombiner-load-slicing.ll
index 7e72e8de01f4f..3bada9d5b3bb4 100644
--- a/llvm/test/CodeGen/AArch64/arm64-dagcombiner-load-slicing.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-dagcombiner-load-slicing.ll
@@ -7,8 +7,8 @@
 
 ; CHECK-LABEL: @test
 ; CHECK: add [[BASE:x[0-9]+]], x0, x1, lsl #3
-; CHECK: ldp [[CPLX1_I:s[0-9]+]], [[CPLX1_R:s[0-9]+]], [[[BASE]]]
-; CHECK: ldp [[CPLX2_I:s[0-9]+]], [[CPLX2_R:s[0-9]+]], [[[BASE]], #64]
+; CHECK-DAG: ldp [[CPLX1_I:s[0-9]+]], [[CPLX1_R:s[0-9]+]], [[[BASE]]]
+; CHECK-DAG: ldp [[CPLX2_I:s[0-9]+]], [[CPLX2_R:s[0-9]+]], [[[BASE]], #64]
 ; CHECK: fadd {{s[0-9]+}}, [[CPLX2_I]], [[CPLX1_I]]
 ; CHECK: fadd {{s[0-9]+}}, [[CPLX2_R]], [[CPLX1_R]]
 ; CHECK: ret
@@ -36,8 +36,8 @@ entry:
 
 ; CHECK-LABEL: @test_int
 ; CHECK: add [[BASE:x[0-9]+]], x0, x1, lsl #3
-; CHECK: ldp [[CPLX1_I:w[0-9]+]], [[CPLX1_R:w[0-9]+]], [[[BASE]]]
-; CHECK: ldp [[CPLX2_I:w[0-9]+]], [[CPLX2_R:w[0-9]+]], [[[BASE]], #64]
+; CHECK-DAG: ldp [[CPLX1_I:w[0-9]+]], [[CPLX1_R:w[0-9]+]], [[[BASE]]]
+; CHECK-DAG: ldp [[CPLX2_I:w[0-9]+]], [[CPLX2_R:w[0-9]+]], [[[BASE]], #64]
 ; CHECK: add {{w[0-9]+}}, [[CPLX2_I]], [[CPLX1_I]]
 ; CHECK: add {{w[0-9]+}}, [[CPLX2_R]], [[CPLX1_R]]
 ; CHECK: ret
@@ -65,8 +65,8 @@ entry:
 
 ; CHECK-LABEL: @test_long
 ; CHECK: add [[BASE:x[0-9]+]], x0, x1, lsl #4
-; CHECK: ldp [[CPLX1_I:x[0-9]+]], [[CPLX1_R:x[0-9]+]], [[[BASE]]]
-; CHECK: ldp [[CPLX2_I:x[0-9]+]], [[CPLX2_R:x[0-9]+]], [[[BASE]], #128]
+; CHECK-DAG: ldp [[CPLX1_I:x[0-9]+]], [[CPLX1_R:x[0-9]+]], [[[BASE]]]
+; CHECK-DAG: ldp [[CPLX2_I:x[0-9]+]], [[CPLX2_R:x[0-9]+]], [[[BASE]], #128]
 ; CHECK: add {{x[0-9]+}}, [[CPLX2_I]], [[CPLX1_I]]
 ; CHECK: add {{x[0-9]+}}, [[CPLX2_R]], [[CPLX1_R]]
 ; CHECK: ret
diff --git a/llvm/test/CodeGen/AArch64/bcmp.ll b/llvm/test/CodeGen/AArch64/bcmp.ll
index fee52ead98962..e70ddc3415cac 100644
--- a/llvm/test/CodeGen/AArch64/bcmp.ll
+++ b/llvm/test/CodeGen/AArch64/bcmp.ll
@@ -494,13 +494,14 @@ define i1 @bcmp_i128(i128 %a0, i128 %b0, i128 %a1, i128 %b1, i128 %a2, i128 %b2)
 ; CHECK-LABEL: bcmp_i128:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    cmp x2, x0
-; CHECK-NEXT:    ldp x8, x10, [sp]
+; CHECK-NEXT:    ldp x10, x8, [sp, #8]
 ; CHECK-NEXT:    ccmp x3, x1, #0, eq
-; CHECK-NEXT:    ldp x9, x11, [sp, #16]
+; CHECK-NEXT:    ldr x9, [sp]
+; CHECK-NEXT:    ldr x11, [sp, #24]
 ; CHECK-NEXT:    ccmp x6, x4, #0, eq
 ; CHECK-NEXT:    ccmp x7, x5, #0, eq
 ; CHECK-NEXT:    cset w12, ne
-; CHECK-NEXT:    cmp x9, x8
+; CHECK-NEXT:    cmp x8, x9
 ; CHECK-NEXT:    ccmp x11, x10, #0, eq
 ; CHECK-NEXT:    csinc w0, w12, wzr, eq
 ; CHECK-NEXT:    ret
diff --git a/llvm/test/CodeGen/AArch64/extbinopload.ll b/llvm/test/CodeGen/AArch64/extbinopload.ll
index 75f3ffc9515e5..cabb0e7278e40 100644
--- a/llvm/test/CodeGen/AArch64/extbinopload.ll
+++ b/llvm/test/CodeGen/AArch64/extbinopload.ll
@@ -861,7 +861,7 @@ define <16 x i32> @extrause_shuffle(ptr %p, ptr %q, ptr %r, ptr %s, ptr %z) {
 define <16 x i32> @extrause_ext(ptr %p, ptr %q, ptr %r, ptr %s, ptr %z) {
 ; CHECK-LABEL: extrause_ext:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp s0, s3, [x2]
+; CHECK-NEXT:    ldp s0, s5, [x2]
 ; CHECK-NEXT:    add x8, x3, #8
 ; CHECK-NEXT:    add x9, x3, #12
 ; CHECK-NEXT:    add x10, x1, #8
@@ -871,26 +871,26 @@ define <16 x i32> @extrause_ext(ptr %p, ptr %q, ptr %r, ptr %s, ptr %z) {
 ; CHECK-NEXT:    ld1 { v1.s }[1], [x1], #4
 ; CHECK-NEXT:    ld1 { v2.s }[1], [x1]
 ; CHECK-NEXT:    ldp s7, s4, [x0, #8]
-; CHECK-NEXT:    ld1 { v3.s }[1], [x3]
-; CHECK-NEXT:    ldp s6, s5, [x2, #8]
+; CHECK-NEXT:    ld1 { v5.s }[1], [x3]
+; CHECK-NEXT:    ldp s6, s3, [x2, #8]
 ; CHECK-NEXT:    ld1 { v4.s }[1], [x11]
 ; CHECK-NEXT:    ld1 { v7.s }[1], [x10]
-; CHECK-NEXT:    ld1 { v5.s }[1], [x9]
+; CHECK-NEXT:    ld1 { v3.s }[1], [x9]
 ; CHECK-NEXT:    ld1 { v6.s }[1], [x8]
 ; CHECK-NEXT:    uaddl v2.8h, v2.8b, v4.8b
 ; CHECK-NEXT:    uaddl v1.8h, v1.8b, v7.8b
 ; CHECK-NEXT:    ushll v4.8h, v4.8b, #0
-; CHECK-NEXT:    uaddl v3.8h, v3.8b, v5.8b
+; CHECK-NEXT:    uaddl v5.8h, v5.8b, v3.8b
 ; CHECK-NEXT:    uaddl v6.8h, v0.8b, v6.8b
-; CHECK-NEXT:    ushll v5.8h, v5.8b, #0
+; CHECK-NEXT:    ushll v16.8h, v3.8b, #0
 ; CHECK-NEXT:    ushll v0.4s, v2.4h, #3
 ; CHECK-NEXT:    ushll2 v2.4s, v2.8h, #3
-; CHECK-NEXT:    ushll v7.4s, v3.4h, #3
-; CHECK-NEXT:    ushll2 v3.4s, v3.8h, #3
-; CHECK-NEXT:    stp q4, q5, [x4]
+; CHECK-NEXT:    ushll v7.4s, v5.4h, #3
+; CHECK-NEXT:    ushll2 v5.4s, v5.8h, #3
+; CHECK-NEXT:    stp q4, q16, [x4]
 ; CHECK-NEXT:    uaddw v0.4s, v0.4s, v1.4h
 ; CHECK-NEXT:    uaddw2 v1.4s, v2.4s, v1.8h
-; CHECK-NEXT:    uaddw2 v3.4s, v3.4s, v6.8h
+; CHECK-NEXT:    uaddw2 v3.4s, v5.4s, v6.8h
 ; CHECK-NEXT:    uaddw v2.4s, v7.4s, v6.4h
 ; CHECK-NEXT:    ret
   %lp1 = load <4 x i8>, ptr %p
@@ -960,7 +960,7 @@ define <16 x i32> @extrause_ext(ptr %p, ptr %q, ptr %r, ptr %s, ptr %z) {
 define <16 x i32> @extrause_add(ptr %p, ptr %q, ptr %r, ptr %s, ptr %z) {
 ; CHECK-LABEL: extrause_add:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp s0, s4, [x2]
+; CHECK-NEXT:    ldp s0, s5, [x2]
 ; CHECK-NEXT:    add x8, x3, #8
 ; CHECK-NEXT:    add x9, x3, #12
 ; CHECK-NEXT:    add x10, x1, #8
@@ -970,15 +970,15 @@ define <16 x i32> @extrause_add(ptr %p, ptr %q, ptr %r, ptr %s, ptr %z) {
 ; CHECK-NEXT:    ld1 { v1.s }[1], [x1], #4
 ; CHECK-NEXT:    ld1 { v2.s }[1], [x1]
 ; CHECK-NEXT:    ldp s7, s3, [x0, #8]
-; CHECK-NEXT:    ld1 { v4.s }[1], [x3]
-; CHECK-NEXT:    ldp s6, s5, [x2, #8]
+; CHECK-NEXT:    ld1 { v5.s }[1], [x3]
+; CHECK-NEXT:    ldp s6, s4, [x2, #8]
 ; CHECK-NEXT:    ld1 { v3.s }[1], [x11]
 ; CHECK-NEXT:    ld1 { v7.s }[1], [x10]
-; CHECK-NEXT:    ld1 { v5.s }[1], [x9]
+; CHECK-NEXT:    ld1 { v4.s }[1], [x9]
 ; CHECK-NEXT:    ld1 { v6.s }[1], [x8]
 ; CHECK-NEXT:    uaddl v16.8h, v2.8b, v3.8b
 ; CHECK-NEXT:    uaddl v1.8h, v1.8b, v7.8b
-; CHECK-NEXT:    uaddl v4.8h, v4.8b, v5.8b
+; CHECK-NEXT:    uaddl v4.8h, v5.8b, v4.8b
 ; CHECK-NEXT:    uaddl v2.8h, v0.8b, v6.8b
 ; CHECK-NEXT:    ushll v0.4s, v16.4h, #3
 ; CHECK-NEXT:    ushll2 v6.4s, v16.8h, #3
diff --git a/llvm/test/CodeGen/AArch64/fp-conversion-to-tbl.ll b/llvm/test/CodeGen/AArch64/fp-conversion-to-tbl.ll
index d9d80f1cb50ee..1fbca7ca2c27c 100644
--- a/llvm/test/CodeGen/AArch64/fp-conversion-to-tbl.ll
+++ b/llvm/test/CodeGen/AArch64/fp-conversion-to-tbl.ll
@@ -118,12 +118,12 @@ define void @fptoui_2x_v8f32_to_v8i8_in_loop(ptr %A, ptr %B, ptr %dst) {
 ; CHECK-NEXT:    add x10, x0, x9
 ; CHECK-NEXT:    add x9, x1, x9
 ; CHECK-NEXT:    ldp q2, q1, [x10]
-; CHECK-NEXT:    fcvtzu.4s v5, v1
-; CHECK-NEXT:    ldp q1, q3, [x9]
-; CHECK-NEXT:    fcvtzu.4s v4, v2
-; CHECK-NEXT:    fcvtzu.4s v7, v3
+; CHECK-NEXT:    fcvtzu.4s v4, v1
+; CHECK-NEXT:    ldp q7, q1, [x9]
+; CHECK-NEXT:    fcvtzu.4s v3, v2
 ; CHECK-NEXT:    fcvtzu.4s v6, v1
-; CHECK-NEXT:    tbl.16b v1, { v4, v5, v6, v7 }, v0
+; CHECK-NEXT:    fcvtzu.4s v5, v7
+; CHECK-NEXT:    tbl.16b v1, { v3, v4, v5, v6 }, v0
 ; CHECK-NEXT:    str q1, [x2, x8, lsl #4]
 ; CHECK-NEXT:    add x8, x8, #1
 ; CHECK-NEXT:    cmp x8, #1000
@@ -185,12 +185,12 @@ define void @fptoui_2x_v8f32_to_v8i8_in_loop_no_concat_shuffle(ptr %A, ptr %B, p
 ; CHECK-NEXT:    add x10, x0, x9
 ; CHECK-NEXT:    add x9, x1, x9
 ; CHECK-NEXT:    ldp q2, q1, [x10]
-; CHECK-NEXT:    fcvtzu.4s v5, v1
-; CHECK-NEXT:    ldp q1, q3, [x9]
-; CHECK-NEXT:    fcvtzu.4s v4, v2
-; CHECK-NEXT:    fcvtzu.4s v7, v3
+; CHECK-NEXT:    fcvtzu.4s v4, v1
+; CHECK-NEXT:    ldp q7, q1, [x9]
+; CHECK-NEXT:    fcvtzu.4s v3, v2
 ; CHECK-NEXT:    fcvtzu.4s v6, v1
-; CHECK-NEXT:    tbl.16b v1, { v4, v5, v6, v7 }, v0
+; CHECK-NEXT:    fcvtzu.4s v5, v7
+; CHECK-NEXT:    tbl.16b v1, { v3, v4, v5, v6 }, v0
 ; CHECK-NEXT:    str q1, [x2, x8, lsl #4]
 ; CHECK-NEXT:    add x8, x8, #1
 ; CHECK-NEXT:    cmp x8, #1000
@@ -252,12 +252,12 @@ define void @fptoui_v16f32_to_v16i8_in_loop(ptr %A, ptr %dst) {
 ; CHECK-NEXT:    add x8, x8, #1
 ; CHECK-NEXT:    cmp x8, #1000
 ; CHECK-NEXT:    ldp q2, q1, [x9, #32]
-; CHECK-NEXT:    fcvtzu.4s v7, v1
-; CHECK-NEXT:    ldp q1, q3, [x9]
-; CHECK-NEXT:    fcvtzu.4s v6, v2
-; CHECK-NEXT:    fcvtzu.4s v5, v3
+; CHECK-NEXT:    fcvtzu.4s v6, v1
+; CHECK-NEXT:    ldp q7, q1, [x9]
+; CHECK-NEXT:    fcvtzu.4s v5, v2
 ; CHECK-NEXT:    fcvtzu.4s v4, v1
-; CHECK-NEXT:    tbl.16b v1, { v4, v5, v6, v7 }, v0
+; CHECK-NEXT:    fcvtzu.4s v3, v7
+; CHECK-NEXT:    tbl.16b v1, { v3, v4, v5, v6 }, v0
 ; CHECK-NEXT:    str q1, [x1], #32
 ; CHECK-NEXT:    b.eq LBB4_1
 ; CHECK-NEXT:  ; %bb.2: ; %exit
diff --git a/llvm/test/CodeGen/AArch64/mul.ll b/llvm/test/CodeGen/AArch64/mul.ll
index 8d9a6e6b92914..39ce70f88d6ba 100644
--- a/llvm/test/CodeGen/AArch64/mul.ll
+++ b/llvm/test/CodeGen/AArch64/mul.ll
@@ -576,17 +576,17 @@ define <3 x i128> @v3i128(<3 x i128> %d, <3 x i128> %e) {
 ; CHECK-GI-NEXT:    mul x9, x2, x10
 ; CHECK-GI-NEXT:    umulh x14, x2, x10
 ; CHECK-GI-NEXT:    madd x10, x3, x10, x13
-; CHECK-GI-NEXT:    ldp x13, x15, [sp, #16]
+; CHECK-GI-NEXT:    ldp x15, x13, [sp, #16]
 ; CHECK-GI-NEXT:    mov x2, x9
 ; CHECK-GI-NEXT:    umulh x11, x0, x6
 ; CHECK-GI-NEXT:    mov x0, x8
-; CHECK-GI-NEXT:    mul x15, x4, x15
+; CHECK-GI-NEXT:    mul x13, x4, x13
 ; CHECK-GI-NEXT:    add x3, x10, x14
-; CHECK-GI-NEXT:    umulh x16, x4, x13
+; CHECK-GI-NEXT:    umulh x16, x4, x15
 ; CHECK-GI-NEXT:    add x1, x12, x11
-; CHECK-GI-NEXT:    madd x15, x5, x13, x15
-; CHECK-GI-NEXT:    mul x4, x4, x13
-; CHECK-GI-NEXT:    add x5, x15, x16
+; CHECK-GI-NEXT:    madd x13, x5, x15, x13
+; CHECK-GI-NEXT:    mul x4, x4, x15
+; CHECK-GI-NEXT:    add x5, x13, x16
 ; CHECK-GI-NEXT:    ret
 entry:
   %s = mul <3 x i128> %d, %e
@@ -638,14 +638,14 @@ define <4 x i128> @v4i128(<4 x i128> %d, <4 x i128> %e) {
 ; CHECK-GI-NEXT:    umulh x17, x4, x15
 ; CHECK-GI-NEXT:    add x3, x13, x14
 ; CHECK-GI-NEXT:    madd x15, x5, x15, x16
-; CHECK-GI-NEXT:    ldp x16, x18, [sp, #48]
+; CHECK-GI-NEXT:    ldp x18, x16, [sp, #48]
 ; CHECK-GI-NEXT:    mov x4, x10
-; CHECK-GI-NEXT:    mul x18, x6, x18
-; CHECK-GI-NEXT:    umulh x0, x6, x16
+; CHECK-GI-NEXT:    mul x16, x6, x16
+; CHECK-GI-NEXT:    umulh x0, x6, x18
 ; CHECK-GI-NEXT:    add x5, x15, x17
-; CHECK-GI-NEXT:    madd x18, x7, x16, x18
-; CHECK-GI-NEXT:    mul x6, x6, x16
-; CHECK-GI-NEXT:    add x7, x18, x0
+; CHECK-GI-NEXT:    madd x16, x7, x18, x16
+; CHECK-GI-NEXT:    mul x6, x6, x18
+; CHECK-GI-NEXT:    add x7, x16, x0
 ; CHECK-GI-NEXT:    mov x0, x8
 ; CHECK-GI-NEXT:    ret
 entry:
diff --git a/llvm/test/CodeGen/AArch64/nzcv-save.ll b/llvm/test/CodeGen/AArch64/nzcv-save.ll
index c40e529ccab1b..cc666dd8d34e6 100644
--- a/llvm/test/CodeGen/AArch64/nzcv-save.ll
+++ b/llvm/test/CodeGen/AArch64/nzcv-save.ll
@@ -6,19 +6,19 @@
 define void @f(ptr nocapture %a, ptr nocapture %b, ptr nocapture %cc, ptr nocapture %dd) nounwind uwtable noinline ssp {
 ; CHECK-LABEL: f:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    ldp x8, x10, [x2]
-; CHECK-NEXT:    ldp x9, x11, [x3]
+; CHECK-NEXT:    ldp x8, x11, [x3]
+; CHECK-NEXT:    ldp x9, x10, [x2]
 ; CHECK-NEXT:    ldp x13, x12, [x2, #16]
-; CHECK-NEXT:    adds x8, x8, x9
-; CHECK-NEXT:    ldp x14, x9, [x3, #16]
+; CHECK-NEXT:    adds x8, x9, x8
+; CHECK-NEXT:    ldp x9, x14, [x3, #16]
 ; CHECK-NEXT:    adcs x10, x10, x11
 ; CHECK-NEXT:    stp x8, x10, [x0]
-; CHECK-NEXT:    adcs x11, x13, x14
-; CHECK-NEXT:    adc x13, x12, x9
+; CHECK-NEXT:    adcs x9, x13, x9
+; CHECK-NEXT:    adc x11, x12, x14
 ; CHECK-NEXT:    orr x12, x12, #0x100
-; CHECK-NEXT:    adc x9, x12, x9
-; CHECK-NEXT:    stp x11, x13, [x0, #16]
-; CHECK-NEXT:    stp x11, x9, [x1, #16]
+; CHECK-NEXT:    stp x9, x11, [x0, #16]
+; CHECK-NEXT:    adc x11, x12, x14
+; CHECK-NEXT:    stp x9, x11, [x1, #16]
 ; CHECK-NEXT:    stp x8, x10, [x1]
 ; CHECK-NEXT:    ret
 entry:
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-convert.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-convert.ll
index 5f6b60a767f9d..4fe303b9bbf46 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-convert.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-convert.ll
@@ -26,18 +26,18 @@ define void @fp_convert_combine_crash(ptr %a, ptr %b) {
 ; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 64
 ; NONEON-NOSVE-NEXT:    ldp s1, s0, [sp, #24]
 ; NONEON-NOSVE-NEXT:    fcvtzs w8, s0, #3
-; NONEON-NOSVE-NEXT:    ldp s0, s2, [sp, #16]
+; NONEON-NOSVE-NEXT:    ldp s2, s0, [sp, #16]
 ; NONEON-NOSVE-NEXT:    fcvtzs w9, s1, #3
-; NONEON-NOSVE-NEXT:    fcvtzs w10, s2, #3
-; NONEON-NOSVE-NEXT:    fcvtzs w11, s0, #3
-; NONEON-NOSVE-NEXT:    ldp s2, s1, [sp, #8]
-; NONEON-NOSVE-NEXT:    ldp s0, s3, [sp]
+; NONEON-NOSVE-NEXT:    fcvtzs w10, s0, #3
+; NONEON-NOSVE-NEXT:    fcvtzs w11, s2, #3
+; NONEON-NOSVE-NEXT:    ldp s1, s0, [sp, #8]
+; NONEON-NOSVE-NEXT:    ldp s2, s3, [sp]
 ; NONEON-NOSVE-NEXT:    stp w9, w8, [sp, #56]
-; NONEON-NOSVE-NEXT:    fcvtzs w12, s1, #3
-; NONEON-NOSVE-NEXT:    fcvtzs w8, s2, #3
+; NONEON-NOSVE-NEXT:    fcvtzs w12, s0, #3
+; NONEON-NOSVE-NEXT:    fcvtzs w8, s1, #3
 ; NONEON-NOSVE-NEXT:    stp w11, w10, [sp, #48]
 ; NONEON-NOSVE-NEXT:    fcvtzs w9, s3, #3
-; NONEON-NOSVE-NEXT:    fcvtzs w10, s0, #3
+; NONEON-NOSVE-NEXT:    fcvtzs w10, s2, #3
 ; NONEON-NOSVE-NEXT:    stp w8, w12, [sp, #40]
 ; NONEON-NOSVE-NEXT:    stp w10, w9, [sp, #32]
 ; NONEON-NOSVE-NEXT:    ldp q0, q1, [sp, #32]
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-vselect.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-vselect.ll
index ad5f91a5f39a4..ec0693a541e44 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-vselect.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-vselect.ll
@@ -194,12 +194,12 @@ define <8 x half> @select_v8f16(<8 x half> %op1, <8 x half> %op2, <8 x i1> %mask
 define void @select_v16f16(ptr %a, ptr %b) {
 ; CHECK-LABEL: select_v16f16:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
+; CHECK-NEXT:    ldp q0, q3, [x1]
 ; CHECK-NEXT:    ptrue p0.h, vl8
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    fcmeq p1.h, p0/z, z0.h, z1.h
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    fcmeq p1.h, p0/z, z1.h, z0.h
 ; CHECK-NEXT:    fcmeq p0.h, p0/z, z2.h, z3.h
-; CHECK-NEXT:    sel z0.h, p1, z0.h, z1.h
+; CHECK-NEXT:    mov z0.h, p1/m, z1.h
 ; CHECK-NEXT:    sel z1.h, p0, z2.h, z3.h
 ; CHECK-NEXT:    stp q0, q1, [x0]
 ; CHECK-NEXT:    ret
@@ -429,12 +429,12 @@ define <4 x float> @select_v4f32(<4 x float> %op1, <4 x float> %op2, <4 x i1> %m
 define void @select_v8f32(ptr %a, ptr %b) {
 ; CHECK-LABEL: select_v8f32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
+; CHECK-NEXT:    ldp q0, q3, [x1]
 ; CHECK-NEXT:    ptrue p0.s, vl4
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    fcmeq p1.s, p0/z, z0.s, z1.s
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    fcmeq p1.s, p0/z, z1.s, z0.s
 ; CHECK-NEXT:    fcmeq p0.s, p0/z, z2.s, z3.s
-; CHECK-NEXT:    sel z0.s, p1, z0.s, z1.s
+; CHECK-NEXT:    mov z0.s, p1/m, z1.s
 ; CHECK-NEXT:    sel z1.s, p0, z2.s, z3.s
 ; CHECK-NEXT:    stp q0, q1, [x0]
 ; CHECK-NEXT:    ret
@@ -553,12 +553,12 @@ define <2 x double> @select_v2f64(<2 x double> %op1, <2 x double> %op2, <2 x i1>
 define void @select_v4f64(ptr %a, ptr %b) {
 ; CHECK-LABEL: select_v4f64:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
+; CHECK-NEXT:    ldp q0, q3, [x1]
 ; CHECK-NEXT:    ptrue p0.d, vl2
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    fcmeq p1.d, p0/z, z0.d, z1.d
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    fcmeq p1.d, p0/z, z1.d, z0.d
 ; CHECK-NEXT:    fcmeq p0.d, p0/z, z2.d, z3.d
-; CHECK-NEXT:    sel z0.d, p1, z0.d, z1.d
+; CHECK-NEXT:    mov z0.d, p1/m, z1.d
 ; CHECK-NEXT:    sel z1.d, p0, z2.d, z3.d
 ; CHECK-NEXT:    stp q0, q1, [x0]
 ; CHECK-NEXT:    ret
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-mulh.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-mulh.ll
index 97f2e7a1e66cb..0c97eedd4362d 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-mulh.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-mulh.ll
@@ -973,11 +973,11 @@ define <4 x i32> @smulh_v4i32(<4 x i32> %op1, <4 x i32> %op2) {
 ; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #48]
 ; NONEON-NOSVE-NEXT:    ldpsw x13, x12, [sp, #48]
 ; NONEON-NOSVE-NEXT:    smull x11, w11, w12
-; NONEON-NOSVE-NEXT:    ldpsw x12, x14, [sp, #56]
+; NONEON-NOSVE-NEXT:    ldpsw x14, x12, [sp, #56]
 ; NONEON-NOSVE-NEXT:    smull x10, w10, w13
 ; NONEON-NOSVE-NEXT:    lsr x11, x11, #32
-; NONEON-NOSVE-NEXT:    smull x9, w9, w14
-; NONEON-NOSVE-NEXT:    smull x8, w8, w12
+; NONEON-NOSVE-NEXT:    smull x9, w9, w12
+; NONEON-NOSVE-NEXT:    smull x8, w8, w14
 ; NONEON-NOSVE-NEXT:    lsr x10, x10, #32
 ; NONEON-NOSVE-NEXT:    lsr x9, x9, #32
 ; NONEON-NOSVE-NEXT:    stp w10, w11, [sp, #72]
@@ -1038,12 +1038,12 @@ define void @smulh_v8i32(ptr %a, ptr %b) {
 ; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #112]
 ; NONEON-NOSVE-NEXT:    ldpsw x17, x16, [sp, #112]
 ; NONEON-NOSVE-NEXT:    smull x15, w15, w16
-; NONEON-NOSVE-NEXT:    ldpsw x16, x18, [sp, #120]
+; NONEON-NOSVE-NEXT:    ldpsw x18, x16, [sp, #120]
 ; NONEON-NOSVE-NEXT:    smull x14, w14, w17
 ; NONEON-NOSVE-NEXT:    ldpsw x17, x1, [sp, #80]
-; NONEON-NOSVE-NEXT:    smull x13, w13, w18
+; NONEON-NOSVE-NEXT:    smull x13, w13, w16
 ; NONEON-NOSVE-NEXT:    lsr x15, x15, #32
-; NONEON-NOSVE-NEXT:    smull x12, w12, w16
+; NONEON-NOSVE-NEXT:    smull x12, w12, w18
 ; NONEON-NOSVE-NEXT:    lsr x14, x14, #32
 ; NONEON-NOSVE-NEXT:    ldpsw x16, x18, [sp, #88]
 ; NONEON-NOSVE-NEXT:    smull x11, w11, w1
@@ -2172,11 +2172,11 @@ define <4 x i32> @umulh_v4i32(<4 x i32> %op1, <4 x i32> %op2) {
 ; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #48]
 ; NONEON-NOSVE-NEXT:    ldp w13, w12, [sp, #48]
 ; NONEON-NOSVE-NEXT:    umull x11, w11, w12
-; NONEON-NOSVE-NEXT:    ldp w12, w14, [sp, #56]
+; NONEON-NOSVE-NEXT:    ldp w14, w12, [sp, #56]
 ; NONEON-NOSVE-NEXT:    umull x10, w10, w13
 ; NONEON-NOSVE-NEXT:    lsr x11, x11, #32
-; NONEON-NOSVE-NEXT:    umull x9, w9, w14
-; NONEON-NOSVE-NEXT:    umull x8, w8, w12
+; NONEON-NOSVE-NEXT:    umull x9, w9, w12
+; NONEON-NOSVE-NEXT:    umull x8, w8, w14
 ; NONEON-NOSVE-NEXT:    lsr x10, x10, #32
 ; NONEON-NOSVE-NEXT:    lsr x9, x9, #32
 ; NONEON-NOSVE-NEXT:    stp w10, w11, [sp, #72]
@@ -2237,12 +2237,12 @@ define void @umulh_v8i32(ptr %a, ptr %b) {
 ; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #112]
 ; NONEON-NOSVE-NEXT:    ldp w17, w16, [sp, #112]
 ; NONEON-NOSVE-NEXT:    umull x15, w15, w16
-; NONEON-NOSVE-NEXT:    ldp w16, w18, [sp, #120]
+; NONEON-NOSVE-NEXT:    ldp w18, w16, [sp, #120]
 ; NONEON-NOSVE-NEXT:    umull x14, w14, w17
 ; NONEON-NOSVE-NEXT:    ldp w17, w1, [sp, #80]
-; NONEON-NOSVE-NEXT:    umull x13, w13, w18
+; NONEON-NOSVE-NEXT:    umull x13, w13, w16
 ; NONEON-NOSVE-NEXT:    lsr x15, x15, #32
-; NONEON-NOSVE-NEXT:    umull x12, w12, w16
+; NONEON-NOSVE-NEXT:    umull x12, w12, w18
 ; NONEON-NOSVE-NEXT:    lsr x14, x14, #32
 ; NONEON-NOSVE-NEXT:    ldp w16, w18, [sp, #88]
 ; NONEON-NOSVE-NEXT:    umull x11, w11, w1
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-vselect.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-vselect.ll
index 41eb731fd66df..39701131d7db6 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-vselect.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-int-vselect.ll
@@ -288,12 +288,12 @@ define <16 x i8> @select_v16i8(<16 x i8> %op1, <16 x i8> %op2, <16 x i1> %mask)
 define void @select_v32i8(ptr %a, ptr %b) {
 ; CHECK-LABEL: select_v32i8:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
+; CHECK-NEXT:    ldp q0, q3, [x1]
 ; CHECK-NEXT:    ptrue p0.b, vl16
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    cmpeq p1.b, p0/z, z0.b, z1.b
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    cmpeq p1.b, p0/z, z1.b, z0.b
 ; CHECK-NEXT:    cmpeq p0.b, p0/z, z2.b, z3.b
-; CHECK-NEXT:    sel z0.b, p1, z0.b, z1.b
+; CHECK-NEXT:    mov z0.b, p1/m, z1.b
 ; CHECK-NEXT:    sel z1.b, p0, z2.b, z3.b
 ; CHECK-NEXT:    stp q0, q1, [x0]
 ; CHECK-NEXT:    ret
@@ -692,12 +692,12 @@ define <8 x i16> @select_v8i16(<8 x i16> %op1, <8 x i16> %op2, <8 x i1> %mask) {
 define void @select_v16i16(ptr %a, ptr %b) {
 ; CHECK-LABEL: select_v16i16:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
+; CHECK-NEXT:    ldp q0, q3, [x1]
 ; CHECK-NEXT:    ptrue p0.h, vl8
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    cmpeq p1.h, p0/z, z0.h, z1.h
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    cmpeq p1.h, p0/z, z1.h, z0.h
 ; CHECK-NEXT:    cmpeq p0.h, p0/z, z2.h, z3.h
-; CHECK-NEXT:    sel z0.h, p1, z0.h, z1.h
+; CHECK-NEXT:    mov z0.h, p1/m, z1.h
 ; CHECK-NEXT:    sel z1.h, p0, z2.h, z3.h
 ; CHECK-NEXT:    stp q0, q1, [x0]
 ; CHECK-NEXT:    ret
@@ -906,12 +906,12 @@ define <4 x i32> @select_v4i32(<4 x i32> %op1, <4 x i32> %op2, <4 x i1> %mask) {
 define void @select_v8i32(ptr %a, ptr %b) {
 ; CHECK-LABEL: select_v8i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
+; CHECK-NEXT:    ldp q0, q3, [x1]
 ; CHECK-NEXT:    ptrue p0.s, vl4
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    cmpeq p1.s, p0/z, z0.s, z1.s
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    cmpeq p1.s, p0/z, z1.s, z0.s
 ; CHECK-NEXT:    cmpeq p0.s, p0/z, z2.s, z3.s
-; CHECK-NEXT:    sel z0.s, p1, z0.s, z1.s
+; CHECK-NEXT:    mov z0.s, p1/m, z1.s
 ; CHECK-NEXT:    sel z1.s, p0, z2.s, z3.s
 ; CHECK-NEXT:    stp q0, q1, [x0]
 ; CHECK-NEXT:    ret
@@ -930,16 +930,16 @@ define void @select_v8i32(ptr %a, ptr %b) {
 ; NONEON-NOSVE-NEXT:    cmp w10, w9
 ; NONEON-NOSVE-NEXT:    csel w9, w10, w9, eq
 ; NONEON-NOSVE-NEXT:    cmp w13, w12
-; NONEON-NOSVE-NEXT:    ldp w15, w16, [sp, #48]
+; NONEON-NOSVE-NEXT:    ldp w10, w16, [sp, #48]
 ; NONEON-NOSVE-NEXT:    csel w12, w13, w12, eq
 ; NONEON-NOSVE-NEXT:    cmp w14, w11
-; NONEON-NOSVE-NEXT:    ldp w10, w13, [sp, #32]
+; NONEON-NOSVE-NEXT:    ldp w15, w13, [sp, #32]
 ; NONEON-NOSVE-NEXT:    csel w11, w14, w11, eq
 ; NONEON-NOSVE-NEXT:    ldp w17, w14, [sp, #56]
 ; NONEON-NOSVE-NEXT:    ldp w18, w1, [sp, #40]
-; NONEON-NOSVE-NEXT:    cmp w10, w15
+; NONEON-NOSVE-NEXT:    cmp w15, w10
 ; NONEON-NOSVE-NEXT:    stp w12, w11, [sp, #72]
-; NONEON-NOSVE-NEXT:    csel w10, w10, w15, eq
+; NONEON-NOSVE-NEXT:    csel w10, w15, w10, eq
 ; NONEON-NOSVE-NEXT:    cmp w13, w16
 ; NONEON-NOSVE-NEXT:    ldr w15, [sp]
 ; NONEON-NOSVE-NEXT:    csel w13, w13, w16, eq
@@ -1039,12 +1039,12 @@ define <2 x i64> @select_v2i64(<2 x i64> %op1, <2 x i64> %op2, <2 x i1> %mask) {
 define void @select_v4i64(ptr %a, ptr %b) {
 ; CHECK-LABEL: select_v4i64:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
+; CHECK-NEXT:    ldp q0, q3, [x1]
 ; CHECK-NEXT:    ptrue p0.d, vl2
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    cmpeq p1.d, p0/z, z0.d, z1.d
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    cmpeq p1.d, p0/z, z1.d, z0.d
 ; CHECK-NEXT:    cmpeq p0.d, p0/z, z2.d, z3.d
-; CHECK-NEXT:    sel z0.d, p1, z0.d, z1.d
+; CHECK-NEXT:    mov z0.d, p1/m, z1.d
 ; CHECK-NEXT:    sel z1.d, p0, z2.d, z3.d
 ; CHECK-NEXT:    stp q0, q1, [x0]
 ; CHECK-NEXT:    ret
@@ -1057,13 +1057,13 @@ define void @select_v4i64(ptr %a, ptr %b) {
 ; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 96
 ; NONEON-NOSVE-NEXT:    stp q2, q3, [sp, #32]
 ; NONEON-NOSVE-NEXT:    ldr x9, [sp, #8]
-; NONEON-NOSVE-NEXT:    ldp x8, x11, [sp, #24]
+; NONEON-NOSVE-NEXT:    ldp x8, x10, [sp, #24]
 ; NONEON-NOSVE-NEXT:    ldr x13, [sp, #40]
-; NONEON-NOSVE-NEXT:    ldp x10, x12, [sp, #48]
+; NONEON-NOSVE-NEXT:    ldp x11, x12, [sp, #48]
 ; NONEON-NOSVE-NEXT:    cmp x9, x8
 ; NONEON-NOSVE-NEXT:    csel x8, x9, x8, eq
-; NONEON-NOSVE-NEXT:    cmp x11, x10
-; NONEON-NOSVE-NEXT:    csel x9, x11, x10, eq
+; NONEON-NOSVE-NEXT:    cmp x10, x11
+; NONEON-NOSVE-NEXT:    csel x9, x10, x11, eq
 ; NONEON-NOSVE-NEXT:    ldr x10, [sp, #16]
 ; NONEON-NOSVE-NEXT:    ldr x11, [sp]
 ; NONEON-NOSVE-NEXT:    cmp x13, x12
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-permute-zip-uzp-trn.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-permute-zip-uzp-trn.ll
index 3d9f407c3064c..e0e88c47fb55c 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-permute-zip-uzp-trn.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-permute-zip-uzp-trn.ll
@@ -151,20 +151,20 @@ define void @zip_v32i16(ptr %a, ptr %b) {
 ; CHECK-NEXT:    .cfi_offset b13, -48
 ; CHECK-NEXT:    .cfi_offset b14, -56
 ; CHECK-NEXT:    .cfi_offset b15, -64
-; CHECK-NEXT:    ldp q1, q0, [x0]
-; CHECK-NEXT:    ldp q2, q3, [x1]
-; CHECK-NEXT:    mov z5.h, z0.h[7]
-; CHECK-NEXT:    mov z7.h, z0.h[6]
-; CHECK-NEXT:    mov z17.h, z0.h[5]
-; CHECK-NEXT:    mov z4.h, z3.h[7]
-; CHECK-NEXT:    mov z6.h, z3.h[6]
-; CHECK-NEXT:    mov z16.h, z3.h[5]
-; CHECK-NEXT:    mov z18.h, z3.h[4]
-; CHECK-NEXT:    mov z19.h, z0.h[4]
-; CHECK-NEXT:    mov z20.h, z2.h[7]
-; CHECK-NEXT:    mov z21.h, z1.h[7]
-; CHECK-NEXT:    mov z22.h, z2.h[6]
-; CHECK-NEXT:    mov z23.h, z1.h[6]
+; CHECK-NEXT:    ldp q1, q2, [x1]
+; CHECK-NEXT:    ldp q0, q3, [x0]
+; CHECK-NEXT:    mov z4.h, z2.h[7]
+; CHECK-NEXT:    mov z6.h, z2.h[6]
+; CHECK-NEXT:    mov z16.h, z2.h[5]
+; CHECK-NEXT:    mov z5.h, z3.h[7]
+; CHECK-NEXT:    mov z7.h, z3.h[6]
+; CHECK-NEXT:    mov z17.h, z3.h[5]
+; CHECK-NEXT:    mov z18.h, z2.h[4]
+; CHECK-NEXT:    mov z19.h, z3.h[4]
+; CHECK-NEXT:    mov z20.h, z1.h[7]
+; CHECK-NEXT:    mov z21.h, z0.h[7]
+; CHECK-NEXT:    mov z22.h, z1.h[6]
+; CHECK-NEXT:    mov z23.h, z0.h[6]
 ; CHECK-NEXT:    zip1 z24.h, z5.h, z4.h
 ; CHECK-NEXT:    zip1 z25.h, z7.h, z6.h
 ; CHECK-NEXT:    zip1 z16.h, z17.h, z16.h
@@ -174,12 +174,12 @@ define void @zip_v32i16(ptr %a, ptr %b) {
 ; CHECK-NEXT:    zip1 z18.h, z21.h, z20.h
 ; CHECK-NEXT:    zip1 z21.s, z25.s, z24.s
 ; CHECK-NEXT:    zip1 z22.h, z23.h, z22.h
-; CHECK-NEXT:    mov z23.h, z2.h[5]
+; CHECK-NEXT:    mov z23.h, z1.h[5]
 ; CHECK-NEXT:    mov z20.h, z6.h[7]
-; CHECK-NEXT:    mov z24.h, z1.h[5]
-; CHECK-NEXT:    mov z25.h, z2.h[4]
+; CHECK-NEXT:    mov z24.h, z0.h[5]
+; CHECK-NEXT:    mov z25.h, z1.h[4]
 ; CHECK-NEXT:    mov z19.h, z7.h[7]
-; CHECK-NEXT:    mov z26.h, z1.h[4]
+; CHECK-NEXT:    mov z26.h, z0.h[4]
 ; CHECK-NEXT:    mov z27.h, z6.h[6]
 ; CHECK-NEXT:    mov z28.h, z7.h[5]
 ; CHECK-NEXT:    mov z29.h, z6.h[5]
@@ -212,22 +212,22 @@ define void @zip_v32i16(ptr %a, ptr %b) {
 ; CHECK-NEXT:    zip1 z19.s, z28.s, z27.s
 ; CHECK-NEXT:    zip1 z18.s, z22.s, z18.s
 ; CHECK-NEXT:    zip1 z20.s, z24.s, z23.s
-; CHECK-NEXT:    zip1 z0.h, z0.h, z3.h
+; CHECK-NEXT:    zip1 z2.h, z3.h, z2.h
 ; CHECK-NEXT:    zip1 z3.s, z26.s, z25.s
 ; CHECK-NEXT:    zip1 z22.s, z30.s, z29.s
 ; CHECK-NEXT:    zip1 z6.h, z6.h, z7.h
 ; CHECK-NEXT:    zip1 z7.d, z16.d, z21.d
 ; CHECK-NEXT:    zip1 z16.d, z19.d, z17.d
-; CHECK-NEXT:    zip1 z1.h, z1.h, z2.h
-; CHECK-NEXT:    zip1 z2.h, z4.h, z5.h
+; CHECK-NEXT:    zip1 z0.h, z0.h, z1.h
+; CHECK-NEXT:    zip1 z1.h, z4.h, z5.h
 ; CHECK-NEXT:    zip1 z4.d, z20.d, z18.d
 ; CHECK-NEXT:    zip1 z3.d, z22.d, z3.d
-; CHECK-NEXT:    add z0.h, z0.h, z6.h
+; CHECK-NEXT:    add z2.h, z2.h, z6.h
 ; CHECK-NEXT:    add z5.h, z7.h, z16.h
-; CHECK-NEXT:    add z1.h, z1.h, z2.h
-; CHECK-NEXT:    add z2.h, z4.h, z3.h
-; CHECK-NEXT:    stp q0, q5, [x0, #32]
-; CHECK-NEXT:    stp q1, q2, [x0]
+; CHECK-NEXT:    add z0.h, z0.h, z1.h
+; CHECK-NEXT:    add z1.h, z4.h, z3.h
+; CHECK-NEXT:    stp q2, q5, [x0, #32]
+; CHECK-NEXT:    stp q0, q1, [x0]
 ; CHECK-NEXT:    ldp d15, d14, [sp], #64 // 16-byte Folded Reload
 ; CHECK-NEXT:    ret
 ;
@@ -659,10 +659,10 @@ define void @zip1_v8i32_undef(ptr %a) {
 define void @trn_v32i8(ptr %a, ptr %b) {
 ; CHECK-LABEL: trn_v32i8:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    trn1 z4.b, z0.b, z1.b
-; CHECK-NEXT:    trn2 z0.b, z0.b, z1.b
+; CHECK-NEXT:    ldp q0, q3, [x1]
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    trn1 z4.b, z1.b, z0.b
+; CHECK-NEXT:    trn2 z0.b, z1.b, z0.b
 ; CHECK-NEXT:    trn1 z1.b, z2.b, z3.b
 ; CHECK-NEXT:    trn2 z2.b, z2.b, z3.b
 ; CHECK-NEXT:    add z0.b, z4.b, z0.b
@@ -862,10 +862,10 @@ define void @trn_v8i16(ptr %a, ptr %b) {
 define void @trn_v16i16(ptr %a, ptr %b) {
 ; CHECK-LABEL: trn_v16i16:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    trn1 z4.h, z0.h, z1.h
-; CHECK-NEXT:    trn2 z0.h, z0.h, z1.h
+; CHECK-NEXT:    ldp q0, q3, [x1]
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    trn1 z4.h, z1.h, z0.h
+; CHECK-NEXT:    trn2 z0.h, z1.h, z0.h
 ; CHECK-NEXT:    trn1 z1.h, z2.h, z3.h
 ; CHECK-NEXT:    trn2 z2.h, z2.h, z3.h
 ; CHECK-NEXT:    add z0.h, z4.h, z0.h
@@ -961,10 +961,10 @@ define void @trn_v16i16(ptr %a, ptr %b) {
 define void @trn_v8i32(ptr %a, ptr %b) {
 ; CHECK-LABEL: trn_v8i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    zip1 z4.s, z0.s, z1.s
-; CHECK-NEXT:    trn2 z0.s, z0.s, z1.s
+; CHECK-NEXT:    ldp q0, q3, [x1]
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    zip1 z4.s, z1.s, z0.s
+; CHECK-NEXT:    trn2 z0.s, z1.s, z0.s
 ; CHECK-NEXT:    trn1 z1.s, z2.s, z3.s
 ; CHECK-NEXT:    trn2 z2.s, z2.s, z3.s
 ; CHECK-NEXT:    add z0.s, z4.s, z0.s
@@ -1006,11 +1006,11 @@ define void @trn_v8i32(ptr %a, ptr %b) {
 define void @trn_v4f64(ptr %a, ptr %b) {
 ; CHECK-LABEL: trn_v4f64:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldp q0, q2, [x0]
+; CHECK-NEXT:    ldp q0, q3, [x1]
 ; CHECK-NEXT:    ptrue p0.d, vl2
-; CHECK-NEXT:    ldp q1, q3, [x1]
-; CHECK-NEXT:    zip1 z4.d, z0.d, z1.d
-; CHECK-NEXT:    trn2 z0.d, z0.d, z1.d
+; CHECK-NEXT:    ldp q1, q2, [x0]
+; CHECK-NEXT:    zip1 z4.d, z1.d, z0.d
+; CHECK-NEXT:    trn2 z0.d, z1.d, z0.d
 ; CHECK-NEXT:    zip1 z1.d, z2.d, z3.d
 ; CHECK-NEXT:    trn2 z2.d, z2.d, z3.d
 ; CHECK-NEXT:    fadd z0.d, p0/m, z0.d, z4.d
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-ptest.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-ptest.ll
index e07036f2a1acf..90466e3cebd5e 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-ptest.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-ptest.ll
@@ -125,14 +125,14 @@ define i1 @ptest_or_v16i1(ptr %a, ptr %b) {
 ; CHECK-NEXT:    ldp q2, q3, [x1, #32]
 ; CHECK-NEXT:    ldp q4, q5, [x0]
 ; CHECK-NEXT:    fcmne p1.s, p0/z, z1.s, #0.0
-; CHECK-NEXT:    ldp q1, q6, [x1]
+; CHECK-NEXT:    ldp q6, q1, [x1]
 ; CHECK-NEXT:    fcmne p3.s, p0/z, z3.s, #0.0
 ; CHECK-NEXT:    fcmne p2.s, p0/z, z0.s, #0.0
 ; CHECK-NEXT:    fcmne p5.s, p0/z, z2.s, #0.0
 ; CHECK-NEXT:    fcmne p4.s, p0/z, z5.s, #0.0
 ; CHECK-NEXT:    fcmne p7.s, p0/z, z4.s, #0.0
-; CHECK-NEXT:    fcmne p6.s, p0/z, z6.s, #0.0
-; CHECK-NEXT:    fcmne p0.s, p0/z, z1.s, #0.0
+; CHECK-NEXT:    fcmne p6.s, p0/z, z1.s, #0.0
+; CHECK-NEXT:    fcmne p0.s, p0/z, z6.s, #0.0
 ; CHECK-NEXT:    mov z0.s, p1/z, #-1 // =0xffffffffffffffff
 ; CHECK-NEXT:    mov z2.s, p3/z, #-1 // =0xffffffffffffffff
 ; CHECK-NEXT:    mov z1.s, p2/z, #-1 // =0xffffffffffffffff
@@ -334,14 +334,14 @@ define i1 @ptest_and_v16i1(ptr %a, ptr %b) {
 ; CHECK-NEXT:    ldp q2, q3, [x1, #32]
 ; CHECK-NEXT:    ldp q4, q5, [x0]
 ; CHECK-NEXT:    fcmne p1.s, p0/z, z1.s, #0.0
-; CHECK-NEXT:    ldp q1, q6, [x1]
+; CHECK-NEXT:    ldp q6, q1, [x1]
 ; CHECK-NEXT:    fcmne p3.s, p0/z, z3.s, #0.0
 ; CHECK-NEXT:    fcmne p2.s, p0/z, z0.s, #0.0
 ; CHECK-NEXT:    fcmne p5.s, p0/z, z2.s, #0.0
 ; CHECK-NEXT:    fcmne p4.s, p0/z, z5.s, #0.0
 ; CHECK-NEXT:    fcmne p7.s, p0/z, z4.s, #0.0
-; CHECK-NEXT:    fcmne p6.s, p0/z, z6.s, #0.0
-; CHECK-NEXT:    fcmne p0.s, p0/z, z1.s, #0.0
+; CHECK-NEXT:    fcmne p6.s, p0/z, z1.s, #0.0
+; CHECK-NEXT:    fcmne p0.s, p0/z, z6.s, #0.0
 ; CHECK-NEXT:    mov z0.s, p1/z, #-1 // =0xffffffffffffffff
 ; CHECK-NEXT:    mov z2.s, p3/z, #-1 // =0xffffffffffffffff
 ; CHECK-NEXT:    mov z1.s, p2/z, #-1 // =0xffffffffffffffff
diff --git a/llvm/test/CodeGen/AArch64/zext-to-tbl.ll b/llvm/test/CodeGen/AArch64/zext-to-tbl.ll
index 46a2459485987..75c5bee2ae0ab 100644
--- a/llvm/test/CodeGen/AArch64/zext-to-tbl.ll
+++ b/llvm/test/CodeGen/AArch64/zext-to-tbl.ll
@@ -2913,28 +2913,28 @@ define i32 @test_widening_instr_mull_64(ptr %p1, ptr %p2, i32 %h) {
 ; CHECK-NEXT:  LBB25_1: ; %loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    ldr q4, [x0]
-; CHECK-NEXT:    ldp q17, q7, [x1, #32]
+; CHECK-NEXT:    ldp q17, q16, [x1, #32]
 ; CHECK-NEXT:    ldr q18, [x1]
 ; CHECK-NEXT:    subs w2, w2, #1
 ; CHECK-NEXT:    tbl.16b v5, { v4 }, v3
 ; CHECK-NEXT:    tbl.16b v6, { v4 }, v0
-; CHECK-NEXT:    tbl.16b v16, { v4 }, v1
+; CHECK-NEXT:    tbl.16b v7, { v4 }, v1
 ; CHECK-NEXT:    tbl.16b v4, { v4 }, v2
 ; CHECK-NEXT:    ldr q21, [x8, #16]!
 ; CHECK-NEXT:    mov x1, x8
-; CHECK-NEXT:    umull2.2d v19, v5, v7
+; CHECK-NEXT:    umull2.2d v19, v5, v16
 ; CHECK-NEXT:    umull2.2d v20, v6, v17
-; CHECK-NEXT:    umull2.2d v22, v16, v18
-; CHECK-NEXT:    umull.2d v5, v5, v7
-; CHECK-NEXT:    umull2.2d v7, v4, v21
+; CHECK-NEXT:    umull2.2d v22, v7, v18
+; CHECK-NEXT:    umull.2d v5, v5, v16
+; CHECK-NEXT:    umull2.2d v16, v4, v21
 ; CHECK-NEXT:    umull.2d v4, v4, v21
-; CHECK-NEXT:    umull.2d v16, v16, v18
+; CHECK-NEXT:    umull.2d v7, v7, v18
 ; CHECK-NEXT:    umull.2d v6, v6, v17
 ; CHECK-NEXT:    str q20, [x0, #80]
 ; CHECK-NEXT:    stp q22, q4, [x0, #16]
 ; CHECK-NEXT:    stp q5, q19, [x0, #96]
-; CHECK-NEXT:    str q7, [x0, #48]
-; CHECK-NEXT:    str q16, [x0]
+; CHECK-NEXT:    str q16, [x0, #48]
+; CHECK-NEXT:    str q7, [x0]
 ; CHECK-NEXT:    str q6, [x0, #64]!
 ; CHECK-NEXT:    b.ne LBB25_1
 ; CHECK-NEXT:  ; %bb.2: ; %exit
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/add.vni16.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/add.vni16.ll
index 27b93872b9f1d..b67080bd4798d 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/add.vni16.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/add.vni16.ll
@@ -40,17 +40,17 @@ define void @add_v3i16(ptr addrspace(1) %ptra, ptr addrspace(1) %ptrb, ptr addrs
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    global_load_ushort v6, v[0:1], off
-; GFX9-NEXT:    global_load_ushort v7, v[0:1], off offset:4
-; GFX9-NEXT:    global_load_ushort v8, v[2:3], off
-; GFX9-NEXT:    global_load_ushort v9, v[2:3], off offset:4
+; GFX9-NEXT:    global_load_ushort v7, v[2:3], off
+; GFX9-NEXT:    global_load_ushort v8, v[2:3], off offset:4
+; GFX9-NEXT:    global_load_ushort v9, v[0:1], off offset:4
 ; GFX9-NEXT:    global_load_ushort v10, v[0:1], off offset:2
 ; GFX9-NEXT:    global_load_ushort v11, v[2:3], off offset:2
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v6
-; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v8
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v7
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_pk_add_u16 v2, v7, v9
+; GFX9-NEXT:    v_pk_add_u16 v2, v9, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshl_or_b32 v0, v10, 16, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -206,10 +206,10 @@ define void @add_v5i16(ptr addrspace(1) %ptra, ptr addrspace(1) %ptrb, ptr addrs
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    global_load_ushort v6, v[0:1], off
 ; GFX9-NEXT:    global_load_ushort v7, v[0:1], off offset:4
-; GFX9-NEXT:    global_load_ushort v8, v[0:1], off offset:8
-; GFX9-NEXT:    global_load_ushort v9, v[2:3], off
-; GFX9-NEXT:    global_load_ushort v10, v[2:3], off offset:4
-; GFX9-NEXT:    global_load_ushort v11, v[2:3], off offset:8
+; GFX9-NEXT:    global_load_ushort v8, v[2:3], off
+; GFX9-NEXT:    global_load_ushort v9, v[2:3], off offset:4
+; GFX9-NEXT:    global_load_ushort v10, v[2:3], off offset:8
+; GFX9-NEXT:    global_load_ushort v11, v[0:1], off offset:8
 ; GFX9-NEXT:    global_load_ushort v12, v[0:1], off offset:2
 ; GFX9-NEXT:    global_load_ushort v13, v[0:1], off offset:6
 ; GFX9-NEXT:    global_load_ushort v14, v[2:3], off offset:2
@@ -218,12 +218,12 @@ define void @add_v5i16(ptr addrspace(1) %ptra, ptr addrspace(1) %ptrb, ptr addrs
 ; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v6
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
 ; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v7
+; GFX9-NEXT:    s_waitcnt vmcnt(7)
+; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v9
-; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v10
+; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v9
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_pk_add_u16 v6, v8, v11
+; GFX9-NEXT:    v_pk_add_u16 v6, v11, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_lshl_or_b32 v0, v12, 16, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
@@ -421,11 +421,11 @@ define void @addv_7i16(ptr addrspace(1) %ptra, ptr addrspace(1) %ptrb, ptr addrs
 ; GFX9-NEXT:    global_load_ushort v6, v[0:1], off
 ; GFX9-NEXT:    global_load_ushort v7, v[0:1], off offset:4
 ; GFX9-NEXT:    global_load_ushort v8, v[0:1], off offset:8
-; GFX9-NEXT:    global_load_ushort v9, v[0:1], off offset:12
-; GFX9-NEXT:    global_load_ushort v10, v[2:3], off
-; GFX9-NEXT:    global_load_ushort v11, v[2:3], off offset:4
-; GFX9-NEXT:    global_load_ushort v12, v[2:3], off offset:8
-; GFX9-NEXT:    global_load_ushort v13, v[2:3], off offset:12
+; GFX9-NEXT:    global_load_ushort v9, v[2:3], off
+; GFX9-NEXT:    global_load_ushort v10, v[2:3], off offset:4
+; GFX9-NEXT:    global_load_ushort v11, v[2:3], off offset:8
+; GFX9-NEXT:    global_load_ushort v12, v[2:3], off offset:12
+; GFX9-NEXT:    global_load_ushort v13, v[0:1], off offset:12
 ; GFX9-NEXT:    global_load_ushort v14, v[0:1], off offset:2
 ; GFX9-NEXT:    global_load_ushort v15, v[0:1], off offset:6
 ; GFX9-NEXT:    global_load_ushort v16, v[0:1], off offset:10
@@ -438,14 +438,14 @@ define void @addv_7i16(ptr addrspace(1) %ptra, ptr addrspace(1) %ptrb, ptr addrs
 ; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v7
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
 ; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v8
+; GFX9-NEXT:    s_waitcnt vmcnt(10)
+; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v9
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v10
+; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v11
-; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff, v12
+; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff, v11
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_pk_add_u16 v8, v9, v13
+; GFX9-NEXT:    v_pk_add_u16 v8, v13, v12
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_lshl_or_b32 v0, v14, 16, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
@@ -720,8 +720,8 @@ define void @add_v11i16(ptr addrspace(1) %ptra, ptr addrspace(1) %ptrb, ptr addr
 ; GFX9-NEXT:    global_load_ushort v14, v[0:1], off offset:16
 ; GFX9-NEXT:    global_load_ushort v15, v[2:3], off offset:16
 ; GFX9-NEXT:    global_load_dwordx4 v[10:13], v[2:3], off
-; GFX9-NEXT:    global_load_ushort v16, v[0:1], off offset:20
-; GFX9-NEXT:    global_load_ushort v17, v[2:3], off offset:20
+; GFX9-NEXT:    global_load_ushort v16, v[2:3], off offset:20
+; GFX9-NEXT:    global_load_ushort v17, v[0:1], off offset:20
 ; GFX9-NEXT:    global_load_ushort v18, v[0:1], off offset:18
 ; GFX9-NEXT:    global_load_ushort v19, v[2:3], off offset:18
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
@@ -738,7 +738,7 @@ define void @add_v11i16(ptr addrspace(1) %ptra, ptr addrspace(1) %ptrb, ptr addr
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_lshl_or_b32 v8, v19, 16, v15
 ; GFX9-NEXT:    global_store_dwordx4 v[4:5], v[0:3], off
-; GFX9-NEXT:    v_pk_add_u16 v6, v16, v17
+; GFX9-NEXT:    v_pk_add_u16 v6, v17, v16
 ; GFX9-NEXT:    v_pk_add_u16 v0, v7, v8
 ; GFX9-NEXT:    global_store_short v[4:5], v0, off offset:16
 ; GFX9-NEXT:    global_store_short_d16_hi v[4:5], v0, off offset:18
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
index 44abfd272be88..37db1c56022e8 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
@@ -3121,8 +3121,8 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; kill: killed $vgpr36
@@ -3284,7 +3284,7 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; kill: killed $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -3522,7 +3522,6 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB12_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
@@ -4334,8 +4333,8 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; VI-NEXT:    ; implicit-def: $vgpr39
 ; VI-NEXT:    ; kill: killed $vgpr39
@@ -4990,13 +4989,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
@@ -5016,13 +5016,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
@@ -5042,13 +5043,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
@@ -5068,13 +5070,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
@@ -5094,13 +5097,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
@@ -5119,13 +5123,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
@@ -5145,13 +5150,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
@@ -5171,13 +5177,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
@@ -5195,13 +5202,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
@@ -5221,13 +5229,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
@@ -5293,8 +5302,8 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX9-NEXT:    ; implicit-def: $vgpr40
 ; GFX9-NEXT:    ; kill: killed $vgpr40
@@ -5484,7 +5493,7 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(30)
+; GFX9-NEXT:    s_waitcnt vmcnt(29)
 ; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    ; kill: killed $vgpr33
@@ -5493,7 +5502,6 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB12_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
-; GFX9-NEXT:    s_waitcnt vmcnt(29)
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v32
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
@@ -5689,7 +5697,6 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB12_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
-; GFX9-NEXT:    s_waitcnt vmcnt(29)
 ; GFX9-NEXT:    v_add_u32_e32 v32, 3, v32
 ; GFX9-NEXT:    s_waitcnt vmcnt(28)
 ; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
@@ -6106,13 +6113,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
@@ -6131,13 +6139,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
@@ -6156,13 +6165,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
@@ -6181,13 +6191,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
@@ -6206,13 +6217,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
@@ -6231,13 +6243,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
@@ -6256,13 +6269,14 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
@@ -12441,10 +12455,10 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:388
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
@@ -12464,33 +12478,35 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:136
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:144
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:152
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:160
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:168
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:176
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:184
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:124
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v1
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:136
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:144
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:160
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:168
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:176
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:184
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 8, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 24, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 8, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 24, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v29
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 24, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 24, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v6
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -12511,10 +12527,10 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v18
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v20
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v22
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -12525,69 +12541,69 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
-; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:132
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:140
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v33
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v36
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v35
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v36
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v39
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v55
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v40
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v50
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v44
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v51
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v47
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
@@ -12602,14 +12618,14 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:196
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -12631,16 +12647,16 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -12659,19 +12675,19 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -12690,10 +12706,10 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -12762,7 +12778,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:356
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v3
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -12780,121 +12796,119 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:92
 ; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v54, v9
 ; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_or_b32_e32 v0, v0, v43
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v0, v0, v54
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v41
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v1, v42, v1
-; SI-NEXT:    v_or_b32_e32 v3, v53, v3
+; SI-NEXT:    v_or_b32_e32 v3, v40, v3
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v3
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v9, v10, v9
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v32, v4
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
+; SI-NEXT:    v_or_b32_e32 v5, v60, v5
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v6, v51, v6
+; SI-NEXT:    v_or_b32_e32 v6, v58, v6
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr53
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v2, v2, v35
+; SI-NEXT:    v_or_b32_e32 v2, v2, v53
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v34, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v3
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr53
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    v_or_b32_e32 v8, v55, v8
+; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v48
@@ -12904,11 +12918,11 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v4, v61
+; SI-NEXT:    v_or_b32_e32 v4, v4, v62
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v38
@@ -12918,18 +12932,18 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v6, v58
+; SI-NEXT:    v_or_b32_e32 v6, v6, v56
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v46
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v40
-; SI-NEXT:    v_or_b32_e32 v8, v8, v50
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v35
+; SI-NEXT:    v_or_b32_e32 v8, v8, v44
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v33
@@ -12940,24 +12954,24 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v10
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v62
-; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v63
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr52
-; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v63
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v60
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v59
@@ -12966,86 +12980,85 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v57
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v51
 ; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v56
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v57
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v47
-; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr47
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v55
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v50
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v44
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v45
+; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v36
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v18
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v18
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v19
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
@@ -13058,7 +13071,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v20
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
@@ -13079,7 +13092,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v21
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13092,7 +13105,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v22
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13105,7 +13118,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    v_or_b32_e32 v22, v22, v23
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13118,7 +13131,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    v_or_b32_e32 v23, v23, v24
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13131,7 +13144,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v25
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13144,7 +13157,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13214,7 +13227,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; SI-NEXT:    v_or_b32_e32 v31, v31, v36
+; SI-NEXT:    v_or_b32_e32 v31, v31, v43
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; SI-NEXT:    v_or_b32_e32 v32, v33, v32
@@ -13228,249 +13241,249 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:  .LBB14_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    s_movk_i32 s6, 0x300
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v54, v9
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_or_b32_e32 v0, v43, v0
+; SI-NEXT:    v_or_b32_e32 v0, v54, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v42, v1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v53, v2
+; SI-NEXT:    v_or_b32_e32 v2, v40, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
@@ -13479,16 +13492,18 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v32, v4
+; SI-NEXT:    v_or_b32_e32 v9, v10, v9
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
-; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_or_b32_e32 v5, v60, v5
+; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v6, v51, v6
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_or_b32_e32 v6, v58, v6
+; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
@@ -13496,7 +13511,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v41, v1
@@ -13509,11 +13524,11 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
+; SI-NEXT:    v_or_b32_e32 v8, v55, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v2, v35, v2
+; SI-NEXT:    v_or_b32_e32 v2, v53, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
@@ -13531,7 +13546,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v61, v4
+; SI-NEXT:    v_or_b32_e32 v4, v62, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
@@ -13549,7 +13564,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v58, v6
+; SI-NEXT:    v_or_b32_e32 v6, v56, v6
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
@@ -13562,9 +13577,9 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v40
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v35
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_or_b32_e32 v8, v50, v8
+; SI-NEXT:    v_or_b32_e32 v8, v44, v8
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
@@ -13579,7 +13594,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v62
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v63
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
@@ -13587,7 +13602,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v63
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v39
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
@@ -13595,7 +13610,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v60
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13608,12 +13623,12 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v57
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v51
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v56
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
@@ -13626,7 +13641,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v55
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v50
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
@@ -13634,19 +13649,17 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v44
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v36
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
@@ -13656,7 +13669,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13665,7 +13678,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v15
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -13673,7 +13686,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13682,7 +13695,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -13690,7 +13703,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13699,7 +13712,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v17
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -13707,7 +13720,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13724,7 +13737,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v19, v20, v19
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -13751,7 +13764,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    v_or_b32_e32 v20, v21, v20
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
@@ -13768,7 +13781,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v21
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
@@ -13785,7 +13798,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
@@ -13802,7 +13815,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v23
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
@@ -13819,7 +13832,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
@@ -13836,7 +13849,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v25
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
@@ -13928,7 +13941,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
-; SI-NEXT:    v_or_b32_e32 v31, v36, v31
+; SI-NEXT:    v_or_b32_e32 v31, v43, v31
 ; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x300, v31
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
@@ -14009,19 +14022,18 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:128
+; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
 ; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
+; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:152
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:160
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:168
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:176
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:184
 ; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v7
+; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v5
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v7
 ; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v9
 ; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
 ; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v13
@@ -14053,59 +14065,61 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v26
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v54
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v32
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v40
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v42
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v57
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
@@ -14114,25 +14128,25 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:204
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:212
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:232
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:240
@@ -14140,25 +14154,25 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:220
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -14166,25 +14180,25 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:252
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -14192,10 +14206,10 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
@@ -14207,10 +14221,10 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:308
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:336
@@ -14221,22 +14235,22 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:340
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:352
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:368
@@ -14247,14 +14261,14 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
@@ -14266,34 +14280,33 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
 ; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:108
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB14_2
 ; VI-NEXT:  ; %bb.1: ; %cmp.false
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -14301,19 +14314,19 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
@@ -14324,23 +14337,23 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr61
-; VI-NEXT:    ; implicit-def: $vgpr62
 ; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    ; implicit-def: $vgpr56
-; VI-NEXT:    ; implicit-def: $vgpr47
+; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr46
 ; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    ; implicit-def: $vgpr41
 ; VI-NEXT:    ; implicit-def: $vgpr55
+; VI-NEXT:    ; implicit-def: $vgpr54
 ; VI-NEXT:    ; implicit-def: $vgpr50
 ; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr38
@@ -14381,147 +14394,217 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; kill: killed $vgpr33
 ; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr33
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr46
+; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr45
-; VI-NEXT:    ; implicit-def: $vgpr54
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -14531,16 +14614,16 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
@@ -14549,7 +14632,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
@@ -14558,253 +14641,190 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr33
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:  .LBB14_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB14_4
 ; VI-NEXT:  ; %bb.3: ; %cmp.true
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(13)
+; VI-NEXT:    v_add_u16_e32 v9, 3, v61
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v31, 0x300
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_add_u16_e32 v9, 3, v61
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    v_add_u16_e32 v8, 3, v8
+; VI-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v1, 3, v1
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v1, 3, v1
+; VI-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; VI-NEXT:    v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
@@ -14813,31 +14833,25 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v1, v2, v3
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
 ; VI-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v5, 3, v5
 ; VI-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(13)
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
 ; VI-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(11)
+; VI-NEXT:    s_waitcnt vmcnt(10)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
 ; VI-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_add_u16_e32 v8, 3, v8
-; VI-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(3)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
-; VI-NEXT:    s_waitcnt vmcnt(2)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -14874,13 +14888,13 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v7, v7, v8
-; VI-NEXT:    v_add_u16_e32 v8, 3, v63
+; VI-NEXT:    v_add_u16_e32 v8, 3, v62
 ; VI-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v8, v8, v9
-; VI-NEXT:    v_add_u16_e32 v9, 3, v59
+; VI-NEXT:    v_add_u16_e32 v9, 3, v32
 ; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v10, 3, v62
+; VI-NEXT:    v_add_u16_e32 v10, 3, v60
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
@@ -14888,27 +14902,28 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v9, v9, v10
 ; VI-NEXT:    v_add_u16_e32 v10, 3, v58
 ; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v11, 3, v60
+; VI-NEXT:    v_add_u16_e32 v11, 3, v56
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v10, v10, v11
-; VI-NEXT:    v_add_u16_e32 v11, 3, v57
+; VI-NEXT:    v_add_u16_e32 v11, 3, v59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v12, 3, v56
+; VI-NEXT:    v_add_u16_e32 v12, 3, v57
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v11, v11, v12
-; VI-NEXT:    v_add_u16_e32 v12, 3, v46
+; VI-NEXT:    v_add_u16_e32 v12, 3, v47
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v13, 3, v47
+; VI-NEXT:    v_add_u16_e32 v13, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
@@ -14919,148 +14934,149 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 3, v44
 ; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v13, v13, v14
-; VI-NEXT:    v_add_u16_e32 v14, 3, v54
+; VI-NEXT:    v_add_u16_e32 v14, 3, v42
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v15, 3, v40
-; VI-NEXT:    v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v14, v14, v15
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v15, 3, v15
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v15, 3, v15
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v15, v15, v16
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v16, v16, v17
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v18, 3, v18
 ; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v17, v17, v18
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v18, 3, v18
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v18, 0x300, v18
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v18, v18, v19
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v19, 0x300, v19
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v19, v19, v20
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v20, 0x300, v20
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
 ; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v20, v20, v21
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v21, v21, v22
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v23, 3, v23
 ; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v22, v22, v23
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v23, 3, v23
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v24, 3, v24
 ; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v23, v23, v24
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v24, 3, v24
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v25, 3, v25
 ; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v24, v24, v25
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v25, 3, v25
@@ -15074,19 +15090,19 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v26, 3, v26
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v27, 3, v27
 ; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v26, v26, v27
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -15100,20 +15116,20 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v27, v27, v28
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v28, 3, v28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v29, 3, v29
 ; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v28, v28, v29
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v29, 3, v29
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -15126,7 +15142,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v29, v29, v30
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v30, 3, v30
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -15139,7 +15155,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v30, v30, v32
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v32, 3, v32
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -15148,7 +15164,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v32, 0x300, v32
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v33, 3, v33
-; VI-NEXT:    v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v33, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v31, v32, v31
 ; VI-NEXT:  .LBB14_4: ; %end
@@ -15225,19 +15241,18 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
 ; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
+; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:152
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:160
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:168
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:176
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:184
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT:    v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v7
+; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v3
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v5
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v7
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v9
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v13
@@ -15249,93 +15264,95 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v25
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v35, 8, v29
-; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v0
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT:    s_nop 0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v54
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v32
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v40
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v42
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v57
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; GFX9-NEXT:    s_nop 0
@@ -15345,25 +15362,25 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:204
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:212
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:232
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:240
@@ -15372,25 +15389,25 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:220
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -15399,25 +15416,25 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:252
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -15426,10 +15443,10 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
@@ -15441,10 +15458,10 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:308
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:336
@@ -15456,22 +15473,22 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:340
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:352
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:368
@@ -15483,14 +15500,14 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
@@ -15502,34 +15519,33 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
 ; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB14_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -15537,20 +15553,20 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
@@ -15561,23 +15577,23 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr61
-; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr43
-; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    ; implicit-def: $vgpr41
 ; GFX9-NEXT:    ; implicit-def: $vgpr55
+; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr38
@@ -15618,147 +15634,217 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; kill: killed $vgpr33
 ; GFX9-NEXT:    ; implicit-def: $vgpr35
+; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr46
+; GFX9-NEXT:    ; implicit-def: $vgpr47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
-; GFX9-NEXT:    ; implicit-def: $vgpr54
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -15768,16 +15854,16 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
@@ -15786,7 +15872,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
@@ -15795,258 +15881,196 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr33
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:  .LBB14_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB14_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
+; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; GFX9-NEXT:    s_waitcnt vmcnt(13)
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_movk_i32 s6, 0x300
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(33)
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(30)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(28)
-; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
+; GFX9-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    s_waitcnt vmcnt(27)
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT:    v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(23)
+; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
+; GFX9-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; GFX9-NEXT:    v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
@@ -16055,32 +16079,26 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
 ; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
 ; GFX9-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
+; GFX9-NEXT:    s_waitcnt vmcnt(17)
 ; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(12)
+; GFX9-NEXT:    s_waitcnt vmcnt(11)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -16117,41 +16135,41 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v63
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v62
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v32
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v60
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
 ; GFX9-NEXT:    v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v60
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v56
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v57
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v56
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v57
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v46
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v46
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
@@ -16162,148 +16180,149 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 3, v44
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v54
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v40
-; GFX9-NEXT:    v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v17, v17, v18
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v18, v18, v19
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v19
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v19, v19, v20
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v20, v20, v21
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v21, v21, v22
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v22, v22, v23
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v23
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v23, v23, v24
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v24, v24, v25
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
@@ -16317,19 +16336,19 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v26, v26, v27
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -16343,20 +16362,20 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v27, v27, v28
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v28, v28, v29
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -16369,7 +16388,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v29, v29, v30
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -16382,7 +16401,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v30, v30, v31
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -16391,7 +16410,7 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v32, 3, v32
-; GFX9-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v32, v63, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v31, v31, v32
 ; GFX9-NEXT:  .LBB14_4: ; %end
@@ -19122,8 +19141,8 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    s_cbranch_vccnz .LBB15_5
 ; SI-NEXT:  ; %bb.4: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_add_i32 s28, s28, 3
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v40
@@ -19178,7 +19197,7 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
@@ -19300,12 +19319,12 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 0x3000000, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
@@ -20220,8 +20239,8 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; VI-NEXT:    s_cbranch_vccnz .LBB15_5
 ; VI-NEXT:  ; %bb.4: ; %cmp.true
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_add_i32 s28, s28, 3
 ; VI-NEXT:    s_and_b32 s4, s28, 0xff
 ; VI-NEXT:    s_lshl_b32 s5, s29, 8
@@ -20273,7 +20292,7 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; VI-NEXT:    s_add_i32 s6, s6, 0x3000000
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x300, v1
@@ -20368,11 +20387,11 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
 ; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
@@ -20495,11 +20514,11 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
 ; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
@@ -21284,11 +21303,11 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -21500,11 +21519,11 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -23572,9 +23591,9 @@ define <64 x bfloat> @bitcast_v32i32_to_v64bf16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -23607,7 +23626,7 @@ define <64 x bfloat> @bitcast_v32i32_to_v64bf16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; kill: killed $vgpr31
@@ -23726,7 +23745,6 @@ define <64 x bfloat> @bitcast_v32i32_to_v64bf16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v62
 ; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -23831,7 +23849,6 @@ define <64 x bfloat> @bitcast_v32i32_to_v64bf16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB16_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v62
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v63
@@ -25530,6 +25547,8 @@ define <32 x i32> @bitcast_v64bf16_to_v32i32(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; kill: killed $vgpr33
@@ -25606,8 +25625,6 @@ define <32 x i32> @bitcast_v64bf16_to_v32i32(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
@@ -31871,9 +31888,9 @@ define <64 x half> @bitcast_v32i32_to_v64f16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -31906,7 +31923,7 @@ define <64 x half> @bitcast_v32i32_to_v64f16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; kill: killed $vgpr31
@@ -31979,7 +31996,6 @@ define <64 x half> @bitcast_v32i32_to_v64f16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB20_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v62
 ; SI-NEXT:    v_cvt_f32_f16_e32 v32, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -35061,9 +35077,9 @@ define inreg <32 x i32> @bitcast_v64f16_to_v32i32_scalar(<64 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
@@ -35113,9 +35129,8 @@ define inreg <32 x i32> @bitcast_v64f16_to_v32i32_scalar(<64 x half> inreg %a, i
 ; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
@@ -35847,8 +35862,8 @@ define <64 x i16> @bitcast_v32i32_to_v64i16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -35882,7 +35897,7 @@ define <64 x i16> @bitcast_v32i32_to_v64i16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; kill: killed $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -35933,7 +35948,6 @@ define <64 x i16> @bitcast_v32i32_to_v64i16(<32 x i32> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB24_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
@@ -37049,26 +37063,25 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:132
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:20
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v5
@@ -37081,102 +37094,107 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v6
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v14
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v20
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v22
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v24
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v26
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v30
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v33
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v33
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:100
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v31
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:68
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:100
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:28
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
+; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB26_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v41
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
@@ -37193,9 +37211,7 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v50
-; SI-NEXT:    v_or_b32_e32 v23, v23, v32
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v32
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
@@ -37225,83 +37241,71 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v48
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v35
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    v_or_b32_e32 v16, v16, v53
+; SI-NEXT:    v_or_b32_e32 v17, v17, v51
+; SI-NEXT:    v_or_b32_e32 v18, v18, v50
 ; SI-NEXT:    ; kill: killed $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v43
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v55
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v53
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    v_or_b32_e32 v19, v19, v48
-; SI-NEXT:    v_or_b32_e32 v21, v21, v36
-; SI-NEXT:    v_or_b32_e32 v22, v22, v34
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v18, v18, v49
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_or_b32_e32 v22, v22, v23
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v54
+; SI-NEXT:    v_or_b32_e32 v19, v19, v39
+; SI-NEXT:    v_or_b32_e32 v23, v23, v24
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v52
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v25
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v39
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v49
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
-; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v35
+; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v37
 ; SI-NEXT:    v_or_b32_e32 v26, v26, v27
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v33
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v28
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v63
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v62
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v61
@@ -37314,13 +37318,15 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v46
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v45
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v44
-; SI-NEXT:    v_or_b32_e32 v12, v12, v42
-; SI-NEXT:    v_or_b32_e32 v13, v13, v41
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v42
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v40
-; SI-NEXT:    v_or_b32_e32 v15, v15, v54
-; SI-NEXT:    v_or_b32_e32 v20, v20, v38
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    v_or_b32_e32 v15, v15, v55
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
@@ -37334,63 +37340,79 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_or_b32_e32 v28, v28, v29
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v16, v16, v52
-; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v20, v20, v36
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    v_or_b32_e32 v29, v29, v30
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
 ; SI-NEXT:    v_or_b32_e32 v30, v30, v31
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; SI-NEXT:    v_or_b32_e32 v17, v17, v51
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v21, v21, v34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; SI-NEXT:    v_or_b32_e32 v31, v31, v37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    v_or_b32_e32 v31, v31, v38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:  .LBB26_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB26_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v41
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
@@ -37407,69 +37429,68 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v48
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v35
+; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v32
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT:    s_mov_b32 s6, 0x30000
+; SI-NEXT:    v_or_b32_e32 v16, v53, v16
+; SI-NEXT:    v_or_b32_e32 v17, v51, v17
+; SI-NEXT:    v_or_b32_e32 v18, v50, v18
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v43
-; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v55
-; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v53
-; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v50
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
+; SI-NEXT:    v_or_b32_e32 v22, v23, v22
+; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v54
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v19, v48, v19
-; SI-NEXT:    v_or_b32_e32 v21, v36, v21
-; SI-NEXT:    v_or_b32_e32 v22, v34, v22
-; SI-NEXT:    v_or_b32_e32 v23, v32, v23
-; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
-; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
-; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
-; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    v_or_b32_e32 v23, v24, v23
+; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v52
 ; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
-; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v39
+; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v49
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
-; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v35
+; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v37
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v33
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    v_or_b32_e32 v27, v28, v27
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -37486,8 +37507,7 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    v_or_b32_e32 v0, v63, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v62, v1
 ; SI-NEXT:    v_or_b32_e32 v2, v61, v2
@@ -37500,28 +37520,25 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v9, v46, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v45, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v44, v11
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_or_b32_e32 v13, v41, v13
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v42, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v40, v14
-; SI-NEXT:    v_or_b32_e32 v15, v54, v15
-; SI-NEXT:    v_or_b32_e32 v18, v49, v18
-; SI-NEXT:    v_or_b32_e32 v20, v38, v20
+; SI-NEXT:    v_or_b32_e32 v15, v55, v15
+; SI-NEXT:    v_or_b32_e32 v19, v39, v19
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
+; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_or_b32_e32 v28, v29, v28
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
-; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
-; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
@@ -37531,34 +37548,37 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
-; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
+; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
+; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
+; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v20, v36, v20
+; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    v_or_b32_e32 v29, v30, v29
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x30000, v29
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
 ; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
 ; SI-NEXT:    v_or_b32_e32 v30, v31, v30
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; SI-NEXT:    v_or_b32_e32 v17, v51, v17
-; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v21, v34, v21
+; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x30000, v30
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; SI-NEXT:    v_or_b32_e32 v31, v37, v31
+; SI-NEXT:    v_or_b32_e32 v31, v38, v31
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x30000, v31
 ; SI-NEXT:  .LBB26_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
@@ -37578,7 +37598,7 @@ define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: bitcast_v64i16_to_v32i32:
@@ -40846,8 +40866,8 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; kill: killed $vgpr36
@@ -41009,7 +41029,7 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; kill: killed $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -41247,7 +41267,6 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB36_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_f32_e32 v32, 1.0, v32
@@ -42059,8 +42078,8 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; VI-NEXT:    ; implicit-def: $vgpr39
 ; VI-NEXT:    ; kill: killed $vgpr39
@@ -42715,13 +42734,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
@@ -42741,13 +42761,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
@@ -42767,13 +42788,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
@@ -42793,13 +42815,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
@@ -42819,13 +42842,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
@@ -42844,13 +42868,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
@@ -42870,13 +42895,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
@@ -42896,13 +42922,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
@@ -42920,13 +42947,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
@@ -42946,13 +42974,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
@@ -43018,8 +43047,8 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX9-NEXT:    ; implicit-def: $vgpr40
 ; GFX9-NEXT:    ; kill: killed $vgpr40
@@ -43209,7 +43238,7 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(30)
+; GFX9-NEXT:    s_waitcnt vmcnt(29)
 ; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    ; kill: killed $vgpr33
@@ -43218,7 +43247,6 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB36_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
-; GFX9-NEXT:    s_waitcnt vmcnt(29)
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v32
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
@@ -43414,7 +43442,6 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB36_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
-; GFX9-NEXT:    s_waitcnt vmcnt(29)
 ; GFX9-NEXT:    v_add_f32_e32 v32, 1.0, v32
 ; GFX9-NEXT:    s_waitcnt vmcnt(28)
 ; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31
@@ -43831,13 +43858,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
@@ -43856,13 +43884,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
@@ -43881,13 +43910,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
@@ -43906,13 +43936,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
@@ -43931,13 +43962,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
@@ -43956,13 +43988,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
@@ -43981,13 +44014,14 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
@@ -51153,10 +51187,10 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:388
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
@@ -51176,33 +51210,35 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:136
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:144
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:152
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:160
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:168
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:176
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:184
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:124
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v1
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:136
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:144
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:160
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:168
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:176
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:184
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 8, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 24, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 8, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 24, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v29
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 24, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 24, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v6
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -51223,10 +51259,10 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v18
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v20
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v22
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -51237,69 +51273,69 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
-; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:132
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:140
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v33
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v36
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v35
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v36
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v39
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v55
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v40
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v50
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v44
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v51
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v47
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
@@ -51314,14 +51350,14 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:196
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -51343,16 +51379,16 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -51371,19 +51407,19 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -51402,10 +51438,10 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -51474,7 +51510,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:356
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v3
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -51492,121 +51528,119 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:92
 ; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB38_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v54, v9
 ; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_or_b32_e32 v0, v0, v43
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v0, v0, v54
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v41
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v1, v42, v1
-; SI-NEXT:    v_or_b32_e32 v3, v53, v3
+; SI-NEXT:    v_or_b32_e32 v3, v40, v3
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v3
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v9, v10, v9
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v32, v4
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
+; SI-NEXT:    v_or_b32_e32 v5, v60, v5
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v6, v51, v6
+; SI-NEXT:    v_or_b32_e32 v6, v58, v6
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr53
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v2, v2, v35
+; SI-NEXT:    v_or_b32_e32 v2, v2, v53
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v34, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v3
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr53
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    v_or_b32_e32 v8, v55, v8
+; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v48
@@ -51616,11 +51650,11 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v4, v61
+; SI-NEXT:    v_or_b32_e32 v4, v4, v62
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v38
@@ -51630,18 +51664,18 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v6, v58
+; SI-NEXT:    v_or_b32_e32 v6, v6, v56
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v46
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v40
-; SI-NEXT:    v_or_b32_e32 v8, v8, v50
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v35
+; SI-NEXT:    v_or_b32_e32 v8, v8, v44
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v33
@@ -51652,24 +51686,24 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v10
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v62
-; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v63
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr52
-; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v63
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v60
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v59
@@ -51678,86 +51712,85 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v57
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v51
 ; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v56
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v57
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v47
-; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr47
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v55
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v50
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v44
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v45
+; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v36
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v18
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v18
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v19
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
@@ -51770,7 +51803,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v20
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
@@ -51791,7 +51824,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v21
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -51804,7 +51837,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v22
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -51817,7 +51850,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    v_or_b32_e32 v22, v22, v23
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -51830,7 +51863,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    v_or_b32_e32 v23, v23, v24
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -51843,7 +51876,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v25
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -51856,7 +51889,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -51926,7 +51959,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; SI-NEXT:    v_or_b32_e32 v31, v31, v36
+; SI-NEXT:    v_or_b32_e32 v31, v31, v43
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; SI-NEXT:    v_or_b32_e32 v32, v33, v32
@@ -51940,249 +51973,249 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:  .LBB38_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB38_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    s_movk_i32 s6, 0x300
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v54, v9
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_or_b32_e32 v0, v43, v0
+; SI-NEXT:    v_or_b32_e32 v0, v54, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v42, v1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v53, v2
+; SI-NEXT:    v_or_b32_e32 v2, v40, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
@@ -52191,16 +52224,18 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v32, v4
+; SI-NEXT:    v_or_b32_e32 v9, v10, v9
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
-; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_or_b32_e32 v5, v60, v5
+; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v6, v51, v6
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_or_b32_e32 v6, v58, v6
+; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
@@ -52208,7 +52243,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v41, v1
@@ -52221,11 +52256,11 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
+; SI-NEXT:    v_or_b32_e32 v8, v55, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v2, v35, v2
+; SI-NEXT:    v_or_b32_e32 v2, v53, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
@@ -52243,7 +52278,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v61, v4
+; SI-NEXT:    v_or_b32_e32 v4, v62, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
@@ -52261,7 +52296,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v58, v6
+; SI-NEXT:    v_or_b32_e32 v6, v56, v6
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
@@ -52274,9 +52309,9 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v40
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v35
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_or_b32_e32 v8, v50, v8
+; SI-NEXT:    v_or_b32_e32 v8, v44, v8
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
@@ -52291,7 +52326,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v62
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v63
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
@@ -52299,7 +52334,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v63
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v39
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
@@ -52307,7 +52342,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v60
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -52320,12 +52355,12 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v57
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v51
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v56
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
@@ -52338,7 +52373,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v55
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v50
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
@@ -52346,19 +52381,17 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v44
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v36
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
@@ -52368,7 +52401,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -52377,7 +52410,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v15
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -52385,7 +52418,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -52394,7 +52427,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -52402,7 +52435,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -52411,7 +52444,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v17
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -52419,7 +52452,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -52436,7 +52469,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v19, v20, v19
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -52463,7 +52496,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    v_or_b32_e32 v20, v21, v20
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
@@ -52480,7 +52513,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v21
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
@@ -52497,7 +52530,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
@@ -52514,7 +52547,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v23
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
@@ -52531,7 +52564,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
@@ -52548,7 +52581,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v25
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
@@ -52640,7 +52673,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
-; SI-NEXT:    v_or_b32_e32 v31, v36, v31
+; SI-NEXT:    v_or_b32_e32 v31, v43, v31
 ; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x300, v31
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
@@ -52721,19 +52754,18 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:128
+; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
 ; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
+; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:152
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:160
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:168
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:176
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:184
 ; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v7
+; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v5
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v7
 ; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v9
 ; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
 ; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v13
@@ -52765,59 +52797,61 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v26
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v54
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v32
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v40
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v42
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v57
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
@@ -52826,25 +52860,25 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:204
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:212
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:232
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:240
@@ -52852,25 +52886,25 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:220
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -52878,25 +52912,25 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:252
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -52904,10 +52938,10 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
@@ -52919,10 +52953,10 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:308
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:336
@@ -52933,22 +52967,22 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:340
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:352
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:368
@@ -52959,14 +52993,14 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
@@ -52978,34 +53012,33 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
 ; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:108
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB38_2
 ; VI-NEXT:  ; %bb.1: ; %cmp.false
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -53013,19 +53046,19 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
@@ -53036,23 +53069,23 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr61
-; VI-NEXT:    ; implicit-def: $vgpr62
 ; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    ; implicit-def: $vgpr56
-; VI-NEXT:    ; implicit-def: $vgpr47
+; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr46
 ; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    ; implicit-def: $vgpr41
 ; VI-NEXT:    ; implicit-def: $vgpr55
+; VI-NEXT:    ; implicit-def: $vgpr54
 ; VI-NEXT:    ; implicit-def: $vgpr50
 ; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr38
@@ -53093,147 +53126,217 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; kill: killed $vgpr33
 ; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr33
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr46
+; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr45
-; VI-NEXT:    ; implicit-def: $vgpr54
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -53243,16 +53346,16 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
@@ -53261,7 +53364,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
@@ -53270,253 +53373,190 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr33
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:  .LBB38_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB38_4
 ; VI-NEXT:  ; %bb.3: ; %cmp.true
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(13)
+; VI-NEXT:    v_add_u16_e32 v9, 3, v61
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v31, 0x300
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_add_u16_e32 v9, 3, v61
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    v_add_u16_e32 v8, 3, v8
+; VI-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v1, 3, v1
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v1, 3, v1
+; VI-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; VI-NEXT:    v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
@@ -53525,31 +53565,25 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v1, v2, v3
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
 ; VI-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v5, 3, v5
 ; VI-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(13)
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
 ; VI-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(11)
+; VI-NEXT:    s_waitcnt vmcnt(10)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
 ; VI-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_add_u16_e32 v8, 3, v8
-; VI-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(3)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
-; VI-NEXT:    s_waitcnt vmcnt(2)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -53586,13 +53620,13 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v7, v7, v8
-; VI-NEXT:    v_add_u16_e32 v8, 3, v63
+; VI-NEXT:    v_add_u16_e32 v8, 3, v62
 ; VI-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v8, v8, v9
-; VI-NEXT:    v_add_u16_e32 v9, 3, v59
+; VI-NEXT:    v_add_u16_e32 v9, 3, v32
 ; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v10, 3, v62
+; VI-NEXT:    v_add_u16_e32 v10, 3, v60
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
@@ -53600,27 +53634,28 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v9, v9, v10
 ; VI-NEXT:    v_add_u16_e32 v10, 3, v58
 ; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v11, 3, v60
+; VI-NEXT:    v_add_u16_e32 v11, 3, v56
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v10, v10, v11
-; VI-NEXT:    v_add_u16_e32 v11, 3, v57
+; VI-NEXT:    v_add_u16_e32 v11, 3, v59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v12, 3, v56
+; VI-NEXT:    v_add_u16_e32 v12, 3, v57
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v11, v11, v12
-; VI-NEXT:    v_add_u16_e32 v12, 3, v46
+; VI-NEXT:    v_add_u16_e32 v12, 3, v47
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v13, 3, v47
+; VI-NEXT:    v_add_u16_e32 v13, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
@@ -53631,148 +53666,149 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 3, v44
 ; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v13, v13, v14
-; VI-NEXT:    v_add_u16_e32 v14, 3, v54
+; VI-NEXT:    v_add_u16_e32 v14, 3, v42
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v15, 3, v40
-; VI-NEXT:    v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v14, v14, v15
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v15, 3, v15
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v15, 3, v15
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v15, v15, v16
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v16, v16, v17
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v18, 3, v18
 ; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v17, v17, v18
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v18, 3, v18
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v18, 0x300, v18
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v18, v18, v19
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v19, 0x300, v19
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v19, v19, v20
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v20, 0x300, v20
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
 ; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v20, v20, v21
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v21, v21, v22
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v23, 3, v23
 ; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v22, v22, v23
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v23, 3, v23
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v24, 3, v24
 ; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v23, v23, v24
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v24, 3, v24
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v25, 3, v25
 ; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v24, v24, v25
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v25, 3, v25
@@ -53786,19 +53822,19 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v26, 3, v26
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v27, 3, v27
 ; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v26, v26, v27
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -53812,20 +53848,20 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v27, v27, v28
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v28, 3, v28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v29, 3, v29
 ; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v28, v28, v29
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v29, 3, v29
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -53838,7 +53874,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v29, v29, v30
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v30, 3, v30
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -53851,7 +53887,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v30, v30, v32
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v32, 3, v32
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -53860,7 +53896,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v32, 0x300, v32
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v33, 3, v33
-; VI-NEXT:    v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v33, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v31, v32, v31
 ; VI-NEXT:  .LBB38_4: ; %end
@@ -53937,19 +53973,18 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
 ; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
+; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:152
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:160
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:168
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:176
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:184
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT:    v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v7
+; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v3
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v5
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v7
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v9
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v13
@@ -53961,93 +53996,95 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v25
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v35, 8, v29
-; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v0
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT:    s_nop 0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v54
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v32
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v40
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v42
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v57
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; GFX9-NEXT:    s_nop 0
@@ -54057,25 +54094,25 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:204
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:212
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:232
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:240
@@ -54084,25 +54121,25 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:220
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -54111,25 +54148,25 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:252
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -54138,10 +54175,10 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
@@ -54153,10 +54190,10 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:308
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:336
@@ -54168,22 +54205,22 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:340
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:352
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:368
@@ -54195,14 +54232,14 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
@@ -54214,34 +54251,33 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
 ; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB38_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -54249,20 +54285,20 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
@@ -54273,23 +54309,23 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr61
-; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr43
-; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    ; implicit-def: $vgpr41
 ; GFX9-NEXT:    ; implicit-def: $vgpr55
+; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr38
@@ -54330,147 +54366,217 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; kill: killed $vgpr33
 ; GFX9-NEXT:    ; implicit-def: $vgpr35
+; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr46
+; GFX9-NEXT:    ; implicit-def: $vgpr47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
-; GFX9-NEXT:    ; implicit-def: $vgpr54
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -54480,16 +54586,16 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
@@ -54498,7 +54604,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
@@ -54507,258 +54613,196 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr33
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:  .LBB38_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB38_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
+; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; GFX9-NEXT:    s_waitcnt vmcnt(13)
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_movk_i32 s6, 0x300
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(33)
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(30)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(28)
-; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
+; GFX9-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    s_waitcnt vmcnt(27)
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT:    v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(23)
+; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
+; GFX9-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; GFX9-NEXT:    v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
@@ -54767,32 +54811,26 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
 ; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
 ; GFX9-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
+; GFX9-NEXT:    s_waitcnt vmcnt(17)
 ; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(12)
+; GFX9-NEXT:    s_waitcnt vmcnt(11)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -54829,41 +54867,41 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v63
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v62
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v32
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v60
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
 ; GFX9-NEXT:    v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v60
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v56
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v57
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v56
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v57
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v46
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v46
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
@@ -54874,148 +54912,149 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 3, v44
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v54
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v40
-; GFX9-NEXT:    v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v17, v17, v18
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v18, v18, v19
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v19
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v19, v19, v20
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v20, v20, v21
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v21, v21, v22
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v22, v22, v23
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v23
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v23, v23, v24
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v24, v24, v25
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
@@ -55029,19 +55068,19 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v26, v26, v27
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -55055,20 +55094,20 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v27, v27, v28
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v28, v28, v29
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -55081,7 +55120,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v29, v29, v30
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -55094,7 +55133,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v30, v30, v31
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -55103,7 +55142,7 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v32, 3, v32
-; GFX9-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v32, v63, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v31, v31, v32
 ; GFX9-NEXT:  .LBB38_4: ; %end
@@ -57834,8 +57873,8 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; SI-NEXT:    s_cbranch_vccnz .LBB39_5
 ; SI-NEXT:  ; %bb.4: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_add_i32 s28, s28, 3
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v40
@@ -57890,7 +57929,7 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
@@ -58012,12 +58051,12 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 0x3000000, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
@@ -58932,8 +58971,8 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; VI-NEXT:    s_cbranch_vccnz .LBB39_5
 ; VI-NEXT:  ; %bb.4: ; %cmp.true
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_add_i32 s28, s28, 3
 ; VI-NEXT:    s_and_b32 s4, s28, 0xff
 ; VI-NEXT:    s_lshl_b32 s5, s29, 8
@@ -58985,7 +59024,7 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; VI-NEXT:    s_add_i32 s6, s6, 0x3000000
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x300, v1
@@ -59080,11 +59119,11 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
 ; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
@@ -59207,11 +59246,11 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
 ; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
@@ -59996,11 +60035,11 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -60212,11 +60251,11 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -62284,9 +62323,9 @@ define <64 x bfloat> @bitcast_v32f32_to_v64bf16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -62319,7 +62358,7 @@ define <64 x bfloat> @bitcast_v32f32_to_v64bf16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; kill: killed $vgpr31
@@ -62438,7 +62477,6 @@ define <64 x bfloat> @bitcast_v32f32_to_v64bf16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v62
 ; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -62543,7 +62581,6 @@ define <64 x bfloat> @bitcast_v32f32_to_v64bf16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB40_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_f32_e32 v32, 1.0, v62
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_f32_e32 v31, 1.0, v63
@@ -64288,6 +64325,8 @@ define <32 x float> @bitcast_v64bf16_to_v32f32(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; kill: killed $vgpr33
@@ -64364,8 +64403,6 @@ define <32 x float> @bitcast_v64bf16_to_v32f32(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
@@ -70629,9 +70666,9 @@ define <64 x half> @bitcast_v32f32_to_v64f16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -70664,7 +70701,7 @@ define <64 x half> @bitcast_v32f32_to_v64f16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; kill: killed $vgpr31
@@ -70737,7 +70774,6 @@ define <64 x half> @bitcast_v32f32_to_v64f16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB44_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v62
 ; SI-NEXT:    v_cvt_f32_f16_e32 v32, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -73792,9 +73828,9 @@ define inreg <32 x float> @bitcast_v64f16_to_v32f32_scalar(<64 x half> inreg %a,
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
@@ -73844,9 +73880,8 @@ define inreg <32 x float> @bitcast_v64f16_to_v32f32_scalar(<64 x half> inreg %a,
 ; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
@@ -74578,8 +74613,8 @@ define <64 x i16> @bitcast_v32f32_to_v64i16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -74613,7 +74648,7 @@ define <64 x i16> @bitcast_v32f32_to_v64i16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; kill: killed $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -74664,7 +74699,6 @@ define <64 x i16> @bitcast_v32f32_to_v64i16(<32 x float> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB48_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_f32_e32 v32, 1.0, v32
@@ -75760,26 +75794,25 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:132
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:20
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v5
@@ -75792,102 +75825,107 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v6
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v14
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v20
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v22
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v24
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v26
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v30
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v33
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v33
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:100
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v31
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:68
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:100
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:28
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
+; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v41
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
@@ -75904,9 +75942,7 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v50
-; SI-NEXT:    v_or_b32_e32 v23, v23, v32
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v32
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
@@ -75936,83 +75972,71 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v48
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v35
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    v_or_b32_e32 v16, v16, v53
+; SI-NEXT:    v_or_b32_e32 v17, v17, v51
+; SI-NEXT:    v_or_b32_e32 v18, v18, v50
 ; SI-NEXT:    ; kill: killed $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v43
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v55
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v53
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    v_or_b32_e32 v19, v19, v48
-; SI-NEXT:    v_or_b32_e32 v21, v21, v36
-; SI-NEXT:    v_or_b32_e32 v22, v22, v34
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v18, v18, v49
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_or_b32_e32 v22, v22, v23
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v54
+; SI-NEXT:    v_or_b32_e32 v19, v19, v39
+; SI-NEXT:    v_or_b32_e32 v23, v23, v24
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v52
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v25
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v39
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v49
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
-; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v35
+; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v37
 ; SI-NEXT:    v_or_b32_e32 v26, v26, v27
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v33
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v28
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v63
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v62
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v61
@@ -76025,13 +76049,15 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v46
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v45
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v44
-; SI-NEXT:    v_or_b32_e32 v12, v12, v42
-; SI-NEXT:    v_or_b32_e32 v13, v13, v41
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v42
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v40
-; SI-NEXT:    v_or_b32_e32 v15, v15, v54
-; SI-NEXT:    v_or_b32_e32 v20, v20, v38
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    v_or_b32_e32 v15, v15, v55
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
@@ -76045,63 +76071,79 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_or_b32_e32 v28, v28, v29
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v16, v16, v52
-; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v20, v20, v36
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    v_or_b32_e32 v29, v29, v30
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
 ; SI-NEXT:    v_or_b32_e32 v30, v30, v31
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; SI-NEXT:    v_or_b32_e32 v17, v17, v51
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v21, v21, v34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; SI-NEXT:    v_or_b32_e32 v31, v31, v37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    v_or_b32_e32 v31, v31, v38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:  .LBB50_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v41
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
@@ -76118,69 +76160,68 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v48
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v35
+; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v32
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT:    s_mov_b32 s6, 0x30000
+; SI-NEXT:    v_or_b32_e32 v16, v53, v16
+; SI-NEXT:    v_or_b32_e32 v17, v51, v17
+; SI-NEXT:    v_or_b32_e32 v18, v50, v18
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v43
-; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v55
-; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v53
-; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v50
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
+; SI-NEXT:    v_or_b32_e32 v22, v23, v22
+; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v54
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v19, v48, v19
-; SI-NEXT:    v_or_b32_e32 v21, v36, v21
-; SI-NEXT:    v_or_b32_e32 v22, v34, v22
-; SI-NEXT:    v_or_b32_e32 v23, v32, v23
-; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
-; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
-; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
-; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    v_or_b32_e32 v23, v24, v23
+; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v52
 ; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
-; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v39
+; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v49
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
-; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v35
+; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v37
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v33
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    v_or_b32_e32 v27, v28, v27
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -76197,8 +76238,7 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    v_or_b32_e32 v0, v63, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v62, v1
 ; SI-NEXT:    v_or_b32_e32 v2, v61, v2
@@ -76211,28 +76251,25 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v9, v46, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v45, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v44, v11
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_or_b32_e32 v13, v41, v13
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v42, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v40, v14
-; SI-NEXT:    v_or_b32_e32 v15, v54, v15
-; SI-NEXT:    v_or_b32_e32 v18, v49, v18
-; SI-NEXT:    v_or_b32_e32 v20, v38, v20
+; SI-NEXT:    v_or_b32_e32 v15, v55, v15
+; SI-NEXT:    v_or_b32_e32 v19, v39, v19
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
+; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_or_b32_e32 v28, v29, v28
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
-; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
-; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
@@ -76242,34 +76279,37 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
-; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
+; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
+; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
+; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v20, v36, v20
+; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    v_or_b32_e32 v29, v30, v29
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x30000, v29
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
 ; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
 ; SI-NEXT:    v_or_b32_e32 v30, v31, v30
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; SI-NEXT:    v_or_b32_e32 v17, v51, v17
-; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v21, v34, v21
+; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x30000, v30
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; SI-NEXT:    v_or_b32_e32 v31, v37, v31
+; SI-NEXT:    v_or_b32_e32 v31, v38, v31
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x30000, v31
 ; SI-NEXT:  .LBB50_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
@@ -76289,7 +76329,7 @@ define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: bitcast_v64i16_to_v32f32:
@@ -78543,8 +78583,8 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; kill: killed $vgpr36
@@ -78706,7 +78746,7 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; kill: killed $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -79755,8 +79795,8 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; VI-NEXT:    ; implicit-def: $vgpr39
 ; VI-NEXT:    ; kill: killed $vgpr39
@@ -80411,13 +80451,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
@@ -80437,13 +80478,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
@@ -80463,13 +80505,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
@@ -80489,13 +80532,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
@@ -80515,13 +80559,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
@@ -80540,13 +80585,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
@@ -80566,13 +80612,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
@@ -80592,13 +80639,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
@@ -80616,13 +80664,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
@@ -80642,13 +80691,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
@@ -80714,8 +80764,8 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX9-NEXT:    ; implicit-def: $vgpr40
 ; GFX9-NEXT:    ; kill: killed $vgpr40
@@ -80905,7 +80955,7 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(30)
+; GFX9-NEXT:    s_waitcnt vmcnt(29)
 ; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    ; kill: killed $vgpr33
@@ -80914,7 +80964,6 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB56_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
-; GFX9-NEXT:    s_waitcnt vmcnt(29)
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v32
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
@@ -81526,13 +81575,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
@@ -81551,13 +81601,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
@@ -81576,13 +81627,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
@@ -81601,13 +81653,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
@@ -81626,13 +81679,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
@@ -81651,13 +81705,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
@@ -81676,13 +81731,14 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
@@ -87884,10 +87940,10 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:388
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
@@ -87907,33 +87963,35 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:136
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:144
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:152
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:160
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:168
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:176
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:184
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:124
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v1
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:136
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:144
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:160
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:168
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:176
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:184
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 8, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 24, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 8, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 24, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v29
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 24, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 24, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v6
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -87954,10 +88012,10 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v18
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v20
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v22
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -87968,69 +88026,69 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
-; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:132
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:140
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v33
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v36
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v35
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v36
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v39
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v55
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v40
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v50
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v44
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v51
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v47
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
@@ -88045,14 +88103,14 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:196
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -88074,16 +88132,16 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -88102,19 +88160,19 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -88133,10 +88191,10 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -88205,7 +88263,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:356
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v3
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -88223,121 +88281,119 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:92
 ; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB58_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v54, v9
 ; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_or_b32_e32 v0, v0, v43
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v0, v0, v54
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v41
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v1, v42, v1
-; SI-NEXT:    v_or_b32_e32 v3, v53, v3
+; SI-NEXT:    v_or_b32_e32 v3, v40, v3
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v3
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v9, v10, v9
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v32, v4
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
+; SI-NEXT:    v_or_b32_e32 v5, v60, v5
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v6, v51, v6
+; SI-NEXT:    v_or_b32_e32 v6, v58, v6
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr53
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v2, v2, v35
+; SI-NEXT:    v_or_b32_e32 v2, v2, v53
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v34, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v3
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr53
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    v_or_b32_e32 v8, v55, v8
+; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v48
@@ -88347,11 +88403,11 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v4, v61
+; SI-NEXT:    v_or_b32_e32 v4, v4, v62
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v38
@@ -88361,18 +88417,18 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v6, v58
+; SI-NEXT:    v_or_b32_e32 v6, v6, v56
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v46
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v40
-; SI-NEXT:    v_or_b32_e32 v8, v8, v50
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v35
+; SI-NEXT:    v_or_b32_e32 v8, v8, v44
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v33
@@ -88383,24 +88439,24 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v10
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v62
-; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v63
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr52
-; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v63
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v60
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v59
@@ -88409,86 +88465,85 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v57
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v51
 ; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v56
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v57
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v47
-; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr47
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v55
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v50
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v44
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v45
+; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v36
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v18
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v18
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v19
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
@@ -88501,7 +88556,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v20
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
@@ -88522,7 +88577,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v21
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -88535,7 +88590,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v22
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -88548,7 +88603,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    v_or_b32_e32 v22, v22, v23
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -88561,7 +88616,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    v_or_b32_e32 v23, v23, v24
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -88574,7 +88629,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v25
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -88587,7 +88642,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -88657,7 +88712,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; SI-NEXT:    v_or_b32_e32 v31, v31, v36
+; SI-NEXT:    v_or_b32_e32 v31, v31, v43
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; SI-NEXT:    v_or_b32_e32 v32, v33, v32
@@ -88671,249 +88726,249 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:  .LBB58_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB58_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    s_movk_i32 s6, 0x300
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v54, v9
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_or_b32_e32 v0, v43, v0
+; SI-NEXT:    v_or_b32_e32 v0, v54, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v42, v1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v53, v2
+; SI-NEXT:    v_or_b32_e32 v2, v40, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
@@ -88922,16 +88977,18 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v32, v4
+; SI-NEXT:    v_or_b32_e32 v9, v10, v9
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
-; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_or_b32_e32 v5, v60, v5
+; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v6, v51, v6
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_or_b32_e32 v6, v58, v6
+; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
@@ -88939,7 +88996,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v41, v1
@@ -88952,11 +89009,11 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
+; SI-NEXT:    v_or_b32_e32 v8, v55, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v2, v35, v2
+; SI-NEXT:    v_or_b32_e32 v2, v53, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
@@ -88974,7 +89031,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v61, v4
+; SI-NEXT:    v_or_b32_e32 v4, v62, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
@@ -88992,7 +89049,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v58, v6
+; SI-NEXT:    v_or_b32_e32 v6, v56, v6
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
@@ -89005,9 +89062,9 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v40
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v35
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_or_b32_e32 v8, v50, v8
+; SI-NEXT:    v_or_b32_e32 v8, v44, v8
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
@@ -89022,7 +89079,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v62
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v63
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
@@ -89030,7 +89087,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v63
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v39
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
@@ -89038,7 +89095,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v60
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -89051,12 +89108,12 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v57
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v51
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v56
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
@@ -89069,7 +89126,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v55
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v50
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
@@ -89077,19 +89134,17 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v44
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v36
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
@@ -89099,7 +89154,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -89108,7 +89163,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v15
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -89116,7 +89171,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -89125,7 +89180,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -89133,7 +89188,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -89142,7 +89197,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v17
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -89150,7 +89205,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -89167,7 +89222,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v19, v20, v19
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -89194,7 +89249,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    v_or_b32_e32 v20, v21, v20
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
@@ -89211,7 +89266,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v21
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
@@ -89228,7 +89283,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
@@ -89245,7 +89300,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v23
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
@@ -89262,7 +89317,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
@@ -89279,7 +89334,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v25
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
@@ -89371,7 +89426,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
-; SI-NEXT:    v_or_b32_e32 v31, v36, v31
+; SI-NEXT:    v_or_b32_e32 v31, v43, v31
 ; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x300, v31
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
@@ -89452,19 +89507,18 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:128
+; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
 ; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
+; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:152
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:160
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:168
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:176
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:184
 ; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v7
+; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v5
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v7
 ; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v9
 ; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
 ; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v13
@@ -89496,59 +89550,61 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v26
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v54
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v32
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v40
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v42
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v57
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
@@ -89557,25 +89613,25 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:204
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:212
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:232
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:240
@@ -89583,25 +89639,25 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:220
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -89609,25 +89665,25 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:252
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -89635,10 +89691,10 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
@@ -89650,10 +89706,10 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:308
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:336
@@ -89664,22 +89720,22 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:340
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:352
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:368
@@ -89690,14 +89746,14 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
@@ -89709,34 +89765,33 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
 ; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:108
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB58_2
 ; VI-NEXT:  ; %bb.1: ; %cmp.false
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -89744,19 +89799,19 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
@@ -89767,23 +89822,23 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr61
-; VI-NEXT:    ; implicit-def: $vgpr62
 ; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    ; implicit-def: $vgpr56
-; VI-NEXT:    ; implicit-def: $vgpr47
+; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr46
 ; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    ; implicit-def: $vgpr41
 ; VI-NEXT:    ; implicit-def: $vgpr55
+; VI-NEXT:    ; implicit-def: $vgpr54
 ; VI-NEXT:    ; implicit-def: $vgpr50
 ; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr38
@@ -89824,147 +89879,217 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; kill: killed $vgpr33
 ; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr33
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr46
+; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr45
-; VI-NEXT:    ; implicit-def: $vgpr54
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -89974,16 +90099,16 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
@@ -89992,7 +90117,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
@@ -90001,253 +90126,190 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr33
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:  .LBB58_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB58_4
 ; VI-NEXT:  ; %bb.3: ; %cmp.true
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(13)
+; VI-NEXT:    v_add_u16_e32 v9, 3, v61
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v31, 0x300
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_add_u16_e32 v9, 3, v61
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    v_add_u16_e32 v8, 3, v8
+; VI-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v1, 3, v1
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v1, 3, v1
+; VI-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; VI-NEXT:    v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
@@ -90256,31 +90318,25 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v1, v2, v3
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
 ; VI-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v5, 3, v5
 ; VI-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(13)
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
 ; VI-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(11)
+; VI-NEXT:    s_waitcnt vmcnt(10)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
 ; VI-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_add_u16_e32 v8, 3, v8
-; VI-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(3)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
-; VI-NEXT:    s_waitcnt vmcnt(2)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -90317,13 +90373,13 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v7, v7, v8
-; VI-NEXT:    v_add_u16_e32 v8, 3, v63
+; VI-NEXT:    v_add_u16_e32 v8, 3, v62
 ; VI-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v8, v8, v9
-; VI-NEXT:    v_add_u16_e32 v9, 3, v59
+; VI-NEXT:    v_add_u16_e32 v9, 3, v32
 ; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v10, 3, v62
+; VI-NEXT:    v_add_u16_e32 v10, 3, v60
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
@@ -90331,27 +90387,28 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v9, v9, v10
 ; VI-NEXT:    v_add_u16_e32 v10, 3, v58
 ; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v11, 3, v60
+; VI-NEXT:    v_add_u16_e32 v11, 3, v56
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v10, v10, v11
-; VI-NEXT:    v_add_u16_e32 v11, 3, v57
+; VI-NEXT:    v_add_u16_e32 v11, 3, v59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v12, 3, v56
+; VI-NEXT:    v_add_u16_e32 v12, 3, v57
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v11, v11, v12
-; VI-NEXT:    v_add_u16_e32 v12, 3, v46
+; VI-NEXT:    v_add_u16_e32 v12, 3, v47
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v13, 3, v47
+; VI-NEXT:    v_add_u16_e32 v13, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
@@ -90362,148 +90419,149 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 3, v44
 ; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v13, v13, v14
-; VI-NEXT:    v_add_u16_e32 v14, 3, v54
+; VI-NEXT:    v_add_u16_e32 v14, 3, v42
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v15, 3, v40
-; VI-NEXT:    v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v14, v14, v15
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v15, 3, v15
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v15, 3, v15
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v15, v15, v16
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v16, v16, v17
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v18, 3, v18
 ; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v17, v17, v18
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v18, 3, v18
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v18, 0x300, v18
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v18, v18, v19
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v19, 0x300, v19
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v19, v19, v20
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v20, 0x300, v20
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
 ; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v20, v20, v21
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v21, v21, v22
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v23, 3, v23
 ; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v22, v22, v23
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v23, 3, v23
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v24, 3, v24
 ; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v23, v23, v24
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v24, 3, v24
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v25, 3, v25
 ; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v24, v24, v25
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v25, 3, v25
@@ -90517,19 +90575,19 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v26, 3, v26
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v27, 3, v27
 ; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v26, v26, v27
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -90543,20 +90601,20 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v27, v27, v28
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v28, 3, v28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v29, 3, v29
 ; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v28, v28, v29
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v29, 3, v29
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -90569,7 +90627,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v29, v29, v30
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v30, 3, v30
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -90582,7 +90640,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v30, v30, v32
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v32, 3, v32
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -90591,7 +90649,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v32, 0x300, v32
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v33, 3, v33
-; VI-NEXT:    v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v33, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v31, v32, v31
 ; VI-NEXT:  .LBB58_4: ; %end
@@ -90668,19 +90726,18 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
 ; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
+; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:152
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:160
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:168
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:176
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:184
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT:    v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v7
+; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v3
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v5
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v7
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v9
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v13
@@ -90692,93 +90749,95 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v25
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v35, 8, v29
-; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v0
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT:    s_nop 0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v54
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v32
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v40
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v42
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v57
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; GFX9-NEXT:    s_nop 0
@@ -90788,25 +90847,25 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:204
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:212
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:232
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:240
@@ -90815,25 +90874,25 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:220
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -90842,25 +90901,25 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:252
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -90869,10 +90928,10 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
@@ -90884,10 +90943,10 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:308
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:336
@@ -90899,22 +90958,22 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:340
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:352
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:368
@@ -90926,14 +90985,14 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
@@ -90945,34 +91004,33 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
 ; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB58_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -90980,20 +91038,20 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
@@ -91004,23 +91062,23 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr61
-; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr43
-; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    ; implicit-def: $vgpr41
 ; GFX9-NEXT:    ; implicit-def: $vgpr55
+; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr38
@@ -91061,147 +91119,217 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; kill: killed $vgpr33
 ; GFX9-NEXT:    ; implicit-def: $vgpr35
+; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr46
+; GFX9-NEXT:    ; implicit-def: $vgpr47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
-; GFX9-NEXT:    ; implicit-def: $vgpr54
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -91211,16 +91339,16 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
@@ -91229,7 +91357,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
@@ -91238,258 +91366,196 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr33
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:  .LBB58_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB58_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
+; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; GFX9-NEXT:    s_waitcnt vmcnt(13)
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_movk_i32 s6, 0x300
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(33)
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(30)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(28)
-; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
+; GFX9-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    s_waitcnt vmcnt(27)
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT:    v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(23)
+; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
+; GFX9-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; GFX9-NEXT:    v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
@@ -91498,32 +91564,26 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
 ; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
 ; GFX9-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
+; GFX9-NEXT:    s_waitcnt vmcnt(17)
 ; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(12)
+; GFX9-NEXT:    s_waitcnt vmcnt(11)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -91560,41 +91620,41 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v63
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v62
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v32
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v60
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
 ; GFX9-NEXT:    v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v60
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v56
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v57
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v56
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v57
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v46
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v46
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
@@ -91605,148 +91665,149 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 3, v44
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v54
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v40
-; GFX9-NEXT:    v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v17, v17, v18
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v18, v18, v19
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v19
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v19, v19, v20
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v20, v20, v21
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v21, v21, v22
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v22, v22, v23
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v23
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v23, v23, v24
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v24, v24, v25
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
@@ -91760,19 +91821,19 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v26, v26, v27
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -91786,20 +91847,20 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v27, v27, v28
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v28, v28, v29
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -91812,7 +91873,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v29, v29, v30
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -91825,7 +91886,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v30, v30, v31
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -91834,7 +91895,7 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v32, 3, v32
-; GFX9-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v32, v63, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v31, v31, v32
 ; GFX9-NEXT:  .LBB58_4: ; %end
@@ -94565,8 +94626,8 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    s_cbranch_vccnz .LBB59_5
 ; SI-NEXT:  ; %bb.4: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_add_i32 s28, s28, 3
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v40
@@ -94621,7 +94682,7 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
@@ -94743,12 +94804,12 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 0x3000000, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
@@ -95663,8 +95724,8 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; VI-NEXT:    s_cbranch_vccnz .LBB59_5
 ; VI-NEXT:  ; %bb.4: ; %cmp.true
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_add_i32 s28, s28, 3
 ; VI-NEXT:    s_and_b32 s4, s28, 0xff
 ; VI-NEXT:    s_lshl_b32 s5, s29, 8
@@ -95716,7 +95777,7 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; VI-NEXT:    s_add_i32 s6, s6, 0x3000000
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x300, v1
@@ -95811,11 +95872,11 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
 ; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
@@ -95938,11 +95999,11 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
 ; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
@@ -96727,11 +96788,11 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -96943,11 +97004,11 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -99015,9 +99076,9 @@ define <64 x bfloat> @bitcast_v16i64_to_v64bf16(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -99050,7 +99111,7 @@ define <64 x bfloat> @bitcast_v16i64_to_v64bf16(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; kill: killed $vgpr31
@@ -99169,7 +99230,6 @@ define <64 x bfloat> @bitcast_v16i64_to_v64bf16(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v62
 ; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -100960,6 +101020,8 @@ define <16 x i64> @bitcast_v64bf16_to_v16i64(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; kill: killed $vgpr33
@@ -101036,8 +101098,6 @@ define <16 x i64> @bitcast_v64bf16_to_v16i64(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
@@ -107301,9 +107361,9 @@ define <64 x half> @bitcast_v16i64_to_v64f16(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -107336,7 +107396,7 @@ define <64 x half> @bitcast_v16i64_to_v64f16(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; kill: killed $vgpr31
@@ -107409,7 +107469,6 @@ define <64 x half> @bitcast_v16i64_to_v64f16(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB64_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v62
 ; SI-NEXT:    v_cvt_f32_f16_e32 v32, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -110506,9 +110565,9 @@ define inreg <16 x i64> @bitcast_v64f16_to_v16i64_scalar(<64 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
@@ -110558,9 +110617,8 @@ define inreg <16 x i64> @bitcast_v64f16_to_v16i64_scalar(<64 x half> inreg %a, i
 ; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
@@ -111292,8 +111350,8 @@ define <64 x i16> @bitcast_v16i64_to_v64i16(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -111327,7 +111385,7 @@ define <64 x i16> @bitcast_v16i64_to_v64i16(<16 x i64> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -112507,26 +112565,25 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:132
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:20
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v5
@@ -112539,102 +112596,107 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v6
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v14
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v20
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v22
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v24
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v26
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v30
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v33
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v33
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:100
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v31
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:68
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:100
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:28
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
+; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB70_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v41
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
@@ -112651,11 +112713,7 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v50
-; SI-NEXT:    v_or_b32_e32 v23, v23, v32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v32
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
@@ -112685,81 +112743,71 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v48
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v35
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    v_or_b32_e32 v16, v16, v53
+; SI-NEXT:    v_or_b32_e32 v17, v17, v51
+; SI-NEXT:    v_or_b32_e32 v18, v18, v50
 ; SI-NEXT:    ; kill: killed $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v43
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v55
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v53
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    v_or_b32_e32 v19, v19, v48
-; SI-NEXT:    v_or_b32_e32 v21, v21, v36
-; SI-NEXT:    v_or_b32_e32 v22, v22, v34
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v18, v18, v49
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_or_b32_e32 v22, v22, v23
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v54
+; SI-NEXT:    v_or_b32_e32 v19, v19, v39
+; SI-NEXT:    v_or_b32_e32 v23, v23, v24
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v52
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v25
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v39
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v49
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
-; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v35
+; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v37
 ; SI-NEXT:    v_or_b32_e32 v26, v26, v27
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v33
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v28
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v63
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v62
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v61
@@ -112772,13 +112820,15 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v46
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v45
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v44
-; SI-NEXT:    v_or_b32_e32 v12, v12, v42
-; SI-NEXT:    v_or_b32_e32 v13, v13, v41
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v42
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v40
-; SI-NEXT:    v_or_b32_e32 v15, v15, v54
-; SI-NEXT:    v_or_b32_e32 v20, v20, v38
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    v_or_b32_e32 v15, v15, v55
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
@@ -112792,63 +112842,79 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_or_b32_e32 v28, v28, v29
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v16, v16, v52
-; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v20, v20, v36
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    v_or_b32_e32 v29, v29, v30
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
 ; SI-NEXT:    v_or_b32_e32 v30, v30, v31
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; SI-NEXT:    v_or_b32_e32 v17, v17, v51
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v21, v21, v34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; SI-NEXT:    v_or_b32_e32 v31, v31, v37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    v_or_b32_e32 v31, v31, v38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:  .LBB70_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB70_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v41
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
@@ -112865,69 +112931,68 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v48
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v35
+; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v32
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT:    s_mov_b32 s6, 0x30000
+; SI-NEXT:    v_or_b32_e32 v16, v53, v16
+; SI-NEXT:    v_or_b32_e32 v17, v51, v17
+; SI-NEXT:    v_or_b32_e32 v18, v50, v18
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v43
-; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v55
-; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v53
-; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v50
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
+; SI-NEXT:    v_or_b32_e32 v22, v23, v22
+; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v54
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v19, v48, v19
-; SI-NEXT:    v_or_b32_e32 v21, v36, v21
-; SI-NEXT:    v_or_b32_e32 v22, v34, v22
-; SI-NEXT:    v_or_b32_e32 v23, v32, v23
-; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
-; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
-; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
-; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    v_or_b32_e32 v23, v24, v23
+; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v52
 ; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
-; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v39
+; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v49
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
-; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v35
+; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v37
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v33
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    v_or_b32_e32 v27, v28, v27
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -112944,8 +113009,7 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    v_or_b32_e32 v0, v63, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v62, v1
 ; SI-NEXT:    v_or_b32_e32 v2, v61, v2
@@ -112958,28 +113022,25 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v9, v46, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v45, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v44, v11
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_or_b32_e32 v13, v41, v13
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v42, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v40, v14
-; SI-NEXT:    v_or_b32_e32 v15, v54, v15
-; SI-NEXT:    v_or_b32_e32 v18, v49, v18
-; SI-NEXT:    v_or_b32_e32 v20, v38, v20
+; SI-NEXT:    v_or_b32_e32 v15, v55, v15
+; SI-NEXT:    v_or_b32_e32 v19, v39, v19
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
+; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_or_b32_e32 v28, v29, v28
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
-; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
-; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
@@ -112989,34 +113050,37 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
-; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
+; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
+; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
+; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v20, v36, v20
+; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    v_or_b32_e32 v29, v30, v29
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x30000, v29
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
 ; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
 ; SI-NEXT:    v_or_b32_e32 v30, v31, v30
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; SI-NEXT:    v_or_b32_e32 v17, v51, v17
-; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v21, v34, v21
+; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x30000, v30
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; SI-NEXT:    v_or_b32_e32 v31, v37, v31
+; SI-NEXT:    v_or_b32_e32 v31, v38, v31
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x30000, v31
 ; SI-NEXT:  .LBB70_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
@@ -113036,7 +113100,7 @@ define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: bitcast_v64i16_to_v16i64:
@@ -114321,8 +114385,8 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; kill: killed $vgpr36
@@ -114484,7 +114548,7 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; kill: killed $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -115517,8 +115581,8 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; VI-NEXT:    ; implicit-def: $vgpr39
 ; VI-NEXT:    ; kill: killed $vgpr39
@@ -116158,13 +116222,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
@@ -116184,13 +116249,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
@@ -116210,13 +116276,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
@@ -116236,13 +116303,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
@@ -116262,13 +116330,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
@@ -116287,13 +116356,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
@@ -116313,13 +116383,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
@@ -116339,13 +116410,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
@@ -116390,13 +116462,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
@@ -116464,8 +116537,8 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX9-NEXT:    ; implicit-def: $vgpr41
 ; GFX9-NEXT:    ; kill: killed $vgpr41
@@ -116659,14 +116732,13 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(32)
+; GFX9-NEXT:    s_waitcnt vmcnt(31)
 ; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB72_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
-; GFX9-NEXT:    s_waitcnt vmcnt(31)
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v32
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
@@ -117268,13 +117340,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
@@ -117293,13 +117366,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
@@ -117318,13 +117392,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
@@ -117343,13 +117418,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
@@ -117368,13 +117444,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
@@ -117393,13 +117470,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
@@ -117418,13 +117496,14 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
@@ -122416,13 +122495,14 @@ define inreg <128 x i8> @bitcast_v16f64_to_v128i8_scalar(<16 x double> inreg %a,
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v12, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
@@ -124558,10 +124638,10 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:388
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
@@ -124581,33 +124661,35 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:136
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:144
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:152
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:160
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:168
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:176
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:184
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:124
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v1
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:136
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:144
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:160
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:168
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:176
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:184
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 8, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 24, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 8, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 24, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v29
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 24, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 24, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v6
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -124628,10 +124710,10 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v18
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v20
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v22
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -124642,69 +124724,69 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
-; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:132
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:140
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v33
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v36
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v35
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v36
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v39
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v55
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v40
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v50
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v44
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v51
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v47
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
@@ -124719,14 +124801,14 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:196
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -124748,16 +124830,16 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -124776,19 +124858,19 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -124807,10 +124889,10 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -124879,7 +124961,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:356
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v3
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -124897,121 +124979,119 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:92
 ; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB74_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v54, v9
 ; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_or_b32_e32 v0, v0, v43
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v0, v0, v54
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v41
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v1, v42, v1
-; SI-NEXT:    v_or_b32_e32 v3, v53, v3
+; SI-NEXT:    v_or_b32_e32 v3, v40, v3
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v3
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v9, v10, v9
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v32, v4
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
+; SI-NEXT:    v_or_b32_e32 v5, v60, v5
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v6, v51, v6
+; SI-NEXT:    v_or_b32_e32 v6, v58, v6
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr53
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v2, v2, v35
+; SI-NEXT:    v_or_b32_e32 v2, v2, v53
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v34, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v3
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr53
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    v_or_b32_e32 v8, v55, v8
+; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v48
@@ -125021,11 +125101,11 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v4, v61
+; SI-NEXT:    v_or_b32_e32 v4, v4, v62
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v38
@@ -125035,18 +125115,18 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v6, v58
+; SI-NEXT:    v_or_b32_e32 v6, v6, v56
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v46
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v40
-; SI-NEXT:    v_or_b32_e32 v8, v8, v50
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v35
+; SI-NEXT:    v_or_b32_e32 v8, v8, v44
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v33
@@ -125057,24 +125137,24 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v10
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v62
-; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v63
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr52
-; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v63
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v60
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v59
@@ -125083,86 +125163,85 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v57
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v51
 ; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v56
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v57
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v47
-; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr47
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v55
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v50
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v44
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v45
+; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v36
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v18
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v18
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v19
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
@@ -125175,7 +125254,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v20
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
@@ -125196,7 +125275,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v21
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125209,7 +125288,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v22
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125222,7 +125301,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    v_or_b32_e32 v22, v22, v23
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125235,7 +125314,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    v_or_b32_e32 v23, v23, v24
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125248,7 +125327,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v25
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125261,7 +125340,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125331,7 +125410,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; SI-NEXT:    v_or_b32_e32 v31, v31, v36
+; SI-NEXT:    v_or_b32_e32 v31, v31, v43
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; SI-NEXT:    v_or_b32_e32 v32, v33, v32
@@ -125345,249 +125424,249 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; kill: killed $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:  .LBB74_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB74_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    s_movk_i32 s6, 0x300
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v54, v9
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_or_b32_e32 v0, v43, v0
+; SI-NEXT:    v_or_b32_e32 v0, v54, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v42, v1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v53, v2
+; SI-NEXT:    v_or_b32_e32 v2, v40, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
@@ -125596,16 +125675,18 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v32, v4
+; SI-NEXT:    v_or_b32_e32 v9, v10, v9
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
-; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_or_b32_e32 v5, v60, v5
+; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v6, v51, v6
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_or_b32_e32 v6, v58, v6
+; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
@@ -125613,7 +125694,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v41, v1
@@ -125626,11 +125707,11 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
+; SI-NEXT:    v_or_b32_e32 v8, v55, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v2, v35, v2
+; SI-NEXT:    v_or_b32_e32 v2, v53, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
@@ -125648,7 +125729,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v61, v4
+; SI-NEXT:    v_or_b32_e32 v4, v62, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
@@ -125666,7 +125747,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v58, v6
+; SI-NEXT:    v_or_b32_e32 v6, v56, v6
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
@@ -125679,9 +125760,9 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v40
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v35
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_or_b32_e32 v8, v50, v8
+; SI-NEXT:    v_or_b32_e32 v8, v44, v8
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
@@ -125696,7 +125777,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v62
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v63
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
@@ -125704,7 +125785,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v63
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v39
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
@@ -125712,7 +125793,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v60
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125725,12 +125806,12 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v57
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v51
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v56
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
@@ -125743,7 +125824,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v55
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v50
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
@@ -125751,19 +125832,17 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v44
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v36
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
@@ -125773,7 +125852,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125782,7 +125861,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v15
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -125790,7 +125869,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125799,7 +125878,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -125807,7 +125886,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125816,7 +125895,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v17
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -125824,7 +125903,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125841,7 +125920,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v19, v20, v19
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -125868,7 +125947,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    v_or_b32_e32 v20, v21, v20
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
@@ -125885,7 +125964,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v21
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
@@ -125902,7 +125981,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
@@ -125919,7 +125998,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    v_or_b32_e32 v23, v24, v23
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v23
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
@@ -125936,7 +126015,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
@@ -125953,7 +126032,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v25
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
@@ -126045,7 +126124,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
-; SI-NEXT:    v_or_b32_e32 v31, v36, v31
+; SI-NEXT:    v_or_b32_e32 v31, v43, v31
 ; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x300, v31
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
@@ -126126,19 +126205,18 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:128
+; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
 ; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
+; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:152
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:160
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:168
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:176
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:184
 ; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v1
-; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v3
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v5
-; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v7
+; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v5
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v7
 ; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v9
 ; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
 ; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v13
@@ -126170,59 +126248,61 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v26
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v54
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v32
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v40
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v42
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v57
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
@@ -126231,25 +126311,25 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:204
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:212
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:232
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:240
@@ -126257,25 +126337,25 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:220
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -126283,25 +126363,25 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:252
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -126309,10 +126389,10 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
@@ -126324,10 +126404,10 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:308
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:336
@@ -126338,22 +126418,22 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:340
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:352
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:368
@@ -126364,14 +126444,14 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
@@ -126383,34 +126463,33 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
 ; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:108
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB74_2
 ; VI-NEXT:  ; %bb.1: ; %cmp.false
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -126418,19 +126497,19 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
@@ -126441,23 +126520,23 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr61
-; VI-NEXT:    ; implicit-def: $vgpr62
 ; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    ; implicit-def: $vgpr56
-; VI-NEXT:    ; implicit-def: $vgpr47
+; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr46
 ; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    ; implicit-def: $vgpr41
 ; VI-NEXT:    ; implicit-def: $vgpr55
+; VI-NEXT:    ; implicit-def: $vgpr54
 ; VI-NEXT:    ; implicit-def: $vgpr50
 ; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr38
@@ -126498,147 +126577,217 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; kill: killed $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; kill: killed $vgpr33
 ; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr33
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr46
+; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr45
-; VI-NEXT:    ; implicit-def: $vgpr54
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -126648,16 +126797,16 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
@@ -126666,7 +126815,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
@@ -126675,253 +126824,190 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr33
-; VI-NEXT:    ; kill: killed $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr32
+; VI-NEXT:    ; kill: killed $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:  .LBB74_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB74_4
 ; VI-NEXT:  ; %bb.3: ; %cmp.true
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(13)
+; VI-NEXT:    v_add_u16_e32 v9, 3, v61
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v31, 0x300
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_add_u16_e32 v9, 3, v61
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    v_add_u16_e32 v8, 3, v8
+; VI-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v1, 3, v1
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v1, 3, v1
+; VI-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; VI-NEXT:    v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
@@ -126930,31 +127016,25 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v1, v2, v3
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
 ; VI-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v5, 3, v5
 ; VI-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(13)
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
 ; VI-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(11)
+; VI-NEXT:    s_waitcnt vmcnt(10)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
 ; VI-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_add_u16_e32 v8, 3, v8
-; VI-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(3)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
-; VI-NEXT:    s_waitcnt vmcnt(2)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -126991,13 +127071,13 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v7, v7, v8
-; VI-NEXT:    v_add_u16_e32 v8, 3, v63
+; VI-NEXT:    v_add_u16_e32 v8, 3, v62
 ; VI-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v8, v8, v9
-; VI-NEXT:    v_add_u16_e32 v9, 3, v59
+; VI-NEXT:    v_add_u16_e32 v9, 3, v32
 ; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v10, 3, v62
+; VI-NEXT:    v_add_u16_e32 v10, 3, v60
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
@@ -127005,27 +127085,28 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v9, v9, v10
 ; VI-NEXT:    v_add_u16_e32 v10, 3, v58
 ; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v11, 3, v60
+; VI-NEXT:    v_add_u16_e32 v11, 3, v56
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v10, v10, v11
-; VI-NEXT:    v_add_u16_e32 v11, 3, v57
+; VI-NEXT:    v_add_u16_e32 v11, 3, v59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v12, 3, v56
+; VI-NEXT:    v_add_u16_e32 v12, 3, v57
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v11, v11, v12
-; VI-NEXT:    v_add_u16_e32 v12, 3, v46
+; VI-NEXT:    v_add_u16_e32 v12, 3, v47
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v13, 3, v47
+; VI-NEXT:    v_add_u16_e32 v13, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
@@ -127036,148 +127117,149 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 3, v44
 ; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v13, v13, v14
-; VI-NEXT:    v_add_u16_e32 v14, 3, v54
+; VI-NEXT:    v_add_u16_e32 v14, 3, v42
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v15, 3, v40
-; VI-NEXT:    v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v14, v14, v15
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v15, 3, v15
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v15, 3, v15
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v15, v15, v16
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v16, v16, v17
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v18, 3, v18
 ; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v17, v17, v18
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v18, 3, v18
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v18, 0x300, v18
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v18, v18, v19
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v19, 0x300, v19
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v19, v19, v20
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v20, 0x300, v20
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
 ; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v20, v20, v21
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v21, v21, v22
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v23, 3, v23
 ; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v22, v22, v23
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v23, 3, v23
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v24, 3, v24
 ; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v23, v23, v24
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v24, 3, v24
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v25, 3, v25
 ; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v24, v24, v25
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v25, 3, v25
@@ -127191,19 +127273,19 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v26, 3, v26
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v27, 3, v27
 ; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v26, v26, v27
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -127217,20 +127299,20 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v27, v27, v28
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v28, 3, v28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v29, 3, v29
 ; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v28, v28, v29
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v29, 3, v29
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -127243,7 +127325,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v29, v29, v30
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v30, 3, v30
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -127256,7 +127338,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v30, v30, v32
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v32, 3, v32
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -127265,7 +127347,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v32, 0x300, v32
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v33, 3, v33
-; VI-NEXT:    v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v33, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v31, v32, v31
 ; VI-NEXT:  .LBB74_4: ; %end
@@ -127342,19 +127424,18 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
 ; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:136
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:144
-; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:152
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:168
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:176
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:184
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
+; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:152
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:160
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:168
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:176
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:184
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v1
-; GFX9-NEXT:    v_lshlrev_b16_e32 v42, 8, v3
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v5
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v7
+; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v3
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v5
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v7
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v9
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v13
@@ -127366,93 +127447,95 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v25
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v35, 8, v29
-; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v0
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT:    s_nop 0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(21)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v54
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v32
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v40
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v42
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v45
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v46
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v47
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v57
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v56
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; GFX9-NEXT:    s_nop 0
@@ -127462,25 +127545,25 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:204
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:212
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:232
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:240
@@ -127489,25 +127572,25 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:220
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -127516,25 +127599,25 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:252
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -127543,10 +127626,10 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
@@ -127558,10 +127641,10 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:308
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:336
@@ -127573,22 +127656,22 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:340
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:352
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:368
@@ -127600,14 +127683,14 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
@@ -127619,34 +127702,33 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
 ; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB74_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -127654,20 +127736,20 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v55 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v41 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
@@ -127678,23 +127760,23 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v57, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr61
-; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr43
-; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    ; implicit-def: $vgpr41
 ; GFX9-NEXT:    ; implicit-def: $vgpr55
+; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr38
@@ -127735,147 +127817,217 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v8, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v62, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; kill: killed $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; kill: killed $vgpr33
 ; GFX9-NEXT:    ; implicit-def: $vgpr35
+; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v32, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr46
+; GFX9-NEXT:    ; implicit-def: $vgpr47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v54, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
-; GFX9-NEXT:    ; implicit-def: $vgpr54
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -127885,16 +128037,16 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
@@ -127903,7 +128055,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
@@ -127912,258 +128064,196 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v63 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; kill: killed $vgpr32
 ; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr33
-; GFX9-NEXT:    ; kill: killed $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; kill: killed $vgpr63
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:  .LBB74_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB74_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
+; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; GFX9-NEXT:    s_waitcnt vmcnt(13)
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_movk_i32 s6, 0x300
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(33)
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(30)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(28)
-; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
+; GFX9-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    s_waitcnt vmcnt(27)
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT:    v_or_b32_sdwa v1, v42, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v2, v41, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v3, v55, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(23)
+; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
+; GFX9-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; GFX9-NEXT:    v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
@@ -128172,32 +128262,26 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
 ; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
 ; GFX9-NEXT:    v_or_b32_sdwa v4, v50, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
+; GFX9-NEXT:    s_waitcnt vmcnt(17)
 ; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v48, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v38, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(12)
+; GFX9-NEXT:    s_waitcnt vmcnt(11)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v36, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT:    v_or_b32_sdwa v8, v34, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v53, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v52, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -128234,41 +128318,41 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v35, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v63
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v62
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v33, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v32
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v60
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
 ; GFX9-NEXT:    v_add_u16_e32 v10, 3, v58
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v60
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v56
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v57
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v56
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v57
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v46
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v46
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
@@ -128279,148 +128363,149 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 3, v44
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v54
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v40
-; GFX9-NEXT:    v_or_b32_sdwa v14, v32, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v17, v17, v18
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v18, v18, v19
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v19
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v19, v19, v20
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v20, v20, v21
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v21, v21, v22
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v22, v22, v23
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v23
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v23, v23, v24
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v24, v24, v25
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
@@ -128434,19 +128519,19 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v26, v26, v27
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -128460,20 +128545,20 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v27, v27, v28
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v28, v28, v29
-; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -128486,7 +128571,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v29, v29, v30
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -128499,7 +128584,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v30, v30, v31
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -128508,7 +128593,7 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v32, 3, v32
-; GFX9-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v32, v63, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_e32 v31, v31, v32
 ; GFX9-NEXT:  .LBB74_4: ; %end
@@ -131239,8 +131324,8 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; SI-NEXT:    s_cbranch_vccnz .LBB75_5
 ; SI-NEXT:  ; %bb.4: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_add_i32 s28, s28, 3
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v40
@@ -131295,7 +131380,7 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
@@ -131417,12 +131502,12 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 0x3000000, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
@@ -132337,8 +132422,8 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; VI-NEXT:    s_cbranch_vccnz .LBB75_5
 ; VI-NEXT:  ; %bb.4: ; %cmp.true
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    s_add_i32 s28, s28, 3
 ; VI-NEXT:    s_and_b32 s4, s28, 0xff
 ; VI-NEXT:    s_lshl_b32 s5, s29, 8
@@ -132390,7 +132475,7 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; VI-NEXT:    s_add_i32 s6, s6, 0x3000000
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x300, v1
@@ -132485,11 +132570,11 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
 ; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
@@ -132612,11 +132697,11 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
 ; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
@@ -133401,11 +133486,11 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -133617,11 +133702,11 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -135689,8 +135774,8 @@ define <64 x bfloat> @bitcast_v16f64_to_v64bf16(<16 x double> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; kill: killed $vgpr35
@@ -135788,14 +135873,13 @@ define <64 x bfloat> @bitcast_v16f64_to_v64bf16(<16 x double> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; kill: killed $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB76_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v32
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
 ; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
@@ -137582,6 +137666,8 @@ define <16 x double> @bitcast_v64bf16_to_v16f64(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; kill: killed $vgpr33
@@ -137658,8 +137744,6 @@ define <16 x double> @bitcast_v64bf16_to_v16f64(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
@@ -143923,8 +144007,8 @@ define <64 x half> @bitcast_v16f64_to_v64f16(<16 x double> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; kill: killed $vgpr35
@@ -144022,7 +144106,7 @@ define <64 x half> @bitcast_v16f64_to_v64f16(<16 x double> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -144069,7 +144153,6 @@ define <64 x half> @bitcast_v16f64_to_v64f16(<16 x double> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v8, v52
-; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v31
@@ -146995,9 +147078,9 @@ define inreg <16 x double> @bitcast_v64f16_to_v16f64_scalar(<64 x half> inreg %a
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
@@ -147047,9 +147130,8 @@ define inreg <16 x double> @bitcast_v64f16_to_v16f64_scalar(<64 x half> inreg %a
 ; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
@@ -147781,8 +147863,8 @@ define <64 x i16> @bitcast_v16f64_to_v64i16(<16 x double> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr60
@@ -147816,7 +147898,7 @@ define <64 x i16> @bitcast_v16f64_to_v64i16(<16 x double> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -148869,26 +148951,25 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:132
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:20
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v5
@@ -148901,102 +148982,107 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v6
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v14
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v20
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v22
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(8) expcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v24
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v26
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v30
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v33
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v33
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:100
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v31
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:68
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:100
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:28
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
+; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB86_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v41
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
@@ -149013,11 +149099,7 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v50
-; SI-NEXT:    v_or_b32_e32 v23, v23, v32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v32
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
@@ -149047,81 +149129,71 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v48
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v35
 ; SI-NEXT:    ; kill: killed $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    v_or_b32_e32 v16, v16, v53
+; SI-NEXT:    v_or_b32_e32 v17, v17, v51
+; SI-NEXT:    v_or_b32_e32 v18, v18, v50
 ; SI-NEXT:    ; kill: killed $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v43
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v55
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v53
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    v_or_b32_e32 v19, v19, v48
-; SI-NEXT:    v_or_b32_e32 v21, v21, v36
-; SI-NEXT:    v_or_b32_e32 v22, v22, v34
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; kill: killed $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v18, v18, v49
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_or_b32_e32 v22, v22, v23
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v54
+; SI-NEXT:    v_or_b32_e32 v19, v19, v39
+; SI-NEXT:    v_or_b32_e32 v23, v23, v24
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v52
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v25
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v39
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v49
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
-; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v35
+; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v37
 ; SI-NEXT:    v_or_b32_e32 v26, v26, v27
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v33
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v28
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v63
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v62
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v61
@@ -149134,13 +149206,15 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v46
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v45
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v44
-; SI-NEXT:    v_or_b32_e32 v12, v12, v42
-; SI-NEXT:    v_or_b32_e32 v13, v13, v41
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v42
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v40
-; SI-NEXT:    v_or_b32_e32 v15, v15, v54
-; SI-NEXT:    v_or_b32_e32 v20, v20, v38
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    v_or_b32_e32 v15, v15, v55
+; SI-NEXT:    ; kill: killed $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
@@ -149154,63 +149228,79 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_or_b32_e32 v28, v28, v29
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v16, v16, v52
-; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v20, v20, v36
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    v_or_b32_e32 v29, v29, v30
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
 ; SI-NEXT:    v_or_b32_e32 v30, v30, v31
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; SI-NEXT:    v_or_b32_e32 v17, v17, v51
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v21, v21, v34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; kill: killed $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; SI-NEXT:    v_or_b32_e32 v31, v31, v37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; kill: killed $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    v_or_b32_e32 v31, v31, v38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; kill: killed $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:  .LBB86_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB86_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v41
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
@@ -149227,69 +149317,68 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v48
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v35
+; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v32
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT:    s_mov_b32 s6, 0x30000
+; SI-NEXT:    v_or_b32_e32 v16, v53, v16
+; SI-NEXT:    v_or_b32_e32 v17, v51, v17
+; SI-NEXT:    v_or_b32_e32 v18, v50, v18
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v43
-; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v55
-; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v53
-; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v50
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
+; SI-NEXT:    v_or_b32_e32 v22, v23, v22
+; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v54
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v19, v48, v19
-; SI-NEXT:    v_or_b32_e32 v21, v36, v21
-; SI-NEXT:    v_or_b32_e32 v22, v34, v22
-; SI-NEXT:    v_or_b32_e32 v23, v32, v23
-; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
-; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
-; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
-; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    v_or_b32_e32 v23, v24, v23
+; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v52
 ; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
-; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v39
+; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v49
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
-; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v35
+; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v37
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v33
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    v_or_b32_e32 v27, v28, v27
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -149306,8 +149395,7 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    v_or_b32_e32 v0, v63, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v62, v1
 ; SI-NEXT:    v_or_b32_e32 v2, v61, v2
@@ -149320,28 +149408,25 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v9, v46, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v45, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v44, v11
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_or_b32_e32 v13, v41, v13
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v42, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v40, v14
-; SI-NEXT:    v_or_b32_e32 v15, v54, v15
-; SI-NEXT:    v_or_b32_e32 v18, v49, v18
-; SI-NEXT:    v_or_b32_e32 v20, v38, v20
+; SI-NEXT:    v_or_b32_e32 v15, v55, v15
+; SI-NEXT:    v_or_b32_e32 v19, v39, v19
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
+; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_or_b32_e32 v28, v29, v28
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
-; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
-; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
@@ -149351,34 +149436,37 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
-; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
+; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
+; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
+; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v20, v36, v20
+; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    v_or_b32_e32 v29, v30, v29
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x30000, v29
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
 ; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
 ; SI-NEXT:    v_or_b32_e32 v30, v31, v30
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; SI-NEXT:    v_or_b32_e32 v17, v51, v17
-; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v21, v34, v21
+; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x30000, v30
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; SI-NEXT:    v_or_b32_e32 v31, v37, v31
+; SI-NEXT:    v_or_b32_e32 v31, v38, v31
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x30000, v31
 ; SI-NEXT:  .LBB86_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
@@ -149398,7 +149486,7 @@ define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: bitcast_v64i16_to_v16f64:
@@ -150727,31 +150815,31 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v6
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v8
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v12
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v14
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v16
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v20
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v22
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v24
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v28
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v30
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
@@ -150763,36 +150851,35 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr24
 ; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr20
-; SI-NEXT:    ; implicit-def: $vgpr14
+; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr6
+; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v18
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v2
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v7
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v9
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v10
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v11
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
@@ -150819,63 +150906,64 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v25
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr11
 ; SI-NEXT:    ; implicit-def: $vgpr10
-; SI-NEXT:    ; implicit-def: $vgpr19
-; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr9
 ; SI-NEXT:    ; implicit-def: $vgpr13
+; SI-NEXT:    ; implicit-def: $vgpr23
 ; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    ; implicit-def: $vgpr17
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:160
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:140
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:136
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:168
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:160
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:96
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v26
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v27
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v29
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:96
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:88
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:80
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196
@@ -150884,27 +150972,27 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:192
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:208
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:204
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:200
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252
@@ -150914,35 +151002,35 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v1
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240
 ; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:256
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v29, 24, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v31, 24, v1
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:272
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -150954,8 +151042,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:316
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:304
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:300
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:296
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:300
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:296
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v1
 ; SI-NEXT:    s_waitcnt vmcnt(5)
@@ -150992,10 +151080,10 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 24, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:40
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -151003,15 +151091,15 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:108
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -151023,44 +151111,44 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:184
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:184
 ; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:216
 ; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:248
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:280
 ; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:312
 ; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:344
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:376
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:376
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:76
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:72
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:64
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:56
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:48
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:40
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:32
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr5
 ; SI-NEXT:    ; kill: killed $vgpr5
 ; SI-NEXT:    ; implicit-def: $vgpr5
@@ -151148,276 +151236,280 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 24, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 24, v54
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v56
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 24, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v54
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v60
 ; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
-; SI-NEXT:    v_or_b32_e32 v5, v6, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v6, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v6, v5, v6
+; SI-NEXT:    v_or_b32_e32 v12, v5, v9
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v12
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v12, v5, v10
+; SI-NEXT:    v_or_b32_e32 v16, v5, v9
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v12
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v14, v5, v10
+; SI-NEXT:    v_or_b32_e32 v20, v5, v9
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v14
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v20, v5, v10
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v20
+; SI-NEXT:    v_or_b32_e32 v22, v5, v9
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v22, v5, v10
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v22
+; SI-NEXT:    v_or_b32_e32 v24, v5, v9
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v24, v5, v10
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v24
+; SI-NEXT:    v_or_b32_e32 v28, v5, v9
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v28
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v28, v5, v10
+; SI-NEXT:    v_or_b32_e32 v29, v5, v9
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v28
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v29
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
@@ -151502,140 +151594,147 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v27
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v27, v5, v10
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
+; SI-NEXT:    v_or_b32_e32 v27, v5, v9
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
 ; SI-NEXT:    ; implicit-def: $vgpr27
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v8
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v6
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v8, v5, v8
+; SI-NEXT:    v_or_b32_e32 v6, v5, v6
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v8
-; SI-NEXT:    ; implicit-def: $vgpr8
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v6
+; SI-NEXT:    ; implicit-def: $vgpr6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v30
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v30, v5, v10
+; SI-NEXT:    v_or_b32_e32 v30, v5, v9
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v30
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v30
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v31
-; SI-NEXT:    ; implicit-def: $vgpr31
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v40
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT:    v_or_b32_e32 v10, v10, v9
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT:    v_or_b32_e32 v9, v57, v9
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v9, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v5, v19, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v53
-; SI-NEXT:    v_or_b32_e32 v9, v5, v26
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v9
+; SI-NEXT:    v_or_b32_e32 v19, v5, v26
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
 ; SI-NEXT:    ; implicit-def: $vgpr53
-; SI-NEXT:    ; implicit-def: $vgpr9
+; SI-NEXT:    ; implicit-def: $vgpr19
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v29, v5
+; SI-NEXT:    v_or_b32_e32 v5, v31, v5
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v15
@@ -151644,43 +151743,15 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr29
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v16, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v42
-; SI-NEXT:    v_or_b32_e32 v7, v5, v38
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v51, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v32
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v40
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_or_b32_e32 v10, v13, v10
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v49
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v19, v57, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v60
-; SI-NEXT:    v_or_b32_e32 v16, v13, v61
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v43
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v23, v62, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v58
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v13, v17, v13
-; SI-NEXT:    v_and_b32_e32 v17, 0xff, v46
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; SI-NEXT:    v_or_b32_e32 v18, v3, v17
-; SI-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v58
+; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; SI-NEXT:    v_or_b32_e32 v23, v17, v15
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v46
+; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; SI-NEXT:    v_or_b32_e32 v18, v3, v15
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v63
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v2
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
@@ -151772,78 +151843,96 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v25
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    v_or_b32_e32 v11, v41, v5
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v47
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    v_or_b32_e32 v15, v5, v44
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v36
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v59
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    v_or_b32_e32 v5, v45, v5
-; SI-NEXT:    v_or_b32_e32 v17, v63, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v16
+; SI-NEXT:    v_or_b32_e32 v17, v4, v3
 ; SI-NEXT:    ; implicit-def: $vgpr15
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    ; implicit-def: $vgpr63
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; kill: killed $vgpr2
+; SI-NEXT:    ; implicit-def: $vgpr31
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr2
+; SI-NEXT:    ; implicit-def: $vgpr4
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_or_b32_e32 v5, v33, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v42
+; SI-NEXT:    v_or_b32_e32 v7, v5, v38
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_or_b32_e32 v5, v51, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_or_b32_e32 v5, v8, v5
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_or_b32_e32 v11, v41, v5
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v47
+; SI-NEXT:    v_or_b32_e32 v8, v5, v44
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v36
+; SI-NEXT:    v_or_b32_e32 v14, v13, v61
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v43
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
+; SI-NEXT:    v_or_b32_e32 v5, v45, v5
+; SI-NEXT:    v_or_b32_e32 v13, v62, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
 ; SI-NEXT:    ; implicit-def: $vgpr7
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr21
-; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr14
+; SI-NEXT:    ; implicit-def: $vgpr8
 ; SI-NEXT:    ; implicit-def: $vgpr25
 ; SI-NEXT:    ; implicit-def: $vgpr47
 ; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr4
-; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    ; kill: killed $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr62
-; SI-NEXT:    ; implicit-def: $vgpr3
-; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:  .LBB88_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB88_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v2, v2, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v59
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v63
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_or_b32_e32 v2, v2, v5
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v59
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x300, v2
-; SI-NEXT:    v_or_b32_e32 v4, v63, v4
+; SI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v4, v2
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v58
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v56
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v46
@@ -151866,359 +151955,358 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v40
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v54
-; SI-NEXT:    v_or_b32_e32 v5, v6, v5
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v49
-; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v54
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
-; SI-NEXT:    v_or_b32_e32 v6, v57, v6
+; SI-NEXT:    v_or_b32_e32 v9, v57, v9
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; SI-NEXT:    v_or_b32_e32 v5, v6, v5
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v47
-; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT:    v_or_b32_e32 v5, v9, v5
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v47
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v36
-; SI-NEXT:    v_or_b32_e32 v6, v44, v6
+; SI-NEXT:    v_or_b32_e32 v9, v44, v9
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v45, v10
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v10, v6
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v32
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_or_b32_e32 v9, v10, v9
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v14
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT:    v_lshlrev_b32_e32 v11, 8, v31
-; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; SI-NEXT:    v_or_b32_e32 v8, v8, v10
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v25
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
+; SI-NEXT:    v_or_b32_e32 v10, v41, v10
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; SI-NEXT:    v_or_b32_e32 v8, v10, v8
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v42
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v21
+; SI-NEXT:    v_or_b32_e32 v10, v38, v10
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
-; SI-NEXT:    v_or_b32_e32 v11, v41, v11
+; SI-NEXT:    v_or_b32_e32 v11, v51, v11
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v42
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v15
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v21
-; SI-NEXT:    v_or_b32_e32 v11, v38, v11
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
-; SI-NEXT:    v_or_b32_e32 v12, v51, v12
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v12, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v15
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT:    v_or_b32_e32 v7, v7, v12
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v7, v7, v11
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
-; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v6
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; SI-NEXT:    v_or_b32_e32 v11, v33, v11
+; SI-NEXT:    v_or_b32_e32 v11, v11, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v53
+; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT:    v_or_b32_e32 v7, v26, v7
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v12, v16, v12
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
+; SI-NEXT:    v_or_b32_e32 v12, v31, v12
+; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v7
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v53
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_or_b32_e32 v7, v13, v7
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
+; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    v_or_b32_e32 v7, v26, v7
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
-; SI-NEXT:    v_or_b32_e32 v13, v29, v13
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_or_b32_e32 v13, v19, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v14, 8, v14
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v30
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v14, v7
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; SI-NEXT:    v_or_b32_e32 v9, v9, v14
-; SI-NEXT:    v_or_b32_e32 v14, v9, v7
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v30
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v7, v9, v7
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v15, v9
-; SI-NEXT:    v_or_b32_e32 v15, v9, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v14, v15, v14
+; SI-NEXT:    v_or_b32_e32 v14, v14, v7
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; SI-NEXT:    v_or_b32_e32 v7, v9, v7
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
+; SI-NEXT:    v_or_b32_e32 v7, v15, v7
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v16, v9
-; SI-NEXT:    v_or_b32_e32 v16, v9, v7
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v8
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_or_b32_e32 v15, v15, v7
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    v_or_b32_e32 v17, v8, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT:    v_or_b32_e32 v7, v16, v7
+; SI-NEXT:    v_or_b32_e32 v16, v7, v6
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    v_or_b32_e32 v18, v8, v7
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT:    v_or_b32_e32 v7, v17, v7
+; SI-NEXT:    v_or_b32_e32 v17, v7, v6
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v27
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    v_or_b32_e32 v19, v8, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT:    v_or_b32_e32 v7, v18, v7
+; SI-NEXT:    v_or_b32_e32 v18, v7, v6
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    v_or_b32_e32 v20, v8, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v13
-; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v17
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v1, v7, v1
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_or_b32_e32 v21, v7, v1
+; SI-NEXT:    v_or_b32_e32 v7, v19, v7
+; SI-NEXT:    v_or_b32_e32 v19, v7, v6
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_or_b32_e32 v1, v6, v1
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    v_or_b32_e32 v20, v6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v21
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT:    v_or_b32_e32 v1, v7, v1
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
+; SI-NEXT:    v_or_b32_e32 v1, v6, v1
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_or_b32_e32 v22, v7, v1
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    v_or_b32_e32 v21, v6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v1, v7, v1
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v1, v6, v1
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_or_b32_e32 v23, v7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    v_or_b32_e32 v22, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT:    v_or_b32_e32 v1, v7, v1
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
+; SI-NEXT:    v_or_b32_e32 v1, v6, v1
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_or_b32_e32 v24, v7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    v_or_b32_e32 v23, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v1, v7, v1
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v1, v6, v1
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_or_b32_e32 v25, v7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v25
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    v_or_b32_e32 v24, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT:    v_or_b32_e32 v1, v7, v1
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
+; SI-NEXT:    v_or_b32_e32 v1, v6, v1
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_or_b32_e32 v26, v7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v12
-; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v16
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v20
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v24
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT:    v_or_b32_e32 v6, v7, v6
+; SI-NEXT:    v_or_b32_e32 v25, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v10
+; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v13
+; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v16
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v19
+; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v22
+; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v25
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v27, v7, v1
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v26, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
-; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; SI-NEXT:    v_or_b32_e32 v24, v25, v24
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v28, v7, v1
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v27, v6, v1
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v9
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v12
+; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v15
+; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v18
+; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v21
+; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v24
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v3
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v10
-; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v14
-; SI-NEXT:    v_add_i32_e32 v14, vcc, s7, v18
-; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v22
-; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v26
+; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v8
+; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v11
+; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v14
+; SI-NEXT:    v_add_i32_e32 v14, vcc, s7, v17
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v20
+; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v23
+; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v26
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
-; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v11
-; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v15
-; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v19
-; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v23
-; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v27
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    v_or_b32_e32 v23, v28, v23
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v23, v27, v23
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; SI-NEXT:    v_or_b32_e32 v24, v25, v24
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
+; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
 ; SI-NEXT:    v_or_b32_e32 v24, v25, v24
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
@@ -152229,7 +152317,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v25, v26, v25
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -152258,7 +152346,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v27, v28, v27
 ; SI-NEXT:    v_or_b32_e32 v26, v27, v26
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, s7, v26
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
@@ -152294,7 +152382,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v29, v30, v29
 ; SI-NEXT:    v_or_b32_e32 v28, v29, v28
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v28
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
@@ -152330,7 +152418,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v31, v32, v31
 ; SI-NEXT:    v_or_b32_e32 v30, v31, v30
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, s7, v30
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
@@ -152368,7 +152456,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v32, vcc, s7, v32
 ; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v32
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v31
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
@@ -152378,27 +152466,27 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
 ; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
 ; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v23
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v21
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
 ; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v18
@@ -152409,78 +152497,78 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v17
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v7
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
-; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v6
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v6
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v2
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v1
+; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
 ; SI-NEXT:  .LBB88_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152498,7 +152586,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -152509,7 +152597,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v52
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -152518,8 +152606,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152529,7 +152617,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v50
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -152538,8 +152626,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152549,7 +152637,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v48
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -152558,8 +152646,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152569,7 +152657,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v39
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -152578,8 +152666,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152589,7 +152677,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v37
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -152599,7 +152687,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152618,7 +152706,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -152629,7 +152717,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v34
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -152638,8 +152726,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152649,8 +152737,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v33
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152658,8 +152746,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152670,7 +152758,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v28
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152678,7 +152766,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -152689,8 +152777,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v24
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v22
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152698,8 +152786,8 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152710,7 +152798,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v22
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152730,7 +152818,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v20
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -152749,35 +152837,35 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v5
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v14
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v12
 ; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v19
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v9
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v10
 ; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v23
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v13
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v12
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v32
 ; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v18
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v13
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v23
 ; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v17
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v6
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v29
 ; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
 ; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
@@ -152852,39 +152940,39 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:136
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:144
-; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:152
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:128
+; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:136
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:144
+; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:152
 ; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:160
 ; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:168
 ; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:176
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
 ; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v1
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v25
 ; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v3
 ; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v5
 ; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v7
-; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v9
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v9
 ; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v13
-; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v15
+; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v15
 ; VI-NEXT:    v_lshlrev_b16_e32 v35, 8, v17
 ; VI-NEXT:    v_lshlrev_b16_e32 v36, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v21
 ; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v23
+; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v27
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
@@ -152898,46 +152986,47 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v27
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
 ; VI-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v37
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v38
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v39
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v48
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v49
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v50
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v51
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
@@ -152954,7 +153043,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v43
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:208
@@ -152962,15 +153051,15 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
@@ -152993,20 +153082,20 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -153019,20 +153108,20 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -153040,17 +153129,17 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
@@ -153066,17 +153155,17 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:316
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
@@ -153091,44 +153180,44 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:376
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:348
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v1
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v1
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v2
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v3
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:372
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:12
 ; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:108
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
 ; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -153154,31 +153243,31 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v4, v4, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v4, v4, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr54
 ; VI-NEXT:    ; implicit-def: $vgpr55
 ; VI-NEXT:    ; implicit-def: $vgpr40
 ; VI-NEXT:    ; implicit-def: $vgpr41
-; VI-NEXT:    ; implicit-def: $vgpr49
+; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr36
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr34
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr50
 ; VI-NEXT:    ; implicit-def: $vgpr52
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr37
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -153197,35 +153286,35 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_or_b32_sdwa v10, v63, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v61, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_or_b32_sdwa v12, v58, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v59, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_or_b32_sdwa v14, v45, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr58
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    ; implicit-def: $vgpr45
 ; VI-NEXT:    s_waitcnt vmcnt(2)
@@ -153234,26 +153323,26 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v63, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr62
-; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v60, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr59
+; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v10, v57, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v56, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr56
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v58, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr56
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -153262,35 +153351,35 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v13, v44, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -153305,20 +153394,20 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -153326,189 +153415,189 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v31, v31, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr48
+; VI-NEXT:    v_or_b32_sdwa v31, v31, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr49
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v32, v32, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr53
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v30, v30, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v30, v30, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v30, v30, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v30, v30, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr39
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v31, v31, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v31, v31, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr32
 ; VI-NEXT:    ; kill: killed $vgpr32
@@ -153613,7 +153702,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; implicit-def: $vgpr32
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr50
+; VI-NEXT:    ; implicit-def: $vgpr51
 ; VI-NEXT:  .LBB88_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB88_4
@@ -153621,53 +153710,51 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v18, 0x300
-; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
 ; VI-NEXT:    v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(12)
 ; VI-NEXT:    v_or_b32_sdwa v29, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(13)
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_sdwa v0, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v2, 0x300, v3
 ; VI-NEXT:    v_or_b32_e32 v0, v2, v0
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v2, 3, v2
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v2, v52, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v2, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
 ; VI-NEXT:    v_or_b32_e32 v2, v3, v2
@@ -153677,19 +153764,19 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v1, 0x300, v1
 ; VI-NEXT:    v_or_b32_e32 v1, v1, v4
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v3, v49, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v48, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v3, v3, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
-; VI-NEXT:    v_or_b32_sdwa v4, v39, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
 ; VI-NEXT:    v_or_b32_e32 v3, v4, v3
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
@@ -153697,14 +153784,13 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
 ; VI-NEXT:    v_or_b32_sdwa v6, v33, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    s_waitcnt vmcnt(2)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v5, 3, v5
 ; VI-NEXT:    v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
 ; VI-NEXT:    v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v4, v4, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
@@ -153721,7 +153807,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v32, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_sdwa v32, v32, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
@@ -153729,7 +153815,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; VI-NEXT:    v_or_b32_e32 v28, v28, v32
 ; VI-NEXT:    s_waitcnt vmcnt(1)
@@ -153737,79 +153823,94 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v33, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v33, v33, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
 ; VI-NEXT:    v_or_b32_e32 v27, v27, v33
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v34, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v34, v34, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; VI-NEXT:    v_or_b32_e32 v26, v26, v34
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v35, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v35, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
+; VI-NEXT:    v_or_b32_e32 v25, v25, v35
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v36, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT:    v_or_b32_e32 v25, v25, v35
+; VI-NEXT:    v_add_u16_sdwa v36, v36, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v36, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_sdwa v36, v36, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v24, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
+; VI-NEXT:    v_or_b32_e32 v24, v24, v36
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v24, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v37, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v23, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
 ; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT:    v_or_b32_e32 v24, v24, v36
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
+; VI-NEXT:    v_or_b32_e32 v23, v23, v37
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v8, 3, v8
 ; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v7, v8, v7
-; VI-NEXT:    v_add_u16_e32 v8, 3, v61
+; VI-NEXT:    v_add_u16_e32 v8, 3, v63
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v9, 3, v62
@@ -153818,30 +153919,30 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; VI-NEXT:    v_or_b32_e32 v8, v9, v8
-; VI-NEXT:    v_add_u16_e32 v9, 3, v63
+; VI-NEXT:    v_add_u16_e32 v9, 3, v61
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v10, 3, v59
+; VI-NEXT:    v_add_u16_e32 v10, 3, v60
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v9, v9, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_or_b32_e32 v9, v10, v9
-; VI-NEXT:    v_add_u16_e32 v10, 3, v60
+; VI-NEXT:    v_add_u16_e32 v10, 3, v57
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v11, 3, v57
+; VI-NEXT:    v_add_u16_e32 v11, 3, v56
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v10, v10, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_or_b32_e32 v10, v11, v10
-; VI-NEXT:    v_add_u16_e32 v11, 3, v58
+; VI-NEXT:    v_add_u16_e32 v11, 3, v59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v12, 3, v56
+; VI-NEXT:    v_add_u16_e32 v12, 3, v58
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; VI-NEXT:    v_or_b32_e32 v11, v12, v11
@@ -153850,7 +153951,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v13, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_or_b32_e32 v12, v13, v12
@@ -153859,7 +153960,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 3, v44
 ; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v13, v13, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_or_b32_e32 v13, v14, v13
@@ -153871,35 +153972,35 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; VI-NEXT:    v_or_b32_e32 v14, v15, v14
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v15, 3, v15
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    v_or_b32_e32 v15, v16, v15
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    v_or_b32_sdwa v17, v19, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -153908,54 +154009,43 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v16, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v19, 0x300, v20
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; VI-NEXT:    v_or_b32_e32 v16, v19, v16
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
-; VI-NEXT:    v_or_b32_sdwa v30, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v30, v39, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
-; VI-NEXT:    v_or_b32_sdwa v31, v50, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v31, v51, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v31, 0x300, v31
-; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
-; VI-NEXT:    v_or_b32_sdwa v21, v37, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v37, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v40, v21, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v29, v29, v40
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v23, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT:    v_or_b32_e32 v23, v23, v37
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v38, v38, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_e32 v20, 3, v20
-; VI-NEXT:    v_or_b32_sdwa v20, v48, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_sdwa v55, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v38, v38, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_or_b32_e32 v30, v30, v55
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v22, v39, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v20, 3, v20
+; VI-NEXT:    v_or_b32_sdwa v20, v49, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v55, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; VI-NEXT:    v_or_b32_e32 v22, v22, v38
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_e32 v30, v30, v55
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v39, 3, v39
 ; VI-NEXT:    v_or_b32_sdwa v39, v48, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
@@ -153977,7 +154067,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v39, 3, v39
 ; VI-NEXT:    v_or_b32_sdwa v39, v49, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -154079,17 +154169,16 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
 ; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:136
+; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:136
 ; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:144
 ; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:152
 ; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:160
 ; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:168
 ; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:176
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v1
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v27
-; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v29
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v3
@@ -154097,81 +154186,81 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v7
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v9
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v11
-; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v13
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v13
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v15
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v17
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v35, 8, v19
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v21
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v23
-; GFX9-NEXT:    s_waitcnt vmcnt(27)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
+; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v25
 ; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v25
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v37
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v39
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v48
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v49
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
@@ -154179,7 +154268,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
@@ -154196,7 +154285,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v43
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
@@ -154205,15 +154294,15 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
@@ -154237,20 +154326,20 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -154264,20 +154353,20 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -154286,17 +154375,17 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
@@ -154323,7 +154412,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
@@ -154343,48 +154432,48 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:372
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:12
 ; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:100
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB88_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -154393,9 +154482,9 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
@@ -154424,10 +154513,10 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; implicit-def: $vgpr51
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr48
+; GFX9-NEXT:    ; implicit-def: $vgpr39
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
@@ -154438,7 +154527,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr34
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -154453,93 +154542,93 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v58, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v15, v43, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr43
+; GFX9-NEXT:    v_or_b32_sdwa v15, v42, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v63, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v62, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr59
+; GFX9-NEXT:    ; implicit-def: $vgpr61
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v57, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr46
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v43, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr42
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -154554,20 +154643,20 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
@@ -154575,58 +154664,58 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
@@ -154644,22 +154733,22 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
 ; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr53
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr37
 ; GFX9-NEXT:    ; kill: killed $vgpr37
 ; GFX9-NEXT:    ; implicit-def: $vgpr37
@@ -154756,7 +154845,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr49
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -154868,27 +154957,27 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_cbranch_execz .LBB88_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
@@ -154896,19 +154985,20 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
 ; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    s_waitcnt vmcnt(20)
+; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(17)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v2
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v3
-; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    v_perm_b32 v0, v2, v0, s6
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(17)
+; GFX9-NEXT:    s_waitcnt vmcnt(16)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
@@ -154933,7 +155023,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT:    v_or_b32_sdwa v3, v48, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
@@ -154959,7 +155049,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_or_b32_sdwa v36, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
@@ -154983,11 +155073,11 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v37, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v37, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT:    v_or_b32_sdwa v21, v39, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v21, v48, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v33, 0x300, v21
 ; GFX9-NEXT:    v_add_u16_e32 v34, 0x300, v23
 ; GFX9-NEXT:    v_perm_b32 v29, v34, v29, s6
@@ -154995,16 +155085,17 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v38, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
@@ -155014,7 +155105,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v39, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
@@ -155024,7 +155115,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v48, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -155037,45 +155128,45 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v62
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v60
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v58
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v57
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v60
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v56
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v57
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v56
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v58
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v46
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v46
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
@@ -155084,63 +155175,63 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 3, v44
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v43
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v15, 3, v43
+; GFX9-NEXT:    v_add_u16_e32 v15, 3, v42
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v16
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v18
 ; GFX9-NEXT:    v_perm_b32 v17, v17, v20, s6
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v19
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v16, v18, v16, s6
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v49, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v49, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v20
 ; GFX9-NEXT:    v_perm_b32 v30, v33, v30, s6
@@ -155148,7 +155239,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v50, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v52, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -155158,14 +155249,14 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v51, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v51
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v52, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -155175,7 +155266,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v53, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v53
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -155188,7 +155279,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v55, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v55
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -155196,7 +155287,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v40, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v50, 0x300, v40
 ; GFX9-NEXT:    v_perm_b32 v21, v50, v21, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -155204,14 +155295,14 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v41, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v41
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v42, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v51, 0x300, v42
 ; GFX9-NEXT:    v_perm_b32 v20, v51, v20, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -155219,7 +155310,7 @@ define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v43, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v43
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -159129,17 +159220,17 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; VI-NEXT:    v_mov_b32_e32 v45, v62
 ; VI-NEXT:    v_or_b32_sdwa v16, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_or_b32_sdwa v3, v51, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v17, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_mov_b32_e32 v54, v22
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_mov_b32_e32 v32, v1
 ; VI-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_mov_b32_e32 v54, v22
 ; VI-NEXT:    v_mov_b32_e32 v41, v24
 ; VI-NEXT:    v_or_b32_sdwa v18, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
@@ -160376,8 +160467,8 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; GFX9-NEXT:    s_cbranch_vccnz .LBB89_5
 ; GFX9-NEXT:  ; %bb.4: ; %cmp.true
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_add_i32 s28, s28, 3
 ; GFX9-NEXT:    s_and_b32 s4, s28, 0xff
 ; GFX9-NEXT:    s_lshl_b32 s5, s29, 8
@@ -160410,7 +160501,7 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(13)
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -160460,9 +160551,9 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; GFX9-NEXT:    s_or_b32 s7, s8, s7
 ; GFX9-NEXT:    s_and_b32 s8, s16, 0xff
 ; GFX9-NEXT:    s_or_b32 s8, s9, s8
+; GFX9-NEXT:    s_and_b32 s9, s18, 0xff
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT:    s_and_b32 s9, s18, 0xff
 ; GFX9-NEXT:    s_or_b32 s9, s10, s9
 ; GFX9-NEXT:    s_addk_i32 s4, 0x300
 ; GFX9-NEXT:    s_addk_i32 s5, 0x300
@@ -160495,11 +160586,11 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; GFX9-NEXT:    v_add_u32_e32 v1, 0x300, v1
 ; GFX9-NEXT:    v_lshl_or_b32 v6, v1, 16, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -160510,11 +160601,11 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; GFX9-NEXT:    v_add_u32_e32 v1, 0x300, v1
 ; GFX9-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -160645,9 +160736,9 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
 ; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
 ; GFX9-NEXT:    v_or_b32_sdwa v51, v34, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -160724,22 +160815,21 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v43, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v0
 ; GFX9-NEXT:    v_add_u32_e32 v52, 0x300, v43
 ; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s8
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v44, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v33
@@ -164579,8 +164669,8 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; VI-NEXT:    ; implicit-def: $vgpr35
 ; VI-NEXT:    ; implicit-def: $vgpr45
@@ -165758,13 +165848,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 8, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 12, v0
@@ -165856,13 +165947,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
@@ -165882,13 +165974,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v16, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
@@ -165908,13 +166001,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v18, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
@@ -165934,13 +166028,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v20, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
@@ -165959,13 +166054,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
@@ -165984,13 +166080,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
@@ -166007,13 +166104,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
@@ -166033,13 +166131,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
@@ -166117,8 +166216,8 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_mov_b32_e32 v46, v15
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    ; kill: killed $vgpr33
@@ -166303,7 +166402,7 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(34)
+; GFX9-NEXT:    s_waitcnt vmcnt(33)
 ; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v15
 ; GFX9-NEXT:    ; implicit-def: $vgpr15
 ; GFX9-NEXT:    ; kill: killed $vgpr15
@@ -166328,7 +166427,6 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(36)
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v32
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
@@ -166757,7 +166855,6 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
 ; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
 ; GFX9-NEXT:    v_cndmask_b32_e32 v15, v16, v17, vcc
 ; GFX9-NEXT:    v_perm_b32 v33, v15, v25, s7
-; GFX9-NEXT:    s_waitcnt vmcnt(52)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v32
 ; GFX9-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
 ; GFX9-NEXT:    v_bfe_u32 v16, v15, 16, 1
@@ -178671,15 +178768,18 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v32, v30
 ; SI-NEXT:    v_mov_b32_e32 v44, v19
 ; SI-NEXT:    v_mov_b32_e32 v43, v17
+; SI-NEXT:    v_mov_b32_e32 v32, v14
+; SI-NEXT:    v_mov_b32_e32 v14, v12
+; SI-NEXT:    v_mov_b32_e32 v12, v10
 ; SI-NEXT:    v_mov_b32_e32 v41, v7
 ; SI-NEXT:    v_mov_b32_e32 v55, v5
 ; SI-NEXT:    v_mov_b32_e32 v54, v3
 ; SI-NEXT:    v_mov_b32_e32 v51, v1
-; SI-NEXT:    v_mov_b32_e32 v30, v0
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:392
+; SI-NEXT:    v_mov_b32_e32 v10, v0
+; SI-NEXT:    s_waitcnt expcnt(1)
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:392
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:20
@@ -178689,13 +178789,13 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:60
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:76
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:92
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:116
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:116
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:124
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:132
 ; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:140
@@ -178706,152 +178806,147 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:180
 ; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:188
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:128
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:136
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v4
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v18
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v20
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v22
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v26
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v28
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v30
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 8, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v14, 8, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 8, v32
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v17
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v49
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v50
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v52
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v42
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v53
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v0
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v25
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:168
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v27
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v29
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v40
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v42
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v33
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:144
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v17
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v34
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:136
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v23
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v35
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v31
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v36
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v31
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    ; implicit-def: $vgpr31
+; SI-NEXT:    ; implicit-def: $vgpr5
+; SI-NEXT:    ; implicit-def: $vgpr27
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:160
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:168
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:176
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v38
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v23
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v39
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v29
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v48
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:176
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v53
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v40
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr38
-; SI-NEXT:    ; implicit-def: $vgpr27
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr23
-; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr29
-; SI-NEXT:    ; implicit-def: $vgpr25
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v4
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v10
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v14
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v16
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:144
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v18
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v20
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v22
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v24
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v26
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v28
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:160
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:220
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:200
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:208
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
@@ -178871,7 +178966,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:224
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:232
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -178882,13 +178977,13 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
@@ -178929,15 +179024,15 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:296
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:304
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
@@ -178980,7 +179075,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v1
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v2
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v3
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -178989,7 +179084,6 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:360
-; SI-NEXT:    ; implicit-def: $vgpr4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -178997,35 +179091,35 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:48
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:384
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:384
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:40
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:120
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:152
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -179040,14 +179134,16 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:312
 ; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:344
 ; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:376
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:56
+; SI-NEXT:    ; implicit-def: $vgpr0
+; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
@@ -179152,75 +179248,83 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB92_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v51
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v2, v17
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v54
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v51
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT:    v_and_b32_e32 v48, 0xff, v50
+; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
+; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
+; SI-NEXT:    v_and_b32_e32 v34, 0xff, v34
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; SI-NEXT:    v_and_b32_e32 v39, 0xff, v50
+; SI-NEXT:    v_and_b32_e32 v48, 0xff, v40
 ; SI-NEXT:    v_and_b32_e32 v49, 0xff, v49
 ; SI-NEXT:    v_and_b32_e32 v52, 0xff, v52
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v42, 0xff, v42
-; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_and_b32_e32 v46, 0xff, v46
 ; SI-NEXT:    v_or_b32_e32 v45, v46, v45
-; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_and_b32_e32 v56, 0xff, v56
 ; SI-NEXT:    v_or_b32_e32 v56, v56, v61
-; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_and_b32_e32 v57, 0xff, v57
+; SI-NEXT:    v_and_b32_e32 v47, 0xff, v47
 ; SI-NEXT:    v_or_b32_e32 v1, v57, v1
+; SI-NEXT:    v_or_b32_e32 v3, v47, v3
 ; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v2, v0
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v54
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v2, v0
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v55
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v2, v6
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v41
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v2, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v9
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v35
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v2, v0
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v11
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v2, v12
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v11
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v0, v2, v14
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v13
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v24
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v17, v2, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; SI-NEXT:    v_mov_b32_e32 v8, v7
+; SI-NEXT:    v_mov_b32_e32 v7, v19
+; SI-NEXT:    v_or_b32_e32 v19, v2, v32
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v15
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT:    v_and_b32_e32 v35, 0xff, v7
+; SI-NEXT:    v_and_b32_e32 v7, 0xff, v53
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr19
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_or_b32_e32 v31, v2, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v17, v2, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v43
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v31
-; SI-NEXT:    v_cvt_f32_f16_e32 v31, v45
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v23, v2, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v44
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v23
@@ -179228,11 +179332,11 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v29, v2, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v31, v2, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v21
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v29
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v31
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:976 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
@@ -179244,231 +179348,227 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v25, v2, v0
+; SI-NEXT:    v_or_b32_e32 v27, v2, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v25
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v27
+; SI-NEXT:    v_cvt_f32_f16_e32 v27, v3
+; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v27, v2, v0
+; SI-NEXT:    v_or_b32_e32 v29, v2, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v27
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v29
 ; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; SI-NEXT:    ; implicit-def: $vgpr17
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v2, v0
+; SI-NEXT:    v_or_b32_e32 v2, v2, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
-; SI-NEXT:    v_or_b32_e32 v2, v2, v32
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v5, v5, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v5
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, v1
-; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr4
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v33, v6, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v33
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v33, v45
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; SI-NEXT:    v_or_b32_e32 v6, v6, v7
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v6, v6, v8
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v6
 ; SI-NEXT:    ; implicit-def: $vgpr6
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v35, v8, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v10
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v35
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v10, v8, v0
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v8, v8, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v10
-; SI-NEXT:    ; implicit-def: $vgpr10
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v8
+; SI-NEXT:    ; implicit-def: $vgpr8
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v34, v8, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v14
-; SI-NEXT:    v_or_b32_e32 v14, v8, v53
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v16
-; SI-NEXT:    v_or_b32_e32 v16, v8, v40
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v20
+; SI-NEXT:    v_or_b32_e32 v12, v12, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v34
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v12
+; SI-NEXT:    ; implicit-def: $vgpr12
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v14, v14, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v14
 ; SI-NEXT:    ; implicit-def: $vgpr14
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v32, 0xff, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v32, v32, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_or_b32_e32 v20, v8, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v32
+; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_or_b32_e32 v18, v18, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v20
-; SI-NEXT:    ; implicit-def: $vgpr20
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v18
+; SI-NEXT:    ; implicit-def: $vgpr18
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v8, v8, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v22, v22, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v8
-; SI-NEXT:    ; implicit-def: $vgpr8
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v24, v12, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v26
+; SI-NEXT:    v_or_b32_e32 v24, v24, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v24
 ; SI-NEXT:    ; implicit-def: $vgpr24
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v26, v12, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v28
+; SI-NEXT:    v_or_b32_e32 v26, v26, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v26
 ; SI-NEXT:    ; implicit-def: $vgpr26
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v28, v12, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v28, v28, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v28
 ; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v12, v12, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v30, v30, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v30
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v18, v18, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v34, v34, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v18
-; SI-NEXT:    ; implicit-def: $vgpr18
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v34
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_or_b32_e32 v16, v16, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v16
+; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v22, v22, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v20, v20, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v22
-; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v20
+; SI-NEXT:    ; implicit-def: $vgpr20
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v32, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v32, v32, v0
+; SI-NEXT:    v_or_b32_e32 v35, v35, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v32
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v35
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v36, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v36, v36, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v36
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v37, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v37, v37, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v37
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v38, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v38, v38, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v38
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v39, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v39, v39, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v39
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v7
 ; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v48, v48, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v48
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v49, v49, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v49
 ; SI-NEXT:    ; implicit-def: $vgpr49
@@ -179478,21 +179578,21 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v50, v50, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v50
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v9
 ; SI-NEXT:    ; implicit-def: $vgpr9
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v54, 0xff, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
@@ -179505,24 +179605,24 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v52, v52, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v52
 ; SI-NEXT:    ; implicit-def: $vgpr52
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v11
 ; SI-NEXT:    ; implicit-def: $vgpr11
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v53, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v53, v53, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
@@ -179531,7 +179631,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr53
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v55, 0xff, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v55, v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
@@ -179553,29 +179653,33 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v40, v40, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v39, v40
+; SI-NEXT:    v_cvt_f32_f16_e32 v48, v40
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v15
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v42, v42, v15
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v42
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v38, v42
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v43, 0xff, v15
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v43, v43, v15
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v43
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v21
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v34, v15
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v36, v15
 ; SI-NEXT:    ; implicit-def: $vgpr15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v44, 0xff, v21
@@ -179583,96 +179687,69 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v44, v44, v21
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v44
-; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v46, 0xff, v21
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v46, v46, v58
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v46
+; SI-NEXT:    v_cvt_f32_f16_e32 v39, v46
 ; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v58, 0xff, v21
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v58, v58, v59
-; SI-NEXT:    v_cvt_f32_f16_e32 v27, v58
+; SI-NEXT:    v_cvt_f32_f16_e32 v29, v58
 ; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v59, 0xff, v21
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v59, v59, v60
-; SI-NEXT:    v_cvt_f32_f16_e32 v36, v59
+; SI-NEXT:    v_cvt_f32_f16_e32 v37, v59
 ; SI-NEXT:    ; implicit-def: $vgpr59
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v60, 0xff, v21
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v60, v60, v62
-; SI-NEXT:    v_cvt_f32_f16_e32 v33, v60
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v60
 ; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v61, 0xff, v21
-; SI-NEXT:    v_or_b32_e32 v21, v61, v47
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v21
-; SI-NEXT:    ; implicit-def: $vgpr21
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v25, v61, v25
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v25
+; SI-NEXT:    ; implicit-def: $vgpr25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v61, 0xff, v47
-; SI-NEXT:    v_and_b32_e32 v47, 0xff, v19
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v61, 0xff, v21
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v61, v61, v63
-; SI-NEXT:    v_or_b32_e32 v3, v47, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v29, v61
-; SI-NEXT:    v_cvt_f32_f16_e32 v25, v3
+; SI-NEXT:    v_cvt_f32_f16_e32 v31, v61
 ; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr47
 ; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
+; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
+; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v55
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v13
-; SI-NEXT:    ; implicit-def: $vgpr13
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v41
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
+; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
+; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, v43
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v44
+; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
@@ -179820,68 +179897,92 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
+; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v55
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; SI-NEXT:    ; implicit-def: $vgpr19
+; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v13
+; SI-NEXT:    ; implicit-def: $vgpr13
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v41
+; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
+; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr21
 ; SI-NEXT:  .LBB92_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB92_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v19
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v47
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v4, v3, v2
+; SI-NEXT:    v_or_b32_e32 v5, v3, v2
 ; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v5, v1, v2
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_or_b32_e32 v1, v1, v2
+; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v19
+; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
+; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_add_i32_e32 v34, vcc, 3, v34
+; SI-NEXT:    v_and_b32_e32 v34, 0xff, v34
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
+; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
 ; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
 ; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
 ; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
-; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v35, vcc, 3, v35
-; SI-NEXT:    v_and_b32_e32 v35, 0xff, v35
-; SI-NEXT:    v_mov_b32_e32 v1, v17
+; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
+; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_mov_b32_e32 v17, v43
 ; SI-NEXT:    v_mov_b32_e32 v19, v44
+; SI-NEXT:    v_add_i32_e32 v47, vcc, 3, v19
+; SI-NEXT:    v_and_b32_e32 v47, 0xff, v47
 ; SI-NEXT:    v_add_i32_e32 v57, vcc, 3, v15
 ; SI-NEXT:    v_and_b32_e32 v57, 0xff, v57
-; SI-NEXT:    s_movk_i32 s6, 0x300
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v51
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    s_movk_i32 s6, 0x300
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
@@ -179893,17 +179994,15 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v25, v47, v2
-; SI-NEXT:    v_add_i32_e32 v47, vcc, 3, v19
-; SI-NEXT:    v_and_b32_e32 v47, 0xff, v47
+; SI-NEXT:    v_or_b32_e32 v25, v25, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v27, v62, v2
+; SI-NEXT:    v_or_b32_e32 v31, v62, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v56
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v29, v61, v2
+; SI-NEXT:    v_or_b32_e32 v33, v61, v2
 ; SI-NEXT:    v_add_i32_e32 v56, vcc, 3, v17
 ; SI-NEXT:    v_and_b32_e32 v56, 0xff, v56
 ; SI-NEXT:    v_add_i32_e32 v61, vcc, 3, v41
@@ -179914,56 +180013,66 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v31, v60, v2
+; SI-NEXT:    v_or_b32_e32 v35, v60, v2
 ; SI-NEXT:    v_add_i32_e32 v60, vcc, 3, v9
 ; SI-NEXT:    v_and_b32_e32 v60, 0xff, v60
+; SI-NEXT:    v_or_b32_e32 v12, v12, v60
+; SI-NEXT:    v_add_i32_e32 v60, vcc, s6, v25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v33, v59, v2
+; SI-NEXT:    v_or_b32_e32 v37, v59, v2
 ; SI-NEXT:    v_add_i32_e32 v59, vcc, 3, v11
 ; SI-NEXT:    v_and_b32_e32 v59, 0xff, v59
-; SI-NEXT:    v_or_b32_e32 v11, v12, v59
-; SI-NEXT:    v_add_i32_e32 v59, vcc, s6, v27
-; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
+; SI-NEXT:    v_or_b32_e32 v14, v14, v59
+; SI-NEXT:    v_add_i32_e32 v59, vcc, s6, v31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v34, v58, v2
+; SI-NEXT:    v_or_b32_e32 v38, v58, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v46
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_or_b32_e32 v36, v45, v2
+; SI-NEXT:    v_or_b32_e32 v39, v45, v2
 ; SI-NEXT:    v_add_i32_e32 v46, vcc, 3, v21
 ; SI-NEXT:    v_and_b32_e32 v46, 0xff, v46
 ; SI-NEXT:    v_add_i32_e32 v58, vcc, 3, v13
 ; SI-NEXT:    v_and_b32_e32 v58, 0xff, v58
+; SI-NEXT:    v_or_b32_e32 v32, v32, v58
+; SI-NEXT:    v_add_i32_e32 v58, vcc, s6, v33
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v32
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:980 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v37, v0, v2
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v48, v0, v2
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v38, v0, v2
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v39, v0, v2
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v42
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v48, v0, v2
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
@@ -179971,7 +180080,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -179980,7 +180089,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -179989,7 +180098,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -179998,618 +180107,622 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
-; SI-NEXT:    v_mov_b32_e32 v2, v8
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v52
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v52
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v49
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v49
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v50
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v40
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v50
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_or_b32_e32 v2, v0, v2
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
+; SI-NEXT:    v_or_b32_e32 v0, v0, v27
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v29
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v22
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v20, v0, v20
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v16, v0, v16
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v0, v34
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v18
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
+; SI-NEXT:    v_or_b32_e32 v0, v0, v30
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v8, v0, v8
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v8
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v28
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v26
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v24
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v49, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v49, 0xff, v49
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v49
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v22
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v20
+; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v20
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v18
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v40, v16
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v36, vcc, 3, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v36, 0xff, v36
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v0, v36
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v53, v14
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v49, vcc, 3, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v49, 0xff, v49
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v0, v49
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v50, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v50, 0xff, v50
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v50
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v10
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v52, vcc, 3, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v52, 0xff, v52
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v35
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v0, v0, v52
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v52, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v52, 0xff, v52
-; SI-NEXT:    v_or_b32_e32 v7, v7, v52
-; SI-NEXT:    v_add_i32_e32 v54, vcc, s6, v7
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v53, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v53, 0xff, v53
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v53
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v53, 0xff, v53
+; SI-NEXT:    v_or_b32_e32 v7, v7, v53
+; SI-NEXT:    v_add_i32_e32 v54, vcc, s6, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v40, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v40, 0xff, v40
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v40
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v0, v4
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v42, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v42, 0xff, v42
-; SI-NEXT:    v_or_b32_e32 v32, v32, v42
-; SI-NEXT:    v_add_i32_e32 v42, vcc, s6, v48
-; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v32
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v0, v42
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v43, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v43
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v43, vcc, s6, v39
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v44, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v44, 0xff, v44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v44
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v44, vcc, s6, v38
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v45, vcc, 3, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v45, 0xff, v45
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v45
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v45, vcc, s6, v37
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v45, vcc, s6, v48
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v46
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v46, vcc, s6, v36
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v19, v0, v47
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v47, vcc, s6, v34
-; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v17, v0, v56
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v56, vcc, s6, v33
-; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v15, v0, v57
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v57, vcc, s6, v31
-; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v13, v0, v58
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v58, vcc, s6, v29
-; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
+; SI-NEXT:    v_add_i32_e32 v46, vcc, s6, v39
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v9, v0, v60
-; SI-NEXT:    v_or_b32_e32 v0, v2, v61
+; SI-NEXT:    v_or_b32_e32 v0, v0, v47
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v47, vcc, s6, v38
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v0, v56
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v56, vcc, s6, v37
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v0, v57
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v8, v61
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v6, v62
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v62, vcc, s6, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v60, vcc, s6, v25
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v57, vcc, s6, v35
 ; SI-NEXT:    v_add_i32_e32 v61, vcc, s6, v23
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
+; SI-NEXT:    v_add_i32_e32 v62, vcc, s6, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v41, vcc, s6, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v63, v0, v63
-; SI-NEXT:    v_or_b32_e32 v0, v1, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v0, v3
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v4
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v32, vcc, s6, v63
-; SI-NEXT:    v_cvt_f32_f16_e32 v32, v32
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_add_i32_e32 v55, vcc, s6, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v40, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v63, vcc, s6, v63
-; SI-NEXT:    v_cvt_f32_f16_e32 v63, v63
-; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v41, vcc, s6, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v55, vcc, s6, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v51, vcc, s6, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v53, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v31, vcc, s6, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v44, vcc, s6, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v52, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v5
+; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v50, vcc, s6, v0
+; SI-NEXT:    v_add_i32_e32 v43, vcc, s6, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v42, vcc, s6, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v40, vcc, s6, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v49, vcc, s6, v0
+; SI-NEXT:    v_add_i32_e32 v53, vcc, s6, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v52, vcc, s6, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v50, vcc, s6, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v49, vcc, s6, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v48, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v39, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v38, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v37, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v36, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v35, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v34, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v33, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v2
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v3
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, v62
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v6
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v8
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, v10
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, v12
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v0, v14
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v16
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v18
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v20
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v22
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v24
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v26
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v27
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v28
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v29
+; SI-NEXT:    v_cvt_f32_f16_e32 v29, v56
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v30
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v33
-; SI-NEXT:    v_cvt_f32_f16_e32 v33, v59
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v33, v46
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v34
-; SI-NEXT:    v_cvt_f32_f16_e32 v34, v44
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v35
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v59
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v36
-; SI-NEXT:    v_cvt_f32_f16_e32 v36, v57
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v36, v44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v37
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v42
+; SI-NEXT:    v_cvt_f32_f16_e32 v37, v57
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v38
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v47
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v38, v42
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v39
-; SI-NEXT:    v_cvt_f32_f16_e32 v39, v53
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v39, v47
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v48
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v45
+; SI-NEXT:    v_cvt_f32_f16_e32 v48, v53
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v49
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v50
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v52
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v40
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v43
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v0, v45
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v27, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v31, vcc, s6, v5
+; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
+; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v5
+; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v12
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v7
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v9
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v11
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v13
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v15
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v17
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v60
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v19
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v31
-; SI-NEXT:    v_cvt_f32_f16_e32 v31, v46
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v21
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v23
 ; SI-NEXT:    v_cvt_f32_f16_e32 v23, v58
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v29
-; SI-NEXT:    v_cvt_f32_f16_e32 v29, v61
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v27
-; SI-NEXT:    v_cvt_f32_f16_e32 v27, v56
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v31
+; SI-NEXT:    v_cvt_f32_f16_e32 v31, v61
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v25
-; SI-NEXT:    v_cvt_f32_f16_e32 v25, v0
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v51
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v54
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v55
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v41
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_add_i32_e32 v32, vcc, s6, v12
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v62
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v12
+; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v63
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v14
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v60
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v32
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_add_i32_e32 v63, vcc, s6, v63
+; SI-NEXT:    v_cvt_f32_f16_e32 v63, v63
+; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; SI-NEXT:  .LBB92_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
-; SI-NEXT:    buffer_store_dword v0, v30, s[0:3], 0 offen
+; SI-NEXT:    buffer_store_dword v0, v10, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
@@ -180619,7 +180732,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 4, v30
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 4, v10
 ; SI-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
@@ -180630,7 +180743,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 8, v30
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 8, v10
 ; SI-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
@@ -180645,7 +180758,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180656,7 +180769,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180667,7 +180780,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180678,7 +180791,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180688,8 +180801,8 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v30
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180698,9 +180811,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v30
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180709,9 +180822,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180721,8 +180834,8 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v30
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180731,9 +180844,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v30
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180742,9 +180855,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180754,8 +180867,8 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v30
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180764,9 +180877,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v30
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180775,9 +180888,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180787,8 +180900,8 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v30
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180797,9 +180910,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v30
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180808,9 +180921,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180820,8 +180933,8 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v30
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180830,9 +180943,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v30
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180841,9 +180954,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180853,8 +180966,8 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v30
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -180863,62 +180976,64 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v39
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v30
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v48
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v37
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v30
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v38
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v48
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v34
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v30
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v36
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v10
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v38
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v31
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v30
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v39
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v33
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v36
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v27
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v30
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v37
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v29
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v33
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v35
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v23
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v30
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v29
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v5
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v30
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v31
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v17
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v25
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v4
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x7c, v30
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v27
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v5
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x7c, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
@@ -180993,39 +181108,39 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:136
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:144
-; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:152
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:128
+; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:136
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:144
+; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:152
 ; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:160
 ; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:168
 ; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:176
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
 ; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v1
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v25
 ; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v3
 ; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v5
 ; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v7
-; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v9
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v9
 ; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v13
-; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v15
+; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v15
 ; VI-NEXT:    v_lshlrev_b16_e32 v35, 8, v17
 ; VI-NEXT:    v_lshlrev_b16_e32 v36, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v21
 ; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v23
+; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v27
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
@@ -181039,46 +181154,47 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v27
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
 ; VI-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v37
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v38
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v39
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v48
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v49
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v50
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v51
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
@@ -181095,7 +181211,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v43
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:208
@@ -181103,15 +181219,15 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
@@ -181134,20 +181250,20 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -181160,20 +181276,20 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -181181,17 +181297,17 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
@@ -181207,17 +181323,17 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:316
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
@@ -181232,44 +181348,44 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:376
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:348
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v1
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v1
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v2
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v3
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:372
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:12
 ; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:108
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
 ; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -181295,31 +181411,31 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v4, v4, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v4, v4, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr54
 ; VI-NEXT:    ; implicit-def: $vgpr55
 ; VI-NEXT:    ; implicit-def: $vgpr40
 ; VI-NEXT:    ; implicit-def: $vgpr41
-; VI-NEXT:    ; implicit-def: $vgpr49
+; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr36
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr34
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr50
 ; VI-NEXT:    ; implicit-def: $vgpr52
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr37
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -181338,35 +181454,35 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_or_b32_sdwa v10, v63, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v61, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_or_b32_sdwa v12, v58, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v59, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_or_b32_sdwa v14, v45, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr58
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    ; implicit-def: $vgpr45
 ; VI-NEXT:    s_waitcnt vmcnt(2)
@@ -181375,26 +181491,26 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v63, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr62
-; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v60, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr59
+; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v10, v57, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v56, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr56
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v58, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr56
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -181403,35 +181519,35 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v13, v44, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -181446,20 +181562,20 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -181467,189 +181583,189 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v31, v31, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr48
+; VI-NEXT:    v_or_b32_sdwa v31, v31, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr49
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v32, v32, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr53
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v30, v30, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v30, v30, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v30, v30, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v30, v30, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr39
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v31, v31, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v31, v31, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr32
 ; VI-NEXT:    ; kill: killed $vgpr32
@@ -181754,7 +181870,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; implicit-def: $vgpr32
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr50
+; VI-NEXT:    ; implicit-def: $vgpr51
 ; VI-NEXT:  .LBB92_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB92_4
@@ -181762,53 +181878,51 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v18, 0x300
-; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
 ; VI-NEXT:    v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(12)
 ; VI-NEXT:    v_or_b32_sdwa v29, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(13)
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_sdwa v0, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v2, 0x300, v3
 ; VI-NEXT:    v_or_b32_e32 v0, v2, v0
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v2, 3, v2
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v2, v52, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v2, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
 ; VI-NEXT:    v_or_b32_e32 v2, v3, v2
@@ -181818,19 +181932,19 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v1, 0x300, v1
 ; VI-NEXT:    v_or_b32_e32 v1, v1, v4
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v3, v49, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v48, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v3, v3, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
-; VI-NEXT:    v_or_b32_sdwa v4, v39, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
 ; VI-NEXT:    v_or_b32_e32 v3, v4, v3
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
@@ -181838,14 +181952,13 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
 ; VI-NEXT:    v_or_b32_sdwa v6, v33, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    s_waitcnt vmcnt(2)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v5, 3, v5
 ; VI-NEXT:    v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
 ; VI-NEXT:    v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v4, v4, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
@@ -181862,7 +181975,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v32, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_sdwa v32, v32, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
@@ -181870,7 +181983,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; VI-NEXT:    v_or_b32_e32 v28, v28, v32
 ; VI-NEXT:    s_waitcnt vmcnt(1)
@@ -181878,79 +181991,94 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v33, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v33, v33, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
 ; VI-NEXT:    v_or_b32_e32 v27, v27, v33
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v34, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v34, v34, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; VI-NEXT:    v_or_b32_e32 v26, v26, v34
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v35, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v35, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
+; VI-NEXT:    v_or_b32_e32 v25, v25, v35
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v36, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT:    v_or_b32_e32 v25, v25, v35
+; VI-NEXT:    v_add_u16_sdwa v36, v36, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v36, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_sdwa v36, v36, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v24, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
+; VI-NEXT:    v_or_b32_e32 v24, v24, v36
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v24, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v37, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v23, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
 ; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT:    v_or_b32_e32 v24, v24, v36
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
+; VI-NEXT:    v_or_b32_e32 v23, v23, v37
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v8, 3, v8
 ; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v7, v8, v7
-; VI-NEXT:    v_add_u16_e32 v8, 3, v61
+; VI-NEXT:    v_add_u16_e32 v8, 3, v63
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v9, 3, v62
@@ -181959,30 +182087,30 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; VI-NEXT:    v_or_b32_e32 v8, v9, v8
-; VI-NEXT:    v_add_u16_e32 v9, 3, v63
+; VI-NEXT:    v_add_u16_e32 v9, 3, v61
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v10, 3, v59
+; VI-NEXT:    v_add_u16_e32 v10, 3, v60
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v9, v9, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_or_b32_e32 v9, v10, v9
-; VI-NEXT:    v_add_u16_e32 v10, 3, v60
+; VI-NEXT:    v_add_u16_e32 v10, 3, v57
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v11, 3, v57
+; VI-NEXT:    v_add_u16_e32 v11, 3, v56
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v10, v10, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_or_b32_e32 v10, v11, v10
-; VI-NEXT:    v_add_u16_e32 v11, 3, v58
+; VI-NEXT:    v_add_u16_e32 v11, 3, v59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v12, 3, v56
+; VI-NEXT:    v_add_u16_e32 v12, 3, v58
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; VI-NEXT:    v_or_b32_e32 v11, v12, v11
@@ -181991,7 +182119,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v13, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_or_b32_e32 v12, v13, v12
@@ -182000,7 +182128,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 3, v44
 ; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v13, v13, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_or_b32_e32 v13, v14, v13
@@ -182012,35 +182140,35 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; VI-NEXT:    v_or_b32_e32 v14, v15, v14
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v15, 3, v15
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    v_or_b32_e32 v15, v16, v15
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    v_or_b32_sdwa v17, v19, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -182049,54 +182177,43 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v16, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v19, 0x300, v20
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; VI-NEXT:    v_or_b32_e32 v16, v19, v16
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
-; VI-NEXT:    v_or_b32_sdwa v30, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v30, v39, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
-; VI-NEXT:    v_or_b32_sdwa v31, v50, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v31, v51, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v31, 0x300, v31
-; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
-; VI-NEXT:    v_or_b32_sdwa v21, v37, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v37, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v40, v21, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v29, v29, v40
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v23, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT:    v_or_b32_e32 v23, v23, v37
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v38, v38, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_e32 v20, 3, v20
-; VI-NEXT:    v_or_b32_sdwa v20, v48, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_sdwa v55, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v38, v38, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_or_b32_e32 v30, v30, v55
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v22, v39, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v20, 3, v20
+; VI-NEXT:    v_or_b32_sdwa v20, v49, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v55, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; VI-NEXT:    v_or_b32_e32 v22, v22, v38
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_e32 v30, v30, v55
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v39, 3, v39
 ; VI-NEXT:    v_or_b32_sdwa v39, v48, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
@@ -182118,7 +182235,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v39, 3, v39
 ; VI-NEXT:    v_or_b32_sdwa v39, v49, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -182220,17 +182337,16 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
 ; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:136
+; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:136
 ; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:144
 ; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:152
 ; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:160
 ; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:168
 ; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:176
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v1
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v27
-; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v29
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v3
@@ -182238,81 +182354,81 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v7
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v9
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v11
-; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v13
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v13
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v15
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v17
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v35, 8, v19
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v21
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v23
-; GFX9-NEXT:    s_waitcnt vmcnt(27)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
+; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v25
 ; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v25
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v37
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v39
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v48
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v49
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
@@ -182320,7 +182436,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
@@ -182337,7 +182453,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v43
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
@@ -182346,15 +182462,15 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
@@ -182378,20 +182494,20 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -182405,20 +182521,20 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -182427,17 +182543,17 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
@@ -182464,7 +182580,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
@@ -182484,48 +182600,48 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:372
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:12
 ; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:100
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB92_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -182534,9 +182650,9 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
@@ -182565,10 +182681,10 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; implicit-def: $vgpr51
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr48
+; GFX9-NEXT:    ; implicit-def: $vgpr39
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
@@ -182579,7 +182695,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr34
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -182594,93 +182710,93 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v58, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v15, v43, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr43
+; GFX9-NEXT:    v_or_b32_sdwa v15, v42, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v63, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v62, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr59
+; GFX9-NEXT:    ; implicit-def: $vgpr61
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v57, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr46
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v43, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr42
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -182695,20 +182811,20 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
@@ -182716,58 +182832,58 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
@@ -182785,22 +182901,22 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
 ; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr53
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr37
 ; GFX9-NEXT:    ; kill: killed $vgpr37
 ; GFX9-NEXT:    ; implicit-def: $vgpr37
@@ -182897,7 +183013,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr49
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -183009,27 +183125,27 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_cbranch_execz .LBB92_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
@@ -183037,19 +183153,20 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
 ; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    s_waitcnt vmcnt(20)
+; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(17)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v2
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v3
-; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    v_perm_b32 v0, v2, v0, s6
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(17)
+; GFX9-NEXT:    s_waitcnt vmcnt(16)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
@@ -183074,7 +183191,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT:    v_or_b32_sdwa v3, v48, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
@@ -183100,7 +183217,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_or_b32_sdwa v36, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
@@ -183124,11 +183241,11 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v37, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v37, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT:    v_or_b32_sdwa v21, v39, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v21, v48, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v33, 0x300, v21
 ; GFX9-NEXT:    v_add_u16_e32 v34, 0x300, v23
 ; GFX9-NEXT:    v_perm_b32 v29, v34, v29, s6
@@ -183136,16 +183253,17 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v38, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
@@ -183155,7 +183273,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v39, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
@@ -183165,7 +183283,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v48, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -183178,45 +183296,45 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v62
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v60
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v58
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v57
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v60
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v56
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v57
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v56
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v58
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v46
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v46
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
@@ -183225,63 +183343,63 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 3, v44
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v43
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v15, 3, v43
+; GFX9-NEXT:    v_add_u16_e32 v15, 3, v42
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v16
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v18
 ; GFX9-NEXT:    v_perm_b32 v17, v17, v20, s6
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v19
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v16, v18, v16, s6
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v49, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v49, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v20
 ; GFX9-NEXT:    v_perm_b32 v30, v33, v30, s6
@@ -183289,7 +183407,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v50, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v52, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -183299,14 +183417,14 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v51, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v51
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v52, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -183316,7 +183434,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v53, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v53
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -183329,7 +183447,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v55, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v55
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -183337,7 +183455,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v40, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v50, 0x300, v40
 ; GFX9-NEXT:    v_perm_b32 v21, v50, v21, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -183345,14 +183463,14 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v41, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v41
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v42, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v51, 0x300, v42
 ; GFX9-NEXT:    v_perm_b32 v20, v51, v20, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -183360,7 +183478,7 @@ define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v43, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v43
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -187174,17 +187292,17 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; VI-NEXT:    v_mov_b32_e32 v45, v62
 ; VI-NEXT:    v_or_b32_sdwa v16, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_or_b32_sdwa v3, v51, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v17, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_mov_b32_e32 v54, v22
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_mov_b32_e32 v32, v1
 ; VI-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_mov_b32_e32 v54, v22
 ; VI-NEXT:    v_mov_b32_e32 v41, v24
 ; VI-NEXT:    v_or_b32_sdwa v18, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
@@ -188421,8 +188539,8 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; GFX9-NEXT:    s_cbranch_vccnz .LBB93_5
 ; GFX9-NEXT:  ; %bb.4: ; %cmp.true
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_add_i32 s28, s28, 3
 ; GFX9-NEXT:    s_and_b32 s4, s28, 0xff
 ; GFX9-NEXT:    s_lshl_b32 s5, s29, 8
@@ -188455,7 +188573,7 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(13)
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -188505,9 +188623,9 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; GFX9-NEXT:    s_or_b32 s7, s8, s7
 ; GFX9-NEXT:    s_and_b32 s8, s16, 0xff
 ; GFX9-NEXT:    s_or_b32 s8, s9, s8
+; GFX9-NEXT:    s_and_b32 s9, s18, 0xff
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT:    s_and_b32 s9, s18, 0xff
 ; GFX9-NEXT:    s_or_b32 s9, s10, s9
 ; GFX9-NEXT:    s_addk_i32 s4, 0x300
 ; GFX9-NEXT:    s_addk_i32 s5, 0x300
@@ -188540,11 +188658,11 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; GFX9-NEXT:    v_add_u32_e32 v1, 0x300, v1
 ; GFX9-NEXT:    v_lshl_or_b32 v6, v1, 16, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -188555,11 +188673,11 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; GFX9-NEXT:    v_add_u32_e32 v1, 0x300, v1
 ; GFX9-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -188690,9 +188808,9 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
 ; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
 ; GFX9-NEXT:    v_or_b32_sdwa v51, v34, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -188769,22 +188887,21 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v43, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v0
 ; GFX9-NEXT:    v_add_u32_e32 v52, 0x300, v43
 ; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s8
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v44, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v33
@@ -191564,8 +191681,8 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v28, v4, v2
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
 ; SI-NEXT:    v_or_b32_e32 v27, v3, v1
@@ -191629,8 +191746,8 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v37, v4, v2
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
 ; SI-NEXT:    v_or_b32_e32 v36, v3, v1
@@ -191694,8 +191811,8 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v51, v4, v2
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
 ; SI-NEXT:    v_or_b32_e32 v50, v3, v1
@@ -193129,8 +193246,8 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; VI-NEXT:    v_lshrrev_b64 v[45:46], 24, v[45:46]
 ; VI-NEXT:    v_lshrrev_b64 v[52:53], 24, v[52:53]
 ; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v49
@@ -193648,8 +193765,8 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    ; kill: killed $vgpr50
@@ -193774,7 +193891,7 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
+; GFX9-NEXT:    s_waitcnt vmcnt(17)
 ; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    ; kill: killed $vgpr33
@@ -193893,7 +194010,6 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(45)
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
@@ -194032,7 +194148,6 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
 ; GFX9-NEXT:    v_pk_add_f16 v32, v32, s6 op_sel_hi:[1,0]
 ; GFX9-NEXT:    s_waitcnt vmcnt(18)
 ; GFX9-NEXT:    v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
@@ -194448,13 +194563,14 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
@@ -194473,13 +194589,14 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
@@ -194498,13 +194615,14 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
@@ -194523,13 +194641,14 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
@@ -194548,13 +194667,14 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
@@ -198826,13 +198946,14 @@ define inreg <128 x i8> @bitcast_v64f16_to_v128i8_scalar(<64 x half> inreg %a, i
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_lshlrev_b32_e32 v2, 8, v10
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v2, v6, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
@@ -199049,12 +199170,12 @@ define inreg <128 x i8> @bitcast_v64f16_to_v128i8_scalar(<64 x half> inreg %a, i
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
 ; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -202519,19 +202640,19 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:392
 ; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:4
@@ -202552,16 +202673,13 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:148
 ; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:188
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:128
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v4
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v14
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v16
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v12
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -202569,266 +202687,254 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v20
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v8
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v28
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v16
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 24, v24
+; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    ; implicit-def: $vgpr20
-; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr53
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr52
-; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr28
+; SI-NEXT:    ; implicit-def: $vgpr52
 ; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; kill: killed $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; kill: killed $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; kill: killed $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:360
+; SI-NEXT:    ; kill: killed $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:360
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v30, 24, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 24, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v7
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v9
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v5
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v13
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v15
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v11
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v19
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v21
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:152
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v17
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v23
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v32
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:332
-; SI-NEXT:    ; implicit-def: $vgpr23
-; SI-NEXT:    ; implicit-def: $vgpr25
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr27
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v32
 ; SI-NEXT:    ; implicit-def: $vgpr21
-; SI-NEXT:    ; kill: killed $vgpr23
-; SI-NEXT:    ; implicit-def: $vgpr23
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:160
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr27
+; SI-NEXT:    ; implicit-def: $vgpr19
+; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:88
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:80
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:144
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:140
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:128
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:152
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v29
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v31
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:108
+; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:160
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:184
-; SI-NEXT:    v_lshlrev_b32_e32 v29, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v18
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:108
-; SI-NEXT:    ; implicit-def: $vgpr18
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:180
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:176
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:140
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v33
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:180
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:176
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v34
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 8, v10
-; SI-NEXT:    ; implicit-def: $vgpr10
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v33
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v34
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v4
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:172
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:196
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:220
-; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:192
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:172
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:192
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 24, v7
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:216
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:208
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:208
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v8
 ; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:252
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:204
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:204
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:252
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:248
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:248
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:244
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:244
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:240
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:240
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v9
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v9
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v7
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v8
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:284
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:256
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:236
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:284
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:256
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v4
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:280
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:276
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:272
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v8
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v11
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v7
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v8
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:292
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:316
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:288
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:268
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:292
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:316
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:288
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 8, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 8, v8
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:312
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:312
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:308
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:308
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:304
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v9
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v9
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v11
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:324
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:348
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:320
+; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:300
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v6
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:348
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:320
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v8
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:344
@@ -202838,123 +202944,135 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:340
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:336
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 24, v9
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:356
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:380
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:352
-; SI-NEXT:    ; implicit-def: $vgpr9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 24, v6
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v9
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:332
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:356
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:380
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:352
 ; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:376
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:372
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:368
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:364
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v4
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v8
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:388
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:384
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:372
-; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:368
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:364
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:388
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:384
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v10
+; SI-NEXT:    ; implicit-def: $vgpr10
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v18
+; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v26
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr26
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v11
+; SI-NEXT:    ; implicit-def: $vgpr11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v14
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v31, 8, v15
+; SI-NEXT:    ; implicit-def: $vgpr15
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v8
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:120
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:44
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:40
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; SI-NEXT:    ; implicit-def: $vgpr3
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:48
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:56
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v3
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr5
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v11
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v6
+; SI-NEXT:    ; implicit-def: $vgpr6
+; SI-NEXT:    ; kill: killed $vgpr6
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v7
+; SI-NEXT:    ; implicit-def: $vgpr6
+; SI-NEXT:    ; kill: killed $vgpr6
+; SI-NEXT:    ; implicit-def: $vgpr7
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v12
+; SI-NEXT:    ; implicit-def: $vgpr6
+; SI-NEXT:    ; implicit-def: $vgpr12
+; SI-NEXT:    ; kill: killed $vgpr6
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr4
-; SI-NEXT:    ; implicit-def: $vgpr11
-; SI-NEXT:    ; kill: killed $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:24
-; SI-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr6
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:80
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:40
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:48
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v13
+; SI-NEXT:    ; implicit-def: $vgpr13
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:16
-; SI-NEXT:    v_lshlrev_b32_e32 v12, 8, v13
-; SI-NEXT:    ; implicit-def: $vgpr13
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr8
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:104
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v14
+; SI-NEXT:    ; implicit-def: $vgpr14
+; SI-NEXT:    ; kill: killed $vgpr14
+; SI-NEXT:    ; implicit-def: $vgpr14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -202964,23 +203082,23 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:168
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:200
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -203024,29 +203142,27 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_or_b32_e32 v20, v1, v2
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_or_b32_e32 v12, v1, v2
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v18, v2, v1
+; SI-NEXT:    v_or_b32_e32 v42, v2, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v3, v2, v1
+; SI-NEXT:    v_or_b32_e32 v5, v2, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -203056,848 +203172,855 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v2, v1
+; SI-NEXT:    v_or_b32_e32 v7, v16, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v9, v2, v1
+; SI-NEXT:    v_or_b32_e32 v11, v2, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v41, v1, v2
+; SI-NEXT:    v_or_b32_e32 v40, v1, v2
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v55, v24, v1
+; SI-NEXT:    v_or_b32_e32 v10, v24, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; SI-NEXT:    v_mov_b32_e32 v24, v33
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v13, v2, v1
+; SI-NEXT:    v_or_b32_e32 v15, v2, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v1, v1, v2
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v26, v1, v2
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v26, v30, v1
+; SI-NEXT:    v_or_b32_e32 v18, v25, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v21, v2, v1
+; SI-NEXT:    v_or_b32_e32 v19, v2, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v32, v1, v2
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v1, v1, v2
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v28, v2, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v30, v2, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v27, v2, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v21, v2, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v1, v1, v2
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v29, v1, v2
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v35, v2, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v33, v2, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v31, v2, v1
+; SI-NEXT:    v_or_b32_e32 v27, v2, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v63, v1, v2
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v48, v2, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v1, v1, v2
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v50, v2, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v11, v1, v2
+; SI-NEXT:    v_or_b32_e32 v38, v1, v2
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v2, v2, v1
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v52, v1, v2
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v1, v2, v1
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v2, v1, v2
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v53, v4, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v6, v1, v4
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v55, v1, v6
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v1, v4, v1
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v43, v10, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v44, v4, v10
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v17, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v16, v10, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v46, v10, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    v_or_b32_e32 v14, v1, v6
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v56, v4, v10
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v17, v10, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v44, v1, v6
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v57, v10, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v61, v4, v10
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v45, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v30, v10, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v49, v10, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    v_or_b32_e32 v25, v1, v6
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v4, v4, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT:    v_or_b32_e32 v47, v1, v6
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v23, v10, v22
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT:    v_or_b32_e32 v58, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v33, v10, v22
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v16, v1, v6
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    v_mov_b32_e32 v10, v29
-; SI-NEXT:    v_or_b32_e32 v29, v22, v34
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT:    v_or_b32_e32 v59, v1, v6
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v62, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v24, v1, v6
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v13, v1, v6
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v6, v6, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v23, v1, v8
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v36, v1, v8
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v8, v8, v35
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT:    v_or_b32_e32 v34, v37, v22
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v35, v37, v22
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
 ; SI-NEXT:    v_or_b32_e32 v37, v51, v22
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT:    v_or_b32_e32 v51, v22, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v51, v22, v63
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT:    v_or_b32_e32 v34, v56, v22
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT:    v_or_b32_e32 v7, v47, v7
-; SI-NEXT:    v_or_b32_e32 v47, v59, v22
+; SI-NEXT:    v_or_b32_e32 v56, v60, v22
 ; SI-NEXT:    v_and_b32_e32 v22, 0xff, v54
-; SI-NEXT:    v_or_b32_e32 v54, v22, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v60
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v19, v8
-; SI-NEXT:    v_and_b32_e32 v19, 0xff, v42
-; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; SI-NEXT:    v_or_b32_e32 v14, v14, v19
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v54, v22, v4
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v41
+; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
+; SI-NEXT:    v_or_b32_e32 v9, v9, v22
+; SI-NEXT:    v_and_b32_e32 v4, 0xff, v57
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v4, v32, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; SI-NEXT:    v_or_b32_e32 v19, v19, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; SI-NEXT:    v_or_b32_e32 v25, v22, v24
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v28, v22, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v32, v22, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v39, v22, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v39, v22, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v42, v22, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v41, v22, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v59, v22, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v57, v22, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v60, v22, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v60, v22, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v36, v22, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v63, v22, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v10, v22, v10
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT:    v_or_b32_e32 v38, v22, v24
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v1, v22, v1
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_or_b32_e32 v1, v1, v55
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT:    v_or_b32_e32 v48, v22, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v58, v22, v24
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v52, v22, v24
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; SI-NEXT:    v_or_b32_e32 v46, v22, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v24, v22, v24
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT:    v_or_b32_e32 v40, v22, v40
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v49, v22, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v50, v22, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v20, v22, v20
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT:    v_or_b32_e32 v12, v22, v12
+; SI-NEXT:    v_or_b32_e32 v53, v22, v53
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT:    v_or_b32_e32 v15, v22, v15
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v45
-; SI-NEXT:    v_or_b32_e32 v45, v22, v62
-; SI-NEXT:    v_mov_b32_e32 v62, v18
-; SI-NEXT:    v_or_b32_e32 v20, v20, v62
-; SI-NEXT:    v_or_b32_e32 v22, v19, v3
-; SI-NEXT:    v_alignbit_b32 v3, v20, v3, 16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v3, v22, v3
+; SI-NEXT:    v_and_b32_e32 v22, 0xff, v61
+; SI-NEXT:    v_mov_b32_e32 v61, v42
+; SI-NEXT:    v_or_b32_e32 v31, v22, v31
+; SI-NEXT:    v_or_b32_e32 v22, v12, v61
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v28
+; SI-NEXT:    v_or_b32_e32 v43, v12, v5
+; SI-NEXT:    v_alignbit_b32 v5, v22, v5, 16
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    v_mov_b32_e32 v19, v5
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; SI-NEXT:    v_mov_b32_e32 v12, v7
+; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v32
+; SI-NEXT:    v_or_b32_e32 v7, v7, v11
+; SI-NEXT:    v_and_b32_e32 v32, 0xffff, v49
+; SI-NEXT:    v_or_b32_e32 v32, v32, v59
+; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v5, v3, v19
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v25
-; SI-NEXT:    v_or_b32_e32 v25, v3, v9
-; SI-NEXT:    v_alignbit_b32 v3, v5, v9, 16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v41
-; SI-NEXT:    v_or_b32_e32 v41, v3, v55
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v39
-; SI-NEXT:    v_or_b32_e32 v18, v3, v13
-; SI-NEXT:    v_alignbit_b32 v3, v41, v13, 16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT:    v_or_b32_e32 v5, v5, v12
+; SI-NEXT:    v_alignbit_b32 v11, v5, v11, 16
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v40
+; SI-NEXT:    v_or_b32_e32 v42, v11, v10
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v39
+; SI-NEXT:    v_or_b32_e32 v40, v11, v15
+; SI-NEXT:    v_alignbit_b32 v11, v42, v15, 16
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v26
+; SI-NEXT:    v_or_b32_e32 v26, v11, v18
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v41
+; SI-NEXT:    v_or_b32_e32 v39, v11, v19
+; SI-NEXT:    v_alignbit_b32 v11, v26, v19, 16
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
+; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v13, v3, v26
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v42
-; SI-NEXT:    v_or_b32_e32 v39, v3, v21
-; SI-NEXT:    v_alignbit_b32 v3, v13, v21, 16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v32
-; SI-NEXT:    v_or_b32_e32 v32, v3, v28
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v59
-; SI-NEXT:    v_or_b32_e32 v42, v3, v27
-; SI-NEXT:    v_alignbit_b32 v3, v32, v27, 16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_or_b32_e32 v28, v11, v30
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v57
+; SI-NEXT:    v_or_b32_e32 v11, v11, v21
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_alignbit_b32 v11, v28, v21, 16
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v29
+; SI-NEXT:    v_or_b32_e32 v29, v11, v33
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v60
+; SI-NEXT:    v_or_b32_e32 v21, v11, v27
+; SI-NEXT:    v_alignbit_b32 v11, v29, v27, 16
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v27, v3, v35
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v60
-; SI-NEXT:    v_or_b32_e32 v3, v3, v31
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v3, v27, v31, 16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v63
-; SI-NEXT:    v_or_b32_e32 v21, v3, v48
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v36
-; SI-NEXT:    v_or_b32_e32 v31, v3, v50
-; SI-NEXT:    v_alignbit_b32 v3, v21, v50, 16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v3, v2
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v10
-; SI-NEXT:    v_or_b32_e32 v3, v3, v53
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v3, v11, v53, 16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v38
-; SI-NEXT:    v_or_b32_e32 v3, v3, v1
-; SI-NEXT:    v_or_b32_e32 v6, v6, v43
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v6, v3, v43, 16
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v44
-; SI-NEXT:    v_or_b32_e32 v9, v6, v16
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v58
-; SI-NEXT:    v_or_b32_e32 v6, v6, v46
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v6, v9, v46, 16
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v56
-; SI-NEXT:    v_or_b32_e32 v63, v6, v17
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v52
-; SI-NEXT:    v_or_b32_e32 v6, v6, v57
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_or_b32_e32 v19, v11, v38
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v63
+; SI-NEXT:    v_or_b32_e32 v27, v11, v52
+; SI-NEXT:    v_alignbit_b32 v11, v19, v52, 16
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v30
+; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v23
+; SI-NEXT:    ; implicit-def: $vgpr63
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_or_b32_e32 v11, v11, v2
+; SI-NEXT:    v_alignbit_b32 v1, v11, v55, 16
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v6, v63, v57, 16
-; SI-NEXT:    v_or_b32_e32 v57, v4, v23
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v40
-; SI-NEXT:    v_or_b32_e32 v52, v4, v33
-; SI-NEXT:    v_alignbit_b32 v4, v57, v33, 16
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v17
+; SI-NEXT:    v_or_b32_e32 v15, v1, v14
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v48
+; SI-NEXT:    v_or_b32_e32 v1, v1, v44
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v29
-; SI-NEXT:    v_or_b32_e32 v46, v4, v34
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v12
-; SI-NEXT:    v_or_b32_e32 v38, v4, v37
-; SI-NEXT:    v_alignbit_b32 v4, v46, v37, 16
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v51
-; SI-NEXT:    v_or_b32_e32 v44, v4, v7
-; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v7
-; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v8
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v15
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    v_or_b32_e32 v36, v4, v47
-; SI-NEXT:    v_alignbit_b32 v4, v44, v47, 16
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v1, v15, v44, 16
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v54
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    v_or_b32_e32 v43, v4, v8
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v45
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    v_or_b32_e32 v12, v4, v14
-; SI-NEXT:    v_alignbit_b32 v4, v43, v14, 16
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v45
+; SI-NEXT:    v_or_b32_e32 v17, v1, v25
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v46
+; SI-NEXT:    v_or_b32_e32 v1, v1, v47
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v28
-; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v23
-; SI-NEXT:    v_mov_b32_e32 v23, v12
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v1, v17, v47, 16
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v61
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    v_or_b32_e32 v61, v6, v30
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v24
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    v_or_b32_e32 v58, v6, v49
-; SI-NEXT:    v_alignbit_b32 v6, v61, v49, 16
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v58
+; SI-NEXT:    v_or_b32_e32 v1, v1, v16
+; SI-NEXT:    v_alignbit_b32 v32, v1, v59, 16
+; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v32, 0xffff, v62
+; SI-NEXT:    v_or_b32_e32 v59, v6, v23
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v62, v32, v24
+; SI-NEXT:    v_and_b32_e32 v32, 0xffff, v50
+; SI-NEXT:    v_or_b32_e32 v50, v6, v36
+; SI-NEXT:    v_alignbit_b32 v6, v59, v36, 16
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v8
+; SI-NEXT:    v_or_b32_e32 v47, v6, v35
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v53
+; SI-NEXT:    v_or_b32_e32 v49, v6, v37
+; SI-NEXT:    v_alignbit_b32 v6, v47, v37, 16
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v51
+; SI-NEXT:    v_or_b32_e32 v45, v6, v34
+; SI-NEXT:    v_or_b32_e32 v48, v3, v56
+; SI-NEXT:    v_alignbit_b32 v3, v45, v56, 16
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v54
+; SI-NEXT:    v_or_b32_e32 v44, v3, v4
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v31
+; SI-NEXT:    v_or_b32_e32 v3, v3, v9
+; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v18
+; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v14
+; SI-NEXT:    v_mov_b32_e32 v14, v3
+; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v4
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    v_or_b32_e32 v46, v32, v13
+; SI-NEXT:    v_alignbit_b32 v13, v62, v13, 16
+; SI-NEXT:    v_alignbit_b32 v6, v44, v9, 16
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v61
+; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v10
+; SI-NEXT:    s_waitcnt expcnt(1)
+; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v33
+; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v38
+; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v2
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v25
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v16
+; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v62
-; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v19
-; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v55
-; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v26
-; SI-NEXT:    v_mov_b32_e32 v26, v42
-; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v35
-; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v48
-; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v2
-; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v16
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v17
-; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v30
-; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v34
-; SI-NEXT:    ; implicit-def: $vgpr17
-; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v35
+; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v34
+; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr24
-; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    ; kill: killed $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr25
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr8
-; SI-NEXT:    ; implicit-def: $vgpr19
-; SI-NEXT:    ; implicit-def: $vgpr14
-; SI-NEXT:    ; implicit-def: $vgpr29
-; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    ; kill: killed $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr15
-; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr4
+; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr9
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:  .LBB96_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB96_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v45
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v42
-; SI-NEXT:    v_or_b32_e32 v1, v62, v1
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v41
+; SI-NEXT:    v_or_b32_e32 v1, v31, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
-; SI-NEXT:    v_or_b32_e32 v2, v14, v2
+; SI-NEXT:    v_or_b32_e32 v2, v9, v2
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v54
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v60
-; SI-NEXT:    v_or_b32_e32 v2, v8, v2
-; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_or_b32_e32 v2, v4, v2
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v57
+; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    s_movk_i32 s6, 0x300
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
-; SI-NEXT:    v_or_b32_e32 v3, v19, v3
+; SI-NEXT:    v_or_b32_e32 v4, v32, v4
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v2, v4, v2
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v34
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_or_b32_e32 v5, v56, v5
+; SI-NEXT:    v_mov_b32_e32 v30, v16
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
-; SI-NEXT:    v_add_i32_e32 v43, vcc, s7, v2
-; SI-NEXT:    v_mov_b32_e32 v28, v24
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
+; SI-NEXT:    v_mov_b32_e32 v31, v24
+; SI-NEXT:    v_add_i32_e32 v44, vcc, s7, v2
+; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v1
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
-; SI-NEXT:    v_or_b32_e32 v3, v15, v3
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
-; SI-NEXT:    v_or_b32_e32 v4, v59, v4
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v4, v3
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v47, v5
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v51, v6
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT:    v_or_b32_e32 v7, v37, v7
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v4, v60, v4
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v32, vcc, s7, v3
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
+; SI-NEXT:    v_mov_b32_e32 v48, v32
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v7, v4
+; SI-NEXT:    v_or_b32_e32 v4, v63, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v44, vcc, s7, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v44
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v45, vcc, s7, v4
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v45
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT:    v_or_b32_e32 v5, v12, v5
+; SI-NEXT:    v_or_b32_e32 v5, v53, v5
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v6, v5
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    v_or_b32_e32 v7, v37, v7
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v35, vcc, s7, v5
-; SI-NEXT:    v_mov_b32_e32 v38, v35
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v33, vcc, s7, v5
+; SI-NEXT:    v_mov_b32_e32 v49, v33
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v34, v6
+; SI-NEXT:    v_or_b32_e32 v6, v35, v6
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v46, vcc, s7, v6
-; SI-NEXT:    v_add_i32_e32 v34, vcc, s7, v3
-; SI-NEXT:    v_mov_b32_e32 v36, v34
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v47, vcc, s7, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -203906,16 +204029,16 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v7
-; SI-NEXT:    v_mov_b32_e32 v52, v6
+; SI-NEXT:    v_mov_b32_e32 v50, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -203924,15 +204047,15 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v57, vcc, s7, v8
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v59, vcc, s7, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -203941,16 +204064,16 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v9
-; SI-NEXT:    v_mov_b32_e32 v58, v7
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v9
+; SI-NEXT:    v_mov_b32_e32 v46, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -203959,9 +204082,9 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v61, vcc, s7, v10
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v62, vcc, s7, v10
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
@@ -203978,7 +204101,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v11
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v11
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
@@ -203993,15 +204116,15 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v63, vcc, s7, v12
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v12
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -204010,50 +204133,45 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v13
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    v_or_b32_e32 v15, v17, v15
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v14
-; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v9
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v17
+; SI-NEXT:    v_or_b32_e32 v15, v17, v15
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v15
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v15
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -204063,9 +204181,8 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v17, v18, v17
 ; SI-NEXT:    v_or_b32_e32 v16, v17, v16
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v16
-; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v46
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
@@ -204079,468 +204196,476 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
-; SI-NEXT:    v_or_b32_e32 v17, v18, v17
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v14, vcc, s7, v17
+; SI-NEXT:    v_or_b32_e32 v18, v18, v17
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v18, v19, v18
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
+; SI-NEXT:    v_or_b32_e32 v17, v19, v17
+; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v23
 ; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
 ; SI-NEXT:    v_or_b32_e32 v19, v20, v19
-; SI-NEXT:    v_or_b32_e32 v18, v19, v18
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_or_b32_e32 v19, v19, v17
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v18
-; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v19
+; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v11
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
-; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v19, v20, v19
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_or_b32_e32 v17, v20, v17
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
 ; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
 ; SI-NEXT:    v_or_b32_e32 v20, v21, v20
-; SI-NEXT:    v_or_b32_e32 v19, v20, v19
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v31, vcc, s7, v19
+; SI-NEXT:    v_or_b32_e32 v20, v20, v17
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
-; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v20, v21, v20
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_or_b32_e32 v17, v21, v17
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
 ; SI-NEXT:    v_or_b32_e32 v21, v22, v21
-; SI-NEXT:    v_or_b32_e32 v20, v21, v20
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v21, v21, v17
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v21
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v21, v22, v21
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v17, v22, v17
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
 ; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
 ; SI-NEXT:    v_or_b32_e32 v22, v23, v22
-; SI-NEXT:    v_or_b32_e32 v22, v22, v21
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v22
+; SI-NEXT:    v_or_b32_e32 v22, v22, v17
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v21, v23, v21
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_or_b32_e32 v17, v23, v17
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
 ; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
-; SI-NEXT:    v_or_b32_e32 v23, v25, v23
-; SI-NEXT:    v_or_b32_e32 v23, v23, v21
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v23
-; SI-NEXT:    v_mov_b32_e32 v23, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v27
+; SI-NEXT:    v_or_b32_e32 v23, v26, v23
+; SI-NEXT:    v_or_b32_e32 v23, v23, v17
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    v_mov_b32_e32 v26, v25
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v29, vcc, s7, v23
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; SI-NEXT:    v_or_b32_e32 v24, v24, v21
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v24
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; SI-NEXT:    v_or_b32_e32 v25, v25, v21
-; SI-NEXT:    v_or_b32_e32 v1, v25, v1
-; SI-NEXT:    v_add_i32_e32 v26, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v20
-; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v21
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; SI-NEXT:    v_or_b32_e32 v24, v24, v17
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v24
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; SI-NEXT:    v_or_b32_e32 v25, v25, v17
+; SI-NEXT:    v_or_b32_e32 v2, v25, v2
+; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v14
+; SI-NEXT:    v_mov_b32_e32 v14, v27
+; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v4, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v32, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v28
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v4, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v39, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    v_add_i32_e32 v39, vcc, s7, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v30, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v13
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v26, v3
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    v_add_i32_e32 v26, vcc, s7, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v26
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v4, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    v_add_i32_e32 v40, vcc, s7, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v28, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v41, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v57
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v31, v3
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    v_add_i32_e32 v42, vcc, s7, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v42
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_or_b32_e32 v1, v33, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v4, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v4, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v30, v3
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v5
+; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v59
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_or_b32_e32 v1, v29, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v4, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    v_add_i32_e32 v43, vcc, s7, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v4, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v1
-; SI-NEXT:    v_alignbit_b32 v1, v20, v22, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v5, v25, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v2
+; SI-NEXT:    v_alignbit_b32 v2, v22, v43, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v41, v18, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v5, v7, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v13, v39, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v42, v40, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v32, v26, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v26, v39, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v1, v27, v15, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v28, v18, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_alignbit_b32 v2, v29, v21, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v21, v31, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v19, v27, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v1, v11, v14, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v11, v16, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v1, v3, v12, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v15, v13, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v1, v9, v10, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v17, v10, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v1, v63, v8, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v1, v9, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v61, v7, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v62, v8, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v57, v6, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v59, v6, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v46, v35, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v47, v33, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v44, v34, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v2, v45, v32, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v1, v43, v23, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v20
-; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v41
-; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v32
-; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v3
+; SI-NEXT:    v_alignbit_b32 v2, v44, v14, 16
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v22
+; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v29
+; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v19
+; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v15
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v17
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v63
-; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v61
-; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v43
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
+; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v62
+; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v47
+; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v44
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
 ; SI-NEXT:  .LBB96_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v22
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v43
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_or_b32_e32 v1, v1, v2
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v7, v7, v8
-; SI-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v20
-; SI-NEXT:    v_or_b32_e32 v6, v7, v6
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 4, v0
-; SI-NEXT:    buffer_store_dword v6, v7, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v22
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v58
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 4, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v25
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    v_or_b32_e32 v6, v6, v7
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 8, v0
-; SI-NEXT:    buffer_store_dword v6, v7, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 8, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v56
-; SI-NEXT:    v_or_b32_e32 v5, v5, v6
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 12, v0
-; SI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v12
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 12, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v18
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v40
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v5, v5, v6
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 16, v0
-; SI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 16, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v41
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v10
-; SI-NEXT:    v_or_b32_e32 v5, v5, v6
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 20, v0
-; SI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v42
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v55
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 20, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v39
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v39
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v5, v5, v6
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 24, v0
-; SI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 24, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v53
-; SI-NEXT:    v_or_b32_e32 v5, v5, v6
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 28, v0
-; SI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v20
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 28, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v26
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v5, v5, v6
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 32, v0
-; SI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 32, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v32
-; SI-NEXT:    v_or_b32_e32 v4, v5, v4
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 36, v0
-; SI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v28
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v52
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 36, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v4, v4, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 40, v0
-; SI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 40, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v50
-; SI-NEXT:    v_or_b32_e32 v4, v4, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 44, v0
-; SI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v29
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v13
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 44, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v31
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v27
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v4, v4, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 48, v0
-; SI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 48, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v49
-; SI-NEXT:    v_or_b32_e32 v4, v4, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 52, v0
-; SI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v8
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 52, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v4, v4, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 56, v0
-; SI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 56, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v48
-; SI-NEXT:    v_or_b32_e32 v4, v4, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 60, v0
-; SI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v38
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 60, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v4, v4, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 64, v0
-; SI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v4
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 64, v0
+; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v35
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x44, v0
 ; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -204549,14 +204674,14 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x48, v0
 ; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v55
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v10
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x4c, v0
 ; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -204564,77 +204689,74 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x50, v0
 ; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v63
-; SI-NEXT:    v_or_b32_e32 v1, v3, v1
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x54, v0
-; SI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v58
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v46
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v1, v1, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x58, v0
-; SI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_or_b32_e32 v1, v1, v2
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
+; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v62
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v33
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v52
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v50
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v57
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v28
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v59
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v30
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v38
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v49
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v46
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v16
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v47
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v36
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v48
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v44
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v40
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v45
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v36
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v23
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v43
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v44
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
@@ -204711,39 +204833,39 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:136
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:144
-; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:152
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:128
+; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:136
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:144
+; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:152
 ; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:160
 ; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:168
 ; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:176
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
 ; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v1
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v25
 ; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v29
-; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v3
 ; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v5
 ; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v7
-; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v9
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v9
 ; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
-; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v13
-; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v15
+; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v15
 ; VI-NEXT:    v_lshlrev_b16_e32 v35, 8, v17
 ; VI-NEXT:    v_lshlrev_b16_e32 v36, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v21
 ; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v23
+; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v27
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
@@ -204757,46 +204879,47 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v27
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
 ; VI-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v37
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v38
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v39
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v48
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v49
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v50
+; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v51
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
@@ -204813,7 +204936,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v43
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:208
@@ -204821,15 +204944,15 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
@@ -204852,20 +204975,20 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -204878,20 +205001,20 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -204899,17 +205022,17 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
@@ -204925,17 +205048,17 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:316
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; VI-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
@@ -204950,44 +205073,44 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:376
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:348
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v1
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v1
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v2
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v3
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:372
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:12
 ; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:4
 ; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:108
+; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
 ; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
 ; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -205013,31 +205136,31 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v4, v4, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v4, v4, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr54
 ; VI-NEXT:    ; implicit-def: $vgpr55
 ; VI-NEXT:    ; implicit-def: $vgpr40
 ; VI-NEXT:    ; implicit-def: $vgpr41
-; VI-NEXT:    ; implicit-def: $vgpr49
+; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr36
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v34 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr34
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr50
 ; VI-NEXT:    ; implicit-def: $vgpr52
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr37
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -205056,35 +205179,35 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_or_b32_sdwa v10, v63, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v61, v10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_or_b32_sdwa v12, v58, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v59, v12 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_or_b32_sdwa v14, v45, v14 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr58
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    ; implicit-def: $vgpr45
 ; VI-NEXT:    s_waitcnt vmcnt(2)
@@ -205093,26 +205216,26 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v63, v9 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr62
-; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v60, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr59
+; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v10, v57, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v56, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr56
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v58, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr56
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -205121,35 +205244,35 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v13, v44, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; VI-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -205164,20 +205287,20 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -205185,189 +205308,189 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v31, v31, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr48
+; VI-NEXT:    v_or_b32_sdwa v31, v31, v49 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr49
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v32, v32, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr53
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v30, v30, v37 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v30, v30, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
-; VI-NEXT:    ; kill: killed $vgpr37
-; VI-NEXT:    ; implicit-def: $vgpr37
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; kill: killed $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v30, v30, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v30, v30, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr39
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v31, v31, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v31, v31, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr32
 ; VI-NEXT:    ; kill: killed $vgpr32
@@ -205472,7 +205595,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; implicit-def: $vgpr32
 ; VI-NEXT:    ; kill: killed $vgpr32
 ; VI-NEXT:    ; implicit-def: $vgpr32
-; VI-NEXT:    ; implicit-def: $vgpr50
+; VI-NEXT:    ; implicit-def: $vgpr51
 ; VI-NEXT:  .LBB96_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB96_4
@@ -205480,53 +205603,51 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v18, 0x300
-; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
 ; VI-NEXT:    v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(12)
 ; VI-NEXT:    v_or_b32_sdwa v29, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(13)
+; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_sdwa v0, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v2, 0x300, v3
 ; VI-NEXT:    v_or_b32_e32 v0, v2, v0
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v2, 3, v2
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v2, 3, v2
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
 ; VI-NEXT:    v_or_b32_sdwa v2, v52, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v3, v51, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v2, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
 ; VI-NEXT:    v_or_b32_e32 v2, v3, v2
@@ -205536,19 +205657,19 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v1, 0x300, v1
 ; VI-NEXT:    v_or_b32_e32 v1, v1, v4
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v3, v49, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v48, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v3, v3, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
-; VI-NEXT:    v_or_b32_sdwa v4, v39, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v4, v37, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
 ; VI-NEXT:    v_or_b32_e32 v3, v4, v3
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
@@ -205556,14 +205677,13 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
 ; VI-NEXT:    v_or_b32_sdwa v6, v33, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
-; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    s_waitcnt vmcnt(2)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v5, 3, v5
 ; VI-NEXT:    v_or_b32_sdwa v5, v35, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v4, 3, v4
 ; VI-NEXT:    v_or_b32_sdwa v4, v36, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v4, v4, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
@@ -205580,7 +205700,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_or_b32_sdwa v32, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_sdwa v32, v32, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
@@ -205588,7 +205708,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v28, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
 ; VI-NEXT:    v_or_b32_e32 v28, v28, v32
 ; VI-NEXT:    s_waitcnt vmcnt(1)
@@ -205596,79 +205716,94 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v33, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v33, v33, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v27, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
 ; VI-NEXT:    v_or_b32_e32 v27, v27, v33
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v34, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v34, v34, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v26, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
 ; VI-NEXT:    v_or_b32_e32 v26, v26, v34
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v35, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v35, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v25, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
+; VI-NEXT:    v_or_b32_e32 v25, v25, v35
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v36, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
-; VI-NEXT:    v_or_b32_e32 v25, v25, v35
+; VI-NEXT:    v_add_u16_sdwa v36, v36, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v36, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_sdwa v36, v36, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v24, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
+; VI-NEXT:    v_or_b32_e32 v24, v24, v36
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v24, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v37, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_sdwa v23, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
 ; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
-; VI-NEXT:    v_or_b32_e32 v24, v24, v36
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
+; VI-NEXT:    v_or_b32_e32 v23, v23, v37
 ; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    v_add_u16_e32 v22, 3, v22
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v8, 3, v8
 ; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v7, v8, v7
-; VI-NEXT:    v_add_u16_e32 v8, 3, v61
+; VI-NEXT:    v_add_u16_e32 v8, 3, v63
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v9, 3, v62
@@ -205677,30 +205812,30 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; VI-NEXT:    v_or_b32_e32 v8, v9, v8
-; VI-NEXT:    v_add_u16_e32 v9, 3, v63
+; VI-NEXT:    v_add_u16_e32 v9, 3, v61
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v10, 3, v59
+; VI-NEXT:    v_add_u16_e32 v10, 3, v60
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v9, v9, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_or_b32_e32 v9, v10, v9
-; VI-NEXT:    v_add_u16_e32 v10, 3, v60
+; VI-NEXT:    v_add_u16_e32 v10, 3, v57
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v11, 3, v57
+; VI-NEXT:    v_add_u16_e32 v11, 3, v56
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v10, v10, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_or_b32_e32 v10, v11, v10
-; VI-NEXT:    v_add_u16_e32 v11, 3, v58
+; VI-NEXT:    v_add_u16_e32 v11, 3, v59
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v12, 3, v56
+; VI-NEXT:    v_add_u16_e32 v12, 3, v58
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; VI-NEXT:    v_or_b32_e32 v11, v12, v11
@@ -205709,7 +205844,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v13, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_or_b32_e32 v12, v13, v12
@@ -205718,7 +205853,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 3, v44
 ; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v13, v13, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_or_b32_e32 v13, v14, v13
@@ -205730,35 +205865,35 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; VI-NEXT:    v_or_b32_e32 v14, v15, v14
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v15, 3, v15
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    v_or_b32_e32 v15, v16, v15
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
 ; VI-NEXT:    v_or_b32_sdwa v17, v19, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
 ; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -205767,54 +205902,43 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v16, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v19, 0x300, v20
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; VI-NEXT:    v_or_b32_e32 v16, v19, v16
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
-; VI-NEXT:    v_or_b32_sdwa v30, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v30, v39, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
-; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v20, 3, v20
-; VI-NEXT:    v_or_b32_sdwa v31, v50, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v31, v51, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v31, 0x300, v31
-; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v21, 3, v21
-; VI-NEXT:    v_or_b32_sdwa v21, v37, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v37, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; VI-NEXT:    v_or_b32_sdwa v21, v38, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v40, v21, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v29, v29, v40
-; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_add_u16_e32 v22, 3, v22
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v23, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
-; VI-NEXT:    v_or_b32_e32 v23, v23, v37
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v38, v38, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_e32 v20, 3, v20
-; VI-NEXT:    v_or_b32_sdwa v20, v48, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
-; VI-NEXT:    v_add_u16_sdwa v55, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_sdwa v38, v38, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_or_b32_e32 v30, v30, v55
-; VI-NEXT:    s_waitcnt vmcnt(1)
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v22, 3, v22
 ; VI-NEXT:    v_or_b32_sdwa v22, v39, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v20, 3, v20
+; VI-NEXT:    v_or_b32_sdwa v20, v49, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_sdwa v55, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
 ; VI-NEXT:    v_or_b32_e32 v22, v22, v38
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_or_b32_e32 v30, v30, v55
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v39, 3, v39
 ; VI-NEXT:    v_or_b32_sdwa v39, v48, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
@@ -205836,7 +205960,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; VI-NEXT:    v_add_u16_e32 v19, 3, v19
 ; VI-NEXT:    v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v39, 3, v39
 ; VI-NEXT:    v_or_b32_sdwa v39, v49, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -205938,17 +206062,16 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:112
 ; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
 ; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:136
+; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:136
 ; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:144
 ; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:152
 ; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:160
 ; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:168
 ; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:176
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:184
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v1
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v27
-; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v29
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v3
@@ -205956,81 +206079,81 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v7
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v9
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v11
-; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v13
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v13
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v15
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v17
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v35, 8, v19
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v21
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v23
-; GFX9-NEXT:    s_waitcnt vmcnt(27)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
+; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v25
 ; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v4
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v6
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v20
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v22
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v24
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v26
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(26)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
-; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v28
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:140
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v31
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
-; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v25
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v37
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v39
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v48
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v49
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
@@ -206038,7 +206161,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
@@ -206055,7 +206178,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v43
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v44
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:200
@@ -206064,15 +206187,15 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:188
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:196
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
@@ -206096,20 +206219,20 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:228
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:236
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:244
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:264
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:272
@@ -206123,20 +206246,20 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:260
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:268
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:276
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:296
 ; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:304
@@ -206145,17 +206268,17 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:284
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:292
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:300
@@ -206182,7 +206305,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v3
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:332
@@ -206202,48 +206325,48 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:356
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:364
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:372
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:384
 ; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:12
 ; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:100
 ; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:92
 ; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB96_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
@@ -206252,9 +206375,9 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
@@ -206283,10 +206406,10 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; implicit-def: $vgpr51
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr48
+; GFX9-NEXT:    ; implicit-def: $vgpr39
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
@@ -206297,7 +206420,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr34
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -206312,93 +206435,93 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v60, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v60, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v56, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v56, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v58, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v46, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v44, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v15, v43, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr43
+; GFX9-NEXT:    v_or_b32_sdwa v15, v42, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v63, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v62, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v61, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr59
+; GFX9-NEXT:    ; implicit-def: $vgpr61
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v58, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v57, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v11, v57, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v59, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v12, v46, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v47, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr46
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    ; implicit-def: $vgpr47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v45, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v43, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr42
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -206413,20 +206536,20 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
@@ -206434,58 +206557,58 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
@@ -206503,22 +206626,22 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
 ; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v32, v32, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr53
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
-; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr37
 ; GFX9-NEXT:    ; kill: killed $vgpr37
 ; GFX9-NEXT:    ; implicit-def: $vgpr37
@@ -206615,7 +206738,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
-; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr49
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
@@ -206727,27 +206850,27 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_cbranch_execz .LBB96_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
@@ -206755,19 +206878,20 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
+; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
 ; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v55, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    s_waitcnt vmcnt(20)
+; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(17)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
 ; GFX9-NEXT:    v_or_b32_sdwa v3, v54, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v2
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v3
-; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
 ; GFX9-NEXT:    v_perm_b32 v0, v2, v0, s6
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(17)
+; GFX9-NEXT:    s_waitcnt vmcnt(16)
 ; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
@@ -206792,7 +206916,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
-; GFX9-NEXT:    v_or_b32_sdwa v3, v48, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
@@ -206818,7 +206942,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_or_b32_sdwa v36, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
@@ -206842,11 +206966,11 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_or_b32_sdwa v23, v37, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v37, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
-; GFX9-NEXT:    v_or_b32_sdwa v21, v39, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v21, v48, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v33, 0x300, v21
 ; GFX9-NEXT:    v_add_u16_e32 v34, 0x300, v23
 ; GFX9-NEXT:    v_perm_b32 v29, v34, v29, s6
@@ -206854,16 +206978,17 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v32, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v38, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
@@ -206873,7 +206998,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v39, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
@@ -206883,7 +207008,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v48, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -206896,45 +207021,45 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v62
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v61
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v60
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v58
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v57
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v60
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v56
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v57
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v56
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v58
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v46
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v47
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v46
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
@@ -206943,63 +207068,63 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 3, v44
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v43
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v15, 3, v43
+; GFX9-NEXT:    v_add_u16_e32 v15, 3, v42
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
 ; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v16
 ; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v18
 ; GFX9-NEXT:    v_perm_b32 v17, v17, v20, s6
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v19
 ; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v16, v18, v16, s6
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
 ; GFX9-NEXT:    v_or_b32_sdwa v20, v49, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_or_b32_sdwa v49, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v20
 ; GFX9-NEXT:    v_perm_b32 v30, v33, v30, s6
@@ -207007,7 +207132,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v50, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
 ; GFX9-NEXT:    v_or_b32_sdwa v18, v52, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -207017,14 +207142,14 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v51, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v51
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v52, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
 ; GFX9-NEXT:    v_or_b32_sdwa v19, v53, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -207034,7 +207159,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v53, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v53
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -207047,7 +207172,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v55, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v55
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -207055,7 +207180,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v40, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v50, 0x300, v40
 ; GFX9-NEXT:    v_perm_b32 v21, v50, v21, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -207063,14 +207188,14 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v41, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v41
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v42, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v51, 0x300, v42
 ; GFX9-NEXT:    v_perm_b32 v20, v51, v20, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
@@ -207078,7 +207203,7 @@ define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v43, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v43
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
@@ -208842,8 +208967,8 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI:       ; %bb.0:
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
-; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
 ; SI-NEXT:    s_mov_b64 exec, s[4:5]
 ; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
@@ -208859,41 +208984,48 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:332
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:328
-; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:324
-; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:320
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:308
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:304
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:332
+; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:328
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:324
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:320
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:308
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:304
 ; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:300
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:296
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:292
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:288
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:276
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:272
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:268
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:264
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:260
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:292
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:288
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:272
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:268
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:264
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:260
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:256
 ; SI-NEXT:    s_waitcnt expcnt(6)
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:256
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:244
 ; SI-NEXT:    s_waitcnt expcnt(5)
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:244
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:240
 ; SI-NEXT:    s_waitcnt expcnt(4)
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:240
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:236
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:232
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:228
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:224
-; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:212
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:236
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:232
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:228
+; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:224
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:212
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:208
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:204
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:200
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:196
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:204
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:200
+; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:196
+; SI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
 ; SI-NEXT:    v_writelane_b32 v63, s30, 0
+; SI-NEXT:    v_writelane_b32 v62, s28, 0
+; SI-NEXT:    v_writelane_b32 v62, s25, 1
+; SI-NEXT:    v_writelane_b32 v62, s24, 2
+; SI-NEXT:    v_writelane_b32 v62, s23, 3
+; SI-NEXT:    v_writelane_b32 v62, s22, 4
+; SI-NEXT:    v_writelane_b32 v62, s21, 5
+; SI-NEXT:    v_writelane_b32 v62, s18, 6
+; SI-NEXT:    v_writelane_b32 v62, s16, 7
 ; SI-NEXT:    v_writelane_b32 v63, s31, 1
 ; SI-NEXT:    v_writelane_b32 v63, s34, 2
 ; SI-NEXT:    v_writelane_b32 v63, s35, 3
@@ -208911,635 +209043,625 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    v_writelane_b32 v63, s55, 15
 ; SI-NEXT:    v_writelane_b32 v63, s64, 16
 ; SI-NEXT:    v_writelane_b32 v63, s65, 17
-; SI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
 ; SI-NEXT:    v_writelane_b32 v63, s66, 18
-; SI-NEXT:    v_writelane_b32 v62, s28, 0
 ; SI-NEXT:    v_writelane_b32 v63, s67, 19
-; SI-NEXT:    v_writelane_b32 v62, s27, 1
 ; SI-NEXT:    v_writelane_b32 v63, s68, 20
-; SI-NEXT:    v_writelane_b32 v62, s25, 2
 ; SI-NEXT:    v_writelane_b32 v63, s69, 21
-; SI-NEXT:    v_writelane_b32 v62, s24, 3
 ; SI-NEXT:    v_writelane_b32 v63, s70, 22
-; SI-NEXT:    v_writelane_b32 v62, s23, 4
 ; SI-NEXT:    v_writelane_b32 v63, s71, 23
-; SI-NEXT:    v_writelane_b32 v62, s22, 5
 ; SI-NEXT:    v_writelane_b32 v63, s80, 24
-; SI-NEXT:    v_writelane_b32 v62, s21, 6
 ; SI-NEXT:    v_writelane_b32 v63, s81, 25
-; SI-NEXT:    v_writelane_b32 v62, s20, 7
 ; SI-NEXT:    v_writelane_b32 v63, s82, 26
-; SI-NEXT:    v_writelane_b32 v62, s19, 8
 ; SI-NEXT:    v_writelane_b32 v63, s83, 27
-; SI-NEXT:    v_writelane_b32 v62, s18, 9
 ; SI-NEXT:    v_writelane_b32 v63, s84, 28
-; SI-NEXT:    v_writelane_b32 v62, s16, 10
 ; SI-NEXT:    v_writelane_b32 v63, s85, 29
 ; SI-NEXT:    v_writelane_b32 v63, s86, 30
 ; SI-NEXT:    v_writelane_b32 v63, s87, 31
 ; SI-NEXT:    v_writelane_b32 v63, s96, 32
 ; SI-NEXT:    v_writelane_b32 v63, s97, 33
 ; SI-NEXT:    v_writelane_b32 v63, s98, 34
-; SI-NEXT:    s_waitcnt expcnt(2)
-; SI-NEXT:    v_mov_b32_e32 v29, v5
-; SI-NEXT:    v_readfirstlane_b32 s76, v18
-; SI-NEXT:    v_readfirstlane_b32 s40, v25
-; SI-NEXT:    v_readfirstlane_b32 s16, v24
-; SI-NEXT:    v_readfirstlane_b32 s42, v23
-; SI-NEXT:    v_readfirstlane_b32 s52, v20
-; SI-NEXT:    v_readfirstlane_b32 s8, v19
+; SI-NEXT:    v_writelane_b32 v63, s99, 35
+; SI-NEXT:    s_waitcnt expcnt(1)
+; SI-NEXT:    v_mov_b32_e32 v29, v26
+; SI-NEXT:    v_readfirstlane_b32 s15, v16
+; SI-NEXT:    v_readfirstlane_b32 s18, v25
+; SI-NEXT:    v_readfirstlane_b32 s43, v15
+; SI-NEXT:    v_readfirstlane_b32 s42, v24
+; SI-NEXT:    v_readfirstlane_b32 s44, v23
+; SI-NEXT:    v_readfirstlane_b32 s49, v12
+; SI-NEXT:    v_readfirstlane_b32 s8, v11
+; SI-NEXT:    v_readfirstlane_b32 s53, v20
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_readfirstlane_b32 s50, v35
-; SI-NEXT:    v_readfirstlane_b32 s31, v36
-; SI-NEXT:    v_readfirstlane_b32 s53, v37
-; SI-NEXT:    v_readfirstlane_b32 s82, v48
-; SI-NEXT:    v_readfirstlane_b32 s7, v49
-; SI-NEXT:    v_readfirstlane_b32 s79, v52
-; SI-NEXT:    v_readfirstlane_b32 s78, v55
-; SI-NEXT:    v_readfirstlane_b32 s4, v41
+; SI-NEXT:    v_readfirstlane_b32 s4, v34
+; SI-NEXT:    v_writelane_b32 v62, s4, 8
+; SI-NEXT:    v_readfirstlane_b32 s4, v38
+; SI-NEXT:    v_writelane_b32 v62, s4, 9
+; SI-NEXT:    v_readfirstlane_b32 s4, v49
+; SI-NEXT:    v_writelane_b32 v62, s4, 10
+; SI-NEXT:    v_readfirstlane_b32 s4, v50
 ; SI-NEXT:    v_writelane_b32 v62, s4, 11
-; SI-NEXT:    v_readfirstlane_b32 s4, v43
+; SI-NEXT:    v_readfirstlane_b32 s79, v52
+; SI-NEXT:    v_readfirstlane_b32 s88, v54
+; SI-NEXT:    v_readfirstlane_b32 s4, v55
+; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:192
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:180
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:176
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:172
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:168
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:164
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:160
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:148
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:144
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:140
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:136
 ; SI-NEXT:    v_writelane_b32 v62, s4, 12
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:192
-; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:180
-; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:176
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:172
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:168
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:164
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:160
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:148
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:144
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:140
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:136
-; SI-NEXT:    v_readfirstlane_b32 s4, v45
-; SI-NEXT:    v_writelane_b32 v62, s4, 13
+; SI-NEXT:    v_readfirstlane_b32 s77, v41
+; SI-NEXT:    v_readfirstlane_b32 s4, v42
+; SI-NEXT:    v_readfirstlane_b32 s94, v31
+; SI-NEXT:    v_readfirstlane_b32 s70, v32
+; SI-NEXT:    v_readfirstlane_b32 s51, v33
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_readfirstlane_b32 s4, v56
-; SI-NEXT:    v_writelane_b32 v62, s4, 14
-; SI-NEXT:    v_readfirstlane_b32 s4, v57
-; SI-NEXT:    v_writelane_b32 v62, s4, 15
-; SI-NEXT:    v_readfirstlane_b32 s4, v60
-; SI-NEXT:    v_readfirstlane_b32 s86, v31
-; SI-NEXT:    v_readfirstlane_b32 s36, v32
-; SI-NEXT:    v_readfirstlane_b32 s71, v33
-; SI-NEXT:    v_readfirstlane_b32 s77, v59
-; SI-NEXT:    v_writelane_b32 v62, s4, 16
-; SI-NEXT:    v_readfirstlane_b32 s4, v61
+; SI-NEXT:    v_readfirstlane_b32 s37, v45
+; SI-NEXT:    v_readfirstlane_b32 s24, v56
+; SI-NEXT:    v_readfirstlane_b32 s7, v57
+; SI-NEXT:    v_readfirstlane_b32 s92, v58
+; SI-NEXT:    v_readfirstlane_b32 s28, v59
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:132
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:128
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:116
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:100
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:84
-; SI-NEXT:    v_readfirstlane_b32 s98, v50
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:80
+; SI-NEXT:    v_readfirstlane_b32 s35, v43
+; SI-NEXT:    v_readfirstlane_b32 s55, v46
+; SI-NEXT:    v_readfirstlane_b32 s68, v35
+; SI-NEXT:    v_readfirstlane_b32 s87, v37
+; SI-NEXT:    v_readfirstlane_b32 s67, v39
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:76
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_readfirstlane_b32 s67, v53
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:64
-; SI-NEXT:    v_writelane_b32 v62, s4, 17
-; SI-NEXT:    v_readfirstlane_b32 s4, v38
-; SI-NEXT:    v_writelane_b32 v62, s4, 18
-; SI-NEXT:    v_readfirstlane_b32 s81, v34
-; SI-NEXT:    v_readfirstlane_b32 s75, v39
-; SI-NEXT:    v_readfirstlane_b32 s68, v42
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:80
-; SI-NEXT:    v_readfirstlane_b32 s49, v40
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:40
-; SI-NEXT:    v_readfirstlane_b32 s51, v54
-; SI-NEXT:    v_readfirstlane_b32 s97, v51
-; SI-NEXT:    v_readfirstlane_b32 s35, v27
-; SI-NEXT:    v_writelane_b32 v63, s99, 35
-; SI-NEXT:    v_readfirstlane_b32 s28, v28
-; SI-NEXT:    v_readfirstlane_b32 s87, v26
+; SI-NEXT:    v_readfirstlane_b32 s74, v53
+; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:64
+; SI-NEXT:    v_readfirstlane_b32 s85, v48
+; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:40
+; SI-NEXT:    v_writelane_b32 v62, s4, 13
+; SI-NEXT:    v_readfirstlane_b32 s98, v40
+; SI-NEXT:    v_readfirstlane_b32 s69, v51
+; SI-NEXT:    v_readfirstlane_b32 s21, v36
+; SI-NEXT:    v_readfirstlane_b32 s40, v19
+; SI-NEXT:    v_readfirstlane_b32 s23, v28
+; SI-NEXT:    v_readfirstlane_b32 s34, v27
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v21, v13
+; SI-NEXT:    v_mov_b32_e32 v13, v5
+; SI-NEXT:    v_readfirstlane_b32 s97, v29
+; SI-NEXT:    v_readfirstlane_b32 s80, v18
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v18, 24, v22
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v30
 ; SI-NEXT:    v_readfirstlane_b32 s96, v17
-; SI-NEXT:    v_readfirstlane_b32 s99, v16
-; SI-NEXT:    v_readfirstlane_b32 s89, v15
-; SI-NEXT:    v_readfirstlane_b32 s88, v12
-; SI-NEXT:    v_readfirstlane_b32 s30, v11
-; SI-NEXT:    v_readfirstlane_b32 s64, v10
-; SI-NEXT:    v_readfirstlane_b32 s55, v9
-; SI-NEXT:    v_readfirstlane_b32 s65, v8
-; SI-NEXT:    v_readfirstlane_b32 s80, v7
-; SI-NEXT:    v_readfirstlane_b32 s21, v2
-; SI-NEXT:    v_readfirstlane_b32 s74, v1
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:156
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:220
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:72
-; SI-NEXT:    v_readfirstlane_b32 s93, v36
-; SI-NEXT:    v_readfirstlane_b32 s24, v37
-; SI-NEXT:    v_readfirstlane_b32 s27, v48
-; SI-NEXT:    v_readfirstlane_b32 s84, v43
-; SI-NEXT:    v_readfirstlane_b32 s83, v44
-; SI-NEXT:    v_readfirstlane_b32 s85, v46
+; SI-NEXT:    v_readfirstlane_b32 s64, v9
+; SI-NEXT:    v_readfirstlane_b32 s25, v8
+; SI-NEXT:    v_readfirstlane_b32 s83, v7
+; SI-NEXT:    v_readfirstlane_b32 s84, v4
+; SI-NEXT:    v_readfirstlane_b32 s93, v3
+; SI-NEXT:    v_readfirstlane_b32 s76, v1
+; SI-NEXT:    v_readfirstlane_b32 s58, v38
+; SI-NEXT:    v_readfirstlane_b32 s65, v49
+; SI-NEXT:    v_readfirstlane_b32 s62, v54
+; SI-NEXT:    v_readfirstlane_b32 s81, v44
+; SI-NEXT:    v_readfirstlane_b32 s71, v47
+; SI-NEXT:    v_readfirstlane_b32 s38, v60
+; SI-NEXT:    v_readfirstlane_b32 s86, v61
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:156
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:220
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_readfirstlane_b32 s20, v47
-; SI-NEXT:    v_readfirstlane_b32 s4, v58
-; SI-NEXT:    v_writelane_b32 v62, s4, 19
-; SI-NEXT:    v_readfirstlane_b32 s23, v49
-; SI-NEXT:    v_readfirstlane_b32 s92, v52
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:336
+; SI-NEXT:    v_readfirstlane_b32 s90, v50
+; SI-NEXT:    v_readfirstlane_b32 s31, v52
+; SI-NEXT:    v_readfirstlane_b32 s4, v55
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:336
 ; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:4
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32
-; SI-NEXT:    v_readfirstlane_b32 s90, v35
-; SI-NEXT:    v_readfirstlane_b32 s4, v31
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:92
-; SI-NEXT:    v_readfirstlane_b32 s38, v32
-; SI-NEXT:    v_readfirstlane_b32 s70, v33
-; SI-NEXT:    v_readfirstlane_b32 s54, v59
-; SI-NEXT:    v_readfirstlane_b32 s57, v60
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32
+; SI-NEXT:    v_readfirstlane_b32 s72, v31
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:316
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:72
+; SI-NEXT:    v_readfirstlane_b32 s82, v56
+; SI-NEXT:    v_readfirstlane_b32 s95, v57
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_readfirstlane_b32 s56, v61
-; SI-NEXT:    v_readfirstlane_b32 s59, v55
-; SI-NEXT:    v_readfirstlane_b32 s61, v41
-; SI-NEXT:    v_readfirstlane_b32 s19, v45
-; SI-NEXT:    v_readfirstlane_b32 s34, v50
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:316
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:284
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:252
+; SI-NEXT:    v_readfirstlane_b32 s39, v58
+; SI-NEXT:    v_readfirstlane_b32 s56, v59
+; SI-NEXT:    v_readfirstlane_b32 s57, v41
+; SI-NEXT:    v_readfirstlane_b32 s36, v42
+; SI-NEXT:    v_readfirstlane_b32 s73, v45
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:284
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:252
 ; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:188
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:124
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:28
-; SI-NEXT:    v_writelane_b32 v62, s4, 20
-; SI-NEXT:    v_readfirstlane_b32 s25, v53
-; SI-NEXT:    v_readfirstlane_b32 s91, v40
-; SI-NEXT:    v_readfirstlane_b32 s37, v34
-; SI-NEXT:    v_readfirstlane_b32 s47, v56
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:124
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:92
+; SI-NEXT:    v_readfirstlane_b32 s16, v34
+; SI-NEXT:    v_readfirstlane_b32 s48, v32
+; SI-NEXT:    v_readfirstlane_b32 s52, v33
+; SI-NEXT:    v_writelane_b32 v62, s4, 14
+; SI-NEXT:    v_readfirstlane_b32 s47, v35
+; SI-NEXT:    v_readfirstlane_b32 s60, v37
+; SI-NEXT:    v_readfirstlane_b32 s61, v39
+; SI-NEXT:    v_readfirstlane_b32 s89, v43
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_readfirstlane_b32 s46, v57
-; SI-NEXT:    v_readfirstlane_b32 s22, v42
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v22
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 24, v5
-; SI-NEXT:    v_readfirstlane_b32 s72, v39
-; SI-NEXT:    v_readfirstlane_b32 s94, v51
-; SI-NEXT:    v_readfirstlane_b32 s48, v54
-; SI-NEXT:    v_readfirstlane_b32 s66, v43
-; SI-NEXT:    v_readfirstlane_b32 s69, v44
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_readfirstlane_b32 s45, v46
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v49
-; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
-; SI-NEXT:    v_readfirstlane_b32 s4, v4
-; SI-NEXT:    v_writelane_b32 v62, s4, 21
-; SI-NEXT:    v_readfirstlane_b32 s4, v3
-; SI-NEXT:    v_writelane_b32 v62, s4, 22
-; SI-NEXT:    v_writelane_b32 v62, s17, 23
-; SI-NEXT:    v_writelane_b32 v62, s40, 24
-; SI-NEXT:    v_writelane_b32 v62, s16, 25
-; SI-NEXT:    v_writelane_b32 v62, s42, 26
-; SI-NEXT:    v_writelane_b32 v62, s46, 27
-; SI-NEXT:    v_writelane_b32 v62, s47, 28
-; SI-NEXT:    v_writelane_b32 v62, s56, 29
-; SI-NEXT:    v_writelane_b32 v62, s57, 30
-; SI-NEXT:    v_writelane_b32 v62, s45, 31
-; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_readfirstlane_b32 s58, v52
-; SI-NEXT:    v_writelane_b32 v62, s49, 32
-; SI-NEXT:    v_writelane_b32 v62, s58, 33
-; SI-NEXT:    v_writelane_b32 v62, s59, 34
-; SI-NEXT:    v_writelane_b32 v62, s52, 35
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_readfirstlane_b32 s60, v38
+; SI-NEXT:    v_readfirstlane_b32 s99, v46
 ; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:312
-; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:280
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:248
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:280
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:248
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:216
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:184
+; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:184
 ; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:152
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:120
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; SI-NEXT:    v_writelane_b32 v62, s60, 36
-; SI-NEXT:    v_writelane_b32 v62, s61, 37
-; SI-NEXT:    v_writelane_b32 v62, s93, 38
-; SI-NEXT:    v_writelane_b32 v62, s8, 39
-; SI-NEXT:    v_readfirstlane_b32 s62, v47
-; SI-NEXT:    v_writelane_b32 v62, s72, 40
-; SI-NEXT:    v_readfirstlane_b32 s73, v58
-; SI-NEXT:    v_writelane_b32 v62, s62, 41
-; SI-NEXT:    v_writelane_b32 v62, s73, 42
-; SI-NEXT:    v_writelane_b32 v62, s35, 43
-; SI-NEXT:    v_writelane_b32 v62, s94, 44
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:24
+; SI-NEXT:    v_readfirstlane_b32 s54, v48
+; SI-NEXT:    v_readfirstlane_b32 s50, v53
+; SI-NEXT:    v_readfirstlane_b32 s78, v49
+; SI-NEXT:    v_readfirstlane_b32 s30, v51
+; SI-NEXT:    v_readfirstlane_b32 s66, v54
+; SI-NEXT:    v_readfirstlane_b32 s91, v40
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_readfirstlane_b32 s6, v44
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v50
+; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
+; SI-NEXT:    v_readfirstlane_b32 s4, v10
+; SI-NEXT:    v_writelane_b32 v62, s4, 15
+; SI-NEXT:    v_readfirstlane_b32 s4, v2
+; SI-NEXT:    v_writelane_b32 v62, s4, 16
+; SI-NEXT:    v_writelane_b32 v62, s17, 17
+; SI-NEXT:    v_writelane_b32 v62, s15, 18
+; SI-NEXT:    v_writelane_b32 v62, s18, 19
+; SI-NEXT:    v_writelane_b32 v62, s43, 20
+; SI-NEXT:    v_writelane_b32 v62, s42, 21
+; SI-NEXT:    v_writelane_b32 v62, s44, 22
+; SI-NEXT:    v_writelane_b32 v62, s16, 23
+; SI-NEXT:    v_writelane_b32 v62, s49, 24
+; SI-NEXT:    v_writelane_b32 v62, s8, 25
+; SI-NEXT:    v_writelane_b32 v62, s6, 26
+; SI-NEXT:    v_readfirstlane_b32 s45, v52
+; SI-NEXT:    v_writelane_b32 v62, s56, 27
+; SI-NEXT:    v_writelane_b32 v62, s45, 28
+; SI-NEXT:    v_writelane_b32 v62, s53, 29
+; SI-NEXT:    v_writelane_b32 v62, s94, 30
+; SI-NEXT:    v_writelane_b32 v62, s57, 31
+; SI-NEXT:    v_writelane_b32 v62, s58, 32
+; SI-NEXT:    v_writelane_b32 v62, s47, 33
+; SI-NEXT:    v_readfirstlane_b32 s46, v55
+; SI-NEXT:    v_writelane_b32 v62, s40, 34
+; SI-NEXT:    v_readfirstlane_b32 s59, v47
+; SI-NEXT:    v_writelane_b32 v62, s46, 35
+; SI-NEXT:    v_writelane_b32 v62, s59, 36
+; SI-NEXT:    v_writelane_b32 v62, s60, 37
+; SI-NEXT:    v_writelane_b32 v62, s36, 38
+; SI-NEXT:    v_writelane_b32 v62, s65, 39
+; SI-NEXT:    v_writelane_b32 v62, s61, 40
+; SI-NEXT:    v_writelane_b32 v62, s73, 41
+; SI-NEXT:    v_writelane_b32 v62, s62, 42
+; SI-NEXT:    v_writelane_b32 v62, s72, 43
+; SI-NEXT:    v_writelane_b32 v62, s23, 44
 ; SI-NEXT:    v_writelane_b32 v62, s48, 45
-; SI-NEXT:    v_writelane_b32 v62, s91, 46
+; SI-NEXT:    v_writelane_b32 v62, s34, 46
+; SI-NEXT:    v_writelane_b32 v62, s78, 47
+; SI-NEXT:    v_writelane_b32 v62, s30, 48
+; SI-NEXT:    v_writelane_b32 v62, s54, 49
+; SI-NEXT:    v_writelane_b32 v62, s50, 50
+; SI-NEXT:    v_writelane_b32 v62, s52, 51
+; SI-NEXT:    v_writelane_b32 v62, s82, 52
+; SI-NEXT:    v_writelane_b32 v62, s66, 53
+; SI-NEXT:    v_readfirstlane_b32 s22, v36
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_lshlrev_b32_e32 v22, 24, v57
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 24, v61
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v60
-; SI-NEXT:    v_lshlrev_b32_e32 v31, 24, v31
-; SI-NEXT:    v_lshlrev_b32_e32 v24, 24, v59
-; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v45
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 24, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 24, v41
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 24, v55
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 24, v50
-; SI-NEXT:    v_writelane_b32 v62, s66, 47
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v18
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; SI-NEXT:    s_cbranch_scc0 .LBB97_2
+; SI-NEXT:    v_lshlrev_b32_e32 v23, 24, v58
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 24, v59
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 24, v56
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 24, v60
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 24, v45
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 24, v61
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v42
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 24, v41
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v31
+; SI-NEXT:    v_writelane_b32 v62, s91, 54
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; SI-NEXT:    s_cbranch_scc0 .LBB97_4
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; SI-NEXT:    v_mov_b32_e32 v21, v29
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v29
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; SI-NEXT:    v_readlane_b32 s4, v62, 7
-; SI-NEXT:    v_readlane_b32 s5, v62, 6
-; SI-NEXT:    s_and_b32 s4, s4, 0xff
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; SI-NEXT:    v_mov_b32_e32 v5, v13
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v13
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; SI-NEXT:    v_readlane_b32 s5, v62, 5
+; SI-NEXT:    s_and_b32 s4, s20, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_or_b32 s4, s4, s5
-; SI-NEXT:    v_writelane_b32 v62, s4, 48
-; SI-NEXT:    v_readlane_b32 s4, v62, 5
+; SI-NEXT:    v_writelane_b32 v62, s4, 55
+; SI-NEXT:    v_readlane_b32 s4, v62, 4
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
-; SI-NEXT:    v_readlane_b32 s5, v62, 4
+; SI-NEXT:    v_readlane_b32 s5, v62, 3
 ; SI-NEXT:    s_lshl_b32 s4, s4, 16
 ; SI-NEXT:    s_lshl_b32 s5, s5, 24
 ; SI-NEXT:    s_or_b32 s63, s5, s4
-; SI-NEXT:    v_readlane_b32 s4, v62, 9
+; SI-NEXT:    v_readlane_b32 s4, v62, 6
 ; SI-NEXT:    s_and_b32 s5, s4, 0xff
-; SI-NEXT:    v_readlane_b32 s4, v62, 8
 ; SI-NEXT:    s_lshl_b32 s5, s5, 16
-; SI-NEXT:    s_lshl_b32 s9, s4, 24
+; SI-NEXT:    s_lshl_b32 s9, s19, 24
 ; SI-NEXT:    v_readlane_b32 s4, v62, 0
 ; SI-NEXT:    s_or_b32 s9, s9, s5
 ; SI-NEXT:    s_and_b32 s5, s4, 0xff
 ; SI-NEXT:    s_lshl_b32 s10, s29, 8
 ; SI-NEXT:    s_or_b32 s4, s5, s10
-; SI-NEXT:    v_writelane_b32 v62, s4, 49
-; SI-NEXT:    s_and_b32 s5, s74, 0xff
+; SI-NEXT:    v_writelane_b32 v62, s4, 56
+; SI-NEXT:    s_and_b32 s5, s76, 0xff
+; SI-NEXT:    v_readlane_b32 s10, v62, 16
 ; SI-NEXT:    s_lshl_b32 s5, s5, 16
-; SI-NEXT:    v_writelane_b32 v62, s54, 50
-; SI-NEXT:    s_lshl_b32 s11, s21, 24
-; SI-NEXT:    s_mov_b32 s18, s22
-; SI-NEXT:    s_mov_b32 s22, s21
-; SI-NEXT:    s_or_b32 s21, s11, s5
+; SI-NEXT:    s_lshl_b32 s11, s10, 24
+; SI-NEXT:    s_or_b32 s5, s11, s5
 ; SI-NEXT:    s_and_b32 s11, s26, 0xff
-; SI-NEXT:    v_readlane_b32 s4, v62, 1
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s12, s4, 24
+; SI-NEXT:    s_lshl_b32 s12, s27, 24
 ; SI-NEXT:    s_or_b32 s14, s12, s11
-; SI-NEXT:    s_and_b32 s11, s80, 0xff
-; SI-NEXT:    s_lshl_b32 s12, s65, 8
-; SI-NEXT:    s_or_b32 s12, s11, s12
-; SI-NEXT:    s_and_b32 s11, s55, 0xff
+; SI-NEXT:    s_and_b32 s11, s83, 0xff
+; SI-NEXT:    s_lshl_b32 s12, s25, 8
+; SI-NEXT:    s_or_b32 s10, s11, s12
+; SI-NEXT:    v_writelane_b32 v62, s10, 57
+; SI-NEXT:    s_and_b32 s11, s64, 0xff
+; SI-NEXT:    v_readlane_b32 s10, v62, 15
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s13, s64, 24
+; SI-NEXT:    s_lshl_b32 s13, s10, 24
 ; SI-NEXT:    s_or_b32 s41, s13, s11
-; SI-NEXT:    s_and_b32 s11, s89, 0xff
-; SI-NEXT:    s_lshl_b32 s13, s99, 8
-; SI-NEXT:    s_or_b32 s13, s11, s13
+; SI-NEXT:    s_and_b32 s11, s43, 0xff
+; SI-NEXT:    s_lshl_b32 s13, s15, 8
+; SI-NEXT:    s_or_b32 s10, s11, s13
 ; SI-NEXT:    s_and_b32 s11, s96, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s76, 24
+; SI-NEXT:    s_lshl_b32 s15, s80, 24
 ; SI-NEXT:    s_or_b32 s43, s15, s11
-; SI-NEXT:    s_and_b32 s11, s42, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s16, 8
-; SI-NEXT:    s_or_b32 s16, s11, s15
-; SI-NEXT:    s_and_b32 s11, s40, 0xff
+; SI-NEXT:    s_and_b32 s11, s44, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s42, 8
+; SI-NEXT:    s_or_b32 s13, s11, s15
+; SI-NEXT:    s_and_b32 s11, s18, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s87, 24
+; SI-NEXT:    s_lshl_b32 s15, s97, 24
 ; SI-NEXT:    s_or_b32 s44, s15, s11
-; SI-NEXT:    s_and_b32 s11, s73, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s62, 8
-; SI-NEXT:    s_or_b32 s62, s11, s15
-; SI-NEXT:    s_and_b32 s11, s58, 0xff
+; SI-NEXT:    s_and_b32 s11, s59, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s46, 8
+; SI-NEXT:    s_or_b32 s12, s11, s15
+; SI-NEXT:    s_and_b32 s11, s45, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s45, 24
+; SI-NEXT:    s_lshl_b32 s15, s6, 24
 ; SI-NEXT:    s_or_b32 s45, s15, s11
-; SI-NEXT:    s_and_b32 s11, s48, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s94, 8
+; SI-NEXT:    s_and_b32 s11, s30, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s78, 8
+; SI-NEXT:    v_writelane_b32 v62, s10, 58
 ; SI-NEXT:    s_or_b32 s10, s11, s15
-; SI-NEXT:    s_and_b32 s11, s46, 0xff
+; SI-NEXT:    s_and_b32 s11, s99, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s47, 24
+; SI-NEXT:    s_lshl_b32 s15, s89, 24
 ; SI-NEXT:    s_or_b32 s46, s15, s11
-; SI-NEXT:    s_and_b32 s11, s25, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s34, 8
-; SI-NEXT:    s_or_b32 s94, s11, s15
-; SI-NEXT:    s_and_b32 s11, s72, 0xff
+; SI-NEXT:    s_and_b32 s11, s61, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s60, 8
+; SI-NEXT:    s_or_b32 s6, s11, s15
+; SI-NEXT:    s_and_b32 s11, s22, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s60, 24
+; SI-NEXT:    s_lshl_b32 s15, s47, 24
 ; SI-NEXT:    s_or_b32 s47, s15, s11
-; SI-NEXT:    s_and_b32 s11, s61, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s59, 8
-; SI-NEXT:    s_or_b32 s73, s11, s15
-; SI-NEXT:    s_and_b32 s11, s56, 0xff
+; SI-NEXT:    s_and_b32 s11, s57, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s56, 8
+; SI-NEXT:    v_writelane_b32 v62, s6, 59
+; SI-NEXT:    s_or_b32 s6, s11, s15
+; SI-NEXT:    s_and_b32 s11, s39, 0xff
+; SI-NEXT:    v_writelane_b32 v62, s6, 60
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s57, 24
-; SI-NEXT:    v_readlane_b32 s4, v62, 20
+; SI-NEXT:    s_lshl_b32 s15, s95, 24
 ; SI-NEXT:    s_or_b32 s56, s15, s11
-; SI-NEXT:    s_and_b32 s11, s38, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s4, 8
+; SI-NEXT:    s_and_b32 s11, s48, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s72, 8
+; SI-NEXT:    v_readlane_b32 s6, v62, 14
 ; SI-NEXT:    s_or_b32 s48, s11, s15
-; SI-NEXT:    s_and_b32 s11, s92, 0xff
+; SI-NEXT:    s_and_b32 s11, s6, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s23, 24
+; SI-NEXT:    s_lshl_b32 s15, s31, 24
 ; SI-NEXT:    s_or_b32 vcc_lo, s15, s11
-; SI-NEXT:    s_and_b32 s11, s20, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s85, 8
+; SI-NEXT:    s_and_b32 s11, s86, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s38, 8
 ; SI-NEXT:    s_or_b32 s72, s11, s15
-; SI-NEXT:    s_and_b32 s11, s83, 0xff
+; SI-NEXT:    s_and_b32 s11, s71, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s84, 24
+; SI-NEXT:    s_lshl_b32 s15, s81, 24
 ; SI-NEXT:    s_or_b32 vcc_hi, s15, s11
-; SI-NEXT:    s_and_b32 s11, s93, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s97, 8
+; SI-NEXT:    s_and_b32 s11, s58, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s85, 8
 ; SI-NEXT:    s_or_b32 s57, s11, s15
-; SI-NEXT:    s_and_b32 s11, s67, 0xff
+; SI-NEXT:    s_and_b32 s11, s69, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s51, 24
-; SI-NEXT:    v_writelane_b32 v62, s67, 51
-; SI-NEXT:    s_mov_b32 s67, s51
-; SI-NEXT:    s_mov_b32 s51, s74
+; SI-NEXT:    s_lshl_b32 s15, s74, 24
+; SI-NEXT:    v_writelane_b32 v62, s74, 61
 ; SI-NEXT:    s_or_b32 s74, s15, s11
-; SI-NEXT:    s_and_b32 s11, s98, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s75, 8
-; SI-NEXT:    v_readlane_b32 s4, v62, 18
-; SI-NEXT:    v_writelane_b32 v62, s87, 52
+; SI-NEXT:    s_and_b32 s11, s87, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s21, 8
 ; SI-NEXT:    s_or_b32 s58, s11, s15
-; SI-NEXT:    s_and_b32 s11, s4, 0xff
-; SI-NEXT:    v_writelane_b32 v62, s25, 53
+; SI-NEXT:    s_and_b32 s11, s68, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s81, 24
-; SI-NEXT:    v_readlane_b32 s4, v62, 15
-; SI-NEXT:    s_mov_b32 s54, s75
+; SI-NEXT:    s_lshl_b32 s15, s28, 24
 ; SI-NEXT:    s_or_b32 s75, s15, s11
-; SI-NEXT:    s_and_b32 s11, s77, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s4, 8
-; SI-NEXT:    v_readlane_b32 s4, v62, 14
+; SI-NEXT:    s_and_b32 s11, s24, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s55, 8
+; SI-NEXT:    v_writelane_b32 v62, s25, 62
 ; SI-NEXT:    s_or_b32 s59, s11, s15
-; SI-NEXT:    s_and_b32 s11, s4, 0xff
-; SI-NEXT:    v_readlane_b32 s4, v62, 13
+; SI-NEXT:    s_and_b32 s11, s37, 0xff
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s4, 24
-; SI-NEXT:    v_readlane_b32 s4, v62, 12
-; SI-NEXT:    s_mov_b32 s95, s69
-; SI-NEXT:    s_mov_b32 s69, s76
+; SI-NEXT:    s_lshl_b32 s15, s51, 24
+; SI-NEXT:    v_readlane_b32 s4, v62, 13
+; SI-NEXT:    s_mov_b32 s18, s21
+; SI-NEXT:    s_mov_b32 s21, s97
+; SI-NEXT:    s_mov_b32 s97, s37
+; SI-NEXT:    s_mov_b32 s37, s76
 ; SI-NEXT:    s_or_b32 s76, s15, s11
-; SI-NEXT:    s_and_b32 s11, s86, 0xff
+; SI-NEXT:    s_and_b32 s11, s35, 0xff
 ; SI-NEXT:    s_lshl_b32 s15, s4, 8
-; SI-NEXT:    v_readlane_b32 s4, v62, 11
 ; SI-NEXT:    s_or_b32 s60, s11, s15
-; SI-NEXT:    s_and_b32 s11, s4, 0xff
+; SI-NEXT:    s_and_b32 s11, s77, 0xff
+; SI-NEXT:    v_readlane_b32 s4, v62, 12
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s78, 24
-; SI-NEXT:    s_mov_b32 s93, s99
-; SI-NEXT:    s_mov_b32 s99, s84
-; SI-NEXT:    s_mov_b32 s84, s77
+; SI-NEXT:    s_lshl_b32 s15, s4, 24
+; SI-NEXT:    v_readlane_b32 s4, v62, 11
+; SI-NEXT:    s_mov_b32 s6, s95
+; SI-NEXT:    s_mov_b32 s95, s39
+; SI-NEXT:    s_mov_b32 s39, s89
+; SI-NEXT:    s_mov_b32 s89, s99
+; SI-NEXT:    s_mov_b32 s99, s83
+; SI-NEXT:    s_mov_b32 s83, s55
+; SI-NEXT:    s_mov_b32 s55, s64
+; SI-NEXT:    s_mov_b32 s64, s35
+; SI-NEXT:    s_mov_b32 s35, s77
 ; SI-NEXT:    s_or_b32 s77, s15, s11
-; SI-NEXT:    s_and_b32 s11, s82, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s53, 8
+; SI-NEXT:    s_and_b32 s11, s4, 0xff
+; SI-NEXT:    v_readlane_b32 s4, v62, 10
+; SI-NEXT:    s_lshl_b32 s15, s4, 8
+; SI-NEXT:    v_readlane_b32 s4, v62, 9
 ; SI-NEXT:    s_or_b32 s61, s11, s15
-; SI-NEXT:    s_and_b32 s11, s31, 0xff
+; SI-NEXT:    s_and_b32 s11, s4, 0xff
+; SI-NEXT:    v_readlane_b32 s4, v62, 8
 ; SI-NEXT:    s_lshl_b32 s11, s11, 16
-; SI-NEXT:    s_lshl_b32 s15, s50, 24
-; SI-NEXT:    s_mov_b32 s4, s85
-; SI-NEXT:    s_mov_b32 s85, s83
-; SI-NEXT:    s_mov_b32 s83, s82
-; SI-NEXT:    s_mov_b32 s82, s53
-; SI-NEXT:    s_mov_b32 s53, s50
-; SI-NEXT:    s_mov_b32 s50, s31
-; SI-NEXT:    s_mov_b32 s31, s78
+; SI-NEXT:    s_lshl_b32 s15, s4, 24
 ; SI-NEXT:    s_or_b32 s78, s15, s11
-; SI-NEXT:    v_readlane_b32 s11, v62, 10
+; SI-NEXT:    v_readlane_b32 s11, v62, 7
 ; SI-NEXT:    s_and_b32 s11, s11, 0xff
 ; SI-NEXT:    s_lshl_b32 s15, s17, 8
 ; SI-NEXT:    s_or_b32 s11, s11, s15
 ; SI-NEXT:    s_and_b32 s11, s11, 0xffff
 ; SI-NEXT:    v_mov_b32_e32 v51, s9
-; SI-NEXT:    s_or_b32 s6, s11, s9
-; SI-NEXT:    v_readlane_b32 s9, v62, 3
-; SI-NEXT:    v_readlane_b32 s11, v62, 2
+; SI-NEXT:    s_or_b32 s17, s11, s9
+; SI-NEXT:    v_readlane_b32 s9, v62, 2
+; SI-NEXT:    v_readlane_b32 s11, v62, 1
 ; SI-NEXT:    s_and_b32 s9, s9, 0xff
 ; SI-NEXT:    s_lshl_b32 s15, s11, 8
 ; SI-NEXT:    s_or_b32 s9, s9, s15
 ; SI-NEXT:    s_and_b32 s9, s9, 0xffff
-; SI-NEXT:    s_or_b32 s17, s9, s14
-; SI-NEXT:    v_readlane_b32 s9, v62, 22
+; SI-NEXT:    s_mov_b32 s4, s96
+; SI-NEXT:    s_mov_b32 s96, s24
 ; SI-NEXT:    v_mov_b32_e32 v52, s14
-; SI-NEXT:    s_and_b32 s14, s9, 0xff
-; SI-NEXT:    v_readlane_b32 s9, v62, 21
-; SI-NEXT:    s_lshl_b32 s15, s9, 8
+; SI-NEXT:    s_or_b32 s24, s9, s14
+; SI-NEXT:    s_and_b32 s14, s93, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s84, 8
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_or_b32_e32 v53, v6, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
 ; SI-NEXT:    v_or_b32_e32 v50, s14, v53
-; SI-NEXT:    s_and_b32 s14, s30, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s88, 8
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v20
+; SI-NEXT:    s_and_b32 s14, s8, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s49, 8
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v21
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_or_b32_e32 v54, v14, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
 ; SI-NEXT:    v_or_b32_e32 v17, s14, v54
-; SI-NEXT:    s_and_b32 s14, s8, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s52, 8
+; SI-NEXT:    s_and_b32 s14, s40, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s53, 8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v19
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v29
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v55, v40, v1
+; SI-NEXT:    v_or_b32_e32 v55, v18, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
 ; SI-NEXT:    v_or_b32_e32 v16, s14, v55
-; SI-NEXT:    s_and_b32 s14, s35, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s28, 8
+; SI-NEXT:    s_and_b32 s14, s34, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s23, 8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v29
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v13
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    s_waitcnt expcnt(4)
-; SI-NEXT:    v_or_b32_e32 v40, v42, v1
+; SI-NEXT:    v_or_b32_e32 v40, v19, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
 ; SI-NEXT:    v_or_b32_e32 v15, s14, v40
-; SI-NEXT:    s_and_b32 s14, s95, 0xff
+; SI-NEXT:    s_and_b32 s14, s91, 0xff
 ; SI-NEXT:    s_lshl_b32 s15, s66, 8
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v56
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v43
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v41, v61, v1
+; SI-NEXT:    v_or_b32_e32 v41, v22, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
 ; SI-NEXT:    v_or_b32_e32 v12, s14, v41
-; SI-NEXT:    s_and_b32 s14, s18, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s91, 8
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v57
+; SI-NEXT:    s_and_b32 s14, s50, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s54, 8
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    s_waitcnt expcnt(3)
-; SI-NEXT:    v_or_b32_e32 v42, v60, v1
+; SI-NEXT:    v_or_b32_e32 v42, v23, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
 ; SI-NEXT:    v_or_b32_e32 v11, s14, v42
-; SI-NEXT:    s_and_b32 s14, s37, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s19, 8
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v33
+; SI-NEXT:    s_and_b32 s14, s73, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s36, 8
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v46
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    s_mov_b32 s91, s6
-; SI-NEXT:    v_or_b32_e32 v59, v31, v1
+; SI-NEXT:    v_or_b32_e32 v28, v59, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
-; SI-NEXT:    v_readlane_b32 s6, v62, 50
-; SI-NEXT:    v_or_b32_e32 v10, s14, v59
-; SI-NEXT:    s_and_b32 s14, s6, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s70, 8
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v30
+; SI-NEXT:    v_or_b32_e32 v10, s14, v28
+; SI-NEXT:    s_and_b32 s14, s82, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s52, 8
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v26
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v5, v24, v1
+; SI-NEXT:    v_or_b32_e32 v60, v24, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
-; SI-NEXT:    v_readlane_b32 s6, v62, 19
-; SI-NEXT:    v_or_b32_e32 v9, s14, v5
+; SI-NEXT:    v_or_b32_e32 v9, s14, v60
 ; SI-NEXT:    s_and_b32 s14, s90, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s6, 8
+; SI-NEXT:    s_lshl_b32 s15, s16, 8
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v35
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v13, v25, v1
+; SI-NEXT:    v_or_b32_e32 v31, v44, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
-; SI-NEXT:    v_or_b32_e32 v8, s14, v13
-; SI-NEXT:    s_and_b32 s14, s27, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s24, 8
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
+; SI-NEXT:    v_or_b32_e32 v8, s14, v31
+; SI-NEXT:    s_and_b32 s14, s62, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s65, 8
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v37
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_mov_b32_e32 v26, v31
-; SI-NEXT:    v_or_b32_e32 v31, v27, v1
+; SI-NEXT:    v_or_b32_e32 v61, v45, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
-; SI-NEXT:    v_or_b32_e32 v7, s14, v31
-; SI-NEXT:    s_and_b32 s14, s49, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s68, 8
+; SI-NEXT:    v_or_b32_e32 v7, s14, v61
+; SI-NEXT:    s_and_b32 s14, s98, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s67, 8
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v38
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_mov_b32_e32 v28, v25
-; SI-NEXT:    v_mov_b32_e32 v25, v60
-; SI-NEXT:    v_or_b32_e32 v60, v43, v1
+; SI-NEXT:    s_waitcnt expcnt(5)
+; SI-NEXT:    v_or_b32_e32 v6, v47, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
-; SI-NEXT:    v_readlane_b32 s6, v62, 17
-; SI-NEXT:    v_or_b32_e32 v4, s14, v60
-; SI-NEXT:    s_and_b32 s14, s6, 0xff
-; SI-NEXT:    v_readlane_b32 s6, v62, 16
-; SI-NEXT:    s_lshl_b32 s15, s6, 8
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; SI-NEXT:    v_or_b32_e32 v4, s14, v6
+; SI-NEXT:    s_and_b32 s14, s92, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s7, 8
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v33
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_mov_b32_e32 v23, v27
-; SI-NEXT:    v_mov_b32_e32 v27, v24
-; SI-NEXT:    v_mov_b32_e32 v24, v61
-; SI-NEXT:    v_or_b32_e32 v61, v44, v1
+; SI-NEXT:    v_readlane_b32 s8, v62, 55
+; SI-NEXT:    s_waitcnt expcnt(4)
+; SI-NEXT:    v_mov_b32_e32 v22, v14
+; SI-NEXT:    v_or_b32_e32 v14, v56, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
-; SI-NEXT:    v_or_b32_e32 v2, s14, v61
-; SI-NEXT:    s_and_b32 s14, s71, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s36, 8
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v48
+; SI-NEXT:    s_and_b32 s8, s8, 0xffff
+; SI-NEXT:    v_or_b32_e32 v2, s14, v14
+; SI-NEXT:    s_and_b32 s14, s70, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s94, 8
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; SI-NEXT:    s_or_b32 s42, s8, s63
+; SI-NEXT:    v_readlane_b32 s8, v62, 56
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_readlane_b32 s6, v62, 48
-; SI-NEXT:    v_or_b32_e32 v6, v45, v1
+; SI-NEXT:    s_and_b32 s8, s8, 0xffff
+; SI-NEXT:    s_waitcnt expcnt(2)
+; SI-NEXT:    v_mov_b32_e32 v32, v23
+; SI-NEXT:    v_mov_b32_e32 v23, v18
+; SI-NEXT:    v_or_b32_e32 v18, v57, v1
 ; SI-NEXT:    s_and_b32 s14, s14, 0xffff
-; SI-NEXT:    s_and_b32 s8, s6, 0xffff
-; SI-NEXT:    v_readlane_b32 s6, v62, 49
-; SI-NEXT:    v_or_b32_e32 v1, s14, v6
-; SI-NEXT:    s_and_b32 s14, s79, 0xff
-; SI-NEXT:    s_lshl_b32 s15, s7, 8
+; SI-NEXT:    s_or_b32 s40, s8, s5
+; SI-NEXT:    v_readlane_b32 s8, v62, 57
+; SI-NEXT:    v_or_b32_e32 v1, s14, v18
+; SI-NEXT:    s_and_b32 s14, s88, 0xff
+; SI-NEXT:    s_lshl_b32 s15, s79, 8
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v34
-; SI-NEXT:    s_or_b32 s42, s8, s63
-; SI-NEXT:    s_and_b32 s8, s6, 0xffff
+; SI-NEXT:    s_and_b32 s8, s8, 0xffff
+; SI-NEXT:    v_readlane_b32 s9, v62, 60
 ; SI-NEXT:    s_or_b32 s14, s14, s15
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    s_or_b32 s40, s8, s21
-; SI-NEXT:    s_and_b32 s8, s12, 0xffff
-; SI-NEXT:    v_or_b32_e32 v14, v46, v3
-; SI-NEXT:    s_and_b32 s14, s14, 0xffff
 ; SI-NEXT:    s_or_b32 s15, s8, s41
-; SI-NEXT:    s_and_b32 s8, s13, 0xffff
-; SI-NEXT:    v_or_b32_e32 v3, s14, v14
-; SI-NEXT:    s_or_b32 s14, s8, s43
-; SI-NEXT:    s_and_b32 s8, s16, 0xffff
-; SI-NEXT:    s_and_b32 s16, s73, 0xffff
-; SI-NEXT:    s_or_b32 s13, s8, s44
-; SI-NEXT:    s_and_b32 s8, s62, 0xffff
-; SI-NEXT:    s_or_b32 s35, s16, s56
+; SI-NEXT:    v_readlane_b32 s8, v62, 58
+; SI-NEXT:    s_and_b32 s16, s9, 0xffff
+; SI-NEXT:    v_mov_b32_e32 v27, v26
+; SI-NEXT:    v_mov_b32_e32 v26, v24
+; SI-NEXT:    v_mov_b32_e32 v24, v19
+; SI-NEXT:    v_or_b32_e32 v19, v58, v3
+; SI-NEXT:    s_and_b32 s14, s14, 0xffff
+; SI-NEXT:    s_and_b32 s8, s8, 0xffff
+; SI-NEXT:    s_or_b32 s36, s16, s56
 ; SI-NEXT:    s_and_b32 s16, s48, 0xffff
-; SI-NEXT:    s_or_b32 s12, s8, s45
-; SI-NEXT:    s_and_b32 s8, s10, 0xffff
-; SI-NEXT:    s_or_b32 s52, s16, vcc_lo
+; SI-NEXT:    v_or_b32_e32 v3, s14, v19
+; SI-NEXT:    s_or_b32 s14, s8, s43
+; SI-NEXT:    s_and_b32 s8, s13, 0xffff
+; SI-NEXT:    s_or_b32 s53, s16, vcc_lo
 ; SI-NEXT:    s_and_b32 s16, s72, 0xffff
-; SI-NEXT:    s_or_b32 s10, s8, s46
-; SI-NEXT:    s_and_b32 s8, s94, 0xffff
+; SI-NEXT:    s_or_b32 s13, s8, s44
+; SI-NEXT:    s_and_b32 s8, s12, 0xffff
 ; SI-NEXT:    s_or_b32 s94, s16, vcc_hi
 ; SI-NEXT:    s_and_b32 s16, s57, 0xffff
+; SI-NEXT:    s_or_b32 s12, s8, s45
+; SI-NEXT:    s_and_b32 s8, s10, 0xffff
 ; SI-NEXT:    s_or_b32 s49, s16, s74
 ; SI-NEXT:    s_and_b32 s16, s58, 0xffff
+; SI-NEXT:    s_or_b32 s10, s8, s46
+; SI-NEXT:    v_readlane_b32 s8, v62, 59
 ; SI-NEXT:    s_or_b32 s48, s16, s75
 ; SI-NEXT:    s_and_b32 s16, s59, 0xffff
-; SI-NEXT:    s_mov_b32 s25, s23
+; SI-NEXT:    s_and_b32 s8, s8, 0xffff
 ; SI-NEXT:    s_or_b32 s11, s16, s76
 ; SI-NEXT:    s_and_b32 s16, s60, 0xffff
 ; SI-NEXT:    s_and_b32 s23, s61, 0xffff
-; SI-NEXT:    s_mov_b32 s87, s34
-; SI-NEXT:    s_mov_b32 s34, s55
-; SI-NEXT:    s_mov_b32 s55, s22
+; SI-NEXT:    s_mov_b32 s30, s87
+; SI-NEXT:    s_mov_b32 s87, s85
 ; SI-NEXT:    s_or_b32 s8, s8, s47
 ; SI-NEXT:    s_or_b32 s9, s16, s77
 ; SI-NEXT:    s_or_b32 s16, s23, s78
-; SI-NEXT:    s_mov_b32 s22, s18
-; SI-NEXT:    v_mov_b32_e32 v49, v30
 ; SI-NEXT:    v_mov_b32_e32 v36, v35
-; SI-NEXT:    v_mov_b32_e32 v37, v32
-; SI-NEXT:    v_mov_b32_e32 v18, v43
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v33, v44
+; SI-NEXT:    v_mov_b32_e32 v30, v37
 ; SI-NEXT:    v_mov_b32_e32 v35, v45
-; SI-NEXT:    v_mov_b32_e32 v22, v46
+; SI-NEXT:    v_mov_b32_e32 v20, v47
+; SI-NEXT:    v_mov_b32_e32 v49, v56
+; SI-NEXT:    v_mov_b32_e32 v48, v39
+; SI-NEXT:    v_mov_b32_e32 v39, v57
+; SI-NEXT:    v_mov_b32_e32 v25, v58
 ; SI-NEXT:    v_alignbit_b32 v57, s42, v51, 16
 ; SI-NEXT:    v_alignbit_b32 v58, s40, v52, 16
 ; SI-NEXT:    v_alignbit_b32 v56, s15, v53, 16
 ; SI-NEXT:    v_alignbit_b32 v47, s14, v54, 16
+; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    v_alignbit_b32 v46, s13, v55, 16
 ; SI-NEXT:    v_alignbit_b32 v45, s12, v40, 16
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v44, s10, v41, 16
 ; SI-NEXT:    v_alignbit_b32 v43, s8, v42, 16
-; SI-NEXT:    v_alignbit_b32 v42, s35, v59, 16
-; SI-NEXT:    v_alignbit_b32 v41, s52, v5, 16
-; SI-NEXT:    v_alignbit_b32 v40, s94, v13, 16
-; SI-NEXT:    v_alignbit_b32 v55, s49, v31, 16
-; SI-NEXT:    v_mov_b32_e32 v31, v26
-; SI-NEXT:    v_alignbit_b32 v54, s48, v60, 16
-; SI-NEXT:    v_mov_b32_e32 v60, v25
-; SI-NEXT:    v_mov_b32_e32 v25, v28
-; SI-NEXT:    v_alignbit_b32 v53, s11, v61, 16
-; SI-NEXT:    v_mov_b32_e32 v61, v24
-; SI-NEXT:    v_mov_b32_e32 v24, v27
-; SI-NEXT:    v_alignbit_b32 v52, s9, v6, 16
-; SI-NEXT:    v_alignbit_b32 v51, s16, v14, 16
+; SI-NEXT:    v_alignbit_b32 v42, s36, v28, 16
+; SI-NEXT:    v_alignbit_b32 v41, s53, v60, 16
+; SI-NEXT:    v_alignbit_b32 v40, s94, v31, 16
+; SI-NEXT:    v_alignbit_b32 v55, s49, v61, 16
+; SI-NEXT:    v_alignbit_b32 v54, s48, v6, 16
+; SI-NEXT:    v_alignbit_b32 v53, s11, v14, 16
+; SI-NEXT:    v_mov_b32_e32 v14, v22
+; SI-NEXT:    v_alignbit_b32 v52, s9, v18, 16
+; SI-NEXT:    v_mov_b32_e32 v18, v23
+; SI-NEXT:    v_alignbit_b32 v51, s16, v19, 16
+; SI-NEXT:    v_mov_b32_e32 v19, v24
+; SI-NEXT:    v_mov_b32_e32 v24, v26
 ; SI-NEXT:    s_lshr_b32 s73, s63, 16
-; SI-NEXT:    s_lshr_b32 s72, s21, 16
+; SI-NEXT:    s_lshr_b32 s72, s5, 16
 ; SI-NEXT:    s_lshr_b32 s63, s41, 16
 ; SI-NEXT:    s_lshr_b32 s62, s43, 16
 ; SI-NEXT:    s_lshr_b32 s61, s44, 16
@@ -209550,573 +209672,477 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    s_lshr_b32 s56, vcc_lo, 16
 ; SI-NEXT:    s_lshr_b32 s47, vcc_hi, 16
 ; SI-NEXT:    s_lshr_b32 s46, s74, 16
-; SI-NEXT:    s_mov_b32 s74, s51
-; SI-NEXT:    s_mov_b32 s51, s67
-; SI-NEXT:    v_readlane_b32 s67, v62, 51
+; SI-NEXT:    v_readlane_b32 s25, v62, 62
+; SI-NEXT:    v_readlane_b32 s74, v62, 61
 ; SI-NEXT:    s_lshr_b32 s45, s75, 16
-; SI-NEXT:    s_mov_b32 s23, s25
-; SI-NEXT:    s_mov_b32 s21, s55
-; SI-NEXT:    s_mov_b32 s55, s34
-; SI-NEXT:    s_mov_b32 s75, s54
-; SI-NEXT:    s_mov_b32 s34, s87
-; SI-NEXT:    v_readlane_b32 s25, v62, 53
-; SI-NEXT:    v_readlane_b32 s87, v62, 52
 ; SI-NEXT:    s_lshr_b32 s44, s76, 16
-; SI-NEXT:    v_readlane_b32 s54, v62, 50
+; SI-NEXT:    s_mov_b32 s76, s37
+; SI-NEXT:    s_mov_b32 s37, s97
+; SI-NEXT:    s_mov_b32 s97, s21
+; SI-NEXT:    s_mov_b32 s21, s18
+; SI-NEXT:    s_mov_b32 s18, s17
+; SI-NEXT:    s_mov_b32 s85, s87
+; SI-NEXT:    s_mov_b32 s87, s30
+; SI-NEXT:    s_mov_b32 s17, s24
 ; SI-NEXT:    s_lshr_b32 s43, s77, 16
-; SI-NEXT:    s_mov_b32 s76, s69
-; SI-NEXT:    s_mov_b32 s69, s95
-; SI-NEXT:    s_mov_b32 s77, s84
-; SI-NEXT:    s_mov_b32 s84, s99
-; SI-NEXT:    s_mov_b32 s99, s93
+; SI-NEXT:    s_mov_b32 s77, s35
+; SI-NEXT:    s_mov_b32 s35, s64
+; SI-NEXT:    s_mov_b32 s64, s55
+; SI-NEXT:    s_mov_b32 s55, s83
+; SI-NEXT:    s_mov_b32 s83, s99
+; SI-NEXT:    s_mov_b32 s99, s89
+; SI-NEXT:    s_mov_b32 s89, s39
+; SI-NEXT:    s_mov_b32 s39, s95
+; SI-NEXT:    s_mov_b32 s95, s6
 ; SI-NEXT:    s_lshr_b32 s41, s78, 16
-; SI-NEXT:    s_mov_b32 s78, s31
-; SI-NEXT:    s_mov_b32 s31, s50
-; SI-NEXT:    s_mov_b32 s50, s53
-; SI-NEXT:    s_mov_b32 s53, s82
-; SI-NEXT:    s_mov_b32 s82, s83
-; SI-NEXT:    s_mov_b32 s83, s85
-; SI-NEXT:    s_mov_b32 s85, s4
-; SI-NEXT:    s_mov_b64 s[4:5], 0
-; SI-NEXT:    v_mov_b32_e32 v6, v20
-; SI-NEXT:    v_mov_b32_e32 v13, v19
-; SI-NEXT:    v_mov_b32_e32 v14, v29
-; SI-NEXT:    s_branch .LBB97_3
-; SI-NEXT:  .LBB97_2:
-; SI-NEXT:    s_waitcnt expcnt(6)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(5)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
-; SI-NEXT:    v_mov_b32_e32 v22, v46
-; SI-NEXT:    v_mov_b32_e32 v37, v32
-; SI-NEXT:    v_mov_b32_e32 v36, v35
-; SI-NEXT:    v_mov_b32_e32 v35, v45
-; SI-NEXT:    v_mov_b32_e32 v49, v30
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v33, v44
-; SI-NEXT:    v_mov_b32_e32 v18, v43
-; SI-NEXT:    v_mov_b32_e32 v23, v27
-; SI-NEXT:    v_mov_b32_e32 v21, v29
-; SI-NEXT:    s_mov_b64 s[4:5], -1
-; SI-NEXT:    ; implicit-def: $sgpr91
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $sgpr42
-; SI-NEXT:    ; implicit-def: $sgpr73
-; SI-NEXT:    ; implicit-def: $sgpr17
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $sgpr40
-; SI-NEXT:    ; implicit-def: $sgpr72
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    ; implicit-def: $sgpr15
-; SI-NEXT:    ; implicit-def: $sgpr63
-; SI-NEXT:    ; implicit-def: $vgpr17
-; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $sgpr14
-; SI-NEXT:    ; implicit-def: $sgpr62
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $sgpr13
-; SI-NEXT:    ; implicit-def: $sgpr61
-; SI-NEXT:    ; implicit-def: $vgpr15
-; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $sgpr12
-; SI-NEXT:    ; implicit-def: $sgpr60
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $sgpr10
-; SI-NEXT:    ; implicit-def: $sgpr59
-; SI-NEXT:    ; implicit-def: $vgpr11
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $sgpr8
-; SI-NEXT:    ; implicit-def: $sgpr58
-; SI-NEXT:    ; implicit-def: $vgpr10
-; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $sgpr35
-; SI-NEXT:    ; implicit-def: $sgpr57
-; SI-NEXT:    ; implicit-def: $vgpr9
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $sgpr52
-; SI-NEXT:    ; implicit-def: $sgpr56
-; SI-NEXT:    ; implicit-def: $vgpr8
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $sgpr94
-; SI-NEXT:    ; implicit-def: $sgpr47
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $sgpr49
-; SI-NEXT:    ; implicit-def: $sgpr46
-; SI-NEXT:    ; implicit-def: $vgpr4
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $sgpr48
-; SI-NEXT:    ; implicit-def: $sgpr45
-; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr53
-; SI-NEXT:    ; implicit-def: $sgpr11
-; SI-NEXT:    ; implicit-def: $sgpr44
-; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr52
-; SI-NEXT:    ; implicit-def: $sgpr9
-; SI-NEXT:    ; implicit-def: $sgpr43
-; SI-NEXT:    ; implicit-def: $vgpr3
-; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $sgpr16
-; SI-NEXT:    ; implicit-def: $sgpr41
-; SI-NEXT:  .LBB97_3: ; %Flow
-; SI-NEXT:    v_mov_b32_e32 v5, v39
-; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
-; SI-NEXT:    s_cbranch_vccnz .LBB97_5
-; SI-NEXT:  ; %bb.4: ; %cmp.true
-; SI-NEXT:    s_add_i32 s4, s79, 3
+; SI-NEXT:    s_mov_b32 s24, s96
+; SI-NEXT:    s_mov_b32 s96, s4
+; SI-NEXT:    s_cbranch_execnz .LBB97_3
+; SI-NEXT:  .LBB97_2: ; %cmp.true
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v36
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_mov_b32_e32 v6, v5
+; SI-NEXT:    v_mov_b32_e32 v5, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    s_add_i32 s4, s88, 3
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s7, 8
+; SI-NEXT:    s_lshl_b32 s5, s79, 8
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v34
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_or_b32_e32 v1, v22, v1
+; SI-NEXT:    v_or_b32_e32 v1, v25, v1
 ; SI-NEXT:    v_or_b32_e32 v1, s4, v1
-; SI-NEXT:    s_add_i32 s4, s82, 3
+; SI-NEXT:    v_readlane_b32 s4, v62, 11
+; SI-NEXT:    s_add_i32 s4, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 10
+; SI-NEXT:    v_readlane_b32 s6, v62, 9
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s53, 8
-; SI-NEXT:    s_add_i32 s8, s31, 3
+; SI-NEXT:    s_lshl_b32 s5, s5, 8
+; SI-NEXT:    s_add_i32 s8, s6, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
+; SI-NEXT:    v_readlane_b32 s5, v62, 8
 ; SI-NEXT:    s_and_b32 s8, s8, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s50, 24
+; SI-NEXT:    s_lshl_b32 s5, s5, 24
 ; SI-NEXT:    s_lshl_b32 s8, s8, 16
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    s_or_b32 s5, s5, s8
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    s_add_i32 s71, s71, 3
+; SI-NEXT:    s_add_i32 s70, s70, 3
+; SI-NEXT:    v_readlane_b32 s6, v62, 30
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    s_and_b32 s5, s71, 0xff
-; SI-NEXT:    s_lshl_b32 s8, s36, 8
+; SI-NEXT:    s_and_b32 s5, s70, 0xff
+; SI-NEXT:    s_lshl_b32 s8, s6, 8
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v48
 ; SI-NEXT:    s_or_b32 s5, s8, s5
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_addk_i32 s5, 0x300
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_readlane_b32 s7, v62, 12
 ; SI-NEXT:    s_and_b32 s5, s5, 0xffff
-; SI-NEXT:    v_or_b32_e32 v2, v35, v2
-; SI-NEXT:    s_add_i32 s36, s86, 3
-; SI-NEXT:    s_lshl_b32 s8, s7, 8
-; SI-NEXT:    v_readlane_b32 s7, v62, 11
+; SI-NEXT:    v_or_b32_e32 v2, v39, v2
 ; SI-NEXT:    v_or_b32_e32 v2, s5, v2
-; SI-NEXT:    s_and_b32 s5, s36, 0xff
-; SI-NEXT:    s_add_i32 s9, s7, 3
+; SI-NEXT:    s_add_i32 s5, s35, 3
+; SI-NEXT:    v_readlane_b32 s6, v62, 13
+; SI-NEXT:    s_and_b32 s5, s5, 0xff
+; SI-NEXT:    s_lshl_b32 s8, s6, 8
+; SI-NEXT:    s_add_i32 s9, s77, 3
 ; SI-NEXT:    s_or_b32 s5, s8, s5
+; SI-NEXT:    v_readlane_b32 s6, v62, 12
 ; SI-NEXT:    s_and_b32 s9, s9, 0xff
-; SI-NEXT:    s_lshl_b32 s8, s78, 24
+; SI-NEXT:    s_lshl_b32 s8, s6, 24
 ; SI-NEXT:    s_lshl_b32 s9, s9, 16
 ; SI-NEXT:    s_addk_i32 s5, 0x300
 ; SI-NEXT:    s_or_b32 s8, s8, s9
 ; SI-NEXT:    s_and_b32 s5, s5, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s8, s5
-; SI-NEXT:    s_add_i32 s16, s4, 0x3000000
-; SI-NEXT:    v_readlane_b32 s4, v62, 17
-; SI-NEXT:    s_add_i32 s9, s5, 0x3000000
-; SI-NEXT:    s_add_i32 s79, s4, 3
-; SI-NEXT:    v_readlane_b32 s5, v62, 16
+; SI-NEXT:    s_add_i32 s79, s92, 3
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x3000000, v1
+; SI-NEXT:    s_add_i32 s16, s4, 0x3000000
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x3000000, v2
+; SI-NEXT:    s_add_i32 s9, s5, 0x3000000
 ; SI-NEXT:    s_and_b32 s4, s79, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s5, 8
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v5
+; SI-NEXT:    s_lshl_b32 s5, s7, 8
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v33
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_or_b32_e32 v2, v33, v2
+; SI-NEXT:    v_or_b32_e32 v2, v49, v2
 ; SI-NEXT:    v_or_b32_e32 v2, s4, v2
-; SI-NEXT:    s_add_i32 s4, s77, 3
-; SI-NEXT:    v_readlane_b32 s5, v62, 15
-; SI-NEXT:    v_readlane_b32 s6, v62, 14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v8, v9, v8
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; SI-NEXT:    s_add_i32 s4, s24, 3
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s5, 8
-; SI-NEXT:    s_add_i32 s8, s6, 3
+; SI-NEXT:    s_lshl_b32 s5, s55, 8
+; SI-NEXT:    s_add_i32 s8, s37, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 13
 ; SI-NEXT:    s_and_b32 s8, s8, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s5, 24
+; SI-NEXT:    s_lshl_b32 s5, s51, 24
 ; SI-NEXT:    s_lshl_b32 s8, s8, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
+; SI-NEXT:    s_add_i32 s52, s98, 3
 ; SI-NEXT:    s_add_i32 s11, s4, 0x3000000
-; SI-NEXT:    v_readlane_b32 s4, v62, 32
-; SI-NEXT:    s_add_i32 s53, s4, 3
-; SI-NEXT:    s_and_b32 s4, s53, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s68, 8
+; SI-NEXT:    s_and_b32 s4, s52, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s67, 8
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v38
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_or_b32_e32 v4, v18, v4
-; SI-NEXT:    s_add_i32 s93, s98, 3
-; SI-NEXT:    v_readlane_b32 s6, v62, 18
+; SI-NEXT:    v_or_b32_e32 v4, v20, v4
+; SI-NEXT:    s_add_i32 s30, s87, 3
 ; SI-NEXT:    v_or_b32_e32 v4, s4, v4
-; SI-NEXT:    s_and_b32 s4, s93, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s75, 8
-; SI-NEXT:    s_add_i32 s8, s6, 3
+; SI-NEXT:    s_and_b32 s4, s30, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s21, 8
+; SI-NEXT:    s_add_i32 s8, s68, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_and_b32 s8, s8, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s81, 24
+; SI-NEXT:    s_lshl_b32 s5, s28, 24
 ; SI-NEXT:    s_lshl_b32 s8, s8, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    s_add_i32 s86, s27, 3
 ; SI-NEXT:    s_add_i32 s48, s4, 0x3000000
-; SI-NEXT:    s_and_b32 s4, s86, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s24, 8
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v37
+; SI-NEXT:    v_readlane_b32 s4, v62, 42
+; SI-NEXT:    v_mov_b32_e32 v22, v30
+; SI-NEXT:    s_add_i32 s87, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 39
+; SI-NEXT:    s_and_b32 s4, s87, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s5, 8
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v22
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_or_b32_e32 v5, v23, v5
-; SI-NEXT:    v_or_b32_e32 v5, s4, v5
-; SI-NEXT:    v_readlane_b32 s4, v62, 38
-; SI-NEXT:    s_add_i32 s68, s4, 3
-; SI-NEXT:    s_and_b32 s4, s68, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s97, 8
-; SI-NEXT:    s_add_i32 s8, s67, 3
+; SI-NEXT:    v_or_b32_e32 v7, v35, v7
+; SI-NEXT:    v_or_b32_e32 v7, s4, v7
+; SI-NEXT:    v_readlane_b32 s4, v62, 32
+; SI-NEXT:    s_add_i32 s67, s4, 3
+; SI-NEXT:    s_and_b32 s4, s67, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s85, 8
+; SI-NEXT:    s_add_i32 s8, s69, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_and_b32 s8, s8, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s51, 24
+; SI-NEXT:    s_lshl_b32 s5, s74, 24
 ; SI-NEXT:    s_lshl_b32 s8, s8, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_add_i32 s50, s90, 3
-; SI-NEXT:    v_readlane_b32 s5, v62, 19
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 0x3000000, v5
+; SI-NEXT:    v_readlane_b32 s5, v62, 23
 ; SI-NEXT:    s_add_i32 s49, s4, 0x3000000
 ; SI-NEXT:    s_and_b32 s4, s50, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v36
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_or_b32_e32 v5, v25, v5
-; SI-NEXT:    s_add_i32 s94, s20, 3
-; SI-NEXT:    v_or_b32_e32 v5, s4, v5
+; SI-NEXT:    s_add_i32 s94, s86, 3
+; SI-NEXT:    v_or_b32_e32 v8, s4, v8
 ; SI-NEXT:    s_and_b32 s4, s94, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s85, 8
-; SI-NEXT:    s_add_i32 s8, s83, 3
+; SI-NEXT:    s_lshl_b32 s5, s38, 8
+; SI-NEXT:    s_add_i32 s8, s71, 3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_and_b32 s8, s8, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s84, 24
+; SI-NEXT:    s_lshl_b32 s5, s81, 24
 ; SI-NEXT:    s_lshl_b32 s8, s8, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    s_add_i32 s20, s54, 3
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 0x3000000, v5
 ; SI-NEXT:    s_add_i32 s94, s4, 0x3000000
-; SI-NEXT:    s_and_b32 s4, s20, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s70, 8
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v49
+; SI-NEXT:    v_readlane_b32 s4, v62, 52
+; SI-NEXT:    s_add_i32 s18, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 51
+; SI-NEXT:    s_and_b32 s4, s18, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_or_b32_e32 v5, v24, v5
-; SI-NEXT:    v_or_b32_e32 v5, s4, v5
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 0x3000000, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; SI-NEXT:    s_add_i32 s98, s38, 3
-; SI-NEXT:    v_readlane_b32 s5, v62, 20
+; SI-NEXT:    v_or_b32_e32 v9, v24, v9
+; SI-NEXT:    v_or_b32_e32 v9, s4, v9
+; SI-NEXT:    v_readlane_b32 s4, v62, 45
+; SI-NEXT:    s_add_i32 s98, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 43
+; SI-NEXT:    v_readlane_b32 s6, v62, 14
 ; SI-NEXT:    s_and_b32 s4, s98, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
-; SI-NEXT:    s_add_i32 s8, s92, 3
+; SI-NEXT:    s_add_i32 s8, s6, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_and_b32 s8, s8, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s23, 24
+; SI-NEXT:    s_lshl_b32 s5, s31, 24
 ; SI-NEXT:    s_lshl_b32 s8, s8, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    s_add_i32 s17, s37, 3
-; SI-NEXT:    s_add_i32 s52, s4, 0x3000000
-; SI-NEXT:    s_and_b32 s4, s17, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s19, 8
+; SI-NEXT:    s_add_i32 s53, s4, 0x3000000
+; SI-NEXT:    v_readlane_b32 s4, v62, 41
+; SI-NEXT:    s_add_i32 s86, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 38
+; SI-NEXT:    s_and_b32 s4, s86, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    s_addk_i32 s4, 0x300
+; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_readlane_b32 s5, v62, 34
-; SI-NEXT:    v_readlane_b32 s6, v62, 29
+; SI-NEXT:    v_or_b32_e32 v10, v59, v10
+; SI-NEXT:    v_or_b32_e32 v10, s4, v10
+; SI-NEXT:    v_readlane_b32 s4, v62, 31
+; SI-NEXT:    s_add_i32 s66, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 27
+; SI-NEXT:    s_and_b32 s4, s66, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
-; SI-NEXT:    s_add_i32 s38, s6, 3
-; SI-NEXT:    s_and_b32 s8, s38, 0xff
-; SI-NEXT:    s_lshl_b32 s8, s8, 16
-; SI-NEXT:    s_add_i32 s22, s22, 3
-; SI-NEXT:    s_add_i32 s85, s25, 3
-; SI-NEXT:    v_readlane_b32 s6, v62, 40
-; SI-NEXT:    s_add_i32 s70, s6, 3
-; SI-NEXT:    s_and_b32 s7, s70, 0xff
-; SI-NEXT:    s_lshl_b32 s7, s7, 16
-; SI-NEXT:    s_add_i32 s19, s69, 3
-; SI-NEXT:    s_add_i32 s51, s30, 3
-; SI-NEXT:    s_add_i32 s95, s89, 3
-; SI-NEXT:    v_mov_b32_e32 v30, s16
-; SI-NEXT:    v_mov_b32_e32 v39, s9
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x3000000, v2
-; SI-NEXT:    v_mov_b32_e32 v28, s11
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x3000000, v4
-; SI-NEXT:    v_mov_b32_e32 v27, s48
-; SI-NEXT:    v_mov_b32_e32 v26, s49
-; SI-NEXT:    v_mov_b32_e32 v25, s94
-; SI-NEXT:    v_mov_b32_e32 v24, s52
-; SI-NEXT:    v_alignbit_b32 v41, v24, v9, 16
-; SI-NEXT:    v_alignbit_b32 v40, v25, v8, 16
-; SI-NEXT:    v_alignbit_b32 v55, v26, v7, 16
-; SI-NEXT:    v_alignbit_b32 v54, v27, v4, 16
-; SI-NEXT:    v_alignbit_b32 v53, v28, v2, 16
-; SI-NEXT:    v_alignbit_b32 v52, v39, v1, 16
-; SI-NEXT:    v_alignbit_b32 v51, v30, v3, 16
-; SI-NEXT:    s_lshr_b32 s56, s52, 16
-; SI-NEXT:    s_lshr_b32 s47, s94, 16
-; SI-NEXT:    s_lshr_b32 s46, s49, 16
-; SI-NEXT:    s_lshr_b32 s45, s48, 16
-; SI-NEXT:    s_lshr_b32 s44, s11, 16
-; SI-NEXT:    s_lshr_b32 s43, s9, 16
-; SI-NEXT:    s_lshr_b32 s41, s16, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v31, v5
-; SI-NEXT:    v_or_b32_e32 v5, s4, v5
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 0x3000000, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; SI-NEXT:    v_readlane_b32 s4, v62, 37
-; SI-NEXT:    s_add_i32 s67, s4, 3
-; SI-NEXT:    s_and_b32 s4, s67, 0xff
+; SI-NEXT:    s_add_i32 s37, s39, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 30
+; SI-NEXT:    s_and_b32 s8, s37, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s5, 24
+; SI-NEXT:    s_lshl_b32 s5, s95, 24
+; SI-NEXT:    s_lshl_b32 s8, s8, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 46
-; SI-NEXT:    s_add_i32 s35, s4, 0x3000000
-; SI-NEXT:    s_and_b32 s4, s22, 0xff
+; SI-NEXT:    s_add_i32 s36, s4, 0x3000000
+; SI-NEXT:    v_readlane_b32 s4, v62, 50
+; SI-NEXT:    s_add_i32 s21, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 49
+; SI-NEXT:    s_and_b32 s4, s21, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    s_lshl_b32 s5, s34, 8
-; SI-NEXT:    v_mov_b32_e32 v23, s35
-; SI-NEXT:    v_alignbit_b32 v42, v23, v10, 16
-; SI-NEXT:    s_lshr_b32 s57, s35, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v60, v5
-; SI-NEXT:    v_or_b32_e32 v5, s4, v5
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 0x3000000, v5
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; SI-NEXT:    v_readlane_b32 s5, v62, 37
+; SI-NEXT:    s_lshl_b32 s5, s5, 8
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; SI-NEXT:    s_add_i32 s71, s22, 3
+; SI-NEXT:    s_and_b32 s8, s71, 0xff
+; SI-NEXT:    s_lshl_b32 s8, s8, 16
+; SI-NEXT:    s_add_i32 s35, s99, 3
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v5
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; SI-NEXT:    v_or_b32_e32 v11, v32, v11
+; SI-NEXT:    v_or_b32_e32 v11, s4, v11
+; SI-NEXT:    v_readlane_b32 s4, v62, 40
+; SI-NEXT:    s_add_i32 s85, s4, 3
 ; SI-NEXT:    s_and_b32 s4, s85, 0xff
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 36
+; SI-NEXT:    v_readlane_b32 s5, v62, 33
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    s_lshl_b32 s5, s5, 24
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    s_or_b32 s5, s5, s7
+; SI-NEXT:    s_or_b32 s5, s5, s8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 47
 ; SI-NEXT:    s_add_i32 s8, s4, 0x3000000
-; SI-NEXT:    s_and_b32 s4, s19, 0xff
+; SI-NEXT:    v_readlane_b32 s4, v62, 54
+; SI-NEXT:    s_add_i32 s17, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 53
+; SI-NEXT:    s_and_b32 s4, s17, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_readlane_b32 s5, v62, 44
+; SI-NEXT:    v_readlane_b32 s5, v62, 47
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
+; SI-NEXT:    s_and_b32 s6, s35, 0xff
+; SI-NEXT:    s_lshl_b32 s6, s6, 16
+; SI-NEXT:    v_mov_b32_e32 v30, s16
+; SI-NEXT:    v_mov_b32_e32 v39, s9
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x3000000, v2
+; SI-NEXT:    v_mov_b32_e32 v28, s11
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x3000000, v4
+; SI-NEXT:    v_mov_b32_e32 v27, s48
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 0x3000000, v7
+; SI-NEXT:    v_mov_b32_e32 v26, s49
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 0x3000000, v8
+; SI-NEXT:    v_mov_b32_e32 v25, s94
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 0x3000000, v9
+; SI-NEXT:    v_mov_b32_e32 v24, s53
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 0x3000000, v10
+; SI-NEXT:    v_mov_b32_e32 v23, s36
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 0x3000000, v11
 ; SI-NEXT:    v_mov_b32_e32 v22, s8
 ; SI-NEXT:    v_alignbit_b32 v43, v22, v11, 16
+; SI-NEXT:    v_alignbit_b32 v42, v23, v10, 16
+; SI-NEXT:    v_alignbit_b32 v41, v24, v9, 16
+; SI-NEXT:    v_alignbit_b32 v40, v25, v8, 16
+; SI-NEXT:    v_alignbit_b32 v55, v26, v7, 16
+; SI-NEXT:    v_alignbit_b32 v54, v27, v4, 16
+; SI-NEXT:    v_alignbit_b32 v53, v28, v2, 16
+; SI-NEXT:    v_alignbit_b32 v52, v39, v1, 16
+; SI-NEXT:    v_alignbit_b32 v51, v30, v3, 16
 ; SI-NEXT:    s_lshr_b32 s58, s8, 16
+; SI-NEXT:    s_lshr_b32 s57, s36, 16
+; SI-NEXT:    s_lshr_b32 s56, s53, 16
+; SI-NEXT:    s_lshr_b32 s47, s94, 16
+; SI-NEXT:    s_lshr_b32 s46, s49, 16
+; SI-NEXT:    s_lshr_b32 s45, s48, 16
+; SI-NEXT:    s_lshr_b32 s44, s11, 16
+; SI-NEXT:    s_lshr_b32 s43, s9, 16
+; SI-NEXT:    s_lshr_b32 s41, s16, 16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v61, v5
+; SI-NEXT:    v_or_b32_e32 v5, v12, v5
 ; SI-NEXT:    v_or_b32_e32 v5, s4, v5
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 0x3000000, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v14
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; SI-NEXT:    v_readlane_b32 s4, v62, 45
-; SI-NEXT:    s_add_i32 s6, s4, 3
-; SI-NEXT:    s_and_b32 s4, s6, 0xff
-; SI-NEXT:    v_readlane_b32 s6, v62, 27
-; SI-NEXT:    s_add_i32 s34, s6, 3
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v13
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v6
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; SI-NEXT:    v_readlane_b32 s4, v62, 48
+; SI-NEXT:    s_add_i32 s7, s4, 3
+; SI-NEXT:    s_and_b32 s4, s7, 0xff
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 28
-; SI-NEXT:    s_and_b32 s6, s34, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s5, 24
-; SI-NEXT:    s_lshl_b32 s6, s6, 16
+; SI-NEXT:    s_lshl_b32 s5, s89, 24
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s6
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_add_i32 s10, s4, 0x3000000
-; SI-NEXT:    v_readlane_b32 s4, v62, 43
-; SI-NEXT:    s_add_i32 s97, s4, 3
-; SI-NEXT:    s_and_b32 s4, s97, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s28, 8
+; SI-NEXT:    v_readlane_b32 s4, v62, 46
+; SI-NEXT:    s_add_i32 s99, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 44
+; SI-NEXT:    s_and_b32 s4, s99, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_readlane_b32 s5, v62, 41
-; SI-NEXT:    v_readlane_b32 s6, v62, 33
-; SI-NEXT:    s_lshl_b32 s5, s5, 8
-; SI-NEXT:    s_add_i32 s54, s6, 3
-; SI-NEXT:    s_and_b32 s6, s54, 0xff
-; SI-NEXT:    s_lshl_b32 s6, s6, 16
-; SI-NEXT:    v_mov_b32_e32 v20, s10
-; SI-NEXT:    v_alignbit_b32 v44, v20, v12, 16
-; SI-NEXT:    s_lshr_b32 s59, s10, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v14, v5
+; SI-NEXT:    v_or_b32_e32 v5, v19, v5
 ; SI-NEXT:    v_or_b32_e32 v5, s4, v5
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v13
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT:    v_readlane_b32 s4, v62, 42
+; SI-NEXT:    v_readlane_b32 s4, v62, 36
 ; SI-NEXT:    s_add_i32 s81, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 35
+; SI-NEXT:    v_readlane_b32 s6, v62, 28
 ; SI-NEXT:    s_and_b32 s4, s81, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s5, 8
+; SI-NEXT:    s_add_i32 s55, s6, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 31
+; SI-NEXT:    v_readlane_b32 s5, v62, 26
+; SI-NEXT:    s_and_b32 s6, s55, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    s_lshl_b32 s5, s5, 24
+; SI-NEXT:    s_lshl_b32 s6, s6, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s6
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_add_i32 s12, s4, 0x3000000
-; SI-NEXT:    v_readlane_b32 s4, v62, 39
+; SI-NEXT:    v_readlane_b32 s4, v62, 34
 ; SI-NEXT:    s_add_i32 s69, s4, 3
-; SI-NEXT:    v_readlane_b32 s5, v62, 35
+; SI-NEXT:    v_readlane_b32 s5, v62, 29
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v5
 ; SI-NEXT:    s_and_b32 s4, s69, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v29
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_readlane_b32 s5, v62, 25
-; SI-NEXT:    v_readlane_b32 s6, v62, 24
+; SI-NEXT:    v_or_b32_e32 v5, v18, v5
+; SI-NEXT:    v_or_b32_e32 v5, s4, v5
+; SI-NEXT:    v_readlane_b32 s4, v62, 22
+; SI-NEXT:    s_add_i32 s34, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 21
+; SI-NEXT:    v_readlane_b32 s6, v62, 19
+; SI-NEXT:    s_and_b32 s4, s34, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_add_i32 s92, s6, 3
-; SI-NEXT:    s_and_b32 s6, s92, 0xff
-; SI-NEXT:    s_lshl_b32 s6, s6, 16
-; SI-NEXT:    v_mov_b32_e32 v19, s12
-; SI-NEXT:    v_alignbit_b32 v45, v19, v15, 16
-; SI-NEXT:    s_lshr_b32 s60, s12, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v13, v5
-; SI-NEXT:    v_or_b32_e32 v5, s4, v5
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 0x3000000, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v6
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_readlane_b32 s4, v62, 26
-; SI-NEXT:    s_add_i32 s31, s4, 3
-; SI-NEXT:    s_and_b32 s4, s31, 0xff
 ; SI-NEXT:    s_or_b32 s4, s5, s4
+; SI-NEXT:    s_and_b32 s6, s92, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s87, 24
+; SI-NEXT:    s_lshl_b32 s5, s97, 24
+; SI-NEXT:    s_lshl_b32 s6, s6, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s6
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_add_i32 s13, s4, 0x3000000
+; SI-NEXT:    v_readlane_b32 s4, v62, 25
+; SI-NEXT:    s_add_i32 s51, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 24
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 0x3000000, v5
 ; SI-NEXT:    s_and_b32 s4, s51, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s88, 8
+; SI-NEXT:    s_lshl_b32 s5, s5, 8
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v21
 ; SI-NEXT:    s_or_b32 s4, s5, s4
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    s_addk_i32 s4, 0x300
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    s_lshl_b32 s5, s99, 8
-; SI-NEXT:    s_add_i32 s6, s96, 3
-; SI-NEXT:    s_and_b32 s6, s6, 0xff
-; SI-NEXT:    s_lshl_b32 s6, s6, 16
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v21
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_mov_b32_e32 v18, s13
-; SI-NEXT:    v_alignbit_b32 v46, v18, v16, 16
-; SI-NEXT:    s_lshr_b32 s61, s13, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v5, v6, v5
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v5, v14, v5
 ; SI-NEXT:    v_or_b32_e32 v5, s4, v5
+; SI-NEXT:    v_readlane_b32 s4, v62, 20
+; SI-NEXT:    s_add_i32 s95, s4, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 18
 ; SI-NEXT:    s_and_b32 s4, s95, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s5, 8
+; SI-NEXT:    s_add_i32 s6, s96, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
+; SI-NEXT:    s_and_b32 s6, s6, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s76, 24
+; SI-NEXT:    s_lshl_b32 s5, s80, 24
+; SI-NEXT:    s_lshl_b32 s6, s6, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s6
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_add_i32 s14, s4, 0x3000000
-; SI-NEXT:    v_readlane_b32 s4, v62, 22
-; SI-NEXT:    s_add_i32 s4, s4, 3
-; SI-NEXT:    v_readlane_b32 s5, v62, 21
+; SI-NEXT:    s_add_i32 s4, s93, 3
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
-; SI-NEXT:    s_lshl_b32 s5, s5, 8
+; SI-NEXT:    s_lshl_b32 s5, s84, 8
 ; SI-NEXT:    s_or_b32 s4, s5, s4
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    s_addk_i32 s4, 0x300
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    s_lshl_b32 s5, s65, 8
-; SI-NEXT:    s_add_i32 s6, s55, 3
-; SI-NEXT:    s_and_b32 s6, s6, 0xff
-; SI-NEXT:    s_lshl_b32 s6, s6, 16
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 0x3000000, v5
-; SI-NEXT:    v_mov_b32_e32 v5, s14
-; SI-NEXT:    v_alignbit_b32 v47, v5, v17, 16
-; SI-NEXT:    s_lshr_b32 s62, s14, 16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v13
 ; SI-NEXT:    v_or_b32_e32 v6, s4, v6
-; SI-NEXT:    s_add_i32 s4, s80, 3
+; SI-NEXT:    s_add_i32 s4, s83, 3
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
+; SI-NEXT:    s_lshl_b32 s5, s25, 8
+; SI-NEXT:    s_add_i32 s6, s64, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
+; SI-NEXT:    v_readlane_b32 s5, v62, 15
+; SI-NEXT:    s_and_b32 s6, s6, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s64, 24
+; SI-NEXT:    s_lshl_b32 s5, s5, 24
+; SI-NEXT:    s_lshl_b32 s6, s6, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s6
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_add_i32 s15, s4, 0x3000000
-; SI-NEXT:    v_readlane_b32 s4, v62, 3
+; SI-NEXT:    v_readlane_b32 s4, v62, 2
 ; SI-NEXT:    s_add_i32 s4, s4, 3
-; SI-NEXT:    v_readlane_b32 s5, v62, 2
+; SI-NEXT:    v_readlane_b32 s5, v62, 1
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_add_i32 s6, s26, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 1
 ; SI-NEXT:    s_and_b32 s6, s6, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s5, 24
+; SI-NEXT:    s_lshl_b32 s5, s27, 24
 ; SI-NEXT:    s_lshl_b32 s6, s6, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s6
@@ -210126,42 +210152,41 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    s_add_i32 s4, s4, 3
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s29, 8
-; SI-NEXT:    s_add_i32 s6, s74, 3
+; SI-NEXT:    s_add_i32 s6, s76, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
+; SI-NEXT:    v_readlane_b32 s5, v62, 16
 ; SI-NEXT:    s_and_b32 s6, s6, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s21, 24
+; SI-NEXT:    s_lshl_b32 s5, s5, 24
 ; SI-NEXT:    s_lshl_b32 s6, s6, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s6
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_add_i32 s40, s4, 0x3000000
-; SI-NEXT:    v_readlane_b32 s4, v62, 10
+; SI-NEXT:    v_readlane_b32 s4, v62, 7
 ; SI-NEXT:    s_add_i32 s4, s4, 3
-; SI-NEXT:    v_readlane_b32 s5, v62, 23
-; SI-NEXT:    v_readlane_b32 s6, v62, 9
+; SI-NEXT:    v_readlane_b32 s5, v62, 17
+; SI-NEXT:    v_readlane_b32 s6, v62, 6
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_add_i32 s6, s6, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 8
 ; SI-NEXT:    s_and_b32 s6, s6, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
-; SI-NEXT:    s_lshl_b32 s5, s5, 24
+; SI-NEXT:    s_lshl_b32 s5, s19, 24
 ; SI-NEXT:    s_lshl_b32 s6, s6, 16
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
 ; SI-NEXT:    s_or_b32 s5, s5, s6
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    s_add_i32 s91, s4, 0x3000000
-; SI-NEXT:    v_readlane_b32 s4, v62, 7
-; SI-NEXT:    s_add_i32 s4, s4, 3
-; SI-NEXT:    v_readlane_b32 s5, v62, 6
-; SI-NEXT:    v_readlane_b32 s6, v62, 5
+; SI-NEXT:    s_add_i32 s18, s4, 0x3000000
+; SI-NEXT:    s_add_i32 s4, s20, 3
+; SI-NEXT:    v_readlane_b32 s5, v62, 5
+; SI-NEXT:    v_readlane_b32 s6, v62, 4
 ; SI-NEXT:    s_and_b32 s4, s4, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s5, 8
 ; SI-NEXT:    s_add_i32 s6, s6, 3
 ; SI-NEXT:    s_or_b32 s4, s5, s4
-; SI-NEXT:    v_readlane_b32 s5, v62, 4
+; SI-NEXT:    v_readlane_b32 s5, v62, 3
 ; SI-NEXT:    s_and_b32 s6, s6, 0xff
 ; SI-NEXT:    s_addk_i32 s4, 0x300
 ; SI-NEXT:    s_lshl_b32 s5, s5, 24
@@ -210170,18 +210195,33 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    s_or_b32 s5, s5, s6
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    s_add_i32 s42, s4, 0x3000000
-; SI-NEXT:    v_mov_b32_e32 v13, s91
+; SI-NEXT:    v_mov_b32_e32 v13, s18
+; SI-NEXT:    v_mov_b32_e32 v20, s10
+; SI-NEXT:    v_mov_b32_e32 v19, s12
+; SI-NEXT:    v_mov_b32_e32 v18, s13
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 0x3000000, v5
+; SI-NEXT:    v_mov_b32_e32 v5, s14
 ; SI-NEXT:    v_add_i32_e32 v50, vcc, 0x3000000, v6
 ; SI-NEXT:    v_mov_b32_e32 v6, s15
 ; SI-NEXT:    v_alignbit_b32 v57, s42, v13, 16
 ; SI-NEXT:    v_mov_b32_e32 v13, s17
 ; SI-NEXT:    v_alignbit_b32 v58, s40, v13, 16
 ; SI-NEXT:    v_alignbit_b32 v56, v6, v50, 16
+; SI-NEXT:    v_alignbit_b32 v47, v5, v17, 16
+; SI-NEXT:    s_waitcnt expcnt(1)
+; SI-NEXT:    v_alignbit_b32 v46, v18, v16, 16
+; SI-NEXT:    v_alignbit_b32 v45, v19, v15, 16
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_alignbit_b32 v44, v20, v12, 16
 ; SI-NEXT:    s_lshr_b32 s73, s42, 16
 ; SI-NEXT:    s_lshr_b32 s72, s40, 16
 ; SI-NEXT:    s_lshr_b32 s63, s15, 16
-; SI-NEXT:  .LBB97_5: ; %end
-; SI-NEXT:    s_and_b32 s4, s91, 0xffff
+; SI-NEXT:    s_lshr_b32 s62, s14, 16
+; SI-NEXT:    s_lshr_b32 s61, s13, 16
+; SI-NEXT:    s_lshr_b32 s60, s12, 16
+; SI-NEXT:    s_lshr_b32 s59, s10, 16
+; SI-NEXT:  .LBB97_3: ; %end
+; SI-NEXT:    s_and_b32 s4, s18, 0xffff
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v57
 ; SI-NEXT:    v_or_b32_e32 v5, s4, v5
 ; SI-NEXT:    s_and_b32 s4, s42, 0xffff
@@ -210190,7 +210230,6 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 4, v0
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_mov_b32_e32 v6, s4
 ; SI-NEXT:    buffer_store_dword v6, v5, s[0:3], 0 offen
 ; SI-NEXT:    s_and_b32 s4, s17, 0xffff
@@ -210287,7 +210326,7 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v10
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v42
-; SI-NEXT:    s_and_b32 s4, s35, 0xffff
+; SI-NEXT:    s_and_b32 s4, s36, 0xffff
 ; SI-NEXT:    s_lshl_b32 s5, s57, 16
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v6
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 64, v0
@@ -210300,7 +210339,7 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v9
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v41
-; SI-NEXT:    s_and_b32 s4, s52, 0xffff
+; SI-NEXT:    s_and_b32 s4, s53, 0xffff
 ; SI-NEXT:    s_lshl_b32 s5, s56, 16
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v6
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x48, v0
@@ -210436,11 +210475,91 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; SI-NEXT:    v_readlane_b32 s31, v63, 1
 ; SI-NEXT:    v_readlane_b32 s30, v63, 0
 ; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
 ; SI-NEXT:    s_mov_b64 exec, s[4:5]
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
+; SI-NEXT:  .LBB97_4:
+; SI-NEXT:    v_mov_b32_e32 v5, v13
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; SI-NEXT:    v_mov_b32_e32 v25, v58
+; SI-NEXT:    v_mov_b32_e32 v48, v39
+; SI-NEXT:    v_mov_b32_e32 v39, v57
+; SI-NEXT:    v_mov_b32_e32 v49, v56
+; SI-NEXT:    v_mov_b32_e32 v20, v47
+; SI-NEXT:    v_mov_b32_e32 v30, v37
+; SI-NEXT:    v_mov_b32_e32 v36, v35
+; SI-NEXT:    v_mov_b32_e32 v35, v45
+; SI-NEXT:    v_mov_b32_e32 v27, v26
+; SI-NEXT:    s_waitcnt expcnt(2)
+; SI-NEXT:    v_mov_b32_e32 v32, v23
+; SI-NEXT:    ; implicit-def: $sgpr18
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $sgpr42
+; SI-NEXT:    ; implicit-def: $sgpr73
+; SI-NEXT:    ; implicit-def: $sgpr17
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $sgpr40
+; SI-NEXT:    ; implicit-def: $sgpr72
+; SI-NEXT:    ; implicit-def: $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $sgpr15
+; SI-NEXT:    ; implicit-def: $sgpr63
+; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $sgpr14
+; SI-NEXT:    ; implicit-def: $sgpr62
+; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    ; implicit-def: $sgpr13
+; SI-NEXT:    ; implicit-def: $sgpr61
+; SI-NEXT:    ; implicit-def: $vgpr15
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $sgpr12
+; SI-NEXT:    ; implicit-def: $sgpr60
+; SI-NEXT:    ; implicit-def: $vgpr12
+; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $sgpr10
+; SI-NEXT:    ; implicit-def: $sgpr59
+; SI-NEXT:    ; implicit-def: $vgpr11
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $sgpr8
+; SI-NEXT:    ; implicit-def: $sgpr58
+; SI-NEXT:    ; implicit-def: $vgpr10
+; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $sgpr36
+; SI-NEXT:    ; implicit-def: $sgpr57
+; SI-NEXT:    ; implicit-def: $vgpr9
+; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $sgpr53
+; SI-NEXT:    ; implicit-def: $sgpr56
+; SI-NEXT:    ; implicit-def: $vgpr8
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $sgpr94
+; SI-NEXT:    ; implicit-def: $sgpr47
+; SI-NEXT:    ; implicit-def: $vgpr7
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $sgpr49
+; SI-NEXT:    ; implicit-def: $sgpr46
+; SI-NEXT:    ; implicit-def: $vgpr4
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $sgpr48
+; SI-NEXT:    ; implicit-def: $sgpr45
+; SI-NEXT:    ; implicit-def: $vgpr2
+; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $sgpr11
+; SI-NEXT:    ; implicit-def: $sgpr44
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $sgpr9
+; SI-NEXT:    ; implicit-def: $sgpr43
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $sgpr16
+; SI-NEXT:    ; implicit-def: $sgpr41
+; SI-NEXT:    s_branch .LBB97_2
 ;
 ; VI-LABEL: bitcast_v128i8_to_v64i16_scalar:
 ; VI:       ; %bb.0:
@@ -210831,17 +210950,17 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; VI-NEXT:    v_mov_b32_e32 v45, v62
 ; VI-NEXT:    v_or_b32_sdwa v16, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
+; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_or_b32_sdwa v3, v51, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v17, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_mov_b32_e32 v54, v22
-; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_mov_b32_e32 v32, v1
 ; VI-NEXT:    v_or_b32_sdwa v1, v41, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_mov_b32_e32 v54, v22
 ; VI-NEXT:    v_mov_b32_e32 v41, v24
 ; VI-NEXT:    v_or_b32_sdwa v18, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
@@ -212078,8 +212197,8 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    s_cbranch_vccnz .LBB97_5
 ; GFX9-NEXT:  ; %bb.4: ; %cmp.true
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_add_i32 s28, s28, 3
 ; GFX9-NEXT:    s_and_b32 s4, s28, 0xff
 ; GFX9-NEXT:    s_lshl_b32 s5, s29, 8
@@ -212112,7 +212231,7 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(13)
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -212162,9 +212281,9 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    s_or_b32 s7, s8, s7
 ; GFX9-NEXT:    s_and_b32 s8, s16, 0xff
 ; GFX9-NEXT:    s_or_b32 s8, s9, s8
+; GFX9-NEXT:    s_and_b32 s9, s18, 0xff
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT:    s_and_b32 s9, s18, 0xff
 ; GFX9-NEXT:    s_or_b32 s9, s10, s9
 ; GFX9-NEXT:    s_addk_i32 s4, 0x300
 ; GFX9-NEXT:    s_addk_i32 s5, 0x300
@@ -212197,11 +212316,11 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    v_add_u32_e32 v1, 0x300, v1
 ; GFX9-NEXT:    v_lshl_or_b32 v6, v1, 16, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -212212,11 +212331,11 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    v_add_u32_e32 v1, 0x300, v1
 ; GFX9-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
@@ -212347,9 +212466,9 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
 ; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
 ; GFX9-NEXT:    v_or_b32_sdwa v51, v34, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
@@ -212426,22 +212545,21 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v43, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v0
 ; GFX9-NEXT:    v_add_u32_e32 v52, 0x300, v43
 ; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s8
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v44, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_add_u32_e32 v26, 3, v33
@@ -214272,13 +214390,15 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:60
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:76
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:124
 ; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:132
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:128
-; SI-NEXT:    ; implicit-def: $vgpr27
-; SI-NEXT:    ; kill: killed $vgpr27
-; SI-NEXT:    ; implicit-def: $vgpr27
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:128
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
+; SI-NEXT:    ; kill: killed $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr23
 ; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr44
@@ -214303,112 +214423,94 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; kill: killed $vgpr27
-; SI-NEXT:    ; implicit-def: $vgpr27
+; SI-NEXT:    ; kill: killed $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr23
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v11
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v13
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v13
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:116
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:116
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:112
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:80
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v56
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:104
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:88
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:84
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:80
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v57
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:24
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v56
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:20
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:72
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v9
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:56
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:52
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v59
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:48
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:68
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v6
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v10
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v7
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v14
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v9
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v18
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v22
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v26
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v10
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v30
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v14
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v18
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v22
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v7
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:100
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v26
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v30
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
@@ -214522,6 +214624,8 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
@@ -214538,6 +214642,13 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:16
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
@@ -214568,10 +214679,6 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:64
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
@@ -214596,12 +214703,10 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v58
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v63
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v61
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v62
+; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v57
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v58
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v62
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr26
@@ -214613,37 +214718,54 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr6
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v60
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v59
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v63
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:100
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v61
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:64
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:32
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32
-; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v60
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v8
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v8
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v12
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v20
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v16
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v20
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v28
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v24
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v47
@@ -214652,8 +214774,8 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB98_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v47
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v47
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -214664,7 +214786,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v45, v2, v6
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -214675,7 +214797,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v43, v2, v6
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -214686,7 +214808,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v41, v2, v6
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -214697,7 +214819,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v55, v2, v6
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -214708,7 +214830,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v53, v2, v6
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -214719,103 +214841,102 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v51, v2, v6
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v50, v2, v6
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v28
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v49, v2, v6
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT:    v_or_b32_e32 v49, v2, v27
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr27
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v48, v2, v6
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v39, v2, v23
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    v_or_b32_e32 v39, v2, v19
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v38, v2, v6
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v37, v2, v15
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr15
+; SI-NEXT:    v_or_b32_e32 v37, v2, v29
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr29
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v36, v2, v6
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v35, v2, v29
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    v_or_b32_e32 v35, v2, v25
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr25
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v34, v2, v6
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v33, v2, v21
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr21
+; SI-NEXT:    v_or_b32_e32 v33, v2, v11
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr11
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v31, v2, v19
+; SI-NEXT:    v_or_b32_e32 v31, v2, v6
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v32, v2, v17
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    v_or_b32_e32 v32, v2, v7
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr7
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v26, v2, v25
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr25
+; SI-NEXT:    v_or_b32_e32 v26, v2, v15
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v30, v2, v13
+; SI-NEXT:    v_or_b32_e32 v30, v2, v3
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr13
+; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v18, v2, v11
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr11
+; SI-NEXT:    v_or_b32_e32 v18, v2, v21
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v22, v2, v9
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr9
+; SI-NEXT:    v_or_b32_e32 v22, v2, v5
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v14, v2, v7
+; SI-NEXT:    v_or_b32_e32 v14, v2, v17
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr7
+; SI-NEXT:    ; implicit-def: $vgpr17
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v10, v2, v5
+; SI-NEXT:    v_or_b32_e32 v10, v2, v9
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr5
+; SI-NEXT:    ; implicit-def: $vgpr9
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v6, v2, v3
+; SI-NEXT:    v_or_b32_e32 v6, v2, v13
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr13
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v1
@@ -214843,17 +214964,17 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v60
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v58
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v59
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v59
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v63
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v57
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v63
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v58
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_bfe_u32 v1, v4, 8, 8
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
@@ -214879,16 +215000,16 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_bfe_u32 v1, v60, 8, 8
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_bfe_u32 v1, v58, 8, 8
+; SI-NEXT:    v_bfe_u32 v1, v59, 8, 8
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_bfe_u32 v1, v59, 8, 8
+; SI-NEXT:    v_bfe_u32 v1, v63, 8, 8
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_bfe_u32 v1, v57, 8, 8
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_bfe_u32 v1, v63, 8, 8
+; SI-NEXT:    v_bfe_u32 v1, v58, 8, 8
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v56
@@ -214898,22 +215019,22 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_bfe_u32 v1, v61, 8, 8
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v62
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_bfe_u32 v1, v62, 8, 8
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_bfe_u32 v1, v47, 8, 8
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v45, v46, 24
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v45, v46, 16
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
@@ -215030,31 +215151,31 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v22, v18, 24
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v22, v18, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v22, v18, 8
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v10, v14, 24
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v10, v14, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v10, v14, 8
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v2, v6, 24
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v2, v6, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v1, v2, v6, 8
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v45
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
@@ -215206,59 +215327,66 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr1
 ; SI-NEXT:  .LBB98_2: ; %Flow
 ; SI-NEXT:    s_or_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
 ; SI-NEXT:    s_xor_b64 exec, exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB98_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x30000, v2
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_or_b32_e32 v2, v13, v2
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_or_b32_e32 v1, v1, v3
+; SI-NEXT:    v_or_b32_e32 v1, v1, v4
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT:    v_or_b32_e32 v8, v9, v6
+; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x30000, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v8
 ; SI-NEXT:    v_alignbit_b32 v12, v2, v6, 24
 ; SI-NEXT:    v_alignbit_b32 v20, v2, v6, 16
 ; SI-NEXT:    v_alignbit_b32 v47, v2, v6, 8
+; SI-NEXT:    v_lshrrev_b32_e32 v24, 24, v10
+; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v10
 ; SI-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
-; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
+; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v11, v1
+; SI-NEXT:    v_or_b32_e32 v1, v21, v1
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v9, v1
+; SI-NEXT:    v_or_b32_e32 v1, v5, v1
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; SI-NEXT:    v_alignbit_b32 v62, v22, v18, 24
 ; SI-NEXT:    v_alignbit_b32 v63, v22, v18, 16
 ; SI-NEXT:    v_lshrrev_b32_e32 v59, 24, v22
@@ -215266,71 +215394,69 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v25, v1
+; SI-NEXT:    v_or_b32_e32 v1, v15, v1
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v7, v3
-; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v13, v1
+; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v19, v1
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v31, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v17, v1
+; SI-NEXT:    v_or_b32_e32 v1, v7, v1
 ; SI-NEXT:    v_add_i32_e32 v32, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v32
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v34, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v21, v1
+; SI-NEXT:    v_or_b32_e32 v1, v11, v1
 ; SI-NEXT:    v_add_i32_e32 v33, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v33
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v36, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v29, v1
+; SI-NEXT:    v_or_b32_e32 v1, v25, v1
 ; SI-NEXT:    v_add_i32_e32 v35, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v38, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v15, v1
+; SI-NEXT:    v_or_b32_e32 v1, v29, v1
 ; SI-NEXT:    v_add_i32_e32 v37, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -215338,12 +215464,12 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v48, vcc, s6, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v23, v1
+; SI-NEXT:    v_or_b32_e32 v1, v19, v1
 ; SI-NEXT:    v_add_i32_e32 v39, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -215352,32 +215478,27 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v50, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v5, v4
-; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v4
+; SI-NEXT:    v_or_b32_e32 v4, v17, v4
+; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v4
 ; SI-NEXT:    v_alignbit_b32 v56, v10, v14, 24
 ; SI-NEXT:    v_alignbit_b32 v57, v10, v14, 16
 ; SI-NEXT:    v_alignbit_b32 v61, v10, v14, 8
-; SI-NEXT:    v_lshrrev_b32_e32 v24, 24, v10
-; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v10
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v1, v3, v1
+; SI-NEXT:    v_or_b32_e32 v1, v27, v1
 ; SI-NEXT:    v_add_i32_e32 v49, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v52, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215385,7 +215506,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v51, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215393,7 +215514,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v54, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215401,7 +215522,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v53, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215409,7 +215530,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v40, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215417,7 +215538,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v55, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215425,7 +215546,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v42, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215433,7 +215554,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v41, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215441,7 +215562,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v44, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215449,7 +215570,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v43, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -215457,7 +215578,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v46, vcc, s6, v1
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -216239,7 +216360,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v27
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
@@ -217510,8 +217631,8 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    ; kill: killed $vgpr50
@@ -217636,7 +217757,7 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
+; GFX9-NEXT:    s_waitcnt vmcnt(17)
 ; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
 ; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:    ; kill: killed $vgpr33
@@ -217755,7 +217876,6 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_waitcnt vmcnt(45)
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
 ; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
@@ -217893,7 +218013,6 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
 ; GFX9-NEXT:    v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
 ; GFX9-NEXT:    s_waitcnt vmcnt(18)
 ; GFX9-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
@@ -218309,13 +218428,14 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v22, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
@@ -218334,13 +218454,14 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v24, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
@@ -218359,13 +218480,14 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v26, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
@@ -218384,13 +218506,14 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v28, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
@@ -218409,13 +218532,14 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v30, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
@@ -219761,14 +219885,14 @@ define inreg <128 x i8> @bitcast_v64i16_to_v128i8_scalar(<64 x i16> inreg %a, i3
 ; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:4
-; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:68
 ; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:32
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; SI-NEXT:    v_writelane_b32 v63, s30, 0
@@ -219850,14 +219974,14 @@ define inreg <128 x i8> @bitcast_v64i16_to_v128i8_scalar(<64 x i16> inreg %a, i3
 ; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v28
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v40
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v41
-; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v55
+; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v40
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v55
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v42
+; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v41
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v42
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v43
+; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v43
 ; SI-NEXT:    v_writelane_b32 v62, s6, 0
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
@@ -225746,12 +225870,6 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:84
 ; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:92
 ; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:88
-; SI-NEXT:    s_waitcnt expcnt(6)
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:96
-; SI-NEXT:    s_waitcnt expcnt(4)
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:104
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -225788,8 +225906,12 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:116
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:116
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:120
 ; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v32
 ; SI-NEXT:    v_mul_f32_e32 v33, 1.0, v33
@@ -225803,11 +225925,11 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_mul_f32_e32 v49, 1.0, v49
 ; SI-NEXT:    v_mul_f32_e32 v50, 1.0, v50
 ; SI-NEXT:    v_mul_f32_e32 v51, 1.0, v51
+; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_mul_f32_e32 v52, 1.0, v52
 ; SI-NEXT:    v_mul_f32_e32 v53, 1.0, v53
 ; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v54
 ; SI-NEXT:    v_mul_f32_e32 v55, 1.0, v55
-; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v40
 ; SI-NEXT:    v_mul_f32_e32 v41, 1.0, v41
 ; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v42
@@ -225820,21 +225942,22 @@ define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v46
 ; SI-NEXT:    v_mul_f32_e32 v47, 1.0, v47
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_mul_f32_e32 v56, 1.0, v56
-; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v57
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_mul_f32_e32 v56, 1.0, v1
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v58
-; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v59
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_mul_f32_e32 v61, 1.0, v31
+; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v31
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:124
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:128
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v58
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v59
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v60
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v62
+; SI-NEXT:    v_mul_f32_e32 v61, 1.0, v61
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_mul_f32_e32 v62, 1.0, v31
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -233259,27 +233382,23 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:84
 ; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:92
 ; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:88
-; SI-NEXT:    s_waitcnt expcnt(6)
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:96
-; SI-NEXT:    s_waitcnt expcnt(4)
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:104
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v2
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
 ; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
@@ -233298,11 +233417,20 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
 ; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
 ; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; kill: killed $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:116
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:116
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:120
 ; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
 ; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
@@ -233317,35 +233445,36 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v50, v50
 ; SI-NEXT:    v_cvt_f16_f32_e32 v51, v51
 ; SI-NEXT:    v_cvt_f16_f32_e32 v52, v52
+; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v53, v53
 ; SI-NEXT:    v_cvt_f16_f32_e32 v54, v54
 ; SI-NEXT:    v_cvt_f16_f32_e32 v55, v55
-; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v40, v40
 ; SI-NEXT:    v_cvt_f16_f32_e32 v41, v41
 ; SI-NEXT:    v_cvt_f16_f32_e32 v42, v42
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v43, v43
-; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v44, v44
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v45, v45
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v46, v46
 ; SI-NEXT:    v_cvt_f16_f32_e32 v47, v47
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f16_f32_e32 v56, v56
-; SI-NEXT:    v_cvt_f16_f32_e32 v57, v57
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f16_f32_e32 v56, v1
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_cvt_f16_f32_e32 v58, v58
-; SI-NEXT:    v_cvt_f16_f32_e32 v59, v59
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f16_f32_e32 v60, v1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f16_f32_e32 v61, v31
+; SI-NEXT:    v_cvt_f16_f32_e32 v57, v31
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:128
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f16_f32_e32 v58, v58
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f16_f32_e32 v59, v59
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_cvt_f16_f32_e32 v60, v60
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f16_f32_e32 v61, v61
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v62, v62
 ; SI-NEXT:    s_waitcnt vmcnt(2)
@@ -233476,217 +233605,211 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; kill: killed $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    ; kill: killed $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    ; kill: killed $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    ; kill: killed $vgpr2
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB102_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v63
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v31
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    ; kill: killed $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    ; implicit-def: $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v7
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v8
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v9
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v10
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v11
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v12
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v13
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v14
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v15
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v16
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v17
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v18
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v19
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v20
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v21
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v22
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v23
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v25
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v26
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v27
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v28
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v29
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v30
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v32
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v33
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v34
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v35
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v36
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v37
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v38
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v39
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v48
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v49
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v50
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v51
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v52
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v53
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v54
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v55
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v40
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v41
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v42
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v43
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v44
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v45
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v46
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v47
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v56
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v57
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v58
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v59
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v60
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v61
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v5
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v62
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr4
 ; SI-NEXT:    ; implicit-def: $vgpr5
 ; SI-NEXT:    ; implicit-def: $vgpr6
@@ -233754,18 +233877,11 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v31, v63
 ; SI-NEXT:    v_cvt_f32_f16_e32 v63, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v62
-; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
 ; SI-NEXT:    v_add_f32_e32 v62, 0x38000000, v63
 ; SI-NEXT:    v_add_f32_e32 v63, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v61
-; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
-; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
 ; SI-NEXT:    v_cvt_f32_f16_e32 v60, v60
 ; SI-NEXT:    v_cvt_f32_f16_e32 v61, v59
 ; SI-NEXT:    v_cvt_f32_f16_e32 v58, v58
@@ -233773,14 +233889,6 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v57, v57
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v46
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
 ; SI-NEXT:    v_add_f32_e32 v59, 0x38000000, v60
 ; SI-NEXT:    v_add_f32_e32 v60, 0x38000000, v61
 ; SI-NEXT:    v_add_f32_e32 v61, 0x38000000, v58
@@ -233789,16 +233897,6 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v46, 0x38000000, v57
 ; SI-NEXT:    v_add_f32_e32 v57, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v45
-; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v5
-; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v6
-; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v7
-; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v8
 ; SI-NEXT:    v_add_f32_e32 v47, 0x38000000, v56
 ; SI-NEXT:    v_add_f32_e32 v56, 0x38000000, v58
 ; SI-NEXT:    v_cvt_f32_f16_e32 v44, v44
@@ -233806,20 +233904,6 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v58, v42
 ; SI-NEXT:    v_add_f32_e32 v42, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v41
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
-; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
 ; SI-NEXT:    v_add_f32_e32 v43, 0x38000000, v44
 ; SI-NEXT:    v_add_f32_e32 v44, 0x38000000, v45
 ; SI-NEXT:    v_add_f32_e32 v45, 0x38000000, v58
@@ -233828,16 +233912,6 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v58, v54
 ; SI-NEXT:    v_add_f32_e32 v54, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v53
-; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v9
-; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v10
-; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v11
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v12
 ; SI-NEXT:    v_add_f32_e32 v55, 0x38000000, v40
 ; SI-NEXT:    v_add_f32_e32 v40, 0x38000000, v41
 ; SI-NEXT:    v_add_f32_e32 v41, 0x38000000, v58
@@ -233846,20 +233920,6 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v58, v50
 ; SI-NEXT:    v_add_f32_e32 v50, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v49
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
-; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
 ; SI-NEXT:    v_add_f32_e32 v51, 0x38000000, v52
 ; SI-NEXT:    v_add_f32_e32 v52, 0x38000000, v53
 ; SI-NEXT:    v_add_f32_e32 v53, 0x38000000, v58
@@ -233868,16 +233928,6 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v58, v38
 ; SI-NEXT:    v_add_f32_e32 v38, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v37
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v13
-; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v14
-; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v15
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v16
 ; SI-NEXT:    v_add_f32_e32 v39, 0x38000000, v48
 ; SI-NEXT:    v_add_f32_e32 v48, 0x38000000, v49
 ; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v58
@@ -233886,20 +233936,6 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v58, v34
 ; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v33
-; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
-; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
 ; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v36
 ; SI-NEXT:    v_add_f32_e32 v36, 0x38000000, v37
 ; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v58
@@ -233908,242 +233944,330 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v58, v29
 ; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v28
-; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v17
-; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v18
-; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v19
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v20
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v32
 ; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v33
 ; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v58
 ; SI-NEXT:    v_cvt_f32_f16_e32 v58, v25
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v1
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v24
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
+; SI-NEXT:    v_cvt_f32_f16_e32 v28, v26
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v27
+; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v28
+; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v58
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
+; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
+; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
+; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
+; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
+; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
+; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
+; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
+; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
+; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
+; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
+; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
+; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
+; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
+; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
+; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v28
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
+; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v6
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
-; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v9
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v10
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v11
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v13
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v14
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v15
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v17
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v18
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v20
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
-; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
-; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v21
-; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT:    v_cvt_f32_f16_e32 v28, v26
-; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v27
-; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v28
-; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v58
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v28
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v27
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v26
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v33
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v32
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v30
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v37
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v36
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v35
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v34
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v49
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v48
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v39
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v38
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v53
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v52
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v51
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v50
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v41
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v40
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v55
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v54
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v45
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v44
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v43
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v42
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v57
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v56
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v47
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v46
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v61
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v60
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v59
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v31
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_cvt_f32_f16_e32 v58, v58
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v63
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v62
+; SI-NEXT:    v_add_f32_e32 v58, 0x38000000, v58
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v31
+; SI-NEXT:    v_cvt_f16_f32_e32 v58, v58
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v58, v58
-; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
-; SI-NEXT:    v_add_f32_e32 v58, 0x38000000, v58
-; SI-NEXT:    v_cvt_f16_f32_e32 v58, v58
-; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v58
-; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
+; SI-NEXT:    v_mov_b32_e32 v3, v24
+; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
 ; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
 ; SI-NEXT:  .LBB102_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234151,11 +234275,9 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
 ; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
 ; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v3
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
@@ -234163,8 +234285,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234174,8 +234296,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234185,8 +234307,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234196,8 +234318,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234207,8 +234329,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234218,8 +234340,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234229,8 +234351,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234240,8 +234362,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234251,8 +234373,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234262,8 +234384,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234273,8 +234395,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234284,8 +234406,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234295,8 +234417,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234306,8 +234428,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234317,8 +234439,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234328,8 +234450,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234339,8 +234461,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234350,8 +234472,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234361,8 +234483,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234372,8 +234494,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234383,8 +234505,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234394,8 +234516,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234405,8 +234527,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234416,8 +234538,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234427,8 +234549,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234438,8 +234560,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234449,8 +234571,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234460,8 +234582,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234471,8 +234593,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
@@ -234482,8 +234604,8 @@ define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
@@ -236259,145 +236381,173 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:72
 ; SI-NEXT:    s_waitcnt expcnt(3)
 ; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt expcnt(2)
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:80
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:84
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:104
-; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v5
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v15
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v16
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v19
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v20
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v23
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v24
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:80
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v27
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v28
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:88
 ; SI-NEXT:    v_mul_f32_e32 v33, 1.0, v1
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v3
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v4
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v7
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v8
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v11
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v12
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; SI-NEXT:    v_mul_f32_e32 v43, 1.0, v2
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v13
-; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v22
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v14
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v15
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v16
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v19
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v20
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v23
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v24
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v27
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v28
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
+; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v13
+; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v21
+; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v22
+; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v25
+; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v17
+; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v29
+; SI-NEXT:    v_mul_f32_e32 v43, 1.0, v2
+; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v5
+; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v6
+; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v9
+; SI-NEXT:    v_mul_f32_e32 v55, 1.0, v10
+; SI-NEXT:    v_mul_f32_e32 v53, 1.0, v14
 ; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v18
 ; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v26
-; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v21
-; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v6
-; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v9
-; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v10
-; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v17
-; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v25
-; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v29
 ; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v30
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v34
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v35
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v36
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v39
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v35
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v48
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v36
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v51
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v39
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v52
-; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:116
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:120
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; kill: killed $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    v_mul_f32_e32 v34, 1.0, v47
-; SI-NEXT:    ; kill: killed $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; kill: killed $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    v_mul_f32_e32 v47, 1.0, v57
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v48
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v37
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v52
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:116
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:120
+; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v44
+; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v46
 ; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v38
-; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v60
-; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v61
 ; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v49
 ; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v50
+; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v58
+; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v59
+; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v51
 ; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v41
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v42
-; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v44
-; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v55
-; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v31
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; kill: killed $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; kill: killed $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; kill: killed $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; kill: killed $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; kill: killed $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; kill: killed $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v46
-; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v56
-; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v58
-; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v59
-; SI-NEXT:    v_mul_f32_e32 v56, 1.0, v62
-; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v45
-; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v63
-; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v54
-; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v53
-; SI-NEXT:    ; kill: killed $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v45
+; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; kill: killed $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; kill: killed $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v47
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v56
+; SI-NEXT:    v_mul_f32_e32 v47, 1.0, v57
+; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v60
+; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v62
+; SI-NEXT:    v_mul_f32_e32 v56, 1.0, v63
+; SI-NEXT:    ; kill: killed $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr41
@@ -236410,720 +236560,672 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr38
-; SI-NEXT:    ; kill: killed $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; kill: killed $vgpr53
-; SI-NEXT:    ; implicit-def: $vgpr53
-; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; kill: killed $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; kill: killed $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; kill: killed $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v1
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v3
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v20
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_mul_f32_e32 v61, 1.0, v12
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v21
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v20
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v22
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:124
+; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v22
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:124
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:132
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v25
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v12
+; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v21
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
+; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    ; kill: killed $vgpr31
-; SI-NEXT:    ; implicit-def: $vgpr31
+; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v34
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; kill: killed $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB104_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v33
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v33
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v43
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v43
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
+; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v22
-; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr4
-; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    s_waitcnt expcnt(2)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v31
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v31
-; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v40
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v30
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v32
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr3
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
+; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v14
-; SI-NEXT:    ; implicit-def: $vgpr14
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v20
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v31
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v31
-; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v61
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v15
-; SI-NEXT:    ; implicit-def: $vgpr15
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v26
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v13
-; SI-NEXT:    ; implicit-def: $vgpr13
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v25
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v19
-; SI-NEXT:    ; implicit-def: $vgpr19
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v21
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v18
-; SI-NEXT:    ; implicit-def: $vgpr18
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v47
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v17
-; SI-NEXT:    ; implicit-def: $vgpr17
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v46
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v11
-; SI-NEXT:    ; implicit-def: $vgpr11
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v29
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v10
-; SI-NEXT:    ; implicit-def: $vgpr10
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v60
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v9
-; SI-NEXT:    ; implicit-def: $vgpr9
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v57
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v56
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v8
-; SI-NEXT:    ; implicit-def: $vgpr8
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v23
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v7
-; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v27
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v59
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v58
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v6
-; SI-NEXT:    ; implicit-def: $vgpr6
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v28
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v16
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
+; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v30
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v44
+; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v44
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v24
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v3
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr15
+; SI-NEXT:    ; implicit-def: $vgpr10
+; SI-NEXT:    ; implicit-def: $vgpr8
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr6
 ; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
+; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v34
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v29
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr29
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v47
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
+; SI-NEXT:    ; implicit-def: $vgpr4
+; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v46
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v28
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v27
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr27
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v57
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
+; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v56
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
 ; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v24
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
+; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr27
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr24
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v25
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr20
+; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr25
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v59
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; kill: killed $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr21
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v22
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr7
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v12
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v58
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v22
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v40
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v23
-; SI-NEXT:    ; implicit-def: $vgpr23
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v32
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v21
-; SI-NEXT:    ; implicit-def: $vgpr21
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v14
+; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v19
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr14
+; SI-NEXT:    ; implicit-def: $vgpr19
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v12
+; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v13
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v60
-; SI-NEXT:    ; implicit-def: $vgpr60
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v20
-; SI-NEXT:    ; implicit-def: $vgpr20
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v9
+; SI-NEXT:    ; implicit-def: $vgpr13
+; SI-NEXT:    ; implicit-def: $vgpr18
+; SI-NEXT:    ; implicit-def: $vgpr9
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v22
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v12
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v22
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v31
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v26
-; SI-NEXT:    ; implicit-def: $vgpr26
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v55
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v12
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr31
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v12
+; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v17
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    ; kill: killed $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    ; kill: killed $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    ; kill: killed $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    ; kill: killed $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    ; kill: killed $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    ; kill: killed $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v22
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v22
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v54
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v53
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v22
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v22
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v11
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v11
+; SI-NEXT:    ; implicit-def: $vgpr11
 ; SI-NEXT:  .LBB104_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB104_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
-; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v33
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v33
 ; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v43
-; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
-; SI-NEXT:    v_add_f32_e32 v51, 0x40c00000, v1
 ; SI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
-; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
-; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v51
-; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v33
-; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
-; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v14
-; SI-NEXT:    v_alignbit_b32 v31, v35, v31, 16
+; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
+; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v33
+; SI-NEXT:    v_alignbit_b32 v22, v34, v22, 16
 ; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v16
-; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v2
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v40
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v40
 ; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v14
+; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
+; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v32
+; SI-NEXT:    v_alignbit_b32 v22, v34, v22, 16
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v31
+; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v55
 ; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v32
-; SI-NEXT:    v_alignbit_b32 v1, v2, v1, 16
-; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v13
-; SI-NEXT:    v_alignbit_b32 v31, v35, v31, 16
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
+; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v31
+; SI-NEXT:    v_alignbit_b32 v22, v34, v22, 16
+; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v53
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v15
-; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v2
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v54
+; SI-NEXT:    v_add_f32_e32 v51, 0x40c00000, v34
+; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
+; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v51
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
+; SI-NEXT:    v_alignbit_b32 v22, v34, v22, 16
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
+; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v3
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v13
-; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v31
-; SI-NEXT:    v_alignbit_b32 v1, v2, v1, 16
-; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v18
-; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v14
+; SI-NEXT:    v_alignbit_b32 v15, v22, v15, 16
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v19
-; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v19
+; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
+; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
+; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v12
+; SI-NEXT:    v_alignbit_b32 v15, v19, v15, 16
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
-; SI-NEXT:    v_alignbit_b32 v1, v3, v1, 16
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v11
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v18
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
+; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
+; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
+; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v15
+; SI-NEXT:    v_alignbit_b32 v13, v18, v13, 16
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
+; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v17
-; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
-; SI-NEXT:    v_alignbit_b32 v1, v4, v1, 16
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v9
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v17
+; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
+; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
+; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v11
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
+; SI-NEXT:    v_alignbit_b32 v13, v17, v13, 16
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
+; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
+; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v10
-; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v4
-; SI-NEXT:    v_alignbit_b32 v1, v9, v1, 16
+; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v9
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v8
+; SI-NEXT:    v_alignbit_b32 v10, v13, v10, 16
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
 ; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
-; SI-NEXT:    v_alignbit_b32 v1, v8, v1, 16
-; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v6
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v7
+; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
+; SI-NEXT:    v_alignbit_b32 v8, v10, v8, 16
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
 ; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
-; SI-NEXT:    v_alignbit_b32 v1, v6, v1, 16
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v29
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v5
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
+; SI-NEXT:    v_alignbit_b32 v6, v8, v6, 16
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
+; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
+; SI-NEXT:    v_alignbit_b32 v4, v6, v4, 16
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v60
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v34
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v29
 ; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
+; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
 ; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v6
-; SI-NEXT:    v_alignbit_b32 v1, v8, v1, 16
+; SI-NEXT:    v_alignbit_b32 v4, v8, v4, 16
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v27
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v28
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v23
 ; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
+; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
+; SI-NEXT:    v_alignbit_b32 v4, v10, v4, 16
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v24
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
 ; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v8
-; SI-NEXT:    v_alignbit_b32 v1, v9, v1, 16
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v25
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v24
-; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
-; SI-NEXT:    v_alignbit_b32 v1, v10, v1, 16
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v21
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v23
 ; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
-; SI-NEXT:    v_alignbit_b32 v1, v11, v1, 16
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v26
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
+; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v61, v1, v3, 16
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v20
-; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v11
-; SI-NEXT:    v_alignbit_b32 v1, v15, v1, 16
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v10
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v58
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v59
+; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
+; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v22
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
-; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    v_alignbit_b32 v18, v12, v1, 16
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v16
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v19, v10, v3, 16
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v60
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v61
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; SI-NEXT:    v_alignbit_b32 v16, v12, v11, 16
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v56
+; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v8
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v57
+; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
+; SI-NEXT:    v_alignbit_b32 v4, v13, v4, 16
+; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v58
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v59
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT:    v_alignbit_b32 v19, v12, v11, 16
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v20
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v26
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v20, v10, v8, 16
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v56
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v57
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    v_alignbit_b32 v20, v12, v11, 16
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v46
+; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v47
+; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
+; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
+; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v13
+; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
+; SI-NEXT:    v_alignbit_b32 v4, v16, v4, 16
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v21
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v21, v10, v8, 16
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v46
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v47
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; SI-NEXT:    v_alignbit_b32 v21, v12, v11, 16
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v44
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v28
+; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
+; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v23, v10, v8, 16
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v44
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v30
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; SI-NEXT:    v_alignbit_b32 v22, v12, v11, 16
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v30
+; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v25
+; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
+; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
+; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
+; SI-NEXT:    v_alignbit_b32 v18, v16, v4, 16
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v13
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
-; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
+; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
+; SI-NEXT:    v_alignbit_b32 v5, v23, v5, 16
+; SI-NEXT:    v_alignbit_b32 v2, v21, v2, 16
+; SI-NEXT:    v_alignbit_b32 v1, v61, v1, 16
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
-; SI-NEXT:    v_mov_b32_e32 v55, v18
-; SI-NEXT:    v_alignbit_b32 v1, v18, v1, 16
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; SI-NEXT:    v_alignbit_b32 v23, v12, v11, 16
+; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
+; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v24, v10, v8, 16
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
-; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
-; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v15
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; SI-NEXT:    v_alignbit_b32 v7, v24, v7, 16
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
+; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
+; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v13
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
+; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
+; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v16
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
 ; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
 ; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v17
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v12
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
-; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_alignbit_b32 v24, v45, v11, 16
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v12
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
-; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_alignbit_b32 v53, v54, v11, 16
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
+; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
+; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v10
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
+; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
+; SI-NEXT:    v_alignbit_b32 v25, v45, v8, 16
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v9
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; SI-NEXT:    v_alignbit_b32 v8, v25, v8, 16
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
+; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
+; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v13
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
+; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
+; SI-NEXT:    v_alignbit_b32 v62, v63, v16, 16
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v33
+; SI-NEXT:    v_alignbit_b32 v16, v62, v16, 16
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
+; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
+; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v10
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
+; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
+; SI-NEXT:    v_alignbit_b32 v22, v34, v9, 16
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v11
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v12
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_alignbit_b32 v9, v22, v9, 16
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
 ; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
-; SI-NEXT:    v_alignbit_b32 v35, v36, v11, 16
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
-; SI-NEXT:    v_alignbit_b32 v2, v35, v2, 16
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v2, v53, v3, 16
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
-; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
-; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v12
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v11
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
 ; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
 ; SI-NEXT:    v_alignbit_b32 v37, v38, v11, 16
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v13
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v12
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
 ; SI-NEXT:    v_alignbit_b32 v11, v37, v11, 16
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
-; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
-; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v13
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
 ; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
 ; SI-NEXT:    v_alignbit_b32 v39, v48, v12, 16
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v14
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
+; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
+; SI-NEXT:    v_alignbit_b32 v35, v36, v10, 16
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v15
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
 ; SI-NEXT:    v_alignbit_b32 v12, v39, v12, 16
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_alignbit_b32 v10, v35, v10, 16
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
 ; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
 ; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v14
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
+; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
+; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v15
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
 ; SI-NEXT:    v_alignbit_b32 v49, v50, v13, 16
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v51
 ; SI-NEXT:    v_alignbit_b32 v13, v49, v13, 16
-; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
 ; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
 ; SI-NEXT:    v_alignbit_b32 v51, v52, v14, 16
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v31
-; SI-NEXT:    v_mov_b32_e32 v31, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; SI-NEXT:    v_alignbit_b32 v14, v51, v14, 16
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
-; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
-; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v16
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
 ; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
 ; SI-NEXT:    v_alignbit_b32 v41, v42, v15, 16
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v32
 ; SI-NEXT:    v_alignbit_b32 v15, v41, v15, 16
-; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
-; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
-; SI-NEXT:    v_alignbit_b32 v62, v63, v16, 16
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v33
-; SI-NEXT:    v_alignbit_b32 v16, v62, v16, 16
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v2, v24, v4, 16
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v2, v23, v7, 16
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v2, v22, v5, 16
+; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v2, v21, v6, 16
+; SI-NEXT:    v_alignbit_b32 v2, v20, v6, 16
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v2, v20, v8, 16
+; SI-NEXT:    v_alignbit_b32 v2, v19, v3, 16
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_alignbit_b32 v2, v19, v9, 16
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v2, v31, v10, 16
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v1, v18, v4, 16
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
 ; SI-NEXT:  .LBB104_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237137,8 +237239,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237153,8 +237255,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237169,8 +237271,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237185,8 +237287,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237201,8 +237303,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237217,8 +237319,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237233,8 +237335,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237243,14 +237345,14 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v53
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v54
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v22
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v34
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237259,7 +237361,7 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v45
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
@@ -237267,8 +237369,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237277,8 +237379,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237287,8 +237389,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237297,8 +237399,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237307,8 +237409,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237317,8 +237419,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237327,8 +237429,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237337,8 +237439,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237347,8 +237449,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237357,8 +237459,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237367,8 +237469,8 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237376,17 +237478,17 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v31
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -237394,10 +237496,12 @@ define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v55
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
@@ -242750,15 +242854,15 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:116
 ; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:124
 ; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:24
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:120
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:128
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:128
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
@@ -242777,14 +242881,14 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:56
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
 ; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    ; kill: killed $vgpr3
@@ -242857,17 +242961,14 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB106_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v46
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v47
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v56
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v56
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v57
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr1
 ; SI-NEXT:    ; kill: killed $vgpr1
@@ -242900,10 +243001,9 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr1
 ; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr47
 ; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v3
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -242951,111 +243051,118 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v41
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v47
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v43
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v41
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v42
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v57
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v43
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v58
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v44
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v44
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v45
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v59
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v46
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v60
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v58
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v61
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v59
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v45
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v61
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v63
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v62
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v40
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v63
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v62
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v40
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v60
+; SI-NEXT:    ; implicit-def: $vgpr47
 ; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:  .LBB106_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB106_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v62
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v60
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v55, v3
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v46
-; SI-NEXT:    v_add_i32_e32 v46, vcc, 0x30000, v3
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v40
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_or_b32_e32 v3, v55, v3
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_or_b32_e32 v9, v52, v1
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x30000, v3
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v40
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_or_b32_e32 v3, v51, v3
+; SI-NEXT:    v_add_i32_e32 v52, vcc, s6, v3
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v63
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v50, v3
-; SI-NEXT:    v_or_b32_e32 v1, v51, v1
 ; SI-NEXT:    v_add_i32_e32 v51, vcc, s6, v3
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v63
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v62
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v49, v3
 ; SI-NEXT:    v_add_i32_e32 v50, vcc, s6, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v45
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v48, v3
 ; SI-NEXT:    v_add_i32_e32 v49, vcc, s6, v3
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v61
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v59
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v39, v3
 ; SI-NEXT:    v_add_i32_e32 v48, vcc, s6, v3
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v60
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v58
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v38, v3
 ; SI-NEXT:    v_add_i32_e32 v39, vcc, s6, v3
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v59
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v46
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v37, v3
 ; SI-NEXT:    v_add_i32_e32 v38, vcc, s6, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v44
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v36, v3
 ; SI-NEXT:    v_add_i32_e32 v37, vcc, s6, v3
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v58
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v44
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_add_i32_e32 v36, vcc, s6, v3
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v57
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v43
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v34, v3
 ; SI-NEXT:    v_add_i32_e32 v35, vcc, s6, v3
@@ -243063,11 +243170,11 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v33, v3
 ; SI-NEXT:    v_add_i32_e32 v34, vcc, s6, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v43
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v41
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v32, v3
 ; SI-NEXT:    v_add_i32_e32 v33, vcc, s6, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v41
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v47
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
@@ -243078,24 +243185,20 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v11, v18, v11
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v11
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v56
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v54, v5
 ; SI-NEXT:    v_add_i32_e32 v55, vcc, s6, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v47
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v56
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v53, v7
 ; SI-NEXT:    v_add_i32_e32 v54, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT:    v_or_b32_e32 v9, v52, v9
 ; SI-NEXT:    v_add_i32_e32 v53, vcc, s6, v9
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v53
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
@@ -243107,8 +243210,8 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v11, v16, v11
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v31
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
+; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v31
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -243158,6 +243261,22 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v50
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v51
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v52
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v53
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v54
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v55
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
 ; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v32
 ; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v33
 ; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v34
@@ -243169,28 +243288,15 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v48
 ; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v49
 ; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v50
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    v_and_b32_e32 v50, 0xffff0000, v51
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v51
-; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v53
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v54
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v52
+; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v53
 ; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v54
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v55
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v1
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v55
-; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v46
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v46
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
@@ -243269,7 +243375,7 @@ define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; SI-NEXT:  .LBB106_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
-; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v2
 ; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v27
@@ -245096,531 +245202,533 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:136
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:4
-; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:28
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt expcnt(6)
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:48
 ; SI-NEXT:    s_waitcnt expcnt(5)
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:52
 ; SI-NEXT:    s_waitcnt expcnt(4)
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:56
 ; SI-NEXT:    s_waitcnt expcnt(3)
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt expcnt(2)
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:64
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:76
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:80
 ; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:84
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:104
-; SI-NEXT:    v_cvt_f16_f32_e32 v55, v7
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v17
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:88
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT:    v_cvt_f16_f32_e32 v34, v2
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v9
 ; SI-NEXT:    v_cvt_f16_f32_e32 v43, v3
-; SI-NEXT:    v_cvt_f16_f32_e32 v44, v4
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v18
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v2
-; SI-NEXT:    v_cvt_f16_f32_e32 v2, v5
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v6
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v5
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v2, v9
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v13
+; SI-NEXT:    v_cvt_f16_f32_e32 v55, v7
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v30
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v12
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v6
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v22
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f16_f32_e32 v40, v8
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v10
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v2, v13
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v16
-; SI-NEXT:    v_cvt_f16_f32_e32 v16, v19
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v30
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v21
+; SI-NEXT:    v_cvt_f16_f32_e32 v53, v12
+; SI-NEXT:    v_cvt_f16_f32_e32 v9, v14
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f16_f32_e32 v12, v18
+; SI-NEXT:    v_cvt_f16_f32_e32 v18, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v23
 ; SI-NEXT:    v_cvt_f16_f32_e32 v23, v25
-; SI-NEXT:    v_cvt_f16_f32_e32 v40, v8
-; SI-NEXT:    v_cvt_f16_f32_e32 v8, v10
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v11
-; SI-NEXT:    v_cvt_f16_f32_e32 v10, v14
-; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v44, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v52, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v15
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v20
-; SI-NEXT:    v_cvt_f16_f32_e32 v11, v21
+; SI-NEXT:    v_cvt_f16_f32_e32 v48, v16
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v17
+; SI-NEXT:    v_cvt_f16_f32_e32 v13, v20
 ; SI-NEXT:    v_cvt_f16_f32_e32 v20, v24
-; SI-NEXT:    v_cvt_f16_f32_e32 v24, v26
+; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
+; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
-; SI-NEXT:    v_cvt_f16_f32_e32 v25, v39
 ; SI-NEXT:    v_cvt_f16_f32_e32 v31, v27
-; SI-NEXT:    v_cvt_f16_f32_e32 v27, v28
-; SI-NEXT:    v_cvt_f16_f32_e32 v28, v29
-; SI-NEXT:    v_cvt_f16_f32_e32 v26, v51
-; SI-NEXT:    v_cvt_f16_f32_e32 v29, v42
-; SI-NEXT:    v_cvt_f16_f32_e32 v42, v47
-; SI-NEXT:    v_cvt_f16_f32_e32 v21, v38
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v50
-; SI-NEXT:    v_cvt_f16_f32_e32 v38, v54
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v56
-; SI-NEXT:    v_cvt_f16_f32_e32 v47, v41
+; SI-NEXT:    v_cvt_f16_f32_e32 v25, v50
+; SI-NEXT:    v_cvt_f16_f32_e32 v27, v29
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v42
+; SI-NEXT:    v_cvt_f16_f32_e32 v21, v47
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v38
+; SI-NEXT:    v_cvt_f16_f32_e32 v38, v49
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f16_f32_e32 v47, v54
+; SI-NEXT:    v_cvt_f16_f32_e32 v24, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v54, v46
+; SI-NEXT:    v_cvt_f16_f32_e32 v42, v56
+; SI-NEXT:    v_cvt_f16_f32_e32 v29, v51
+; SI-NEXT:    v_cvt_f16_f32_e32 v16, v41
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v59
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v45
-; SI-NEXT:    v_cvt_f16_f32_e32 v12, v59
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    v_cvt_f16_f32_e32 v18, v46
-; SI-NEXT:    v_cvt_f16_f32_e32 v46, v57
-; SI-NEXT:    v_cvt_f16_f32_e32 v41, v58
-; SI-NEXT:    v_cvt_f16_f32_e32 v50, v60
-; SI-NEXT:    v_cvt_f16_f32_e32 v30, v34
-; SI-NEXT:    v_cvt_f16_f32_e32 v51, v61
-; SI-NEXT:    v_cvt_f16_f32_e32 v34, v36
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:116
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:120
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cvt_f16_f32_e32 v39, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v57, v62
-; SI-NEXT:    v_cvt_f16_f32_e32 v13, v35
-; SI-NEXT:    v_cvt_f16_f32_e32 v58, v32
-; SI-NEXT:    v_cvt_f16_f32_e32 v61, v33
+; SI-NEXT:    v_cvt_f16_f32_e32 v11, v57
+; SI-NEXT:    v_cvt_f16_f32_e32 v58, v58
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v62
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_cvt_f16_f32_e32 v49, v32
+; SI-NEXT:    v_cvt_f16_f32_e32 v51, v60
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_cvt_f16_f32_e32 v50, v37
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:116
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:120
+; SI-NEXT:    v_cvt_f16_f32_e32 v61, v61
+; SI-NEXT:    v_cvt_f16_f32_e32 v35, v33
 ; SI-NEXT:    v_cvt_f16_f32_e32 v33, v63
-; SI-NEXT:    v_cvt_f16_f32_e32 v45, v48
-; SI-NEXT:    v_cvt_f16_f32_e32 v14, v49
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f16_f32_e32 v54, v53
-; SI-NEXT:    v_cvt_f16_f32_e32 v32, v52
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f16_f32_e32 v35, v7
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v9
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v59, v36
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:124
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:128
+; SI-NEXT:    v_cvt_f16_f32_e32 v36, v36
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_cvt_f16_f32_e32 v45, v39
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f16_f32_e32 v14, v6
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v7
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f16_f32_e32 v41, v8
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f16_f32_e32 v30, v10
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:124
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:128
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:132
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_cvt_f16_f32_e32 v7, v32
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f16_f32_e32 v8, v46
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f16_f32_e32 v46, v56
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f16_f32_e32 v56, v9
+; SI-NEXT:    v_cvt_f16_f32_e32 v32, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f16_f32_e32 v36, v36
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v9, v37
+; SI-NEXT:    v_cvt_f16_f32_e32 v10, v37
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_or_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT:    v_mov_b32_e32 v60, v29
-; SI-NEXT:    v_mov_b32_e32 v29, v28
-; SI-NEXT:    v_mov_b32_e32 v62, v23
-; SI-NEXT:    v_mov_b32_e32 v63, v11
+; SI-NEXT:    v_mov_b32_e32 v59, v29
+; SI-NEXT:    v_mov_b32_e32 v29, v27
+; SI-NEXT:    v_mov_b32_e32 v57, v23
+; SI-NEXT:    v_mov_b32_e32 v60, v3
+; SI-NEXT:    v_mov_b32_e32 v62, v4
+; SI-NEXT:    v_mov_b32_e32 v63, v49
+; SI-NEXT:    v_mov_b32_e32 v49, v12
 ; SI-NEXT:    s_xor_b64 exec, exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB108_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.true
-; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
-; SI-NEXT:    v_cvt_f32_f16_e32 v36, v36
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    v_cvt_f32_f16_e32 v35, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    v_cvt_f32_f16_e32 v34, v34
-; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
+; SI-NEXT:    v_cvt_f32_f16_e32 v36, v36
+; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f32_f16_e32 v33, v33
-; SI-NEXT:    v_add_f32_e32 v36, 0x38000000, v36
-; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
+; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
+; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
+; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
-; SI-NEXT:    v_cvt_f16_f32_e32 v36, v36
+; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
 ; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
-; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
+; SI-NEXT:    v_add_f32_e32 v36, 0x38000000, v36
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
-; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
-; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
+; SI-NEXT:    v_cvt_f16_f32_e32 v36, v36
 ; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
-; SI-NEXT:    v_or_b32_e32 v36, v36, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
-; SI-NEXT:    v_or_b32_e32 v35, v35, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v14
-; SI-NEXT:    v_or_b32_e32 v13, v13, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v34
+; SI-NEXT:    v_or_b32_e32 v6, v6, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v8
+; SI-NEXT:    v_or_b32_e32 v7, v7, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
+; SI-NEXT:    v_or_b32_e32 v14, v14, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v36
 ; SI-NEXT:    v_or_b32_e32 v33, v33, v37
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v58
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v57
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_cvt_f32_f16_e32 v11, v41
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v58, v37
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v48
-; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
-; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v58
-; SI-NEXT:    v_or_b32_e32 v57, v37, v48
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v12
-; SI-NEXT:    v_or_b32_e32 v41, v11, v37
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v58
+; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
 ; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
+; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
+; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
+; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
+; SI-NEXT:    v_cvt_f16_f32_e32 v58, v12
+; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
+; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
 ; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
-; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
-; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
+; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
+; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
+; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
-; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
+; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
-; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
-; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
-; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
-; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
-; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
+; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
+; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
+; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
+; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
+; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
 ; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
 ; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v18
 ; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
 ; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
-; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
-; SI-NEXT:    v_or_b32_e32 v17, v17, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v22
+; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
-; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
-; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
-; SI-NEXT:    v_or_b32_e32 v21, v21, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v26
-; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
+; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
+; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v39, v52
+; SI-NEXT:    v_cvt_f32_f16_e32 v34, v34
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
 ; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
-; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
-; SI-NEXT:    v_or_b32_e32 v25, v25, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v27
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v3
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v61
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
+; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v4
+; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT:    v_or_b32_e32 v61, v3, v37
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v58
+; SI-NEXT:    v_or_b32_e32 v11, v11, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v17
+; SI-NEXT:    v_or_b32_e32 v16, v16, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v22
+; SI-NEXT:    v_or_b32_e32 v21, v21, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v25
+; SI-NEXT:    v_or_b32_e32 v24, v24, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v28
 ; SI-NEXT:    v_or_b32_e32 v31, v31, v37
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v20
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
-; SI-NEXT:    v_or_b32_e32 v16, v16, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v13
+; SI-NEXT:    v_or_b32_e32 v18, v18, v37
+; SI-NEXT:    v_cvt_f32_f16_e32 v37, v48
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
+; SI-NEXT:    v_cvt_f16_f32_e32 v48, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v48
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v6
-; SI-NEXT:    v_or_b32_e32 v5, v5, v37
+; SI-NEXT:    v_cvt_f32_f16_e32 v37, v53
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
+; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
+; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
+; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
+; SI-NEXT:    v_cvt_f16_f32_e32 v53, v37
+; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v39
+; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v53
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT:    v_or_b32_e32 v52, v37, v39
 ; SI-NEXT:    v_cvt_f32_f16_e32 v37, v40
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v55
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v39, v55
+; SI-NEXT:    v_cvt_f32_f16_e32 v56, v62
+; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
 ; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
 ; SI-NEXT:    v_cvt_f16_f32_e32 v40, v37
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v48
+; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v39
 ; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v40
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT:    v_or_b32_e32 v55, v37, v48
+; SI-NEXT:    v_add_f32_e32 v56, 0x38000000, v56
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v40
+; SI-NEXT:    v_cvt_f16_f32_e32 v56, v56
+; SI-NEXT:    v_or_b32_e32 v55, v37, v39
 ; SI-NEXT:    v_cvt_f32_f16_e32 v37, v44
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v43
-; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v39, v43
+; SI-NEXT:    v_cvt_f32_f16_e32 v57, v57
+; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
 ; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
 ; SI-NEXT:    v_cvt_f16_f32_e32 v44, v37
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v48
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v44
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v43, v37, v48
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v11
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
+; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v39
 ; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
-; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
-; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
-; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_cvt_f32_f16_e32 v52, v62
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
-; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
-; SI-NEXT:    v_cvt_f32_f16_e32 v28, v42
+; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v44
+; SI-NEXT:    v_cvt_f32_f16_e32 v51, v51
+; SI-NEXT:    v_or_b32_e32 v43, v37, v39
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
+; SI-NEXT:    v_cvt_f32_f16_e32 v47, v47
+; SI-NEXT:    v_cvt_f32_f16_e32 v42, v42
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v50, v50
 ; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
-; SI-NEXT:    v_cvt_f32_f16_e32 v39, v39
-; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
-; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
-; SI-NEXT:    v_cvt_f32_f16_e32 v42, v45
-; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
-; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
-; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; SI-NEXT:    v_add_f32_e32 v39, 0x38000000, v39
-; SI-NEXT:    v_cvt_f16_f32_e32 v39, v39
-; SI-NEXT:    v_add_f32_e32 v42, 0x38000000, v42
-; SI-NEXT:    v_cvt_f16_f32_e32 v42, v42
-; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
 ; SI-NEXT:    v_cvt_f32_f16_e32 v32, v32
-; SI-NEXT:    v_cvt_f32_f16_e32 v51, v51
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
 ; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
 ; SI-NEXT:    v_cvt_f32_f16_e32 v38, v38
-; SI-NEXT:    v_cvt_f32_f16_e32 v50, v50
-; SI-NEXT:    v_or_b32_e32 v45, v42, v39
-; SI-NEXT:    v_cvt_f32_f16_e32 v42, v59
-; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
+; SI-NEXT:    v_cvt_f32_f16_e32 v54, v54
 ; SI-NEXT:    v_add_f32_e32 v51, 0x38000000, v51
-; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
-; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
-; SI-NEXT:    v_add_f32_e32 v38, 0x38000000, v38
+; SI-NEXT:    v_cvt_f32_f16_e32 v45, v45
+; SI-NEXT:    v_cvt_f32_f16_e32 v41, v41
+; SI-NEXT:    v_cvt_f32_f16_e32 v46, v46
 ; SI-NEXT:    v_cvt_f16_f32_e32 v51, v51
+; SI-NEXT:    v_add_f32_e32 v47, 0x38000000, v47
+; SI-NEXT:    v_add_f32_e32 v42, 0x38000000, v42
+; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v37, v3
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v50, 0x38000000, v50
+; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
+; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
+; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
+; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
+; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
+; SI-NEXT:    v_cvt_f16_f32_e32 v47, v47
+; SI-NEXT:    v_add_f32_e32 v38, 0x38000000, v38
+; SI-NEXT:    v_cvt_f16_f32_e32 v42, v42
+; SI-NEXT:    v_add_f32_e32 v54, 0x38000000, v54
+; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v50, v50
+; SI-NEXT:    v_add_f32_e32 v45, 0x38000000, v45
+; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
+; SI-NEXT:    v_add_f32_e32 v41, 0x38000000, v41
+; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
+; SI-NEXT:    v_add_f32_e32 v46, 0x38000000, v46
 ; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
 ; SI-NEXT:    v_cvt_f16_f32_e32 v38, v38
-; SI-NEXT:    v_cvt_f16_f32_e32 v50, v50
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
+; SI-NEXT:    v_cvt_f16_f32_e32 v54, v54
 ; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
-; SI-NEXT:    v_alignbit_b32 v39, v13, v39, 16
-; SI-NEXT:    v_or_b32_e32 v50, v50, v51
-; SI-NEXT:    v_alignbit_b32 v51, v57, v51, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v11
-; SI-NEXT:    v_or_b32_e32 v11, v37, v1
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    v_alignbit_b32 v1, v43, v1, 16
-; SI-NEXT:    v_add_f32_e32 v48, 0x38000000, v48
-; SI-NEXT:    v_cvt_f16_f32_e32 v48, v48
+; SI-NEXT:    v_cvt_f16_f32_e32 v45, v45
+; SI-NEXT:    v_cvt_f16_f32_e32 v41, v41
+; SI-NEXT:    v_cvt_f16_f32_e32 v46, v46
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v47
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v42
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
+; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
+; SI-NEXT:    v_or_b32_e32 v38, v38, v47
+; SI-NEXT:    v_or_b32_e32 v54, v54, v42
+; SI-NEXT:    v_or_b32_e32 v45, v45, v50
+; SI-NEXT:    v_or_b32_e32 v41, v41, v30
+; SI-NEXT:    v_or_b32_e32 v46, v46, v32
+; SI-NEXT:    v_alignbit_b32 v47, v16, v47, 16
+; SI-NEXT:    v_alignbit_b32 v42, v11, v42, 16
+; SI-NEXT:    v_alignbit_b32 v50, v14, v50, 16
+; SI-NEXT:    v_alignbit_b32 v30, v7, v30, 16
+; SI-NEXT:    v_alignbit_b32 v32, v6, v32, 16
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v39, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v34
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v34, v43, v34, 16
+; SI-NEXT:    v_add_f32_e32 v39, 0x38000000, v39
+; SI-NEXT:    v_cvt_f16_f32_e32 v39, v39
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_or_b32_e32 v11, v48, v4
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v3, v39, v1
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    v_alignbit_b32 v4, v55, v4, 16
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    v_alignbit_b32 v1, v55, v1, 16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v11
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v37, v3
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
 ; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v11
-; SI-NEXT:    v_or_b32_e32 v11, v37, v8
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v48, 0x38000000, v48
-; SI-NEXT:    v_cvt_f16_f32_e32 v48, v48
-; SI-NEXT:    v_alignbit_b32 v8, v5, v8, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v11
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v39, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v5
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v37, v49
+; SI-NEXT:    v_add_f32_e32 v39, 0x38000000, v39
+; SI-NEXT:    v_cvt_f16_f32_e32 v39, v39
+; SI-NEXT:    v_alignbit_b32 v5, v52, v5, 16
 ; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
 ; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v49, v11
-; SI-NEXT:    v_or_b32_e32 v11, v48, v10
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v49
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
-; SI-NEXT:    v_alignbit_b32 v10, v2, v10, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v11
-; SI-NEXT:    v_or_b32_e32 v11, v49, v37
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    v_or_b32_e32 v3, v39, v9
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v49, v63
-; SI-NEXT:    v_add_f32_e32 v48, 0x38000000, v48
-; SI-NEXT:    v_cvt_f16_f32_e32 v48, v48
-; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v49
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; SI-NEXT:    v_or_b32_e32 v63, v49, v48
-; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v52
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
-; SI-NEXT:    v_cvt_f32_f16_e32 v52, v47
-; SI-NEXT:    v_or_b32_e32 v62, v49, v24
-; SI-NEXT:    v_cvt_f32_f16_e32 v49, v60
-; SI-NEXT:    v_add_f32_e32 v52, 0x38000000, v52
-; SI-NEXT:    v_cvt_f16_f32_e32 v52, v52
-; SI-NEXT:    v_alignbit_b32 v24, v31, v24, 16
-; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v49
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
-; SI-NEXT:    v_or_b32_e32 v60, v49, v28
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v52
-; SI-NEXT:    v_cvt_f32_f16_e32 v52, v46
-; SI-NEXT:    v_or_b32_e32 v38, v38, v49
-; SI-NEXT:    v_add_f32_e32 v52, 0x38000000, v52
-; SI-NEXT:    v_cvt_f16_f32_e32 v52, v52
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
-; SI-NEXT:    v_alignbit_b32 v46, v41, v52, 16
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
+; SI-NEXT:    v_or_b32_e32 v62, v56, v37
+; SI-NEXT:    v_cvt_f32_f16_e32 v56, v60
+; SI-NEXT:    v_alignbit_b32 v9, v2, v9, 16
+; SI-NEXT:    v_add_f32_e32 v56, 0x38000000, v56
+; SI-NEXT:    v_cvt_f16_f32_e32 v56, v56
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v23, v11
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v39, v3
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v39, 0x38000000, v39
+; SI-NEXT:    v_cvt_f16_f32_e32 v39, v39
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
+; SI-NEXT:    v_or_b32_e32 v60, v56, v39
+; SI-NEXT:    v_add_f32_e32 v56, 0x38000000, v57
+; SI-NEXT:    v_cvt_f16_f32_e32 v56, v56
+; SI-NEXT:    v_or_b32_e32 v57, v56, v26
+; SI-NEXT:    v_cvt_f32_f16_e32 v56, v59
+; SI-NEXT:    v_alignbit_b32 v26, v31, v26, 16
+; SI-NEXT:    v_add_f32_e32 v56, 0x38000000, v56
+; SI-NEXT:    v_cvt_f16_f32_e32 v56, v56
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v23, v3
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
 ; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
 ; SI-NEXT:    v_or_b32_e32 v29, v29, v23
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v53, v11
-; SI-NEXT:    v_add_f32_e32 v53, 0x38000000, v53
-; SI-NEXT:    v_cvt_f16_f32_e32 v53, v53
-; SI-NEXT:    v_or_b32_e32 v11, v53, v52
-; SI-NEXT:    v_cvt_f32_f16_e32 v53, v61
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    v_cvt_f32_f16_e32 v27, v3
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
+; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
+; SI-NEXT:    v_or_b32_e32 v59, v56, v27
+; SI-NEXT:    v_cvt_f32_f16_e32 v56, v63
+; SI-NEXT:    v_add_f32_e32 v56, 0x38000000, v56
+; SI-NEXT:    v_cvt_f16_f32_e32 v56, v56
+; SI-NEXT:    v_or_b32_e32 v63, v56, v35
+; SI-NEXT:    v_alignbit_b32 v35, v33, v35, 16
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v49, v3
+; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v49
+; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
+; SI-NEXT:    v_or_b32_e32 v3, v49, v51
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v11, v16, v37, 16
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    v_add_f32_e32 v53, 0x38000000, v53
-; SI-NEXT:    v_cvt_f16_f32_e32 v53, v53
+; SI-NEXT:    v_alignbit_b32 v3, v19, v39, 16
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v11, v19, v48, 16
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    v_alignbit_b32 v3, v24, v23, 16
+; SI-NEXT:    v_alignbit_b32 v49, v18, v37, 16
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v11, v25, v23, 16
-; SI-NEXT:    v_or_b32_e32 v61, v53, v30
-; SI-NEXT:    v_cvt_f32_f16_e32 v53, v54
-; SI-NEXT:    v_cvt_f32_f16_e32 v54, v56
-; SI-NEXT:    v_alignbit_b32 v30, v33, v30, 16
-; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    v_add_f32_e32 v53, 0x38000000, v53
-; SI-NEXT:    v_add_f32_e32 v54, 0x38000000, v54
-; SI-NEXT:    v_cvt_f16_f32_e32 v53, v53
-; SI-NEXT:    v_cvt_f16_f32_e32 v47, v54
-; SI-NEXT:    v_add_f32_e32 v54, 0x38000000, v42
-; SI-NEXT:    v_cvt_f16_f32_e32 v42, v54
-; SI-NEXT:    v_or_b32_e32 v54, v53, v32
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v47
-; SI-NEXT:    v_alignbit_b32 v47, v17, v49, 16
-; SI-NEXT:    v_or_b32_e32 v59, v42, v53
-; SI-NEXT:    v_alignbit_b32 v42, v21, v28, 16
-; SI-NEXT:    v_alignbit_b32 v32, v35, v32, 16
-; SI-NEXT:    v_alignbit_b32 v56, v36, v53, 16
+; SI-NEXT:    v_alignbit_b32 v3, v21, v27, 16
+; SI-NEXT:    v_alignbit_b32 v51, v61, v51, 16
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:  .LBB108_2: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v37, 0xffff, v11
-; SI-NEXT:    v_or_b32_e32 v1, v37, v1
-; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v37, 0xffff, v3
+; SI-NEXT:    v_or_b32_e32 v34, v37, v34
+; SI-NEXT:    buffer_store_dword v34, v0, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v43
+; SI-NEXT:    v_and_b32_e32 v34, 0xffff, v43
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v44
-; SI-NEXT:    v_or_b32_e32 v1, v1, v37
+; SI-NEXT:    v_or_b32_e32 v34, v34, v37
 ; SI-NEXT:    v_add_i32_e32 v37, vcc, 4, v0
-; SI-NEXT:    buffer_store_dword v1, v37, s[0:3], 0 offen
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v1, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 8, v0
-; SI-NEXT:    buffer_store_dword v1, v4, s[0:3], 0 offen
+; SI-NEXT:    buffer_store_dword v34, v37, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v34, 0xffff, v3
+; SI-NEXT:    v_or_b32_e32 v1, v34, v1
+; SI-NEXT:    v_add_i32_e32 v34, vcc, 8, v0
+; SI-NEXT:    buffer_store_dword v1, v34, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v55
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v40
-; SI-NEXT:    v_or_b32_e32 v1, v1, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 12, v0
-; SI-NEXT:    buffer_store_dword v1, v4, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v40
+; SI-NEXT:    v_or_b32_e32 v1, v1, v34
+; SI-NEXT:    v_add_i32_e32 v34, vcc, 12, v0
+; SI-NEXT:    buffer_store_dword v1, v34, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v8
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v1, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 16, v0
-; SI-NEXT:    buffer_store_dword v1, v4, s[0:3], 0 offen
+; SI-NEXT:    v_or_b32_e32 v1, v1, v5
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 16, v0
+; SI-NEXT:    buffer_store_dword v1, v5, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v6
-; SI-NEXT:    v_or_b32_e32 v1, v1, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 20, v0
-; SI-NEXT:    buffer_store_dword v1, v4, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v52
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v53
+; SI-NEXT:    v_or_b32_e32 v1, v1, v5
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 20, v0
+; SI-NEXT:    buffer_store_dword v1, v5, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v10
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v9
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v1, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 24, v0
-; SI-NEXT:    buffer_store_dword v1, v4, s[0:3], 0 offen
+; SI-NEXT:    v_or_b32_e32 v1, v1, v5
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 24, v0
+; SI-NEXT:    buffer_store_dword v1, v5, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v48
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v62
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v49
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v15
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v63
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v60
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
@@ -245633,14 +245741,14 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v62
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v57
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v26
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v31
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v28
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
@@ -245653,14 +245761,16 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v26
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v25
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v60
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v42
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v59
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
@@ -245677,82 +245787,84 @@ define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v18
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v17
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v46
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v54
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v42
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v41
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v12
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v58
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v50
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v51
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v57
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v58
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v30
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v63
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v35
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v33
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v36
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v45
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v39
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v50
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v14
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v15
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v54
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v32
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v41
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v30
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v7
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v8
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v59
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v56
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v46
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v32
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v9
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v10
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
 ; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
index d966d136d75b6..f0fa7e95c75ef 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
@@ -4568,14 +4568,14 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v33, v4
 ; SI-NEXT:    v_mov_b32_e32 v32, v2
 ; SI-NEXT:    v_mov_b32_e32 v31, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
 ; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:4
@@ -4595,18 +4595,18 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v23, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v21, 8, v29
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v15, 24, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v0
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v2
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v4
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v11, 24, v4
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v11, 24, v8
-; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 8, v6
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 8, v44
+; SI-NEXT:    v_lshlrev_b32_e32 v15, 24, v44
+; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_2
@@ -4862,13 +4862,13 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_mov_b32_e32 v33, v4
 ; VI-NEXT:    v_mov_b32_e32 v32, v2
 ; VI-NEXT:    v_mov_b32_e32 v31, v0
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:28
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:16
 ; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:20
 ; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:12
 ; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:4
@@ -4888,18 +4888,18 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v27
 ; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v29
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v2
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v4
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v8
-; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
+; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v6
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v44
+; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB14_2
@@ -5074,13 +5074,13 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_mov_b32_e32 v33, v4
 ; GFX9-NEXT:    v_mov_b32_e32 v32, v2
 ; GFX9-NEXT:    v_mov_b32_e32 v31, v0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:16
 ; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:20
 ; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:12
 ; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:4
@@ -5100,18 +5100,18 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v2
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v8
-; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
+; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v6
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v44
+; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB14_2
@@ -12053,14 +12053,14 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v33, v4
 ; SI-NEXT:    v_mov_b32_e32 v32, v2
 ; SI-NEXT:    v_mov_b32_e32 v31, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
 ; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:4
@@ -12080,18 +12080,18 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v23, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v21, 8, v29
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v15, 24, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v0
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v2
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v4
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v11, 24, v4
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v11, 24, v8
-; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 8, v6
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 8, v44
+; SI-NEXT:    v_lshlrev_b32_e32 v15, 24, v44
+; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB34_2
@@ -12347,13 +12347,13 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_mov_b32_e32 v33, v4
 ; VI-NEXT:    v_mov_b32_e32 v32, v2
 ; VI-NEXT:    v_mov_b32_e32 v31, v0
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:28
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:16
 ; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:20
 ; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:12
 ; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:4
@@ -12373,18 +12373,18 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v27
 ; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v29
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v2
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v4
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v8
-; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
+; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v6
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v44
+; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB34_2
@@ -12559,13 +12559,13 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_mov_b32_e32 v33, v4
 ; GFX9-NEXT:    v_mov_b32_e32 v32, v2
 ; GFX9-NEXT:    v_mov_b32_e32 v31, v0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:16
 ; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:20
 ; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:12
 ; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:4
@@ -12585,18 +12585,18 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v2
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v8
-; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
+; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v6
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v44
+; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB34_2
@@ -19423,13 +19423,13 @@ define <20 x i16> @bitcast_v40i8_to_v20i16(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_mov_b32_e32 v38, v4
 ; VI-NEXT:    v_mov_b32_e32 v32, v2
 ; VI-NEXT:    v_mov_b32_e32 v36, v0
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:28
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:16
 ; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:20
 ; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:12
 ; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:4
@@ -19451,17 +19451,17 @@ define <20 x i16> @bitcast_v40i8_to_v20i16(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
 ; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v44, 8, v2
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v4
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v8
+; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v6
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v44, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
+; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -19646,13 +19646,13 @@ define <20 x i16> @bitcast_v40i8_to_v20i16(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_mov_b32_e32 v35, v4
 ; GFX9-NEXT:    v_mov_b32_e32 v33, v2
 ; GFX9-NEXT:    v_mov_b32_e32 v36, v0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:16
 ; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:20
 ; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:12
 ; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:4
@@ -19674,17 +19674,17 @@ define <20 x i16> @bitcast_v40i8_to_v20i16(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v2
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
+; GFX9-NEXT:    v_lshlrev_b16_e32 v47, 8, v6
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v47, 8, v10
+; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v10
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -25685,14 +25685,14 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v36, v4
 ; SI-NEXT:    v_mov_b32_e32 v31, v2
 ; SI-NEXT:    v_mov_b32_e32 v35, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:28
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:4
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 8, v1
@@ -25700,7 +25700,7 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v39, 8, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v49, 8, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v55, 8, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v17
@@ -25726,23 +25726,23 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr15
 ; SI-NEXT:    ; implicit-def: $vgpr17
 ; SI-NEXT:    ; implicit-def: $vgpr19
-; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v0
+; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v0
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v2
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v4
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
-; SI-NEXT:    s_waitcnt vmcnt(5) expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v33
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; SI-NEXT:    s_waitcnt vmcnt(7) expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v4
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    ; implicit-def: $vgpr4
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v32
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v33
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v34
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v34
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -25786,7 +25786,7 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v51
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v57
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v6
-; SI-NEXT:    v_and_b32_e32 v6, 0xff, v52
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v53
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v58
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v35
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v31
@@ -25794,13 +25794,13 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v8
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v10
 ; SI-NEXT:    v_cvt_f32_f16_e32 v34, v6
-; SI-NEXT:    v_and_b32_e32 v6, 0xff, v54
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v52
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v37
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v38
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v39
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v48
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v49
-; SI-NEXT:    v_or_b32_e32 v5, v5, v53
+; SI-NEXT:    v_or_b32_e32 v5, v5, v54
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v59
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
@@ -25827,14 +25827,14 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr53
 ; SI-NEXT:    ; implicit-def: $vgpr52
-; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr41
@@ -25853,12 +25853,12 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB62_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v54
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v52
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v59, v0
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 0x300, v0
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v52
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v53
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    s_movk_i32 s6, 0x300
 ; SI-NEXT:    v_or_b32_e32 v0, v58, v0
@@ -25915,7 +25915,7 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v0
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v10
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_or_b32_e32 v0, v53, v0
+; SI-NEXT:    v_or_b32_e32 v0, v54, v0
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v0
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v8
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
@@ -25998,13 +25998,13 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_mov_b32_e32 v38, v4
 ; VI-NEXT:    v_mov_b32_e32 v32, v2
 ; VI-NEXT:    v_mov_b32_e32 v36, v0
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:28
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:16
 ; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:20
 ; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:12
 ; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:4
@@ -26026,17 +26026,17 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
 ; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v44, 8, v2
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v4
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v8
+; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v6
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v44, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
+; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -26221,13 +26221,13 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_mov_b32_e32 v35, v4
 ; GFX9-NEXT:    v_mov_b32_e32 v33, v2
 ; GFX9-NEXT:    v_mov_b32_e32 v36, v0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:16
 ; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:20
 ; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:12
 ; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:4
@@ -26249,17 +26249,17 @@ define <20 x half> @bitcast_v40i8_to_v20f16(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v2
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
+; GFX9-NEXT:    v_lshlrev_b16_e32 v47, 8, v6
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v47, 8, v10
+; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v10
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -30050,13 +30050,13 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v33, v4
 ; SI-NEXT:    v_mov_b32_e32 v32, v2
 ; SI-NEXT:    v_mov_b32_e32 v31, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:4
@@ -30079,17 +30079,17 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v27, 8, v29
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v21, 24, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 24, v0
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v23, 8, v2
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 24, v4
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v4
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v6
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v23, 8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v21, 24, v10
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -30355,13 +30355,13 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_mov_b32_e32 v33, v4
 ; VI-NEXT:    v_mov_b32_e32 v32, v2
 ; VI-NEXT:    v_mov_b32_e32 v31, v0
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:28
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:16
 ; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:20
 ; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:12
 ; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:4
@@ -30383,17 +30383,17 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v27
 ; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v29
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v2
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v4
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v8
+; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v6
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v10
+; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v10
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -30578,13 +30578,13 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_mov_b32_e32 v33, v4
 ; GFX9-NEXT:    v_mov_b32_e32 v32, v2
 ; GFX9-NEXT:    v_mov_b32_e32 v31, v0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:16
 ; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:20
 ; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:12
 ; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:4
@@ -30606,17 +30606,17 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v2
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v8
+; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v6
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v10
+; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v10
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -34694,13 +34694,13 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v33, v4
 ; SI-NEXT:    v_mov_b32_e32 v32, v2
 ; SI-NEXT:    v_mov_b32_e32 v31, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:4
@@ -34723,17 +34723,17 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v27, 8, v29
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v21, 24, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 24, v0
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v23, 8, v2
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 24, v4
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v4
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v6
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v23, 8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v21, 24, v10
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -34999,13 +34999,13 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_mov_b32_e32 v33, v4
 ; VI-NEXT:    v_mov_b32_e32 v32, v2
 ; VI-NEXT:    v_mov_b32_e32 v31, v0
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:28
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:16
 ; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:20
 ; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:12
 ; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:4
@@ -35027,17 +35027,17 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v27
 ; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v29
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v2
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v4
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v8
+; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v6
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v10
+; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v10
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -35222,13 +35222,13 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_mov_b32_e32 v33, v4
 ; GFX9-NEXT:    v_mov_b32_e32 v32, v2
 ; GFX9-NEXT:    v_mov_b32_e32 v31, v0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:32
 ; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:16
 ; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:20
 ; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:12
 ; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:4
@@ -35250,17 +35250,17 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v27
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v29
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v2
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v8
+; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v6
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v10
+; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v10
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
index 397955a8a8928..a14008d65743e 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
@@ -11396,170 +11396,171 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v34, v6
 ; SI-NEXT:    v_mov_b32_e32 v33, v4
 ; SI-NEXT:    v_mov_b32_e32 v32, v2
 ; SI-NEXT:    v_mov_b32_e32 v31, v0
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:132
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:120
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:128
 ; SI-NEXT:    v_lshlrev_b32_e32 v20, 24, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 8, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 24, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v28, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v30, 8, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v49, 8, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 24, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 24, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 24, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 8, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v18, 8, v17
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v53
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 24, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v6
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 24, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v10
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 24, v12
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v39
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 24, v6
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v50
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v52
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v10
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v54
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v12
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v41
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 24, v14
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v45
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v50
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v23, 24, v56
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 24, v51
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v21, 8, v58
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 8, v52
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 24, v61
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v21, 24, v53
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v55
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v40
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:68
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v42
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 24, v58
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v47
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 8, v13
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB26_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v31
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v33
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v34
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v49
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v2, v2, v36
+; SI-NEXT:    v_or_b32_e32 v2, v2, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v1, v35, v1
-; SI-NEXT:    v_or_b32_e32 v3, v37, v3
+; SI-NEXT:    v_or_b32_e32 v1, v48, v1
+; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v3
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v55
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v57
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v60, v9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v39
+; SI-NEXT:    v_or_b32_e32 v9, v62, v9
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v42
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    v_or_b32_e32 v10, v56, v10
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v52
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    v_or_b32_e32 v11, v46, v11
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v56
+; SI-NEXT:    v_or_b32_e32 v11, v43, v11
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v58
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v45
+; SI-NEXT:    v_or_b32_e32 v12, v54, v12
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v44
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v13, v53, v13
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v54
+; SI-NEXT:    v_or_b32_e32 v13, v21, v13
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v53
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v27, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_or_b32_e32 v15, v23, v15
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr29
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr21
 ; SI-NEXT:    ; implicit-def: $vgpr27
-; SI-NEXT:    ; implicit-def: $vgpr23
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -11567,18 +11568,18 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v2, v2, v48
+; SI-NEXT:    v_or_b32_e32 v2, v2, v36
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v3, v38, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v18, v4
+; SI-NEXT:    v_or_b32_e32 v4, v16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
@@ -11587,7 +11588,6 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v24, v6
-; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -11595,30 +11595,30 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v28, v7
 ; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v3, v16
+; SI-NEXT:    v_or_b32_e32 v3, v3, v38
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_or_b32_e32 v8, v17, v8
+; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v4, v17
+; SI-NEXT:    v_or_b32_e32 v4, v4, v18
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr17
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    v_or_b32_e32 v8, v63, v8
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v22
@@ -11638,54 +11638,59 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v30
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v43
-; SI-NEXT:    v_or_b32_e32 v8, v8, v62
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v60
+; SI-NEXT:    v_or_b32_e32 v8, v8, v63
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v9
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v51
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v45
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v59
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    ; implicit-def: $vgpr62
-; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT:    v_or_b32_e32 v10, v10, v47
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v55
+; SI-NEXT:    v_or_b32_e32 v10, v10, v46
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v58
-; SI-NEXT:    v_or_b32_e32 v11, v11, v44
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    v_or_b32_e32 v11, v11, v41
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_or_b32_e32 v12, v12, v40
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v47
+; SI-NEXT:    v_or_b32_e32 v12, v12, v51
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v41
-; SI-NEXT:    v_or_b32_e32 v13, v13, v50
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v40
+; SI-NEXT:    v_or_b32_e32 v13, v13, v19
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v23
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v25
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v19
-; SI-NEXT:    v_or_b32_e32 v15, v15, v21
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; implicit-def: $vgpr63
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr19
+; SI-NEXT:    ; implicit-def: $vgpr25
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_or_b32_e32 v15, v15, v16
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT:    v_or_b32_e32 v16, v50, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
 ; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; kill: killed $vgpr16
@@ -11712,21 +11717,20 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; kill: killed $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr19
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr25
-; SI-NEXT:    ; implicit-def: $vgpr21
+; SI-NEXT:    ; kill: killed $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:  .LBB26_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB26_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v31
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v32
@@ -11734,55 +11738,54 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
-; SI-NEXT:    v_or_b32_e32 v1, v35, v1
+; SI-NEXT:    v_or_b32_e32 v1, v48, v1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v33
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v34
-; SI-NEXT:    v_or_b32_e32 v1, v36, v1
+; SI-NEXT:    v_or_b32_e32 v1, v39, v1
 ; SI-NEXT:    s_movk_i32 s6, 0x300
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v37, v2
+; SI-NEXT:    v_or_b32_e32 v2, v35, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v55
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v60, v9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v39
+; SI-NEXT:    v_or_b32_e32 v9, v62, v9
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v42
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
+; SI-NEXT:    v_or_b32_e32 v10, v56, v10
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v52
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    v_or_b32_e32 v11, v46, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v56
+; SI-NEXT:    v_or_b32_e32 v11, v43, v11
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v58
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v45
+; SI-NEXT:    v_or_b32_e32 v12, v54, v12
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v44
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v13, v53, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v54
+; SI-NEXT:    v_or_b32_e32 v13, v21, v13
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v53
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v27, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v29
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_or_b32_e32 v15, v23, v15
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -11791,20 +11794,20 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_or_b32_e32 v2, v48, v2
+; SI-NEXT:    v_or_b32_e32 v2, v36, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
-; SI-NEXT:    v_or_b32_e32 v3, v38, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v18, v4
+; SI-NEXT:    v_or_b32_e32 v4, v16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
@@ -11820,35 +11823,34 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v28, v7
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v38, v3
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v4, v3
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v3
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_or_b32_e32 v8, v17, v8
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v17, v4
+; SI-NEXT:    v_or_b32_e32 v4, v18, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v4
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    v_or_b32_e32 v8, v63, v8
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
@@ -11874,68 +11876,70 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v43
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v60
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_or_b32_e32 v8, v62, v8
+; SI-NEXT:    v_or_b32_e32 v8, v63, v8
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v51
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v59, v9
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v55
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT:    v_or_b32_e32 v10, v47, v10
+; SI-NEXT:    v_or_b32_e32 v10, v46, v10
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v58
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v44, v11
+; SI-NEXT:    v_or_b32_e32 v11, v41, v11
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v47
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_or_b32_e32 v12, v40, v12
+; SI-NEXT:    v_or_b32_e32 v12, v51, v12
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v41
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v40
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT:    v_or_b32_e32 v13, v50, v13
+; SI-NEXT:    v_or_b32_e32 v13, v19, v13
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v23
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v25, v14
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v19
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
+; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
+; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
+; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
+; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
+; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; SI-NEXT:    v_or_b32_e32 v15, v21, v15
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x300, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT:    v_or_b32_e32 v16, v50, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v15
 ; SI-NEXT:  .LBB26_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
@@ -11955,7 +11959,7 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: bitcast_v64i8_to_v16i32:
@@ -11993,38 +11997,31 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:56
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
+; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:72
+; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:80
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:88
+; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:96
+; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:104
+; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:112
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:120
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:128
+; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v21
-; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v23
+; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v23
 ; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v25
-; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v27
-; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v29
+; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v27
+; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v29
 ; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v17
-; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
-; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:124
 ; VI-NEXT:    v_lshlrev_b16_e32 v31, 8, v1
 ; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v3
@@ -12035,38 +12032,52 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
 ; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v15
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v54
-; VI-NEXT:    v_lshlrev_b16_e32 v61, 8, v0
-; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v2
-; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v4
-; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
-; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
-; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v12
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v14
-; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v24
-; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v30
-; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v53
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v4
+; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v42
+; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v44, 8, v12
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v47
+; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v14
 ; VI-NEXT:    s_waitcnt vmcnt(10)
-; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v57
+; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v26
+; VI-NEXT:    s_waitcnt vmcnt(9)
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v38
+; VI-NEXT:    s_waitcnt vmcnt(8)
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v39
+; VI-NEXT:    s_waitcnt vmcnt(7)
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v48
+; VI-NEXT:    s_waitcnt vmcnt(6)
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v49
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v52
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v53
+; VI-NEXT:    s_waitcnt vmcnt(3)
+; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v40
+; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v60
+; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v41
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v63
+; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v45
 ; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:100
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
+; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -12077,28 +12088,28 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr51
-; VI-NEXT:    ; implicit-def: $vgpr24
-; VI-NEXT:    ; implicit-def: $vgpr30
-; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr47
-; VI-NEXT:    ; implicit-def: $vgpr42
-; VI-NEXT:    ; implicit-def: $vgpr53
+; VI-NEXT:    v_or_b32_sdwa v9, v40, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v49, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v23, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_or_b32_sdwa v12, v38, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v58, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v45, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v53, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr40
+; VI-NEXT:    ; implicit-def: $vgpr49
+; VI-NEXT:    ; implicit-def: $vgpr23
+; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    ; implicit-def: $vgpr58
+; VI-NEXT:    ; implicit-def: $vgpr45
+; VI-NEXT:    ; implicit-def: $vgpr53
+; VI-NEXT:    ; implicit-def: $vgpr57
 ; VI-NEXT:    ; implicit-def: $vgpr46
-; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr40
+; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    ; implicit-def: $vgpr54
-; VI-NEXT:    ; implicit-def: $vgpr27
-; VI-NEXT:    ; implicit-def: $vgpr23
+; VI-NEXT:    ; implicit-def: $vgpr50
+; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr21
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -12125,18 +12136,18 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v6, v6, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v16, v17, v60 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr17
 ; VI-NEXT:    ; implicit-def: $vgpr20
-; VI-NEXT:    ; implicit-def: $vgpr26
-; VI-NEXT:    ; implicit-def: $vgpr38
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr24
+; VI-NEXT:    ; implicit-def: $vgpr30
+; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    v_or_b32_sdwa v8, v8, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr62
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -12166,23 +12177,23 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v43, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v52, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v29, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v26, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v61, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v47, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v41, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v19, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
@@ -12216,24 +12227,24 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; kill: killed $vgpr16
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
+; VI-NEXT:    ; implicit-def: $vgpr43
 ; VI-NEXT:    ; implicit-def: $vgpr52
-; VI-NEXT:    ; implicit-def: $vgpr50
-; VI-NEXT:    ; implicit-def: $vgpr49
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr48
-; VI-NEXT:    ; implicit-def: $vgpr44
-; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr26
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr47
+; VI-NEXT:    ; implicit-def: $vgpr41
 ; VI-NEXT:    ; implicit-def: $vgpr19
 ; VI-NEXT:    ; implicit-def: $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    ; implicit-def: $vgpr56
-; VI-NEXT:    ; implicit-def: $vgpr45
-; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    ; implicit-def: $vgpr55
-; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr25
-; VI-NEXT:    ; implicit-def: $vgpr21
+; VI-NEXT:    ; implicit-def: $vgpr27
 ; VI-NEXT:  .LBB26_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB26_4
@@ -12244,27 +12255,27 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v15, 0x300
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_add_u16_e32 v9, 3, v51
-; VI-NEXT:    v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v9, 3, v40
+; VI-NEXT:    v_or_b32_sdwa v9, v57, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_add_u16_e32 v10, 3, v24
+; VI-NEXT:    v_add_u16_e32 v10, 3, v49
 ; VI-NEXT:    v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v11, 3, v30
-; VI-NEXT:    v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v11, 3, v23
+; VI-NEXT:    v_or_b32_sdwa v11, v42, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v12, 3, v60
-; VI-NEXT:    v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_add_u16_e32 v12, 3, v38
+; VI-NEXT:    v_or_b32_sdwa v12, v54, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v13, 3, v47
-; VI-NEXT:    v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v13, 3, v58
+; VI-NEXT:    v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v14, 3, v42
-; VI-NEXT:    v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v14, 3, v45
+; VI-NEXT:    v_or_b32_sdwa v14, v39, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
-; VI-NEXT:    v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v17, v60, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -12300,18 +12311,18 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
-; VI-NEXT:    v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v6, v24, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v7, v30, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v53
-; VI-NEXT:    v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v8, 3, v8
-; VI-NEXT:    v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
@@ -12350,39 +12361,39 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v7, v63, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v7, v7, v8
-; VI-NEXT:    v_add_u16_e32 v8, 3, v52
+; VI-NEXT:    v_add_u16_e32 v8, 3, v43
 ; VI-NEXT:    v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v8, v8, v9
-; VI-NEXT:    v_add_u16_e32 v9, 3, v50
+; VI-NEXT:    v_add_u16_e32 v9, 3, v52
 ; VI-NEXT:    v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; VI-NEXT:    v_or_b32_e32 v9, v9, v10
-; VI-NEXT:    v_add_u16_e32 v10, 3, v49
-; VI-NEXT:    v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v10, 3, v29
+; VI-NEXT:    v_or_b32_sdwa v10, v44, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_or_b32_e32 v10, v10, v11
-; VI-NEXT:    v_add_u16_e32 v11, 3, v63
-; VI-NEXT:    v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v11, 3, v26
+; VI-NEXT:    v_or_b32_sdwa v11, v55, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_or_b32_e32 v11, v11, v12
-; VI-NEXT:    v_add_u16_e32 v12, 3, v48
-; VI-NEXT:    v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v12, 3, v61
+; VI-NEXT:    v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; VI-NEXT:    v_or_b32_e32 v12, v12, v13
-; VI-NEXT:    v_add_u16_e32 v13, 3, v44
-; VI-NEXT:    v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v13, 3, v47
+; VI-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_or_b32_e32 v13, v13, v14
-; VI-NEXT:    v_add_u16_e32 v14, 3, v39
+; VI-NEXT:    v_add_u16_e32 v14, 3, v41
 ; VI-NEXT:    v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_or_b32_e32 v14, v14, v16
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v19
-; VI-NEXT:    v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v16, v27, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    v_or_b32_e32 v15, v16, v15
 ; VI-NEXT:  .LBB26_4: ; %end
@@ -12441,39 +12452,32 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:56
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
+; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:72
+; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:80
+; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:88
+; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:96
+; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:104
+; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:112
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:120
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v21
-; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v23
+; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v23
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v25
-; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v27
-; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v29
+; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v27
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v29
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v17
-; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
-; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v1
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v3
@@ -12483,50 +12487,57 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v11
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v15
-; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v54
-; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v2
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; GFX9-NEXT:    s_waitcnt vmcnt(20)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v12
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(17)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v14
+; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v24
+; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v30
+; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v53
+; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v42
+; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v12
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v47
+; GFX9-NEXT:    v_lshlrev_b16_e32 v42, 8, v14
 ; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v57
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v60
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v38
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v63
+; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v39
+; GFX9-NEXT:    s_waitcnt vmcnt(7)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v48
+; GFX9-NEXT:    s_waitcnt vmcnt(6)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v49
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v52
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v53
+; GFX9-NEXT:    s_waitcnt vmcnt(3)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v40
+; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
+; GFX9-NEXT:    s_waitcnt vmcnt(9)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v41
+; GFX9-NEXT:    s_waitcnt vmcnt(8)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v45
 ; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:100
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
+; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -12536,29 +12547,29 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr51
-; GFX9-NEXT:    ; implicit-def: $vgpr24
-; GFX9-NEXT:    ; implicit-def: $vgpr30
-; GFX9-NEXT:    ; implicit-def: $vgpr60
-; GFX9-NEXT:    ; implicit-def: $vgpr47
-; GFX9-NEXT:    ; implicit-def: $vgpr42
-; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
+; GFX9-NEXT:    v_or_b32_sdwa v9, v40, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v23, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_or_b32_sdwa v12, v38, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v58, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v45, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v53, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr49
+; GFX9-NEXT:    ; implicit-def: $vgpr23
+; GFX9-NEXT:    ; implicit-def: $vgpr38
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    ; implicit-def: $vgpr45
+; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    ; implicit-def: $vgpr46
-; GFX9-NEXT:    ; implicit-def: $vgpr43
-; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    ; implicit-def: $vgpr54
-; GFX9-NEXT:    ; implicit-def: $vgpr27
-; GFX9-NEXT:    ; implicit-def: $vgpr23
+; GFX9-NEXT:    ; implicit-def: $vgpr50
+; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr21
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -12585,18 +12596,18 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v60 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr17
 ; GFX9-NEXT:    ; implicit-def: $vgpr20
-; GFX9-NEXT:    ; implicit-def: $vgpr26
-; GFX9-NEXT:    ; implicit-def: $vgpr38
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr24
+; GFX9-NEXT:    ; implicit-def: $vgpr30
+; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -12626,23 +12637,23 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v43, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v52, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v29, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v26, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v61, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v41, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v19, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
@@ -12676,24 +12687,24 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; kill: killed $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    ; implicit-def: $vgpr52
-; GFX9-NEXT:    ; implicit-def: $vgpr50
-; GFX9-NEXT:    ; implicit-def: $vgpr49
-; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr48
-; GFX9-NEXT:    ; implicit-def: $vgpr44
-; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr26
+; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr41
 ; GFX9-NEXT:    ; implicit-def: $vgpr19
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
-; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr45
-; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr55
-; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr51
+; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr25
-; GFX9-NEXT:    ; implicit-def: $vgpr21
+; GFX9-NEXT:    ; implicit-def: $vgpr27
 ; GFX9-NEXT:  .LBB26_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB26_4
@@ -12703,28 +12714,28 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_movk_i32 s6, 0x300
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v51
-; GFX9-NEXT:    v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v40
+; GFX9-NEXT:    v_or_b32_sdwa v9, v57, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v24
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v49
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v30
-; GFX9-NEXT:    v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v23
+; GFX9-NEXT:    v_or_b32_sdwa v11, v42, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v60
-; GFX9-NEXT:    v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v38
+; GFX9-NEXT:    v_or_b32_sdwa v12, v54, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT:    v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v58
+; GFX9-NEXT:    v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT:    v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v45
+; GFX9-NEXT:    v_or_b32_sdwa v14, v39, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v53
-; GFX9-NEXT:    v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
@@ -12761,18 +12772,18 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT:    v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v6, v24, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT:    v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v7, v30, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v17
-; GFX9-NEXT:    v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v16, v60, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT:    v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
@@ -12810,39 +12821,39 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT:    v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v7, v63, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v52
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v43
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v50
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v52
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v49
-; GFX9-NEXT:    v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v29
+; GFX9-NEXT:    v_or_b32_sdwa v10, v44, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v63
-; GFX9-NEXT:    v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v26
+; GFX9-NEXT:    v_or_b32_sdwa v11, v55, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v48
-; GFX9-NEXT:    v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v61
+; GFX9-NEXT:    v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v44
-; GFX9-NEXT:    v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v39
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v41
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v19
-; GFX9-NEXT:    v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v15, v27, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
 ; GFX9-NEXT:  .LBB26_4: ; %end
@@ -26751,170 +26762,171 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v34, v6
 ; SI-NEXT:    v_mov_b32_e32 v33, v4
 ; SI-NEXT:    v_mov_b32_e32 v32, v2
 ; SI-NEXT:    v_mov_b32_e32 v31, v0
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:132
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:120
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:128
 ; SI-NEXT:    v_lshlrev_b32_e32 v20, 24, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 8, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 24, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v28, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v30, 8, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v49, 8, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 24, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 24, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v53
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 24, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v6
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 24, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v10
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 24, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 24, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 8, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v18, 8, v17
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v39
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 24, v6
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v50
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v52
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v10
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v54
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v12
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v41
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 24, v14
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v45
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v50
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v23, 24, v56
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 24, v51
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v21, 8, v58
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 8, v52
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 24, v61
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v21, 24, v53
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v55
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v40
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:68
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v42
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 24, v58
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v47
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 8, v13
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v31
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v33
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v34
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v49
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v2, v2, v36
+; SI-NEXT:    v_or_b32_e32 v2, v2, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v1, v35, v1
-; SI-NEXT:    v_or_b32_e32 v3, v37, v3
+; SI-NEXT:    v_or_b32_e32 v1, v48, v1
+; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v3
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v55
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v57
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v60, v9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v39
+; SI-NEXT:    v_or_b32_e32 v9, v62, v9
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v42
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    v_or_b32_e32 v10, v56, v10
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v52
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    v_or_b32_e32 v11, v46, v11
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v56
+; SI-NEXT:    v_or_b32_e32 v11, v43, v11
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v58
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v45
+; SI-NEXT:    v_or_b32_e32 v12, v54, v12
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v44
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v13, v53, v13
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v54
+; SI-NEXT:    v_or_b32_e32 v13, v21, v13
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v53
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v27, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_or_b32_e32 v15, v23, v15
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr29
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr21
 ; SI-NEXT:    ; implicit-def: $vgpr27
-; SI-NEXT:    ; implicit-def: $vgpr23
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -26922,18 +26934,18 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v2, v2, v48
+; SI-NEXT:    v_or_b32_e32 v2, v2, v36
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v3, v38, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v18, v4
+; SI-NEXT:    v_or_b32_e32 v4, v16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
@@ -26942,7 +26954,6 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v24, v6
-; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -26950,30 +26961,30 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v28, v7
 ; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v3, v16
+; SI-NEXT:    v_or_b32_e32 v3, v3, v38
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_or_b32_e32 v8, v17, v8
+; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v4, v17
+; SI-NEXT:    v_or_b32_e32 v4, v4, v18
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr17
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    v_or_b32_e32 v8, v63, v8
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v22
@@ -26993,54 +27004,59 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v30
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v43
-; SI-NEXT:    v_or_b32_e32 v8, v8, v62
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v60
+; SI-NEXT:    v_or_b32_e32 v8, v8, v63
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v9
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v51
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v45
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v59
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    ; implicit-def: $vgpr62
-; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT:    v_or_b32_e32 v10, v10, v47
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v55
+; SI-NEXT:    v_or_b32_e32 v10, v10, v46
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v58
-; SI-NEXT:    v_or_b32_e32 v11, v11, v44
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    v_or_b32_e32 v11, v11, v41
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_or_b32_e32 v12, v12, v40
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v47
+; SI-NEXT:    v_or_b32_e32 v12, v12, v51
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v41
-; SI-NEXT:    v_or_b32_e32 v13, v13, v50
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v40
+; SI-NEXT:    v_or_b32_e32 v13, v13, v19
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v23
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v25
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v19
-; SI-NEXT:    v_or_b32_e32 v15, v15, v21
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; implicit-def: $vgpr63
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr19
+; SI-NEXT:    ; implicit-def: $vgpr25
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_or_b32_e32 v15, v15, v16
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT:    v_or_b32_e32 v16, v50, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
 ; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; kill: killed $vgpr16
@@ -27067,21 +27083,20 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; kill: killed $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr19
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr25
-; SI-NEXT:    ; implicit-def: $vgpr21
+; SI-NEXT:    ; kill: killed $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:  .LBB50_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v31
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v32
@@ -27089,55 +27104,54 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
-; SI-NEXT:    v_or_b32_e32 v1, v35, v1
+; SI-NEXT:    v_or_b32_e32 v1, v48, v1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v33
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v34
-; SI-NEXT:    v_or_b32_e32 v1, v36, v1
+; SI-NEXT:    v_or_b32_e32 v1, v39, v1
 ; SI-NEXT:    s_movk_i32 s6, 0x300
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v37, v2
+; SI-NEXT:    v_or_b32_e32 v2, v35, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v55
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v60, v9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v39
+; SI-NEXT:    v_or_b32_e32 v9, v62, v9
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v42
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
+; SI-NEXT:    v_or_b32_e32 v10, v56, v10
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v52
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    v_or_b32_e32 v11, v46, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v56
+; SI-NEXT:    v_or_b32_e32 v11, v43, v11
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v58
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v45
+; SI-NEXT:    v_or_b32_e32 v12, v54, v12
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v44
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v13, v53, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v54
+; SI-NEXT:    v_or_b32_e32 v13, v21, v13
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v53
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v27, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v29
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_or_b32_e32 v15, v23, v15
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -27146,20 +27160,20 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_or_b32_e32 v2, v48, v2
+; SI-NEXT:    v_or_b32_e32 v2, v36, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
-; SI-NEXT:    v_or_b32_e32 v3, v38, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v18, v4
+; SI-NEXT:    v_or_b32_e32 v4, v16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
@@ -27175,35 +27189,34 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v28, v7
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v38, v3
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v4, v3
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v3
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_or_b32_e32 v8, v17, v8
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v17, v4
+; SI-NEXT:    v_or_b32_e32 v4, v18, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v4
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    v_or_b32_e32 v8, v63, v8
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
@@ -27229,68 +27242,70 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v43
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v60
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_or_b32_e32 v8, v62, v8
+; SI-NEXT:    v_or_b32_e32 v8, v63, v8
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v51
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v59, v9
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v55
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT:    v_or_b32_e32 v10, v47, v10
+; SI-NEXT:    v_or_b32_e32 v10, v46, v10
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v58
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v44, v11
+; SI-NEXT:    v_or_b32_e32 v11, v41, v11
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v47
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_or_b32_e32 v12, v40, v12
+; SI-NEXT:    v_or_b32_e32 v12, v51, v12
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v41
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v40
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT:    v_or_b32_e32 v13, v50, v13
+; SI-NEXT:    v_or_b32_e32 v13, v19, v13
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v23
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v25, v14
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v19
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
+; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
+; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
+; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
+; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
+; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; SI-NEXT:    v_or_b32_e32 v15, v21, v15
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x300, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT:    v_or_b32_e32 v16, v50, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v15
 ; SI-NEXT:  .LBB50_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
@@ -27310,7 +27325,7 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: bitcast_v64i8_to_v16f32:
@@ -27348,38 +27363,31 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:56
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
+; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:72
+; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:80
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:88
+; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:96
+; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:104
+; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:112
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:120
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:128
+; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v21
-; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v23
+; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v23
 ; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v25
-; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v27
-; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v29
+; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v27
+; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v29
 ; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v17
-; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
-; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:124
 ; VI-NEXT:    v_lshlrev_b16_e32 v31, 8, v1
 ; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v3
@@ -27390,38 +27398,52 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
 ; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v15
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v54
-; VI-NEXT:    v_lshlrev_b16_e32 v61, 8, v0
-; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v2
-; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v4
-; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
-; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
-; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v12
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v14
-; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v24
-; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v30
-; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v53
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v4
+; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v42
+; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v44, 8, v12
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v47
+; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v14
 ; VI-NEXT:    s_waitcnt vmcnt(10)
-; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v57
+; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v26
+; VI-NEXT:    s_waitcnt vmcnt(9)
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v38
+; VI-NEXT:    s_waitcnt vmcnt(8)
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v39
+; VI-NEXT:    s_waitcnt vmcnt(7)
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v48
+; VI-NEXT:    s_waitcnt vmcnt(6)
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v49
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v52
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v53
+; VI-NEXT:    s_waitcnt vmcnt(3)
+; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v40
+; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v60
+; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v41
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v63
+; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v45
 ; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:100
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
+; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -27432,28 +27454,28 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr51
-; VI-NEXT:    ; implicit-def: $vgpr24
-; VI-NEXT:    ; implicit-def: $vgpr30
-; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr47
-; VI-NEXT:    ; implicit-def: $vgpr42
-; VI-NEXT:    ; implicit-def: $vgpr53
+; VI-NEXT:    v_or_b32_sdwa v9, v40, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v49, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v23, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_or_b32_sdwa v12, v38, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v58, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v45, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v53, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr40
+; VI-NEXT:    ; implicit-def: $vgpr49
+; VI-NEXT:    ; implicit-def: $vgpr23
+; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    ; implicit-def: $vgpr58
+; VI-NEXT:    ; implicit-def: $vgpr45
+; VI-NEXT:    ; implicit-def: $vgpr53
+; VI-NEXT:    ; implicit-def: $vgpr57
 ; VI-NEXT:    ; implicit-def: $vgpr46
-; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr40
+; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    ; implicit-def: $vgpr54
-; VI-NEXT:    ; implicit-def: $vgpr27
-; VI-NEXT:    ; implicit-def: $vgpr23
+; VI-NEXT:    ; implicit-def: $vgpr50
+; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr21
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -27480,18 +27502,18 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v6, v6, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v16, v17, v60 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr17
 ; VI-NEXT:    ; implicit-def: $vgpr20
-; VI-NEXT:    ; implicit-def: $vgpr26
-; VI-NEXT:    ; implicit-def: $vgpr38
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr24
+; VI-NEXT:    ; implicit-def: $vgpr30
+; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    v_or_b32_sdwa v8, v8, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr62
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -27521,23 +27543,23 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v43, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v52, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v29, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v26, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v61, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v47, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v41, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v19, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
@@ -27571,24 +27593,24 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; kill: killed $vgpr16
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
+; VI-NEXT:    ; implicit-def: $vgpr43
 ; VI-NEXT:    ; implicit-def: $vgpr52
-; VI-NEXT:    ; implicit-def: $vgpr50
-; VI-NEXT:    ; implicit-def: $vgpr49
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr48
-; VI-NEXT:    ; implicit-def: $vgpr44
-; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr26
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr47
+; VI-NEXT:    ; implicit-def: $vgpr41
 ; VI-NEXT:    ; implicit-def: $vgpr19
 ; VI-NEXT:    ; implicit-def: $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    ; implicit-def: $vgpr56
-; VI-NEXT:    ; implicit-def: $vgpr45
-; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    ; implicit-def: $vgpr55
-; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr25
-; VI-NEXT:    ; implicit-def: $vgpr21
+; VI-NEXT:    ; implicit-def: $vgpr27
 ; VI-NEXT:  .LBB50_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB50_4
@@ -27599,27 +27621,27 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v15, 0x300
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_add_u16_e32 v9, 3, v51
-; VI-NEXT:    v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v9, 3, v40
+; VI-NEXT:    v_or_b32_sdwa v9, v57, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_add_u16_e32 v10, 3, v24
+; VI-NEXT:    v_add_u16_e32 v10, 3, v49
 ; VI-NEXT:    v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v11, 3, v30
-; VI-NEXT:    v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v11, 3, v23
+; VI-NEXT:    v_or_b32_sdwa v11, v42, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v12, 3, v60
-; VI-NEXT:    v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_add_u16_e32 v12, 3, v38
+; VI-NEXT:    v_or_b32_sdwa v12, v54, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v13, 3, v47
-; VI-NEXT:    v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v13, 3, v58
+; VI-NEXT:    v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v14, 3, v42
-; VI-NEXT:    v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v14, 3, v45
+; VI-NEXT:    v_or_b32_sdwa v14, v39, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
-; VI-NEXT:    v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v17, v60, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -27655,18 +27677,18 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
-; VI-NEXT:    v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v6, v24, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v7, v30, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v53
-; VI-NEXT:    v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v8, 3, v8
-; VI-NEXT:    v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
@@ -27705,39 +27727,39 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v7, v63, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v7, v7, v8
-; VI-NEXT:    v_add_u16_e32 v8, 3, v52
+; VI-NEXT:    v_add_u16_e32 v8, 3, v43
 ; VI-NEXT:    v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v8, v8, v9
-; VI-NEXT:    v_add_u16_e32 v9, 3, v50
+; VI-NEXT:    v_add_u16_e32 v9, 3, v52
 ; VI-NEXT:    v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; VI-NEXT:    v_or_b32_e32 v9, v9, v10
-; VI-NEXT:    v_add_u16_e32 v10, 3, v49
-; VI-NEXT:    v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v10, 3, v29
+; VI-NEXT:    v_or_b32_sdwa v10, v44, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_or_b32_e32 v10, v10, v11
-; VI-NEXT:    v_add_u16_e32 v11, 3, v63
-; VI-NEXT:    v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v11, 3, v26
+; VI-NEXT:    v_or_b32_sdwa v11, v55, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_or_b32_e32 v11, v11, v12
-; VI-NEXT:    v_add_u16_e32 v12, 3, v48
-; VI-NEXT:    v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v12, 3, v61
+; VI-NEXT:    v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; VI-NEXT:    v_or_b32_e32 v12, v12, v13
-; VI-NEXT:    v_add_u16_e32 v13, 3, v44
-; VI-NEXT:    v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v13, 3, v47
+; VI-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_or_b32_e32 v13, v13, v14
-; VI-NEXT:    v_add_u16_e32 v14, 3, v39
+; VI-NEXT:    v_add_u16_e32 v14, 3, v41
 ; VI-NEXT:    v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_or_b32_e32 v14, v14, v16
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v19
-; VI-NEXT:    v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v16, v27, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    v_or_b32_e32 v15, v16, v15
 ; VI-NEXT:  .LBB50_4: ; %end
@@ -27796,39 +27818,32 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:56
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
+; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:72
+; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:80
+; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:88
+; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:96
+; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:104
+; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:112
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:120
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v21
-; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v23
+; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v23
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v25
-; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v27
-; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v29
+; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v27
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v29
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v17
-; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
-; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v1
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v3
@@ -27838,50 +27853,57 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v11
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v15
-; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v54
-; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v2
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; GFX9-NEXT:    s_waitcnt vmcnt(20)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v12
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(17)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v14
+; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v24
+; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v30
+; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v53
+; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v42
+; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v12
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v47
+; GFX9-NEXT:    v_lshlrev_b16_e32 v42, 8, v14
 ; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v57
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v60
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v38
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v63
+; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v39
+; GFX9-NEXT:    s_waitcnt vmcnt(7)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v48
+; GFX9-NEXT:    s_waitcnt vmcnt(6)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v49
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v52
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v53
+; GFX9-NEXT:    s_waitcnt vmcnt(3)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v40
+; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
+; GFX9-NEXT:    s_waitcnt vmcnt(9)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v41
+; GFX9-NEXT:    s_waitcnt vmcnt(8)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v45
 ; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:100
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
+; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -27891,29 +27913,29 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr51
-; GFX9-NEXT:    ; implicit-def: $vgpr24
-; GFX9-NEXT:    ; implicit-def: $vgpr30
-; GFX9-NEXT:    ; implicit-def: $vgpr60
-; GFX9-NEXT:    ; implicit-def: $vgpr47
-; GFX9-NEXT:    ; implicit-def: $vgpr42
-; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
+; GFX9-NEXT:    v_or_b32_sdwa v9, v40, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v23, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_or_b32_sdwa v12, v38, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v58, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v45, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v53, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr49
+; GFX9-NEXT:    ; implicit-def: $vgpr23
+; GFX9-NEXT:    ; implicit-def: $vgpr38
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    ; implicit-def: $vgpr45
+; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    ; implicit-def: $vgpr46
-; GFX9-NEXT:    ; implicit-def: $vgpr43
-; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    ; implicit-def: $vgpr54
-; GFX9-NEXT:    ; implicit-def: $vgpr27
-; GFX9-NEXT:    ; implicit-def: $vgpr23
+; GFX9-NEXT:    ; implicit-def: $vgpr50
+; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr21
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -27940,18 +27962,18 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v60 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr17
 ; GFX9-NEXT:    ; implicit-def: $vgpr20
-; GFX9-NEXT:    ; implicit-def: $vgpr26
-; GFX9-NEXT:    ; implicit-def: $vgpr38
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr24
+; GFX9-NEXT:    ; implicit-def: $vgpr30
+; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -27981,23 +28003,23 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v43, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v52, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v29, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v26, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v61, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v41, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v19, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
@@ -28031,24 +28053,24 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; kill: killed $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    ; implicit-def: $vgpr52
-; GFX9-NEXT:    ; implicit-def: $vgpr50
-; GFX9-NEXT:    ; implicit-def: $vgpr49
-; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr48
-; GFX9-NEXT:    ; implicit-def: $vgpr44
-; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr26
+; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr41
 ; GFX9-NEXT:    ; implicit-def: $vgpr19
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
-; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr45
-; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr55
-; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr51
+; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr25
-; GFX9-NEXT:    ; implicit-def: $vgpr21
+; GFX9-NEXT:    ; implicit-def: $vgpr27
 ; GFX9-NEXT:  .LBB50_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB50_4
@@ -28058,28 +28080,28 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_movk_i32 s6, 0x300
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v51
-; GFX9-NEXT:    v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v40
+; GFX9-NEXT:    v_or_b32_sdwa v9, v57, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v24
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v49
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v30
-; GFX9-NEXT:    v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v23
+; GFX9-NEXT:    v_or_b32_sdwa v11, v42, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v60
-; GFX9-NEXT:    v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v38
+; GFX9-NEXT:    v_or_b32_sdwa v12, v54, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT:    v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v58
+; GFX9-NEXT:    v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT:    v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v45
+; GFX9-NEXT:    v_or_b32_sdwa v14, v39, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v53
-; GFX9-NEXT:    v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
@@ -28116,18 +28138,18 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT:    v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v6, v24, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT:    v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v7, v30, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v17
-; GFX9-NEXT:    v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v16, v60, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT:    v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
@@ -28165,39 +28187,39 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT:    v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v7, v63, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v52
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v43
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v50
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v52
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v49
-; GFX9-NEXT:    v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v29
+; GFX9-NEXT:    v_or_b32_sdwa v10, v44, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v63
-; GFX9-NEXT:    v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v26
+; GFX9-NEXT:    v_or_b32_sdwa v11, v55, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v48
-; GFX9-NEXT:    v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v61
+; GFX9-NEXT:    v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v44
-; GFX9-NEXT:    v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v39
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v41
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v19
-; GFX9-NEXT:    v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v15, v27, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
 ; GFX9-NEXT:  .LBB50_4: ; %end
@@ -41377,170 +41399,171 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v34, v6
 ; SI-NEXT:    v_mov_b32_e32 v33, v4
 ; SI-NEXT:    v_mov_b32_e32 v32, v2
 ; SI-NEXT:    v_mov_b32_e32 v31, v0
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:132
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:120
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:128
 ; SI-NEXT:    v_lshlrev_b32_e32 v20, 24, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 8, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 24, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v28, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v30, 8, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v49, 8, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 24, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 24, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v53
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 24, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v6
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 24, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v10
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 24, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 24, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 8, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v18, 8, v17
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v39
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 24, v6
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v50
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v52
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v10
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v54
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v12
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v41
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 24, v14
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v45
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v50
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v23, 24, v56
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 24, v51
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v21, 8, v58
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 8, v52
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 24, v61
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v21, 24, v53
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v55
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v40
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:68
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v42
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 24, v58
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v47
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 8, v13
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB70_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v31
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v33
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v34
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v49
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v2, v2, v36
+; SI-NEXT:    v_or_b32_e32 v2, v2, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v1, v35, v1
-; SI-NEXT:    v_or_b32_e32 v3, v37, v3
+; SI-NEXT:    v_or_b32_e32 v1, v48, v1
+; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v3
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v55
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v57
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v60, v9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v39
+; SI-NEXT:    v_or_b32_e32 v9, v62, v9
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v42
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    v_or_b32_e32 v10, v56, v10
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v52
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    v_or_b32_e32 v11, v46, v11
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v56
+; SI-NEXT:    v_or_b32_e32 v11, v43, v11
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v58
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v45
+; SI-NEXT:    v_or_b32_e32 v12, v54, v12
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v44
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v13, v53, v13
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v54
+; SI-NEXT:    v_or_b32_e32 v13, v21, v13
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v53
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v27, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_or_b32_e32 v15, v23, v15
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr29
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr21
 ; SI-NEXT:    ; implicit-def: $vgpr27
-; SI-NEXT:    ; implicit-def: $vgpr23
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -41548,18 +41571,18 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v2, v2, v48
+; SI-NEXT:    v_or_b32_e32 v2, v2, v36
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v3, v38, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v18, v4
+; SI-NEXT:    v_or_b32_e32 v4, v16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
@@ -41568,7 +41591,6 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v24, v6
-; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -41576,30 +41598,30 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v28, v7
 ; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v3, v16
+; SI-NEXT:    v_or_b32_e32 v3, v3, v38
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_or_b32_e32 v8, v17, v8
+; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v4, v17
+; SI-NEXT:    v_or_b32_e32 v4, v4, v18
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr17
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    v_or_b32_e32 v8, v63, v8
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v22
@@ -41619,54 +41641,59 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v30
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v43
-; SI-NEXT:    v_or_b32_e32 v8, v8, v62
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v60
+; SI-NEXT:    v_or_b32_e32 v8, v8, v63
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v9
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v51
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v45
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v59
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    ; implicit-def: $vgpr62
-; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT:    v_or_b32_e32 v10, v10, v47
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v55
+; SI-NEXT:    v_or_b32_e32 v10, v10, v46
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v58
-; SI-NEXT:    v_or_b32_e32 v11, v11, v44
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    v_or_b32_e32 v11, v11, v41
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_or_b32_e32 v12, v12, v40
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v47
+; SI-NEXT:    v_or_b32_e32 v12, v12, v51
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v41
-; SI-NEXT:    v_or_b32_e32 v13, v13, v50
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v40
+; SI-NEXT:    v_or_b32_e32 v13, v13, v19
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v23
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v25
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v19
-; SI-NEXT:    v_or_b32_e32 v15, v15, v21
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; implicit-def: $vgpr63
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr19
+; SI-NEXT:    ; implicit-def: $vgpr25
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_or_b32_e32 v15, v15, v16
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT:    v_or_b32_e32 v16, v50, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
 ; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; kill: killed $vgpr16
@@ -41693,21 +41720,20 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; kill: killed $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr19
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr25
-; SI-NEXT:    ; implicit-def: $vgpr21
+; SI-NEXT:    ; kill: killed $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:  .LBB70_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB70_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v31
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v32
@@ -41715,55 +41741,54 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
-; SI-NEXT:    v_or_b32_e32 v1, v35, v1
+; SI-NEXT:    v_or_b32_e32 v1, v48, v1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v33
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v34
-; SI-NEXT:    v_or_b32_e32 v1, v36, v1
+; SI-NEXT:    v_or_b32_e32 v1, v39, v1
 ; SI-NEXT:    s_movk_i32 s6, 0x300
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v37, v2
+; SI-NEXT:    v_or_b32_e32 v2, v35, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v55
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v60, v9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v39
+; SI-NEXT:    v_or_b32_e32 v9, v62, v9
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v42
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
+; SI-NEXT:    v_or_b32_e32 v10, v56, v10
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v52
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    v_or_b32_e32 v11, v46, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v56
+; SI-NEXT:    v_or_b32_e32 v11, v43, v11
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v58
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v45
+; SI-NEXT:    v_or_b32_e32 v12, v54, v12
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v44
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v13, v53, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v54
+; SI-NEXT:    v_or_b32_e32 v13, v21, v13
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v53
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v27, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v29
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_or_b32_e32 v15, v23, v15
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -41772,20 +41797,20 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_or_b32_e32 v2, v48, v2
+; SI-NEXT:    v_or_b32_e32 v2, v36, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
-; SI-NEXT:    v_or_b32_e32 v3, v38, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v18, v4
+; SI-NEXT:    v_or_b32_e32 v4, v16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
@@ -41801,35 +41826,34 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v28, v7
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v38, v3
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v4, v3
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v3
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_or_b32_e32 v8, v17, v8
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v17, v4
+; SI-NEXT:    v_or_b32_e32 v4, v18, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v4
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    v_or_b32_e32 v8, v63, v8
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
@@ -41855,68 +41879,70 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v43
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v60
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_or_b32_e32 v8, v62, v8
+; SI-NEXT:    v_or_b32_e32 v8, v63, v8
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v51
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v59, v9
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v55
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT:    v_or_b32_e32 v10, v47, v10
+; SI-NEXT:    v_or_b32_e32 v10, v46, v10
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v58
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v44, v11
+; SI-NEXT:    v_or_b32_e32 v11, v41, v11
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v47
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_or_b32_e32 v12, v40, v12
+; SI-NEXT:    v_or_b32_e32 v12, v51, v12
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v41
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v40
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT:    v_or_b32_e32 v13, v50, v13
+; SI-NEXT:    v_or_b32_e32 v13, v19, v13
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v23
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v25, v14
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v19
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
+; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
+; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
+; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
+; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
+; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; SI-NEXT:    v_or_b32_e32 v15, v21, v15
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x300, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT:    v_or_b32_e32 v16, v50, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v15
 ; SI-NEXT:  .LBB70_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
@@ -41936,7 +41962,7 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: bitcast_v64i8_to_v8i64:
@@ -41974,38 +42000,31 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:56
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
+; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:72
+; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:80
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:88
+; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:96
+; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:104
+; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:112
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:120
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:128
+; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v21
-; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v23
+; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v23
 ; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v25
-; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v27
-; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v29
+; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v27
+; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v29
 ; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v17
-; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
-; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:124
 ; VI-NEXT:    v_lshlrev_b16_e32 v31, 8, v1
 ; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v3
@@ -42016,38 +42035,52 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
 ; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v15
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v54
-; VI-NEXT:    v_lshlrev_b16_e32 v61, 8, v0
-; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v2
-; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v4
-; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
-; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
-; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v12
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v14
-; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v24
-; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v30
-; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v53
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v4
+; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v42
+; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v44, 8, v12
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v47
+; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v14
 ; VI-NEXT:    s_waitcnt vmcnt(10)
-; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v57
+; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v26
+; VI-NEXT:    s_waitcnt vmcnt(9)
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v38
+; VI-NEXT:    s_waitcnt vmcnt(8)
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v39
+; VI-NEXT:    s_waitcnt vmcnt(7)
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v48
+; VI-NEXT:    s_waitcnt vmcnt(6)
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v49
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v52
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v53
+; VI-NEXT:    s_waitcnt vmcnt(3)
+; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v40
+; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v60
+; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v41
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v63
+; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v45
 ; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:100
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
+; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -42058,28 +42091,28 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr51
-; VI-NEXT:    ; implicit-def: $vgpr24
-; VI-NEXT:    ; implicit-def: $vgpr30
-; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr47
-; VI-NEXT:    ; implicit-def: $vgpr42
-; VI-NEXT:    ; implicit-def: $vgpr53
+; VI-NEXT:    v_or_b32_sdwa v9, v40, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v49, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v23, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_or_b32_sdwa v12, v38, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v58, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v45, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v53, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr40
+; VI-NEXT:    ; implicit-def: $vgpr49
+; VI-NEXT:    ; implicit-def: $vgpr23
+; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    ; implicit-def: $vgpr58
+; VI-NEXT:    ; implicit-def: $vgpr45
+; VI-NEXT:    ; implicit-def: $vgpr53
+; VI-NEXT:    ; implicit-def: $vgpr57
 ; VI-NEXT:    ; implicit-def: $vgpr46
-; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr40
+; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    ; implicit-def: $vgpr54
-; VI-NEXT:    ; implicit-def: $vgpr27
-; VI-NEXT:    ; implicit-def: $vgpr23
+; VI-NEXT:    ; implicit-def: $vgpr50
+; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr21
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -42106,18 +42139,18 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v6, v6, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v16, v17, v60 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr17
 ; VI-NEXT:    ; implicit-def: $vgpr20
-; VI-NEXT:    ; implicit-def: $vgpr26
-; VI-NEXT:    ; implicit-def: $vgpr38
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr24
+; VI-NEXT:    ; implicit-def: $vgpr30
+; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    v_or_b32_sdwa v8, v8, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr62
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -42147,23 +42180,23 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v43, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v52, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v29, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v26, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v61, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v47, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v41, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v19, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
@@ -42197,24 +42230,24 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; kill: killed $vgpr16
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
+; VI-NEXT:    ; implicit-def: $vgpr43
 ; VI-NEXT:    ; implicit-def: $vgpr52
-; VI-NEXT:    ; implicit-def: $vgpr50
-; VI-NEXT:    ; implicit-def: $vgpr49
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr48
-; VI-NEXT:    ; implicit-def: $vgpr44
-; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr26
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr47
+; VI-NEXT:    ; implicit-def: $vgpr41
 ; VI-NEXT:    ; implicit-def: $vgpr19
 ; VI-NEXT:    ; implicit-def: $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    ; implicit-def: $vgpr56
-; VI-NEXT:    ; implicit-def: $vgpr45
-; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    ; implicit-def: $vgpr55
-; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr25
-; VI-NEXT:    ; implicit-def: $vgpr21
+; VI-NEXT:    ; implicit-def: $vgpr27
 ; VI-NEXT:  .LBB70_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB70_4
@@ -42225,27 +42258,27 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v15, 0x300
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_add_u16_e32 v9, 3, v51
-; VI-NEXT:    v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v9, 3, v40
+; VI-NEXT:    v_or_b32_sdwa v9, v57, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_add_u16_e32 v10, 3, v24
+; VI-NEXT:    v_add_u16_e32 v10, 3, v49
 ; VI-NEXT:    v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v11, 3, v30
-; VI-NEXT:    v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v11, 3, v23
+; VI-NEXT:    v_or_b32_sdwa v11, v42, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v12, 3, v60
-; VI-NEXT:    v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_add_u16_e32 v12, 3, v38
+; VI-NEXT:    v_or_b32_sdwa v12, v54, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v13, 3, v47
-; VI-NEXT:    v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v13, 3, v58
+; VI-NEXT:    v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v14, 3, v42
-; VI-NEXT:    v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v14, 3, v45
+; VI-NEXT:    v_or_b32_sdwa v14, v39, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
-; VI-NEXT:    v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v17, v60, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -42281,18 +42314,18 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
-; VI-NEXT:    v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v6, v24, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v7, v30, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v53
-; VI-NEXT:    v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v8, 3, v8
-; VI-NEXT:    v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
@@ -42331,39 +42364,39 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v7, v63, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v7, v7, v8
-; VI-NEXT:    v_add_u16_e32 v8, 3, v52
+; VI-NEXT:    v_add_u16_e32 v8, 3, v43
 ; VI-NEXT:    v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v8, v8, v9
-; VI-NEXT:    v_add_u16_e32 v9, 3, v50
+; VI-NEXT:    v_add_u16_e32 v9, 3, v52
 ; VI-NEXT:    v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; VI-NEXT:    v_or_b32_e32 v9, v9, v10
-; VI-NEXT:    v_add_u16_e32 v10, 3, v49
-; VI-NEXT:    v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v10, 3, v29
+; VI-NEXT:    v_or_b32_sdwa v10, v44, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_or_b32_e32 v10, v10, v11
-; VI-NEXT:    v_add_u16_e32 v11, 3, v63
-; VI-NEXT:    v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v11, 3, v26
+; VI-NEXT:    v_or_b32_sdwa v11, v55, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_or_b32_e32 v11, v11, v12
-; VI-NEXT:    v_add_u16_e32 v12, 3, v48
-; VI-NEXT:    v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v12, 3, v61
+; VI-NEXT:    v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; VI-NEXT:    v_or_b32_e32 v12, v12, v13
-; VI-NEXT:    v_add_u16_e32 v13, 3, v44
-; VI-NEXT:    v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v13, 3, v47
+; VI-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_or_b32_e32 v13, v13, v14
-; VI-NEXT:    v_add_u16_e32 v14, 3, v39
+; VI-NEXT:    v_add_u16_e32 v14, 3, v41
 ; VI-NEXT:    v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_or_b32_e32 v14, v14, v16
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v19
-; VI-NEXT:    v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v16, v27, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    v_or_b32_e32 v15, v16, v15
 ; VI-NEXT:  .LBB70_4: ; %end
@@ -42422,39 +42455,32 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:56
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
+; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:72
+; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:80
+; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:88
+; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:96
+; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:104
+; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:112
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:120
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v21
-; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v23
+; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v23
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v25
-; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v27
-; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v29
+; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v27
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v29
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v17
-; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
-; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v1
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v3
@@ -42464,50 +42490,57 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v11
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v15
-; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v54
-; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v2
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; GFX9-NEXT:    s_waitcnt vmcnt(20)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v12
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(17)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v14
+; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v24
+; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v30
+; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v53
+; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v42
+; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v12
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v47
+; GFX9-NEXT:    v_lshlrev_b16_e32 v42, 8, v14
 ; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v57
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v26
+; GFX9-NEXT:    s_waitcnt vmcnt(9)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v38
+; GFX9-NEXT:    s_waitcnt vmcnt(8)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v39
+; GFX9-NEXT:    s_waitcnt vmcnt(7)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v48
+; GFX9-NEXT:    s_waitcnt vmcnt(6)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v49
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v52
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v53
+; GFX9-NEXT:    s_waitcnt vmcnt(3)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v40
+; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v60
+; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v41
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v63
+; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v45
 ; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:100
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
+; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -42517,29 +42550,29 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr51
-; GFX9-NEXT:    ; implicit-def: $vgpr24
-; GFX9-NEXT:    ; implicit-def: $vgpr30
-; GFX9-NEXT:    ; implicit-def: $vgpr60
-; GFX9-NEXT:    ; implicit-def: $vgpr47
-; GFX9-NEXT:    ; implicit-def: $vgpr42
-; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
+; GFX9-NEXT:    v_or_b32_sdwa v9, v40, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v23, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_or_b32_sdwa v12, v38, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v58, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v45, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v53, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr49
+; GFX9-NEXT:    ; implicit-def: $vgpr23
+; GFX9-NEXT:    ; implicit-def: $vgpr38
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    ; implicit-def: $vgpr45
+; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    ; implicit-def: $vgpr46
-; GFX9-NEXT:    ; implicit-def: $vgpr43
-; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    ; implicit-def: $vgpr54
-; GFX9-NEXT:    ; implicit-def: $vgpr27
-; GFX9-NEXT:    ; implicit-def: $vgpr23
+; GFX9-NEXT:    ; implicit-def: $vgpr50
+; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr21
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -42566,18 +42599,18 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v60 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr17
 ; GFX9-NEXT:    ; implicit-def: $vgpr20
-; GFX9-NEXT:    ; implicit-def: $vgpr26
-; GFX9-NEXT:    ; implicit-def: $vgpr38
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr24
+; GFX9-NEXT:    ; implicit-def: $vgpr30
+; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -42607,23 +42640,23 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v43, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v52, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v29, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v26, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v61, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v41, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v19, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
@@ -42657,24 +42690,24 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; kill: killed $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    ; implicit-def: $vgpr52
-; GFX9-NEXT:    ; implicit-def: $vgpr50
-; GFX9-NEXT:    ; implicit-def: $vgpr49
-; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr48
-; GFX9-NEXT:    ; implicit-def: $vgpr44
-; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr26
+; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr41
 ; GFX9-NEXT:    ; implicit-def: $vgpr19
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
-; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr45
-; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr55
-; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr51
+; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr25
-; GFX9-NEXT:    ; implicit-def: $vgpr21
+; GFX9-NEXT:    ; implicit-def: $vgpr27
 ; GFX9-NEXT:  .LBB70_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB70_4
@@ -42684,28 +42717,28 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_movk_i32 s6, 0x300
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v51
-; GFX9-NEXT:    v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v40
+; GFX9-NEXT:    v_or_b32_sdwa v9, v57, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v24
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v49
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v30
-; GFX9-NEXT:    v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v23
+; GFX9-NEXT:    v_or_b32_sdwa v11, v42, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v60
-; GFX9-NEXT:    v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v38
+; GFX9-NEXT:    v_or_b32_sdwa v12, v54, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT:    v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v58
+; GFX9-NEXT:    v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT:    v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v45
+; GFX9-NEXT:    v_or_b32_sdwa v14, v39, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v53
-; GFX9-NEXT:    v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
@@ -42742,18 +42775,18 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT:    v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v6, v24, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT:    v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v7, v30, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v17
-; GFX9-NEXT:    v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v16, v60, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT:    v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
@@ -42791,39 +42824,39 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT:    v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v7, v63, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v52
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v43
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v50
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v52
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v49
-; GFX9-NEXT:    v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v29
+; GFX9-NEXT:    v_or_b32_sdwa v10, v44, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v63
-; GFX9-NEXT:    v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v26
+; GFX9-NEXT:    v_or_b32_sdwa v11, v55, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v48
-; GFX9-NEXT:    v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v61
+; GFX9-NEXT:    v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v44
-; GFX9-NEXT:    v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v39
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v41
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v19
-; GFX9-NEXT:    v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v15, v27, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
 ; GFX9-NEXT:  .LBB70_4: ; %end
@@ -55157,170 +55190,171 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v34, v6
 ; SI-NEXT:    v_mov_b32_e32 v33, v4
 ; SI-NEXT:    v_mov_b32_e32 v32, v2
 ; SI-NEXT:    v_mov_b32_e32 v31, v0
-; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:132
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:120
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:128
 ; SI-NEXT:    v_lshlrev_b32_e32 v20, 24, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 8, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 24, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v28, 24, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v30, 8, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v49, 8, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 24, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 24, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 24, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 8, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v18, 8, v17
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v53
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 24, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v6
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 24, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v10
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 24, v12
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v39
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 24, v6
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v50
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v52
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v10
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v54
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v12
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v41
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 24, v14
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v45
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v50
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v23, 24, v56
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 24, v51
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v21, 8, v58
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 8, v52
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 24, v61
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v21, 24, v53
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v55
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v40
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:68
+; SI-NEXT:    v_lshlrev_b32_e32 v25, 8, v42
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 24, v58
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 8, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v47
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB86_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v31
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v33
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v34
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v49
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v2, v2, v36
+; SI-NEXT:    v_or_b32_e32 v2, v2, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v1, v35, v1
-; SI-NEXT:    v_or_b32_e32 v3, v37, v3
+; SI-NEXT:    v_or_b32_e32 v1, v48, v1
+; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v3
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v55
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v57
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v60, v9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v39
+; SI-NEXT:    v_or_b32_e32 v9, v62, v9
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v42
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    v_or_b32_e32 v10, v56, v10
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v52
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    v_or_b32_e32 v11, v46, v11
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v56
+; SI-NEXT:    v_or_b32_e32 v11, v43, v11
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v58
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v45
+; SI-NEXT:    v_or_b32_e32 v12, v54, v12
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v44
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v13, v53, v13
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v54
+; SI-NEXT:    v_or_b32_e32 v13, v21, v13
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v53
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v27, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_or_b32_e32 v15, v23, v15
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr29
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr46
 ; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    ; implicit-def: $vgpr49
+; SI-NEXT:    ; implicit-def: $vgpr48
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr21
 ; SI-NEXT:    ; implicit-def: $vgpr27
-; SI-NEXT:    ; implicit-def: $vgpr23
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -55328,18 +55362,18 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v2, v2, v48
+; SI-NEXT:    v_or_b32_e32 v2, v2, v36
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v3, v38, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v18, v4
+; SI-NEXT:    v_or_b32_e32 v4, v16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
@@ -55348,7 +55382,6 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v24, v6
-; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -55356,30 +55389,30 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v28, v7
 ; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v3, v16
+; SI-NEXT:    v_or_b32_e32 v3, v3, v38
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v4
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_or_b32_e32 v8, v17, v8
+; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v4, v17
+; SI-NEXT:    v_or_b32_e32 v4, v4, v18
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr17
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    v_or_b32_e32 v8, v63, v8
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v22
@@ -55399,54 +55432,59 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v30
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v8
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v43
-; SI-NEXT:    v_or_b32_e32 v8, v8, v62
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v60
+; SI-NEXT:    v_or_b32_e32 v8, v8, v63
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v9
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v51
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v45
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v59
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v9, v10
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    ; implicit-def: $vgpr62
-; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT:    v_or_b32_e32 v10, v10, v47
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v55
+; SI-NEXT:    v_or_b32_e32 v10, v10, v46
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v58
-; SI-NEXT:    v_or_b32_e32 v11, v11, v44
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
+; SI-NEXT:    v_or_b32_e32 v11, v11, v41
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_or_b32_e32 v12, v12, v40
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v47
+; SI-NEXT:    v_or_b32_e32 v12, v12, v51
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v13
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v41
-; SI-NEXT:    v_or_b32_e32 v13, v13, v50
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v40
+; SI-NEXT:    v_or_b32_e32 v13, v13, v19
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v14
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v23
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v25
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v15
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v19
-; SI-NEXT:    v_or_b32_e32 v15, v15, v21
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; implicit-def: $vgpr63
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr19
+; SI-NEXT:    ; implicit-def: $vgpr25
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_or_b32_e32 v15, v15, v16
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT:    v_or_b32_e32 v16, v50, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v16
 ; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; kill: killed $vgpr16
@@ -55473,21 +55511,20 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; kill: killed $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr16
 ; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr19
-; SI-NEXT:    ; kill: killed $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr25
-; SI-NEXT:    ; implicit-def: $vgpr21
+; SI-NEXT:    ; kill: killed $vgpr50
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:  .LBB86_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB86_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v31
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v32
@@ -55495,55 +55532,54 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
-; SI-NEXT:    v_or_b32_e32 v1, v35, v1
+; SI-NEXT:    v_or_b32_e32 v1, v48, v1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v33
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v34
-; SI-NEXT:    v_or_b32_e32 v1, v36, v1
+; SI-NEXT:    v_or_b32_e32 v1, v39, v1
 ; SI-NEXT:    s_movk_i32 s6, 0x300
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v37, v2
+; SI-NEXT:    v_or_b32_e32 v2, v35, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v55
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v57
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v9, v60, v9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v39
+; SI-NEXT:    v_or_b32_e32 v9, v62, v9
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v42
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
+; SI-NEXT:    v_or_b32_e32 v10, v56, v10
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v52
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    v_or_b32_e32 v11, v46, v11
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v56
+; SI-NEXT:    v_or_b32_e32 v11, v43, v11
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v58
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v12, v42, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v45
+; SI-NEXT:    v_or_b32_e32 v12, v54, v12
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v44
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_or_b32_e32 v13, v53, v13
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v54
+; SI-NEXT:    v_or_b32_e32 v13, v21, v13
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v53
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v27, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v29
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; SI-NEXT:    v_or_b32_e32 v15, v23, v15
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
@@ -55552,20 +55588,20 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_or_b32_e32 v2, v48, v2
+; SI-NEXT:    v_or_b32_e32 v2, v36, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
-; SI-NEXT:    v_or_b32_e32 v3, v38, v3
+; SI-NEXT:    v_or_b32_e32 v3, v37, v3
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v18, v4
+; SI-NEXT:    v_or_b32_e32 v4, v16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
@@ -55581,35 +55617,34 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v28, v7
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v3, v16, v3
+; SI-NEXT:    v_or_b32_e32 v3, v38, v3
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v4, v3
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v3
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_or_b32_e32 v8, v17, v8
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v4, v17, v4
+; SI-NEXT:    v_or_b32_e32 v4, v18, v4
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v4
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    v_or_b32_e32 v8, v63, v8
-; SI-NEXT:    v_or_b32_e32 v16, v52, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
@@ -55635,68 +55670,70 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v43
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v60
 ; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_or_b32_e32 v8, v62, v8
+; SI-NEXT:    v_or_b32_e32 v8, v63, v8
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v51
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v59, v9
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v55
 ; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; SI-NEXT:    v_or_b32_e32 v10, v47, v10
+; SI-NEXT:    v_or_b32_e32 v10, v46, v10
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v58
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v44, v11
+; SI-NEXT:    v_or_b32_e32 v11, v41, v11
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v12, v11
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
-; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
+; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v47
 ; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_or_b32_e32 v12, v40, v12
+; SI-NEXT:    v_or_b32_e32 v12, v51, v12
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v41
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v40
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT:    v_or_b32_e32 v13, v50, v13
+; SI-NEXT:    v_or_b32_e32 v13, v19, v13
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v13, v14, v13
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
-; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v23
 ; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v25, v14
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v14, v15, v14
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v19
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
+; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
+; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
+; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
+; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
+; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; SI-NEXT:    v_or_b32_e32 v15, v21, v15
+; SI-NEXT:    v_or_b32_e32 v15, v16, v15
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x300, v15
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; SI-NEXT:    v_or_b32_e32 v16, v50, v16
 ; SI-NEXT:    v_or_b32_e32 v15, v16, v15
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v15
 ; SI-NEXT:  .LBB86_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
@@ -55716,7 +55753,7 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: bitcast_v64i8_to_v8f64:
@@ -55754,38 +55791,31 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:56
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
+; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:72
+; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:80
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:88
+; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:96
+; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:104
+; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:112
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:120
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:128
+; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v21
-; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v23
+; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v23
 ; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v25
-; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v27
-; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v29
+; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v27
+; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v29
 ; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v17
-; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
-; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:124
 ; VI-NEXT:    v_lshlrev_b16_e32 v31, 8, v1
 ; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v3
@@ -55796,38 +55826,52 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
 ; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v15
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v54
-; VI-NEXT:    v_lshlrev_b16_e32 v61, 8, v0
-; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v2
-; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v4
-; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
-; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
-; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v12
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v14
-; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v24
-; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v30
-; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v53
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v4
+; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v42
+; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v44
+; VI-NEXT:    v_lshlrev_b16_e32 v44, 8, v12
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v47
+; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v14
 ; VI-NEXT:    s_waitcnt vmcnt(10)
-; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v57
+; VI-NEXT:    v_lshlrev_b16_e32 v55, 8, v26
+; VI-NEXT:    s_waitcnt vmcnt(9)
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v38
+; VI-NEXT:    s_waitcnt vmcnt(8)
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v39
+; VI-NEXT:    s_waitcnt vmcnt(7)
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v48
+; VI-NEXT:    s_waitcnt vmcnt(6)
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v49
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v52
+; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v53
+; VI-NEXT:    s_waitcnt vmcnt(3)
+; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v40
+; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v60
+; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v41
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v63
+; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v45
 ; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:100
+; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
+; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
+; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:52
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -55838,28 +55882,28 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr51
-; VI-NEXT:    ; implicit-def: $vgpr24
-; VI-NEXT:    ; implicit-def: $vgpr30
-; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr47
-; VI-NEXT:    ; implicit-def: $vgpr42
-; VI-NEXT:    ; implicit-def: $vgpr53
+; VI-NEXT:    v_or_b32_sdwa v9, v40, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v49, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v23, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_or_b32_sdwa v12, v38, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v58, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v45, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v53, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr40
+; VI-NEXT:    ; implicit-def: $vgpr49
+; VI-NEXT:    ; implicit-def: $vgpr23
+; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    ; implicit-def: $vgpr58
+; VI-NEXT:    ; implicit-def: $vgpr45
+; VI-NEXT:    ; implicit-def: $vgpr53
+; VI-NEXT:    ; implicit-def: $vgpr57
 ; VI-NEXT:    ; implicit-def: $vgpr46
-; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr40
+; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    ; implicit-def: $vgpr54
-; VI-NEXT:    ; implicit-def: $vgpr27
-; VI-NEXT:    ; implicit-def: $vgpr23
+; VI-NEXT:    ; implicit-def: $vgpr50
+; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr21
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -55886,18 +55930,18 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v6, v6, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v16, v17, v60 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr17
 ; VI-NEXT:    ; implicit-def: $vgpr20
-; VI-NEXT:    ; implicit-def: $vgpr26
-; VI-NEXT:    ; implicit-def: $vgpr38
-; VI-NEXT:    ; implicit-def: $vgpr57
+; VI-NEXT:    ; implicit-def: $vgpr24
+; VI-NEXT:    ; implicit-def: $vgpr30
+; VI-NEXT:    ; implicit-def: $vgpr60
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    v_or_b32_sdwa v8, v8, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr62
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -55927,23 +55971,23 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr28
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v43, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v52, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v29, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v26, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v61, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v47, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v41, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v19, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
@@ -55977,24 +56021,24 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; kill: killed $vgpr16
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
+; VI-NEXT:    ; implicit-def: $vgpr43
 ; VI-NEXT:    ; implicit-def: $vgpr52
-; VI-NEXT:    ; implicit-def: $vgpr50
-; VI-NEXT:    ; implicit-def: $vgpr49
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr48
-; VI-NEXT:    ; implicit-def: $vgpr44
-; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr26
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr47
+; VI-NEXT:    ; implicit-def: $vgpr41
 ; VI-NEXT:    ; implicit-def: $vgpr19
 ; VI-NEXT:    ; implicit-def: $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; implicit-def: $vgpr63
 ; VI-NEXT:    ; implicit-def: $vgpr59
 ; VI-NEXT:    ; implicit-def: $vgpr56
-; VI-NEXT:    ; implicit-def: $vgpr45
-; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr44
 ; VI-NEXT:    ; implicit-def: $vgpr55
-; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr48
 ; VI-NEXT:    ; implicit-def: $vgpr25
-; VI-NEXT:    ; implicit-def: $vgpr21
+; VI-NEXT:    ; implicit-def: $vgpr27
 ; VI-NEXT:  .LBB86_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB86_4
@@ -56005,27 +56049,27 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    v_mov_b32_e32 v15, 0x300
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_add_u16_e32 v9, 3, v51
-; VI-NEXT:    v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v9, 3, v40
+; VI-NEXT:    v_or_b32_sdwa v9, v57, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v9, v9, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_add_u16_e32 v10, 3, v24
+; VI-NEXT:    v_add_u16_e32 v10, 3, v49
 ; VI-NEXT:    v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v10, v10, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v11, 3, v30
-; VI-NEXT:    v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v11, 3, v23
+; VI-NEXT:    v_or_b32_sdwa v11, v42, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v11, v11, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v12, 3, v60
-; VI-NEXT:    v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_add_u16_e32 v12, 3, v38
+; VI-NEXT:    v_or_b32_sdwa v12, v54, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v12, v12, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v13, 3, v47
-; VI-NEXT:    v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v13, 3, v58
+; VI-NEXT:    v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v13, v13, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v14, 3, v42
-; VI-NEXT:    v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v14, 3, v45
+; VI-NEXT:    v_or_b32_sdwa v14, v39, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v14, v14, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v17, 3, v17
-; VI-NEXT:    v_or_b32_sdwa v17, v57, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v17, v60, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -56061,18 +56105,18 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_add_u16_sdwa v5, v5, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_add_u16_e32 v6, 3, v6
-; VI-NEXT:    v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v6, v24, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v6, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v7, v30, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v7, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v53
-; VI-NEXT:    v_or_b32_sdwa v16, v23, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v16, v16, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_u16_e32 v8, 3, v8
-; VI-NEXT:    v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v8, v15 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_u16_e32 v2, 3, v2
@@ -56111,39 +56155,39 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v7, 3, v7
-; VI-NEXT:    v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v7, v63, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; VI-NEXT:    v_or_b32_e32 v7, v7, v8
-; VI-NEXT:    v_add_u16_e32 v8, 3, v52
+; VI-NEXT:    v_add_u16_e32 v8, 3, v43
 ; VI-NEXT:    v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; VI-NEXT:    v_or_b32_e32 v8, v8, v9
-; VI-NEXT:    v_add_u16_e32 v9, 3, v50
+; VI-NEXT:    v_add_u16_e32 v9, 3, v52
 ; VI-NEXT:    v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; VI-NEXT:    v_or_b32_e32 v9, v9, v10
-; VI-NEXT:    v_add_u16_e32 v10, 3, v49
-; VI-NEXT:    v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v10, 3, v29
+; VI-NEXT:    v_or_b32_sdwa v10, v44, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; VI-NEXT:    v_or_b32_e32 v10, v10, v11
-; VI-NEXT:    v_add_u16_e32 v11, 3, v63
-; VI-NEXT:    v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v11, 3, v26
+; VI-NEXT:    v_or_b32_sdwa v11, v55, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; VI-NEXT:    v_or_b32_e32 v11, v11, v12
-; VI-NEXT:    v_add_u16_e32 v12, 3, v48
-; VI-NEXT:    v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v12, 3, v61
+; VI-NEXT:    v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; VI-NEXT:    v_or_b32_e32 v12, v12, v13
-; VI-NEXT:    v_add_u16_e32 v13, 3, v44
-; VI-NEXT:    v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v13, 3, v47
+; VI-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; VI-NEXT:    v_or_b32_e32 v13, v13, v14
-; VI-NEXT:    v_add_u16_e32 v14, 3, v39
+; VI-NEXT:    v_add_u16_e32 v14, 3, v41
 ; VI-NEXT:    v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_or_b32_e32 v14, v14, v16
 ; VI-NEXT:    v_add_u16_e32 v16, 3, v19
-; VI-NEXT:    v_or_b32_sdwa v16, v21, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v16, v27, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
 ; VI-NEXT:    v_or_b32_e32 v15, v16, v15
 ; VI-NEXT:  .LBB86_4: ; %end
@@ -56202,39 +56246,32 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:64
-; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:72
-; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:80
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:88
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:104
-; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:112
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:56
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
+; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:72
+; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:80
+; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:88
+; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:96
+; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:104
+; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:112
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:120
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v21
-; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v23
+; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v23
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v25
-; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v27
-; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v29
+; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v27
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v29
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v17
-; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v19
-; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:124
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v1
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v3
@@ -56244,50 +56281,57 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v11
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v37, 8, v13
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v15
-; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v54
-; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v2
-; GFX9-NEXT:    s_waitcnt vmcnt(22)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; GFX9-NEXT:    s_waitcnt vmcnt(20)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v8
-; GFX9-NEXT:    s_waitcnt vmcnt(19)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v12
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(17)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v14
+; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v24
+; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v30
+; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v53
+; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v42
+; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v44
+; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v12
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v47
+; GFX9-NEXT:    v_lshlrev_b16_e32 v42, 8, v14
 ; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v57
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v26
+; GFX9-NEXT:    s_waitcnt vmcnt(9)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v38
+; GFX9-NEXT:    s_waitcnt vmcnt(8)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v39
+; GFX9-NEXT:    s_waitcnt vmcnt(7)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v48
+; GFX9-NEXT:    s_waitcnt vmcnt(6)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v49
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v52
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v53
+; GFX9-NEXT:    s_waitcnt vmcnt(3)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v40
+; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:116
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v60
+; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v41
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v63
+; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v45
 ; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:108
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:92
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:100
+; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
+; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -56297,29 +56341,29 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v51, v58 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v24, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v30, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v60, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v42, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v53, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr51
-; GFX9-NEXT:    ; implicit-def: $vgpr24
-; GFX9-NEXT:    ; implicit-def: $vgpr30
-; GFX9-NEXT:    ; implicit-def: $vgpr60
-; GFX9-NEXT:    ; implicit-def: $vgpr47
-; GFX9-NEXT:    ; implicit-def: $vgpr42
-; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
+; GFX9-NEXT:    v_or_b32_sdwa v9, v40, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v46 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v23, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_or_b32_sdwa v12, v38, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v58, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v45, v39 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v53, v21 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr49
+; GFX9-NEXT:    ; implicit-def: $vgpr23
+; GFX9-NEXT:    ; implicit-def: $vgpr38
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
+; GFX9-NEXT:    ; implicit-def: $vgpr45
+; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    ; implicit-def: $vgpr46
-; GFX9-NEXT:    ; implicit-def: $vgpr43
-; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr42
 ; GFX9-NEXT:    ; implicit-def: $vgpr54
-; GFX9-NEXT:    ; implicit-def: $vgpr27
-; GFX9-NEXT:    ; implicit-def: $vgpr23
+; GFX9-NEXT:    ; implicit-def: $vgpr50
+; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr21
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
@@ -56346,18 +56390,18 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v20 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v26 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v30 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v60 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr17
 ; GFX9-NEXT:    ; implicit-def: $vgpr20
-; GFX9-NEXT:    ; implicit-def: $vgpr26
-; GFX9-NEXT:    ; implicit-def: $vgpr38
-; GFX9-NEXT:    ; implicit-def: $vgpr57
+; GFX9-NEXT:    ; implicit-def: $vgpr24
+; GFX9-NEXT:    ; implicit-def: $vgpr30
+; GFX9-NEXT:    ; implicit-def: $vgpr60
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr62
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -56387,23 +56431,23 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr28
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v8, v52, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v43, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v9, v50, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v52, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v29, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v63, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v26, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v48, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v61, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v44, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v47, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v39, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v41, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v19, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v19, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
@@ -56437,24 +56481,24 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; kill: killed $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    ; implicit-def: $vgpr52
-; GFX9-NEXT:    ; implicit-def: $vgpr50
-; GFX9-NEXT:    ; implicit-def: $vgpr49
-; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr48
-; GFX9-NEXT:    ; implicit-def: $vgpr44
-; GFX9-NEXT:    ; implicit-def: $vgpr39
+; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr26
+; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr47
+; GFX9-NEXT:    ; implicit-def: $vgpr41
 ; GFX9-NEXT:    ; implicit-def: $vgpr19
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
-; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
-; GFX9-NEXT:    ; implicit-def: $vgpr45
-; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr55
-; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr51
+; GFX9-NEXT:    ; implicit-def: $vgpr48
 ; GFX9-NEXT:    ; implicit-def: $vgpr25
-; GFX9-NEXT:    ; implicit-def: $vgpr21
+; GFX9-NEXT:    ; implicit-def: $vgpr27
 ; GFX9-NEXT:  .LBB86_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB86_4
@@ -56464,28 +56508,28 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_movk_i32 s6, 0x300
-; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v51
-; GFX9-NEXT:    v_or_b32_sdwa v9, v58, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(14)
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v40
+; GFX9-NEXT:    v_or_b32_sdwa v9, v57, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v24
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v49
 ; GFX9-NEXT:    v_or_b32_sdwa v10, v46, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v30
-; GFX9-NEXT:    v_or_b32_sdwa v11, v43, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v23
+; GFX9-NEXT:    v_or_b32_sdwa v11, v42, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v60
-; GFX9-NEXT:    v_or_b32_sdwa v12, v40, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(5)
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v38
+; GFX9-NEXT:    v_or_b32_sdwa v12, v54, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
-; GFX9-NEXT:    v_or_b32_sdwa v13, v54, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v58
+; GFX9-NEXT:    v_or_b32_sdwa v13, v50, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v42
-; GFX9-NEXT:    v_or_b32_sdwa v14, v27, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v45
+; GFX9-NEXT:    v_or_b32_sdwa v14, v39, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v53
-; GFX9-NEXT:    v_or_b32_sdwa v15, v23, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
@@ -56522,18 +56566,18 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
-; GFX9-NEXT:    v_or_b32_sdwa v6, v26, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v6, v24, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT:    v_or_b32_sdwa v7, v38, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v7, v30, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    v_add_u16_e32 v16, 3, v17
-; GFX9-NEXT:    v_or_b32_sdwa v16, v57, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v16, v60, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
-; GFX9-NEXT:    v_or_b32_sdwa v8, v61, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v8, v62, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
@@ -56571,39 +56615,39 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
-; GFX9-NEXT:    v_or_b32_sdwa v7, v62, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v7, v63, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
 ; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX9-NEXT:    v_add_u16_e32 v8, 3, v52
+; GFX9-NEXT:    v_add_u16_e32 v8, 3, v43
 ; GFX9-NEXT:    v_or_b32_sdwa v8, v59, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
 ; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX9-NEXT:    v_add_u16_e32 v9, 3, v50
+; GFX9-NEXT:    v_add_u16_e32 v9, 3, v52
 ; GFX9-NEXT:    v_or_b32_sdwa v9, v56, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
 ; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX9-NEXT:    v_add_u16_e32 v10, 3, v49
-; GFX9-NEXT:    v_or_b32_sdwa v10, v45, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v10, 3, v29
+; GFX9-NEXT:    v_or_b32_sdwa v10, v44, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
 ; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX9-NEXT:    v_add_u16_e32 v11, 3, v63
-; GFX9-NEXT:    v_or_b32_sdwa v11, v41, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v11, 3, v26
+; GFX9-NEXT:    v_or_b32_sdwa v11, v55, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
 ; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX9-NEXT:    v_add_u16_e32 v12, 3, v48
-; GFX9-NEXT:    v_or_b32_sdwa v12, v55, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v12, 3, v61
+; GFX9-NEXT:    v_or_b32_sdwa v12, v51, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
 ; GFX9-NEXT:    v_or_b32_e32 v12, v12, v13
-; GFX9-NEXT:    v_add_u16_e32 v13, 3, v44
-; GFX9-NEXT:    v_or_b32_sdwa v13, v29, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
+; GFX9-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
 ; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX9-NEXT:    v_add_u16_e32 v14, 3, v39
+; GFX9-NEXT:    v_add_u16_e32 v14, 3, v41
 ; GFX9-NEXT:    v_or_b32_sdwa v14, v25, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
 ; GFX9-NEXT:    v_add_u16_e32 v15, 3, v19
-; GFX9-NEXT:    v_or_b32_sdwa v15, v21, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v15, v27, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
 ; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
 ; GFX9-NEXT:  .LBB86_4: ; %end
@@ -69144,9 +69188,9 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
 ; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:56
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:64
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:48
 ; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:88
@@ -69159,394 +69203,389 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:40
 ; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:72
 ; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:92
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 8, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v29
-; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 24, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 24, v27
 ; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr17
 ; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr19
 ; SI-NEXT:    ; implicit-def: $vgpr21
 ; SI-NEXT:    ; implicit-def: $vgpr53
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr27
-; SI-NEXT:    ; implicit-def: $vgpr29
 ; SI-NEXT:    ; implicit-def: $vgpr55
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:124
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:124
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v4
 ; SI-NEXT:    v_lshlrev_b32_e32 v20, 8, v10
-; SI-NEXT:    v_lshlrev_b32_e32 v22, 24, v18
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 24, v26
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v30
+; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v30, 24, v22
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v31
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v32
-; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v33
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v34
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v35
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v36
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v35
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 24, v37
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v36
+; SI-NEXT:    v_lshlrev_b32_e32 v61, 24, v38
+; SI-NEXT:    v_lshlrev_b32_e32 v22, 8, v25
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v30, 8, v25
-; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v39
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v48
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v49
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v50
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v48
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr25
 ; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:84
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v9
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT:    ; implicit-def: $vgpr1
-; SI-NEXT:    ; implicit-def: $vgpr9
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v13
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v7
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v15
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v11
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v13
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v9
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v15
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v7
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v11
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v23
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v23
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:28
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:12
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v17
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:12
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v29
+; SI-NEXT:    s_waitcnt expcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v49
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v38
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v50
+; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    ; implicit-def: $vgpr7
+; SI-NEXT:    ; kill: killed $vgpr3
 ; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    ; kill: killed $vgpr7
+; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr7
+; SI-NEXT:    ; implicit-def: $vgpr9
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr11
 ; SI-NEXT:    ; implicit-def: $vgpr13
 ; SI-NEXT:    ; implicit-def: $vgpr15
-; SI-NEXT:    ; implicit-def: $vgpr19
+; SI-NEXT:    ; implicit-def: $vgpr17
 ; SI-NEXT:    ; implicit-def: $vgpr23
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr29
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB98_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v5, v5, v12
-; SI-NEXT:    v_and_b32_e32 v23, 0xff, v46
-; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
-; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_or_b32_e32 v1, v1, v3
+; SI-NEXT:    v_and_b32_e32 v21, 0xff, v58
+; SI-NEXT:    v_or_b32_e32 v21, v21, v26
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    v_or_b32_e32 v0, v0, v44
+; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_or_b32_e32 v7, v3, v7
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
+; SI-NEXT:    v_or_b32_e32 v26, v26, v27
+; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v3, v7, v3
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    v_or_b32_e32 v48, v1, v3
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v3, v9, v3
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v48, v7, v3
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT:    v_or_b32_e32 v11, v11, v9
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    v_or_b32_e32 v11, v9, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v37, v26, v11
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v9, v7, v9
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    v_or_b32_e32 v9, v9, v13
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v7, v7, v1
+; SI-NEXT:    v_or_b32_e32 v7, v13, v7
 ; SI-NEXT:    v_or_b32_e32 v49, v9, v7
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    v_alignbit_b32 v1, v48, v11, 16
+; SI-NEXT:    v_alignbit_b32 v13, v48, v11, 16
+; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v15, v13, v9
 ; SI-NEXT:    v_alignbit_b32 v9, v49, v15, 16
 ; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; SI-NEXT:    v_or_b32_e32 v11, v11, v26
+; SI-NEXT:    ; implicit-def: $vgpr26
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_or_b32_e32 v35, v11, v15
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_or_b32_e32 v8, v9, v8
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; SI-NEXT:    v_or_b32_e32 v9, v9, v40
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_or_b32_e32 v19, v17, v13
+; SI-NEXT:    v_or_b32_e32 v50, v9, v19
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_or_b32_e32 v11, v11, v15
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; SI-NEXT:    v_or_b32_e32 v6, v13, v6
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; SI-NEXT:    v_or_b32_e32 v17, v17, v20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v19, v13, v9
-; SI-NEXT:    v_or_b32_e32 v50, v8, v19
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    v_alignbit_b32 v9, v50, v0, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_or_b32_e32 v6, v8, v6
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; SI-NEXT:    v_or_b32_e32 v8, v16, v8
-; SI-NEXT:    v_or_b32_e32 v51, v6, v8
+; SI-NEXT:    v_or_b32_e32 v23, v63, v9
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_alignbit_b32 v9, v50, v23, 16
+; SI-NEXT:    ; implicit-def: $vgpr63
+; SI-NEXT:    v_or_b32_e32 v33, v11, v23
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
+; SI-NEXT:    v_or_b32_e32 v16, v16, v13
+; SI-NEXT:    v_or_b32_e32 v51, v6, v16
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v16, 0xff, v43
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; SI-NEXT:    v_or_b32_e32 v16, v22, v16
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v28
-; SI-NEXT:    v_or_b32_e32 v22, v22, v60
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; SI-NEXT:    v_or_b32_e32 v14, v14, v20
+; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v16
+; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    v_or_b32_e32 v52, v17, v14
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v46
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; SI-NEXT:    v_or_b32_e32 v20, v30, v17
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    v_alignbit_b32 v17, v52, v20, 16
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; SI-NEXT:    v_or_b32_e32 v11, v11, v22
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT:    v_or_b32_e32 v6, v14, v6
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v57
-; SI-NEXT:    v_or_b32_e32 v14, v14, v20
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    v_or_b32_e32 v52, v14, v16
-; SI-NEXT:    v_and_b32_e32 v14, 0xff, v45
-; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; SI-NEXT:    v_or_b32_e32 v14, v18, v14
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v6, v24, v6
 ; SI-NEXT:    v_alignbit_b32 v13, v51, v6, 16
-; SI-NEXT:    v_alignbit_b32 v17, v52, v14, 16
-; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; SI-NEXT:    v_or_b32_e32 v18, v18, v26
-; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v10, v10, v20
-; SI-NEXT:    v_or_b32_e32 v53, v18, v10
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v20, 0xff, v40
-; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; SI-NEXT:    v_or_b32_e32 v20, v42, v20
-; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v20
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr26
-; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; SI-NEXT:    v_or_b32_e32 v2, v2, v18
-; SI-NEXT:    v_and_b32_e32 v18, 0xff, v24
-; SI-NEXT:    v_or_b32_e32 v24, v58, v23
-; SI-NEXT:    v_or_b32_e32 v55, v22, v24
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    v_or_b32_e32 v18, v18, v44
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v54, v18, v20
-; SI-NEXT:    v_and_b32_e32 v18, 0xff, v61
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; SI-NEXT:    v_or_b32_e32 v18, v56, v18
+; SI-NEXT:    v_and_b32_e32 v24, 0xff, v47
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
+; SI-NEXT:    v_or_b32_e32 v10, v10, v24
+; SI-NEXT:    v_or_b32_e32 v53, v21, v10
+; SI-NEXT:    v_and_b32_e32 v21, 0xff, v57
+; SI-NEXT:    v_and_b32_e32 v24, 0xff, v62
+; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
+; SI-NEXT:    v_or_b32_e32 v2, v2, v21
+; SI-NEXT:    v_or_b32_e32 v12, v12, v24
 ; SI-NEXT:    v_alignbit_b32 v21, v53, v2, 16
-; SI-NEXT:    v_alignbit_b32 v25, v54, v18, 16
-; SI-NEXT:    v_alignbit_b32 v29, v55, v5, 16
+; SI-NEXT:    v_or_b32_e32 v54, v0, v12
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v8
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v28
+; SI-NEXT:    v_and_b32_e32 v24, 0xff, v43
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; SI-NEXT:    v_or_b32_e32 v8, v8, v60
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
+; SI-NEXT:    v_or_b32_e32 v0, v56, v0
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; SI-NEXT:    v_or_b32_e32 v24, v59, v24
+; SI-NEXT:    v_alignbit_b32 v25, v54, v0, 16
+; SI-NEXT:    v_or_b32_e32 v55, v8, v24
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v41
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_or_b32_e32 v8, v61, v8
+; SI-NEXT:    v_alignbit_b32 v29, v55, v8, 16
 ; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v10
+; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
 ; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v24
-; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr28
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr24
-; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr10
-; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v12, v12, v22
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; SI-NEXT:    v_or_b32_e32 v37, v12, v11
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; SI-NEXT:    v_or_b32_e32 v35, v11, v15
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v8
 ; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr8
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v11, v63
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; SI-NEXT:    v_or_b32_e32 v33, v11, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    v_or_b32_e32 v32, v11, v6
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v19
-; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr63
+; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
+; SI-NEXT:    ; implicit-def: $vgpr14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_or_b32_e32 v0, v0, v30
-; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; SI-NEXT:    v_or_b32_e32 v32, v0, v6
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; SI-NEXT:    v_or_b32_e32 v4, v6, v4
 ; SI-NEXT:    ; implicit-def: $vgpr6
-; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_or_b32_e32 v34, v4, v20
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_or_b32_e32 v0, v0, v4
-; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; SI-NEXT:    v_or_b32_e32 v34, v0, v14
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr4
-; SI-NEXT:    ; kill: killed $vgpr4
-; SI-NEXT:    ; implicit-def: $vgpr4
-; SI-NEXT:    ; implicit-def: $vgpr14
-; SI-NEXT:    ; kill: killed $vgpr4
+; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    v_or_b32_e32 v4, v4, v45
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; SI-NEXT:    v_or_b32_e32 v36, v4, v2
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v42
+; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr4
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_or_b32_e32 v0, v0, v47
-; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; SI-NEXT:    v_or_b32_e32 v36, v0, v2
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v2, v2, v5
+; SI-NEXT:    ; implicit-def: $vgpr5
+; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; SI-NEXT:    v_or_b32_e32 v38, v2, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v18
+; SI-NEXT:    ; implicit-def: $vgpr18
 ; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_or_b32_e32 v0, v0, v59
-; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; SI-NEXT:    v_or_b32_e32 v38, v0, v18
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v41
-; SI-NEXT:    v_or_b32_e32 v0, v0, v62
+; SI-NEXT:    v_or_b32_e32 v0, v0, v1
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; SI-NEXT:    v_or_b32_e32 v39, v0, v5
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    v_or_b32_e32 v39, v0, v8
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
+; SI-NEXT:    ; implicit-def: $vgpr8
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    ; kill: killed $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
@@ -69568,25 +69607,21 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr18
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:  .LBB98_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB98_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v41
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v12
-; SI-NEXT:    v_or_b32_e32 v1, v62, v1
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v18
+; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; SI-NEXT:    v_or_b32_e32 v1, v1, v3
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v41
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
-; SI-NEXT:    v_or_b32_e32 v3, v5, v3
+; SI-NEXT:    v_or_b32_e32 v3, v61, v3
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v28
@@ -69596,156 +69631,155 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
 ; SI-NEXT:    v_add_i32_e32 v39, vcc, s7, v1
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v46
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v43
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v3, v58, v3
+; SI-NEXT:    v_or_b32_e32 v3, v59, v3
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
 ; SI-NEXT:    v_add_i32_e32 v55, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v61
-; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_or_b32_e32 v3, v56, v3
-; SI-NEXT:    v_alignbit_b32 v29, v55, v39, 16
-; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v55
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v59, v1
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v3, v1
-; SI-NEXT:    v_add_i32_e32 v38, vcc, s7, v1
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v24
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v42
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v40
-; SI-NEXT:    v_or_b32_e32 v1, v44, v1
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v8
+; SI-NEXT:    v_or_b32_e32 v1, v5, v1
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v3, v42, v3
+; SI-NEXT:    v_or_b32_e32 v3, v56, v3
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v1
-; SI-NEXT:    v_add_i32_e32 v54, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
+; SI-NEXT:    v_add_i32_e32 v38, vcc, s7, v1
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v62
+; SI-NEXT:    v_or_b32_e32 v0, v44, v0
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    v_or_b32_e32 v1, v12, v1
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
+; SI-NEXT:    v_add_i32_e32 v54, vcc, s7, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v57
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v1, v2, v1
 ; SI-NEXT:    v_alignbit_b32 v25, v54, v38, 16
+; SI-NEXT:    v_alignbit_b32 v29, v55, v39, 16
 ; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v54
+; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v55
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    v_or_b32_e32 v0, v45, v0
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
+; SI-NEXT:    v_add_i32_e32 v36, vcc, s7, v0
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v58
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v47
+; SI-NEXT:    v_or_b32_e32 v0, v26, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_or_b32_e32 v1, v47, v1
-; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v2, v3
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v36, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    v_or_b32_e32 v1, v10, v1
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
+; SI-NEXT:    v_add_i32_e32 v53, vcc, s7, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v46
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_or_b32_e32 v1, v26, v1
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v10, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v53, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v45
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v2, v18, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v1, v30, v1
 ; SI-NEXT:    v_alignbit_b32 v21, v53, v36, 16
 ; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v53
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    v_or_b32_e32 v0, v4, v0
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
+; SI-NEXT:    v_add_i32_e32 v34, vcc, s7, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_or_b32_e32 v0, v20, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v4, v1
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v34, vcc, s7, v1
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v57
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v43
-; SI-NEXT:    v_or_b32_e32 v1, v20, v1
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v22, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v52, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    v_or_b32_e32 v1, v14, v1
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
+; SI-NEXT:    v_add_i32_e32 v52, vcc, s7, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; SI-NEXT:    v_alignbit_b32 v17, v52, v34, 16
 ; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v52
 ; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_or_b32_e32 v0, v22, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_or_b32_e32 v1, v30, v1
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v14, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v32, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    v_or_b32_e32 v1, v24, v1
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
+; SI-NEXT:    v_add_i32_e32 v32, vcc, s7, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_or_b32_e32 v0, v6, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_or_b32_e32 v1, v6, v1
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v2, v16, v2
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    v_add_i32_e32 v51, vcc, s7, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    v_or_b32_e32 v1, v16, v1
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
+; SI-NEXT:    v_add_i32_e32 v51, vcc, s7, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; SI-NEXT:    v_alignbit_b32 v13, v51, v32, 16
 ; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v51
 ; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v63, v1
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v0, v0, v2
-; SI-NEXT:    v_or_b32_e32 v0, v0, v1
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v33, vcc, s7, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
-; SI-NEXT:    v_or_b32_e32 v0, v8, v0
+; SI-NEXT:    v_or_b32_e32 v0, v40, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
@@ -69774,8 +69808,8 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v35, vcc, s7, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
@@ -69792,8 +69826,8 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v49, vcc, s7, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v49
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
@@ -69811,7 +69845,7 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v37, vcc, s7, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
@@ -69828,12 +69862,15 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v2, v1
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_add_i32_e32 v48, vcc, s7, v0
-; SI-NEXT:    v_alignbit_b32 v1, v48, v37, 16
+; SI-NEXT:    v_alignbit_b32 v0, v48, v37, 16
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_alignbit_b32 v0, v49, v35, 16
 ; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v48
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
 ; SI-NEXT:  .LBB98_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
@@ -69861,6 +69898,7 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v12, v32
 ; SI-NEXT:    v_mov_b32_e32 v14, v51
 ; SI-NEXT:    v_mov_b32_e32 v16, v34
+; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_mov_b32_e32 v18, v52
 ; SI-NEXT:    v_mov_b32_e32 v20, v36
 ; SI-NEXT:    v_mov_b32_e32 v22, v53
@@ -69906,78 +69944,78 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:88
+; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:56
+; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:64
+; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:72
+; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:80
+; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:88
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:96
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
 ; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:4
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:116
-; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:124
-; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v1
-; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v3
-; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v5
-; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v7
-; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v9
-; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v11
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v13
-; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v15
+; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v27
+; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v29
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:116
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v1
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v5
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v7
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v9
+; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
+; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v13
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v15
 ; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
 ; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
 ; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
 ; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
-; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v24
-; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v0
-; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v2
-; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v4
-; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v8
-; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v10
-; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v12
-; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v14
-; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v16
-; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v18
-; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v20
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v4
+; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
+; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v10
+; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v12
+; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v14
+; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v16
+; VI-NEXT:    v_lshlrev_b16_e32 v61, 8, v18
+; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v20
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v22
+; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v22
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v26
+; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v24
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v28
+; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v26
+; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:124
 ; VI-NEXT:    s_waitcnt vmcnt(10)
-; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v30
+; VI-NEXT:    v_lshlrev_b16_e32 v36, 8, v30
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v31
+; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v31
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:108
-; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:92
+; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:100
+; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:84
 ; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:68
 ; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:52
+; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
@@ -69989,79 +70027,80 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT:    v_or_b32_sdwa v9, v39, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v10, v20, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v28, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v22, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v30, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v16, v38, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr39
-; VI-NEXT:    ; implicit-def: $vgpr20
-; VI-NEXT:    ; implicit-def: $vgpr28
+; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_or_b32_sdwa v9, v29, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v41, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v46, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_or_b32_sdwa v12, v22, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v30, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v34, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v31, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v16, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr46
 ; VI-NEXT:    ; implicit-def: $vgpr22
 ; VI-NEXT:    ; implicit-def: $vgpr30
-; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr34
 ; VI-NEXT:    ; implicit-def: $vgpr31
-; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    ; implicit-def: $vgpr57
 ; VI-NEXT:    ; implicit-def: $vgpr59
-; VI-NEXT:    ; implicit-def: $vgpr62
-; VI-NEXT:    ; implicit-def: $vgpr18
-; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr24
+; VI-NEXT:    ; implicit-def: $vgpr36
 ; VI-NEXT:    ; implicit-def: $vgpr32
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_or_b32_sdwa v0, v0, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_or_b32_sdwa v1, v1, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v1, v1, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr48
+; VI-NEXT:    ; implicit-def: $vgpr49
 ; VI-NEXT:    ; implicit-def: $vgpr50
-; VI-NEXT:    ; implicit-def: $vgpr51
-; VI-NEXT:    ; implicit-def: $vgpr52
-; VI-NEXT:    ; implicit-def: $vgpr53
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr19
 ; VI-NEXT:    ; implicit-def: $vgpr23
-; VI-NEXT:    ; implicit-def: $vgpr27
+; VI-NEXT:    ; implicit-def: $vgpr40
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_or_b32_sdwa v8, v8, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr45
+; VI-NEXT:    v_or_b32_sdwa v8, v8, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr43
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_or_b32_sdwa v4, v4, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v4, v4, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr52
 ; VI-NEXT:    ; implicit-def: $vgpr54
-; VI-NEXT:    ; implicit-def: $vgpr40
-; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr53
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -70078,23 +70117,23 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr25
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v49, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v48, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v55, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v55, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v44, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v36, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v18, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v43, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v26, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v37, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v37, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v44, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v35, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v61, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
@@ -70128,98 +70167,92 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; kill: killed $vgpr16
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr49
-; VI-NEXT:    ; implicit-def: $vgpr48
+; VI-NEXT:    ; implicit-def: $vgpr27
 ; VI-NEXT:    ; implicit-def: $vgpr55
-; VI-NEXT:    ; implicit-def: $vgpr36
-; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr37
 ; VI-NEXT:    ; implicit-def: $vgpr44
-; VI-NEXT:    ; implicit-def: $vgpr61
-; VI-NEXT:    ; implicit-def: $vgpr29
-; VI-NEXT:    ; implicit-def: $vgpr46
+; VI-NEXT:    ; implicit-def: $vgpr18
+; VI-NEXT:    ; implicit-def: $vgpr26
+; VI-NEXT:    ; implicit-def: $vgpr37
+; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr16
+; VI-NEXT:    ; implicit-def: $vgpr42
+; VI-NEXT:    ; implicit-def: $vgpr45
 ; VI-NEXT:    ; implicit-def: $vgpr56
 ; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr24
-; VI-NEXT:    ; implicit-def: $vgpr26
-; VI-NEXT:    ; implicit-def: $vgpr34
+; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; implicit-def: $vgpr20
+; VI-NEXT:    ; implicit-def: $vgpr28
+; VI-NEXT:    ; implicit-def: $vgpr33
 ; VI-NEXT:  .LBB98_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB98_4
 ; VI-NEXT:  ; %bb.3: ; %cmp.true
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v38
-; VI-NEXT:    v_add_u16_e32 v2, 3, v44
+; VI-NEXT:    v_add_u16_e32 v0, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v0, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v14, v26, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_add_u16_e32 v2, 3, v35
 ; VI-NEXT:    v_mov_b32_e32 v3, 0x300
-; VI-NEXT:    v_or_b32_sdwa v2, v18, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_sdwa v18, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v37
-; VI-NEXT:    v_or_b32_sdwa v24, v24, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v16, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_add_u16_e32 v0, 3, v37
+; VI-NEXT:    v_or_b32_sdwa v20, v20, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v30
-; VI-NEXT:    v_or_b32_sdwa v0, v16, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v38, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v12, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v43
-; VI-NEXT:    v_or_b32_sdwa v16, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v22
-; VI-NEXT:    v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v36
-; VI-NEXT:    v_or_b32_sdwa v22, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v26
+; VI-NEXT:    v_or_b32_sdwa v26, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v28
+; VI-NEXT:    v_add_u16_e32 v0, 3, v22
+; VI-NEXT:    v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_add_u16_e32 v0, 3, v18
+; VI-NEXT:    v_or_b32_sdwa v18, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v10, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v55
-; VI-NEXT:    v_or_b32_sdwa v28, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v20
+; VI-NEXT:    v_add_u16_e32 v0, 3, v44
+; VI-NEXT:    v_or_b32_sdwa v22, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v41
 ; VI-NEXT:    v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v2, 3, v35
 ; VI-NEXT:    v_add_u16_sdwa v9, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v48
-; VI-NEXT:    v_or_b32_sdwa v20, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v0, 3, v39
+; VI-NEXT:    v_add_u16_e32 v0, 3, v55
+; VI-NEXT:    v_or_b32_sdwa v14, v28, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v28, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v29
 ; VI-NEXT:    v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v49
-; VI-NEXT:    v_or_b32_sdwa v30, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v27
+; VI-NEXT:    v_or_b32_sdwa v27, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v2, 3, v34
+; VI-NEXT:    v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v13, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v1, 3, v61
-; VI-NEXT:    v_or_b32_sdwa v15, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v1, 3, v63
+; VI-NEXT:    v_or_b32_sdwa v15, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v1, 3, v31
-; VI-NEXT:    v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_sdwa v26, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT:    v_or_b32_e32 v12, v16, v12
-; VI-NEXT:    v_add_u16_e32 v16, 0x300, v24
+; VI-NEXT:    v_or_b32_sdwa v1, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v24, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT:    v_or_b32_e32 v13, v16, v13
-; VI-NEXT:    v_or_b32_e32 v14, v14, v26
-; VI-NEXT:    v_or_b32_e32 v15, v15, v18
+; VI-NEXT:    v_or_b32_e32 v14, v14, v24
+; VI-NEXT:    v_or_b32_e32 v15, v15, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v29, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v29, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -70252,54 +70285,58 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v6, v17, v6
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v29
 ; VI-NEXT:    v_or_b32_e32 v7, v17, v7
-; VI-NEXT:    v_add_u16_e32 v17, 0x300, v30
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v27
 ; VI-NEXT:    v_or_b32_e32 v8, v17, v8
-; VI-NEXT:    v_add_u16_e32 v17, 0x300, v20
-; VI-NEXT:    v_or_b32_e32 v9, v17, v9
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v28
-; VI-NEXT:    v_or_b32_e32 v10, v17, v10
+; VI-NEXT:    v_or_b32_e32 v9, v17, v9
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v22
+; VI-NEXT:    v_or_b32_e32 v10, v17, v10
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v18
 ; VI-NEXT:    v_or_b32_e32 v11, v17, v11
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v26
+; VI-NEXT:    v_or_b32_e32 v12, v17, v12
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v20
+; VI-NEXT:    v_or_b32_e32 v13, v17, v13
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v19, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v23, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v23, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v2, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v27, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v30, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v1, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v31, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v31, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
 ; VI-NEXT:    v_or_b32_e32 v0, v3, v0
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v31
 ; VI-NEXT:    v_or_b32_e32 v1, v3, v1
-; VI-NEXT:    v_add_u16_e32 v3, 0x300, v27
+; VI-NEXT:    v_add_u16_e32 v3, 0x300, v30
 ; VI-NEXT:    v_or_b32_e32 v2, v3, v2
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v23
 ; VI-NEXT:    v_or_b32_e32 v3, v3, v19
@@ -70359,99 +70396,100 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:80
+; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:56
+; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:64
+; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:72
+; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:80
 ; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:88
 ; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:96
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:104
-; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:112
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v1
-; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v3
-; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v5
-; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v7
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v9
-; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v11
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v13
-; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v15
-; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v17
+; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:112
+; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v17
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v19
-; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v23
-; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
+; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v25
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
-; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v29
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:116
+; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v1
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v3
+; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v5
+; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v7
+; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v9
+; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v11
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v13
+; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v15
+; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v22
+; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v6
 ; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v4
+; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(20)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
+; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(19)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
+; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v12
 ; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v10
+; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v14
 ; GFX9-NEXT:    s_waitcnt vmcnt(17)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v12
+; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v14
+; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v16
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v18
+; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v20
+; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v24
-; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:116
+; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v28
+; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v31
+; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v32
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:108
-; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:92
-; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
+; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v31
+; GFX9-NEXT:    s_waitcnt vmcnt(8)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v32
+; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:100
+; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:92
+; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:84
+; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB98_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
@@ -70459,108 +70497,107 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v39, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v63, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v35, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v34, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v16, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v32, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v16, v38, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr39
-; GFX9-NEXT:    ; implicit-def: $vgpr62
-; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr35
-; GFX9-NEXT:    ; implicit-def: $vgpr34
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr38
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
+; GFX9-NEXT:    v_or_b32_sdwa v9, v25, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v41, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(10)
+; GFX9-NEXT:    v_or_b32_sdwa v12, v26, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v37, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v20, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr25
+; GFX9-NEXT:    ; implicit-def: $vgpr27
+; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr26
+; GFX9-NEXT:    ; implicit-def: $vgpr37
+; GFX9-NEXT:    ; implicit-def: $vgpr20
+; GFX9-NEXT:    ; implicit-def: $vgpr31
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
-; GFX9-NEXT:    ; implicit-def: $vgpr36
-; GFX9-NEXT:    ; implicit-def: $vgpr26
-; GFX9-NEXT:    ; implicit-def: $vgpr20
-; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; implicit-def: $vgpr18
+; GFX9-NEXT:    ; implicit-def: $vgpr30
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
 ; GFX9-NEXT:    v_perm_b32 v1, v3, v2, s6
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr49
 ; GFX9-NEXT:    ; implicit-def: $vgpr48
-; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    ; implicit-def: $vgpr39
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
-; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr49
+; GFX9-NEXT:    ; implicit-def: $vgpr53
 ; GFX9-NEXT:    ; implicit-def: $vgpr17
 ; GFX9-NEXT:    ; implicit-def: $vgpr19
 ; GFX9-NEXT:    ; implicit-def: $vgpr23
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr27
+; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr55
-; GFX9-NEXT:    ; implicit-def: $vgpr54
+; GFX9-NEXT:    ; implicit-def: $vgpr52
+; GFX9-NEXT:    ; implicit-def: $vgpr51
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr43
+; GFX9-NEXT:    ; implicit-def: $vgpr55
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr25
+; GFX9-NEXT:    ; implicit-def: $vgpr40
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT:    v_or_b32_sdwa v8, v51, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v29, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT:    v_or_b32_sdwa v9, v52, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v42, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT:    v_or_b32_sdwa v10, v42, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v47, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT:    v_or_b32_sdwa v11, v37, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v35, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT:    v_or_b32_sdwa v12, v44, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v16, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT:    v_or_b32_sdwa v13, v28, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v24, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT:    v_or_b32_sdwa v14, v47, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v32, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT:    v_or_b32_sdwa v15, v18, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v38, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v16, v36, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
@@ -70594,111 +70631,110 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; kill: killed $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
-; GFX9-NEXT:    ; implicit-def: $vgpr51
-; GFX9-NEXT:    ; implicit-def: $vgpr52
+; GFX9-NEXT:    ; implicit-def: $vgpr29
 ; GFX9-NEXT:    ; implicit-def: $vgpr42
-; GFX9-NEXT:    ; implicit-def: $vgpr37
-; GFX9-NEXT:    ; implicit-def: $vgpr44
-; GFX9-NEXT:    ; implicit-def: $vgpr28
-; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr47
-; GFX9-NEXT:    ; implicit-def: $vgpr18
-; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr35
+; GFX9-NEXT:    ; implicit-def: $vgpr16
+; GFX9-NEXT:    ; implicit-def: $vgpr24
+; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; implicit-def: $vgpr38
+; GFX9-NEXT:    ; implicit-def: $vgpr36
+; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; implicit-def: $vgpr22
-; GFX9-NEXT:    ; implicit-def: $vgpr30
-; GFX9-NEXT:    ; implicit-def: $vgpr24
-; GFX9-NEXT:    ; implicit-def: $vgpr31
+; GFX9-NEXT:    ; implicit-def: $vgpr34
+; GFX9-NEXT:    ; implicit-def: $vgpr28
+; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:  .LBB98_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB98_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
-; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v18
-; GFX9-NEXT:    v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(9)
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v38
+; GFX9-NEXT:    v_or_b32_sdwa v0, v28, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v28
-; GFX9-NEXT:    v_or_b32_sdwa v0, v30, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v24
+; GFX9-NEXT:    v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v20
+; GFX9-NEXT:    v_or_b32_sdwa v0, v18, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v0
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v16
-; GFX9-NEXT:    v_or_b32_sdwa v0, v26, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v44
-; GFX9-NEXT:    v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v34
-; GFX9-NEXT:    v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v37
+; GFX9-NEXT:    v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v35
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v35
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v26
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v42
+; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v47
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v63
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v41
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v52
+; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v42
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v27
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v51
+; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v29
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v39
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v25
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT:    v_add_u16_e32 v3, 3, v32
-; GFX9-NEXT:    v_or_b32_sdwa v3, v20, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v3
+; GFX9-NEXT:    v_add_u16_e32 v3, 3, v31
+; GFX9-NEXT:    v_or_b32_sdwa v3, v30, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v3
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v32
+; GFX9-NEXT:    v_or_b32_sdwa v2, v34, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v2
-; GFX9-NEXT:    v_add_u16_e32 v1, 3, v38
+; GFX9-NEXT:    v_add_u16_e32 v1, 3, v36
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v1
+; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v1
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
-; GFX9-NEXT:    v_perm_b32 v8, v30, v8, s6
-; GFX9-NEXT:    v_perm_b32 v9, v28, v9, s6
-; GFX9-NEXT:    v_perm_b32 v10, v26, v10, s6
-; GFX9-NEXT:    v_perm_b32 v11, v24, v11, s6
-; GFX9-NEXT:    v_perm_b32 v12, v22, v12, s6
-; GFX9-NEXT:    v_perm_b32 v13, v16, v13, s6
-; GFX9-NEXT:    v_perm_b32 v14, v20, v14, s6
-; GFX9-NEXT:    v_perm_b32 v15, v18, v15, s6
+; GFX9-NEXT:    v_perm_b32 v8, v25, v8, s6
+; GFX9-NEXT:    v_perm_b32 v9, v24, v9, s6
+; GFX9-NEXT:    v_perm_b32 v10, v22, v10, s6
+; GFX9-NEXT:    v_perm_b32 v11, v20, v11, s6
+; GFX9-NEXT:    v_perm_b32 v12, v16, v12, s6
+; GFX9-NEXT:    v_perm_b32 v13, v18, v13, s6
+; GFX9-NEXT:    v_perm_b32 v14, v30, v14, s6
+; GFX9-NEXT:    v_perm_b32 v15, v28, v15, s6
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v44, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT:    v_or_b32_sdwa v31, v48, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v31, v39, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT:    v_perm_b32 v7, v27, v7, s6
+; GFX9-NEXT:    v_perm_b32 v7, v26, v7, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -70720,7 +70756,7 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v5, v19, v5, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -70731,40 +70767,40 @@ define <32 x i16> @bitcast_v64i8_to_v32i16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v4, v17, v4, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v3, v21, v3, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT:    v_perm_b32 v2, v25, v2, s6
+; GFX9-NEXT:    v_perm_b32 v2, v27, v2, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v1, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v1, v29, v1, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; GFX9-NEXT:    v_perm_b32 v0, v31, v0, s6
 ; GFX9-NEXT:  .LBB98_4: ; %end
@@ -71630,13 +71666,13 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt expcnt(4)
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt expcnt(2)
 ; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:52
 ; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:68
 ; SI-NEXT:    v_readfirstlane_b32 s15, v27
 ; SI-NEXT:    v_readfirstlane_b32 s40, v26
 ; SI-NEXT:    v_readfirstlane_b32 s12, v19
@@ -71668,14 +71704,14 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v37
 ; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v38
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v38
 ; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v39
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_lshlrev_b32_e32 v47, 24, v48
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v59
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 24, v61
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v59, 24, v62
@@ -71768,10 +71804,9 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v43
 ; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
 ; SI-NEXT:    v_or_b32_e32 v33, v58, v32
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v32, 0xff, v46
+; SI-NEXT:    v_and_b32_e32 v32, 0xff, v45
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v34, 0xff, v45
+; SI-NEXT:    v_and_b32_e32 v34, 0xff, v46
 ; SI-NEXT:    s_or_b32 s4, s4, s56
 ; SI-NEXT:    v_or_b32_e32 v25, v54, v17
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
@@ -71789,13 +71824,13 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; SI-NEXT:    s_lshl_b32 s56, s15, 8
 ; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
 ; SI-NEXT:    v_or_b32_e32 v48, v32, v63
-; SI-NEXT:    v_and_b32_e32 v32, 0xff, v56
+; SI-NEXT:    v_and_b32_e32 v32, 0xff, v57
 ; SI-NEXT:    s_or_b32 s4, s4, s56
 ; SI-NEXT:    v_or_b32_e32 v29, v44, v21
 ; SI-NEXT:    v_and_b32_e32 v26, 0xff, v40
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
 ; SI-NEXT:    s_and_b32 s4, s4, 0xffff
-; SI-NEXT:    v_or_b32_e32 v26, v26, v57
+; SI-NEXT:    v_or_b32_e32 v26, v26, v56
 ; SI-NEXT:    v_or_b32_e32 v34, v61, v32
 ; SI-NEXT:    v_or_b32_e32 v32, s4, v29
 ; SI-NEXT:    s_and_b32 s4, s43, 0xff
@@ -71833,7 +71868,8 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; SI-NEXT:    s_add_i32 s45, s45, 3
 ; SI-NEXT:    s_and_b32 s4, s45, 0xff
 ; SI-NEXT:    s_lshl_b32 s5, s44, 8
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v56
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v57
 ; SI-NEXT:    s_or_b32 s4, s5, s4
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_addk_i32 s4, 0x300
@@ -71842,11 +71878,10 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; SI-NEXT:    v_or_b32_e32 v1, v61, v1
 ; SI-NEXT:    v_or_b32_e32 v1, s4, v1
 ; SI-NEXT:    v_add_i32_e32 v34, vcc, 0x3000000, v1
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v46
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v45
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v46
 ; SI-NEXT:    v_or_b32_e32 v1, v60, v1
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
@@ -71870,7 +71905,7 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v40
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v55
-; SI-NEXT:    v_or_b32_e32 v1, v57, v1
+; SI-NEXT:    v_or_b32_e32 v1, v56, v1
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
@@ -82280,29 +82315,53 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(6)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:132
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:120
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:128
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:128
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v7
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v3
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v5
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v7
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v9
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v11
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v13
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v15
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v23
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124
 ; SI-NEXT:    ; implicit-def: $vgpr7
 ; SI-NEXT:    ; kill: killed $vgpr7
 ; SI-NEXT:    ; implicit-def: $vgpr7
@@ -82314,21 +82373,19 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr7
 ; SI-NEXT:    ; kill: killed $vgpr7
 ; SI-NEXT:    ; implicit-def: $vgpr7
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 8, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v24, 8, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v22, 8, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 8, v29
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v61, 8, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 8, v29
 ; SI-NEXT:    ; kill: killed $vgpr7
 ; SI-NEXT:    ; implicit-def: $vgpr7
 ; SI-NEXT:    ; kill: killed $vgpr7
 ; SI-NEXT:    ; implicit-def: $vgpr7
+; SI-NEXT:    ; implicit-def: $vgpr9
 ; SI-NEXT:    ; implicit-def: $vgpr11
+; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr13
 ; SI-NEXT:    ; implicit-def: $vgpr53
 ; SI-NEXT:    ; implicit-def: $vgpr15
@@ -82336,205 +82393,104 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr19
 ; SI-NEXT:    ; implicit-def: $vgpr21
 ; SI-NEXT:    ; implicit-def: $vgpr23
+; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr25
 ; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr27
 ; SI-NEXT:    ; implicit-def: $vgpr52
 ; SI-NEXT:    ; implicit-def: $vgpr29
 ; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:124
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
-; SI-NEXT:    v_lshlrev_b32_e32 v28, 8, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v20, 8, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 8, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 8, v6
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:116
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_lshlrev_b32_e32 v12, 8, v10
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 8, v31
-; SI-NEXT:    ; implicit-def: $vgpr31
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v28, 8, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v30, 8, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v20, 8, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 8, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v14, 8, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v12, 8, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v10, 8, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 8, v22
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 8, v26
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:116
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v31
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v32
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v33
-; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v34
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v35
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v55, 8, v36
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v37
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v36
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 8, v38
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v39
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v48
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v49
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:12
-; SI-NEXT:    ; implicit-def: $vgpr49
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:100
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v3
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v5
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v9
-; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v38
-; SI-NEXT:    ; implicit-def: $vgpr9
 ; SI-NEXT:    ; implicit-def: $vgpr38
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v14, 8, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 8, v32
-; SI-NEXT:    ; implicit-def: $vgpr32
+; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB106_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v6, v14
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v6
-; SI-NEXT:    ; implicit-def: $vgpr14
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_or_b32_e32 v7, v7, v9
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v6, v12
-; SI-NEXT:    v_cvt_f32_f16_e32 v36, v6
-; SI-NEXT:    v_and_b32_e32 v6, 0xff, v39
-; SI-NEXT:    v_or_b32_e32 v6, v6, v10
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v6
-; SI-NEXT:    v_and_b32_e32 v6, 0xff, v35
-; SI-NEXT:    v_or_b32_e32 v6, v6, v8
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v6
-; SI-NEXT:    v_and_b32_e32 v6, 0xff, v44
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr10
-; SI-NEXT:    ; implicit-def: $vgpr8
-; SI-NEXT:    v_or_b32_e32 v0, v6, v0
-; SI-NEXT:    ; implicit-def: $vgpr6
-; SI-NEXT:    v_cvt_f32_f16_e32 v23, v0
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_and_b32_e32 v19, 0xff, v55
+; SI-NEXT:    v_or_b32_e32 v16, v19, v16
+; SI-NEXT:    v_cvt_f32_f16_e32 v34, v16
+; SI-NEXT:    v_and_b32_e32 v16, 0xff, v22
+; SI-NEXT:    v_or_b32_e32 v14, v16, v14
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v14
+; SI-NEXT:    v_and_b32_e32 v14, 0xff, v26
+; SI-NEXT:    v_or_b32_e32 v12, v14, v12
+; SI-NEXT:    v_cvt_f32_f16_e32 v36, v12
+; SI-NEXT:    v_and_b32_e32 v12, 0xff, v18
+; SI-NEXT:    v_or_b32_e32 v10, v12, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v10
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; SI-NEXT:    v_or_b32_e32 v17, v17, v20
-; SI-NEXT:    v_or_b32_e32 v15, v15, v26
-; SI-NEXT:    ; implicit-def: $vgpr26
-; SI-NEXT:    ; implicit-def: $vgpr20
-; SI-NEXT:    v_cvt_f32_f16_e32 v32, v17
-; SI-NEXT:    v_cvt_f32_f16_e32 v53, v15
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_or_b32_e32 v9, v9, v11
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v9
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; SI-NEXT:    v_or_b32_e32 v17, v17, v18
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_or_b32_e32 v0, v0, v2
-; SI-NEXT:    v_or_b32_e32 v15, v15, v28
-; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    ; implicit-def: $vgpr2
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v45
-; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
-; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    v_or_b32_e32 v0, v0, v41
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    v_cvt_f32_f16_e32 v25, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v3
-; SI-NEXT:    ; implicit-def: $vgpr3
-; SI-NEXT:    v_or_b32_e32 v0, v0, v55
-; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v41
+; SI-NEXT:    v_or_b32_e32 v8, v10, v8
+; SI-NEXT:    v_cvt_f32_f16_e32 v38, v8
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v42
+; SI-NEXT:    v_or_b32_e32 v8, v8, v39
+; SI-NEXT:    v_cvt_f32_f16_e32 v23, v8
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v43
+; SI-NEXT:    v_or_b32_e32 v8, v8, v40
+; SI-NEXT:    v_cvt_f32_f16_e32 v48, v8
+; SI-NEXT:    v_and_b32_e32 v8, 0xff, v44
+; SI-NEXT:    v_or_b32_e32 v4, v8, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v25, v4
+; SI-NEXT:    v_and_b32_e32 v4, 0xff, v45
+; SI-NEXT:    v_or_b32_e32 v0, v4, v0
 ; SI-NEXT:    v_cvt_f32_f16_e32 v50, v0
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5
-; SI-NEXT:    ; implicit-def: $vgpr5
-; SI-NEXT:    v_or_b32_e32 v0, v0, v43
-; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    v_or_b32_e32 v0, v0, v2
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v30
-; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    v_or_b32_e32 v0, v0, v1
-; SI-NEXT:    ; implicit-def: $vgpr1
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v47
+; SI-NEXT:    v_or_b32_e32 v0, v0, v35
 ; SI-NEXT:    v_cvt_f32_f16_e32 v52, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v51
-; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    v_or_b32_e32 v0, v0, v46
-; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v56
+; SI-NEXT:    v_or_b32_e32 v0, v0, v3
 ; SI-NEXT:    v_cvt_f32_f16_e32 v29, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v4
-; SI-NEXT:    ; implicit-def: $vgpr4
-; SI-NEXT:    v_or_b32_e32 v0, v0, v47
-; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v6
+; SI-NEXT:    v_or_b32_e32 v0, v0, v46
 ; SI-NEXT:    v_cvt_f32_f16_e32 v54, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v42
-; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    v_or_b32_e32 v0, v0, v56
-; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v1
+; SI-NEXT:    v_or_b32_e32 v0, v0, v51
 ; SI-NEXT:    v_cvt_f32_f16_e32 v31, v0
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
@@ -82577,332 +82533,389 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v60
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
+; SI-NEXT:    v_or_b32_e32 v17, v17, v30
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
+; SI-NEXT:    v_cvt_f32_f16_e32 v32, v17
+; SI-NEXT:    v_and_b32_e32 v17, 0xff, v62
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
+; SI-NEXT:    v_or_b32_e32 v17, v17, v20
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr55
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr26
+; SI-NEXT:    ; implicit-def: $vgpr18
+; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr5
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr6
+; SI-NEXT:    ; implicit-def: $vgpr1
 ; SI-NEXT:    ; kill: killed $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; implicit-def: $vgpr20
+; SI-NEXT:    ; implicit-def: $vgpr16
+; SI-NEXT:    ; implicit-def: $vgpr14
+; SI-NEXT:    ; implicit-def: $vgpr12
+; SI-NEXT:    ; implicit-def: $vgpr10
+; SI-NEXT:    ; implicit-def: $vgpr8
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr4
 ; SI-NEXT:    ; implicit-def: $vgpr0
+; SI-NEXT:    ; implicit-def: $vgpr2
+; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr3
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_or_b32_e32 v7, v7, v9
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    v_or_b32_e32 v9, v9, v11
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v9
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v13
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v11
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; SI-NEXT:    v_or_b32_e32 v16, v18, v16
-; SI-NEXT:    ; implicit-def: $vgpr18
-; SI-NEXT:    v_cvt_f32_f16_e32 v34, v16
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT:    v_or_b32_e32 v13, v13, v58
+; SI-NEXT:    v_or_b32_e32 v13, v13, v15
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v13
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v11, v24
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v11
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_or_b32_e32 v15, v15, v24
 ; SI-NEXT:    ; implicit-def: $vgpr24
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v53, v15
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v9
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT:    v_or_b32_e32 v13, v13, v63
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v49, v13
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v11, v60
-; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
-; SI-NEXT:    ; implicit-def: $vgpr60
-; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_or_b32_e32 v9, v9, v62
-; SI-NEXT:    v_cvt_f32_f16_e32 v33, v9
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; SI-NEXT:    v_or_b32_e32 v15, v15, v28
+; SI-NEXT:    ; implicit-def: $vgpr28
+; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v7, v57
+; SI-NEXT:    v_or_b32_e32 v7, v7, v9
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; SI-NEXT:    v_or_b32_e32 v13, v13, v22
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr62
-; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_or_b32_e32 v9, v9, v40
-; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
-; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v7, v59
+; SI-NEXT:    v_or_b32_e32 v7, v7, v9
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; SI-NEXT:    v_or_b32_e32 v11, v11, v63
+; SI-NEXT:    v_cvt_f32_f16_e32 v37, v11
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v7, v61
+; SI-NEXT:    v_or_b32_e32 v7, v7, v9
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    ; implicit-def: $vgpr63
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    v_or_b32_e32 v9, v9, v57
+; SI-NEXT:    v_cvt_f32_f16_e32 v33, v9
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; SI-NEXT:    v_or_b32_e32 v11, v11, v13
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; SI-NEXT:    v_or_b32_e32 v13, v13, v58
+; SI-NEXT:    v_cvt_f32_f16_e32 v49, v13
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    v_or_b32_e32 v9, v9, v59
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; SI-NEXT:    v_or_b32_e32 v13, v13, v61
 ; SI-NEXT:    ; implicit-def: $vgpr61
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:  .LBB106_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB106_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v42
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v30
-; SI-NEXT:    v_or_b32_e32 v7, v56, v7
-; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_or_b32_e32 v11, v1, v11
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v7
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v45
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    s_movk_i32 s6, 0x300
-; SI-NEXT:    v_or_b32_e32 v7, v41, v7
-; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v56
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; SI-NEXT:    v_or_b32_e32 v6, v14, v6
-; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v11
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v51
-; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; SI-NEXT:    v_or_b32_e32 v9, v46, v9
-; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v9
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_or_b32_e32 v7, v3, v7
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v47
+; SI-NEXT:    v_or_b32_e32 v6, v46, v6
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v4, v47, v4
-; SI-NEXT:    v_or_b32_e32 v5, v43, v5
-; SI-NEXT:    v_or_b32_e32 v3, v55, v3
-; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
-; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
-; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
-; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
-; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    v_cvt_f32_f16_e32 v50, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v52, v30
-; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
-; SI-NEXT:    v_cvt_f32_f16_e32 v54, v4
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v2, v2, v7
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v44
+; SI-NEXT:    s_movk_i32 s6, 0x300
+; SI-NEXT:    v_or_b32_e32 v9, v35, v3
+; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v6
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v7
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v42
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v0, v0, v7
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v35
+; SI-NEXT:    v_or_b32_e32 v7, v39, v7
+; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v7
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v41
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v7
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v39
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v18
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v10, v7
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
-; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v8
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v2
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
-; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
-; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
-; SI-NEXT:    v_or_b32_e32 v23, v57, v23
-; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
-; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
-; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v23, v0
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v26
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v12, v7
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v36, v12
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v22
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v16, v7
+; SI-NEXT:    v_or_b32_e32 v7, v14, v7
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v34, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v55
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v18, v7
+; SI-NEXT:    v_or_b32_e32 v7, v16, v7
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v7
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v62
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v20, v7
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v60
+; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT:    v_or_b32_e32 v7, v30, v7
+; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v9
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_or_b32_e32 v2, v2, v5
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v45
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_or_b32_e32 v0, v0, v5
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v44
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_or_b32_e32 v4, v4, v5
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v43
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_or_b32_e32 v1, v51, v1
+; SI-NEXT:    v_or_b32_e32 v5, v40, v5
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
+; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
+; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT:    v_cvt_f32_f16_e32 v34, v16
+; SI-NEXT:    v_cvt_f32_f16_e32 v36, v12
+; SI-NEXT:    v_cvt_f32_f16_e32 v38, v8
+; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
+; SI-NEXT:    v_cvt_f32_f16_e32 v48, v5
+; SI-NEXT:    v_cvt_f32_f16_e32 v50, v0
+; SI-NEXT:    v_cvt_f32_f16_e32 v52, v29
+; SI-NEXT:    v_cvt_f32_f16_e32 v29, v6
+; SI-NEXT:    v_cvt_f32_f16_e32 v54, v3
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v28, v7
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v27, v28, v27
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v26, v7
-; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v7
+; SI-NEXT:    v_or_b32_e32 v7, v24, v7
+; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
-; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT:    v_cvt_f32_f16_e32 v53, v18
-; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v27, v5
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
-; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; SI-NEXT:    v_or_b32_e32 v28, v31, v28
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v53, v19
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v14
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v22, v7
+; SI-NEXT:    v_or_b32_e32 v7, v61, v7
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
 ; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
-; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; SI-NEXT:    v_or_b32_e32 v31, v32, v31
-; SI-NEXT:    v_cvt_f32_f16_e32 v32, v17
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v16
-; SI-NEXT:    v_add_i32_e32 v31, vcc, s6, v31
-; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v63, v7
-; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v7
+; SI-NEXT:    v_or_b32_e32 v7, v58, v7
+; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
-; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; SI-NEXT:    v_or_b32_e32 v22, v59, v22
-; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
-; SI-NEXT:    v_cvt_f32_f16_e32 v49, v19
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v6
-; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v31, v1
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v49, v20
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v60, v7
+; SI-NEXT:    v_or_b32_e32 v7, v9, v7
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v24, v7
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v7
+; SI-NEXT:    v_or_b32_e32 v7, v63, v7
+; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v20
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v37, v21
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v10
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v40, v7
+; SI-NEXT:    v_or_b32_e32 v7, v59, v7
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
-; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; SI-NEXT:    v_or_b32_e32 v24, v26, v24
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
-; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
 ; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v62, v7
-; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v7
+; SI-NEXT:    v_or_b32_e32 v7, v57, v7
+; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v7
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v33, v22
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
+; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; SI-NEXT:    v_or_b32_e32 v7, v24, v7
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; SI-NEXT:    v_or_b32_e32 v24, v25, v24
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
+; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
+; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
+; SI-NEXT:    v_or_b32_e32 v25, v26, v25
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
+; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
+; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v25, v4
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
 ; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
-; SI-NEXT:    v_or_b32_e32 v26, v58, v26
+; SI-NEXT:    v_or_b32_e32 v26, v27, v26
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
 ; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
-; SI-NEXT:    v_cvt_f32_f16_e32 v33, v21
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v10
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; SI-NEXT:    v_or_b32_e32 v7, v61, v7
-; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
+; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
+; SI-NEXT:    v_or_b32_e32 v27, v28, v27
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
+; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
+; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v27, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
+; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; SI-NEXT:    v_or_b32_e32 v28, v30, v28
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
+; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
+; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
+; SI-NEXT:    v_or_b32_e32 v30, v31, v30
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v30
+; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
+; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
+; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
+; SI-NEXT:    v_or_b32_e32 v31, v32, v31
+; SI-NEXT:    v_add_i32_e32 v31, vcc, s6, v31
+; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    v_cvt_f32_f16_e32 v32, v18
+; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v31, v1
 ; SI-NEXT:  .LBB106_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
@@ -82927,11 +82940,9 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v16, v32
 ; SI-NEXT:    v_mov_b32_e32 v18, v34
 ; SI-NEXT:    v_mov_b32_e32 v20, v36
-; SI-NEXT:    s_waitcnt expcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_mov_b32_e32 v22, v38
-; SI-NEXT:    s_waitcnt expcnt(1)
 ; SI-NEXT:    v_mov_b32_e32 v24, v48
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_mov_b32_e32 v26, v50
 ; SI-NEXT:    v_mov_b32_e32 v28, v52
 ; SI-NEXT:    v_mov_b32_e32 v30, v54
@@ -82973,78 +82984,78 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:88
+; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:56
+; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:64
+; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:72
+; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:80
+; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:88
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:96
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
 ; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:4
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:116
-; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:124
-; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v1
-; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v3
-; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v5
-; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v7
-; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v9
-; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v11
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v13
-; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v15
+; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v27
+; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v29
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:116
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v1
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v5
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v7
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v9
+; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
+; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v13
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v15
 ; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
 ; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
 ; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
 ; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
-; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v24
-; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v0
-; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v2
-; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v4
-; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v8
-; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v10
-; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v12
-; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v14
-; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v16
-; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v18
-; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v20
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v4
+; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
+; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v10
+; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v12
+; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v14
+; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v16
+; VI-NEXT:    v_lshlrev_b16_e32 v61, 8, v18
+; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v20
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v22
+; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v22
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v26
+; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v24
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v28
+; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v26
+; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:124
 ; VI-NEXT:    s_waitcnt vmcnt(10)
-; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v30
+; VI-NEXT:    v_lshlrev_b16_e32 v36, 8, v30
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v31
+; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v31
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:108
-; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:92
+; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:100
+; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:84
 ; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:68
 ; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:52
+; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
@@ -83056,79 +83067,80 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT:    v_or_b32_sdwa v9, v39, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v10, v20, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v28, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v22, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v30, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v16, v38, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr39
-; VI-NEXT:    ; implicit-def: $vgpr20
-; VI-NEXT:    ; implicit-def: $vgpr28
+; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_or_b32_sdwa v9, v29, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v41, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v46, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_or_b32_sdwa v12, v22, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v30, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v34, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v31, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v16, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr46
 ; VI-NEXT:    ; implicit-def: $vgpr22
 ; VI-NEXT:    ; implicit-def: $vgpr30
-; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr34
 ; VI-NEXT:    ; implicit-def: $vgpr31
-; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    ; implicit-def: $vgpr57
 ; VI-NEXT:    ; implicit-def: $vgpr59
-; VI-NEXT:    ; implicit-def: $vgpr62
-; VI-NEXT:    ; implicit-def: $vgpr18
-; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr24
+; VI-NEXT:    ; implicit-def: $vgpr36
 ; VI-NEXT:    ; implicit-def: $vgpr32
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_or_b32_sdwa v0, v0, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_or_b32_sdwa v1, v1, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v1, v1, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr48
+; VI-NEXT:    ; implicit-def: $vgpr49
 ; VI-NEXT:    ; implicit-def: $vgpr50
-; VI-NEXT:    ; implicit-def: $vgpr51
-; VI-NEXT:    ; implicit-def: $vgpr52
-; VI-NEXT:    ; implicit-def: $vgpr53
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr19
 ; VI-NEXT:    ; implicit-def: $vgpr23
-; VI-NEXT:    ; implicit-def: $vgpr27
+; VI-NEXT:    ; implicit-def: $vgpr40
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_or_b32_sdwa v8, v8, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr45
+; VI-NEXT:    v_or_b32_sdwa v8, v8, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr43
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_or_b32_sdwa v4, v4, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v4, v4, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr52
 ; VI-NEXT:    ; implicit-def: $vgpr54
-; VI-NEXT:    ; implicit-def: $vgpr40
-; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr53
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -83145,23 +83157,23 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr25
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v49, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v48, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v55, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v55, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v44, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v36, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v18, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v43, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v26, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v37, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v37, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v44, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v35, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v61, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
@@ -83195,98 +83207,92 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; kill: killed $vgpr16
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr49
-; VI-NEXT:    ; implicit-def: $vgpr48
+; VI-NEXT:    ; implicit-def: $vgpr27
 ; VI-NEXT:    ; implicit-def: $vgpr55
-; VI-NEXT:    ; implicit-def: $vgpr36
-; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr37
 ; VI-NEXT:    ; implicit-def: $vgpr44
-; VI-NEXT:    ; implicit-def: $vgpr61
-; VI-NEXT:    ; implicit-def: $vgpr29
-; VI-NEXT:    ; implicit-def: $vgpr46
+; VI-NEXT:    ; implicit-def: $vgpr18
+; VI-NEXT:    ; implicit-def: $vgpr26
+; VI-NEXT:    ; implicit-def: $vgpr37
+; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr16
+; VI-NEXT:    ; implicit-def: $vgpr42
+; VI-NEXT:    ; implicit-def: $vgpr45
 ; VI-NEXT:    ; implicit-def: $vgpr56
 ; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr24
-; VI-NEXT:    ; implicit-def: $vgpr26
-; VI-NEXT:    ; implicit-def: $vgpr34
+; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; implicit-def: $vgpr20
+; VI-NEXT:    ; implicit-def: $vgpr28
+; VI-NEXT:    ; implicit-def: $vgpr33
 ; VI-NEXT:  .LBB106_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB106_4
 ; VI-NEXT:  ; %bb.3: ; %cmp.true
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v38
-; VI-NEXT:    v_add_u16_e32 v2, 3, v44
+; VI-NEXT:    v_add_u16_e32 v0, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v0, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v14, v26, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_add_u16_e32 v2, 3, v35
 ; VI-NEXT:    v_mov_b32_e32 v3, 0x300
-; VI-NEXT:    v_or_b32_sdwa v2, v18, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_sdwa v18, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v37
-; VI-NEXT:    v_or_b32_sdwa v24, v24, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v16, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_add_u16_e32 v0, 3, v37
+; VI-NEXT:    v_or_b32_sdwa v20, v20, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v30
-; VI-NEXT:    v_or_b32_sdwa v0, v16, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v38, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v12, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v43
-; VI-NEXT:    v_or_b32_sdwa v16, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v22
-; VI-NEXT:    v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v36
-; VI-NEXT:    v_or_b32_sdwa v22, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v26
+; VI-NEXT:    v_or_b32_sdwa v26, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v28
+; VI-NEXT:    v_add_u16_e32 v0, 3, v22
+; VI-NEXT:    v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_add_u16_e32 v0, 3, v18
+; VI-NEXT:    v_or_b32_sdwa v18, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v10, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v55
-; VI-NEXT:    v_or_b32_sdwa v28, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v20
+; VI-NEXT:    v_add_u16_e32 v0, 3, v44
+; VI-NEXT:    v_or_b32_sdwa v22, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v41
 ; VI-NEXT:    v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v2, 3, v35
 ; VI-NEXT:    v_add_u16_sdwa v9, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v48
-; VI-NEXT:    v_or_b32_sdwa v20, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v0, 3, v39
+; VI-NEXT:    v_add_u16_e32 v0, 3, v55
+; VI-NEXT:    v_or_b32_sdwa v14, v28, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v28, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v29
 ; VI-NEXT:    v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v49
-; VI-NEXT:    v_or_b32_sdwa v30, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v27
+; VI-NEXT:    v_or_b32_sdwa v27, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v2, 3, v34
+; VI-NEXT:    v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v13, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v1, 3, v61
-; VI-NEXT:    v_or_b32_sdwa v15, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v1, 3, v63
+; VI-NEXT:    v_or_b32_sdwa v15, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v1, 3, v31
-; VI-NEXT:    v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_sdwa v26, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT:    v_or_b32_e32 v12, v16, v12
-; VI-NEXT:    v_add_u16_e32 v16, 0x300, v24
+; VI-NEXT:    v_or_b32_sdwa v1, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v24, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT:    v_or_b32_e32 v13, v16, v13
-; VI-NEXT:    v_or_b32_e32 v14, v14, v26
-; VI-NEXT:    v_or_b32_e32 v15, v15, v18
+; VI-NEXT:    v_or_b32_e32 v14, v14, v24
+; VI-NEXT:    v_or_b32_e32 v15, v15, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v29, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v29, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -83319,54 +83325,58 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v6, v17, v6
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v29
 ; VI-NEXT:    v_or_b32_e32 v7, v17, v7
-; VI-NEXT:    v_add_u16_e32 v17, 0x300, v30
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v27
 ; VI-NEXT:    v_or_b32_e32 v8, v17, v8
-; VI-NEXT:    v_add_u16_e32 v17, 0x300, v20
-; VI-NEXT:    v_or_b32_e32 v9, v17, v9
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v28
-; VI-NEXT:    v_or_b32_e32 v10, v17, v10
+; VI-NEXT:    v_or_b32_e32 v9, v17, v9
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v22
+; VI-NEXT:    v_or_b32_e32 v10, v17, v10
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v18
 ; VI-NEXT:    v_or_b32_e32 v11, v17, v11
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v26
+; VI-NEXT:    v_or_b32_e32 v12, v17, v12
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v20
+; VI-NEXT:    v_or_b32_e32 v13, v17, v13
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v19, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v23, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v23, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v2, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v27, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v30, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v1, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v31, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v31, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
 ; VI-NEXT:    v_or_b32_e32 v0, v3, v0
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v31
 ; VI-NEXT:    v_or_b32_e32 v1, v3, v1
-; VI-NEXT:    v_add_u16_e32 v3, 0x300, v27
+; VI-NEXT:    v_add_u16_e32 v3, 0x300, v30
 ; VI-NEXT:    v_or_b32_e32 v2, v3, v2
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v23
 ; VI-NEXT:    v_or_b32_e32 v3, v3, v19
@@ -83426,99 +83436,100 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:80
+; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:56
+; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:64
+; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:72
+; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:80
 ; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:88
 ; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:96
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:104
-; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:112
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v1
-; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v3
-; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v5
-; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v7
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v9
-; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v11
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v13
-; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v15
-; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v17
+; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:112
+; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v17
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v19
-; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v23
-; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
+; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v25
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
-; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v29
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:116
+; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v1
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v3
+; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v5
+; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v7
+; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v9
+; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v11
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v13
+; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v15
+; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v22
+; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v6
 ; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v4
+; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(20)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
+; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(19)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
+; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v12
 ; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v10
+; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v14
 ; GFX9-NEXT:    s_waitcnt vmcnt(17)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v12
+; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v14
+; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v16
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v18
+; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v20
+; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v24
-; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:116
+; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v28
+; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v31
+; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v32
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:108
-; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:92
-; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
+; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v31
+; GFX9-NEXT:    s_waitcnt vmcnt(8)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v32
+; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:100
+; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:92
+; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:84
+; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB106_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
@@ -83526,108 +83537,107 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v39, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v63, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v35, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v34, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v16, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v32, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v16, v38, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr39
-; GFX9-NEXT:    ; implicit-def: $vgpr62
-; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr35
-; GFX9-NEXT:    ; implicit-def: $vgpr34
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr38
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
+; GFX9-NEXT:    v_or_b32_sdwa v9, v25, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v41, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(10)
+; GFX9-NEXT:    v_or_b32_sdwa v12, v26, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v37, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v20, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr25
+; GFX9-NEXT:    ; implicit-def: $vgpr27
+; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr26
+; GFX9-NEXT:    ; implicit-def: $vgpr37
+; GFX9-NEXT:    ; implicit-def: $vgpr20
+; GFX9-NEXT:    ; implicit-def: $vgpr31
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
-; GFX9-NEXT:    ; implicit-def: $vgpr36
-; GFX9-NEXT:    ; implicit-def: $vgpr26
-; GFX9-NEXT:    ; implicit-def: $vgpr20
-; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; implicit-def: $vgpr18
+; GFX9-NEXT:    ; implicit-def: $vgpr30
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
 ; GFX9-NEXT:    v_perm_b32 v1, v3, v2, s6
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr49
 ; GFX9-NEXT:    ; implicit-def: $vgpr48
-; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    ; implicit-def: $vgpr39
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
-; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr49
+; GFX9-NEXT:    ; implicit-def: $vgpr53
 ; GFX9-NEXT:    ; implicit-def: $vgpr17
 ; GFX9-NEXT:    ; implicit-def: $vgpr19
 ; GFX9-NEXT:    ; implicit-def: $vgpr23
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr27
+; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr55
-; GFX9-NEXT:    ; implicit-def: $vgpr54
+; GFX9-NEXT:    ; implicit-def: $vgpr52
+; GFX9-NEXT:    ; implicit-def: $vgpr51
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr43
+; GFX9-NEXT:    ; implicit-def: $vgpr55
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr25
+; GFX9-NEXT:    ; implicit-def: $vgpr40
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT:    v_or_b32_sdwa v8, v51, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v29, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT:    v_or_b32_sdwa v9, v52, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v42, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT:    v_or_b32_sdwa v10, v42, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v47, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT:    v_or_b32_sdwa v11, v37, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v35, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT:    v_or_b32_sdwa v12, v44, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v16, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT:    v_or_b32_sdwa v13, v28, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v24, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT:    v_or_b32_sdwa v14, v47, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v32, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT:    v_or_b32_sdwa v15, v18, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v38, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v16, v36, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
@@ -83661,111 +83671,110 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; kill: killed $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
-; GFX9-NEXT:    ; implicit-def: $vgpr51
-; GFX9-NEXT:    ; implicit-def: $vgpr52
+; GFX9-NEXT:    ; implicit-def: $vgpr29
 ; GFX9-NEXT:    ; implicit-def: $vgpr42
-; GFX9-NEXT:    ; implicit-def: $vgpr37
-; GFX9-NEXT:    ; implicit-def: $vgpr44
-; GFX9-NEXT:    ; implicit-def: $vgpr28
-; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr47
-; GFX9-NEXT:    ; implicit-def: $vgpr18
-; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr35
+; GFX9-NEXT:    ; implicit-def: $vgpr16
+; GFX9-NEXT:    ; implicit-def: $vgpr24
+; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; implicit-def: $vgpr38
+; GFX9-NEXT:    ; implicit-def: $vgpr36
+; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; implicit-def: $vgpr22
-; GFX9-NEXT:    ; implicit-def: $vgpr30
-; GFX9-NEXT:    ; implicit-def: $vgpr24
-; GFX9-NEXT:    ; implicit-def: $vgpr31
+; GFX9-NEXT:    ; implicit-def: $vgpr34
+; GFX9-NEXT:    ; implicit-def: $vgpr28
+; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:  .LBB106_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB106_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
-; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v18
-; GFX9-NEXT:    v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(9)
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v38
+; GFX9-NEXT:    v_or_b32_sdwa v0, v28, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v28
-; GFX9-NEXT:    v_or_b32_sdwa v0, v30, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v24
+; GFX9-NEXT:    v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v20
+; GFX9-NEXT:    v_or_b32_sdwa v0, v18, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v0
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v16
-; GFX9-NEXT:    v_or_b32_sdwa v0, v26, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v44
-; GFX9-NEXT:    v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v34
-; GFX9-NEXT:    v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v37
+; GFX9-NEXT:    v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v35
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v35
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v26
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v42
+; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v47
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v63
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v41
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v52
+; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v42
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v27
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v51
+; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v29
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v39
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v25
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT:    v_add_u16_e32 v3, 3, v32
-; GFX9-NEXT:    v_or_b32_sdwa v3, v20, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v3
+; GFX9-NEXT:    v_add_u16_e32 v3, 3, v31
+; GFX9-NEXT:    v_or_b32_sdwa v3, v30, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v3
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v32
+; GFX9-NEXT:    v_or_b32_sdwa v2, v34, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v2
-; GFX9-NEXT:    v_add_u16_e32 v1, 3, v38
+; GFX9-NEXT:    v_add_u16_e32 v1, 3, v36
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v1
+; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v1
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
-; GFX9-NEXT:    v_perm_b32 v8, v30, v8, s6
-; GFX9-NEXT:    v_perm_b32 v9, v28, v9, s6
-; GFX9-NEXT:    v_perm_b32 v10, v26, v10, s6
-; GFX9-NEXT:    v_perm_b32 v11, v24, v11, s6
-; GFX9-NEXT:    v_perm_b32 v12, v22, v12, s6
-; GFX9-NEXT:    v_perm_b32 v13, v16, v13, s6
-; GFX9-NEXT:    v_perm_b32 v14, v20, v14, s6
-; GFX9-NEXT:    v_perm_b32 v15, v18, v15, s6
+; GFX9-NEXT:    v_perm_b32 v8, v25, v8, s6
+; GFX9-NEXT:    v_perm_b32 v9, v24, v9, s6
+; GFX9-NEXT:    v_perm_b32 v10, v22, v10, s6
+; GFX9-NEXT:    v_perm_b32 v11, v20, v11, s6
+; GFX9-NEXT:    v_perm_b32 v12, v16, v12, s6
+; GFX9-NEXT:    v_perm_b32 v13, v18, v13, s6
+; GFX9-NEXT:    v_perm_b32 v14, v30, v14, s6
+; GFX9-NEXT:    v_perm_b32 v15, v28, v15, s6
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v44, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT:    v_or_b32_sdwa v31, v48, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v31, v39, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT:    v_perm_b32 v7, v27, v7, s6
+; GFX9-NEXT:    v_perm_b32 v7, v26, v7, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -83787,7 +83796,7 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v5, v19, v5, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -83798,40 +83807,40 @@ define <32 x half> @bitcast_v64i8_to_v32f16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v4, v17, v4, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v3, v21, v3, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT:    v_perm_b32 v2, v25, v2, s6
+; GFX9-NEXT:    v_perm_b32 v2, v27, v2, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v1, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v1, v29, v1, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; GFX9-NEXT:    v_perm_b32 v0, v31, v0, s6
 ; GFX9-NEXT:  .LBB106_4: ; %end
@@ -93581,129 +93590,123 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(5)
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:132
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:120
-; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:128
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:128
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:28
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v10, 8, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v14, 8, v29
 ; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr51
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr13
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr21
 ; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; implicit-def: $vgpr23
 ; SI-NEXT:    ; implicit-def: $vgpr52
 ; SI-NEXT:    ; implicit-def: $vgpr53
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr27
 ; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr29
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v22, 24, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v18, 24, v17
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v17, 8, v20
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 24, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 24, v24
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_lshlrev_b32_e32 v8, 8, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v28
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_lshlrev_b32_e32 v18, 24, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v31
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_lshlrev_b32_e32 v30, 24, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 24, v32
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_lshlrev_b32_e32 v12, 8, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v33
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 24, v31
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 8, v34
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v32
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 8, v33
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 24, v34
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v35
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v36
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 24, v37
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:12
+; SI-NEXT:    v_lshlrev_b32_e32 v61, 24, v36
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr36
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:116
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:124
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v5
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v7
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v11
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v15
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v19
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:28
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:76
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:4
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v3
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:116
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v5
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v7
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v11
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v15
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v19
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:76
+; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:40
 ; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v29
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v23
 ; SI-NEXT:    ; kill: killed $vgpr3
 ; SI-NEXT:    ; implicit-def: $vgpr3
 ; SI-NEXT:    ; kill: killed $vgpr3
@@ -93713,7 +93716,7 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr11
 ; SI-NEXT:    ; implicit-def: $vgpr15
 ; SI-NEXT:    ; implicit-def: $vgpr19
-; SI-NEXT:    ; implicit-def: $vgpr29
+; SI-NEXT:    ; implicit-def: $vgpr23
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB110_2
@@ -93729,7 +93732,7 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v5
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
@@ -93740,7 +93743,7 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
@@ -93763,7 +93766,7 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 24, v7
 ; SI-NEXT:    v_or_b32_e32 v33, v7, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
@@ -93795,105 +93798,101 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; SI-NEXT:    v_or_b32_e32 v1, v11, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v10, v11, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v10
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v11, v2, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; SI-NEXT:    v_or_b32_e32 v11, v2, v10
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
+; SI-NEXT:    v_or_b32_e32 v32, v10, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr10
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_or_b32_e32 v13, v6, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT:    v_or_b32_e32 v2, v2, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v2
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    ; implicit-def: $vgpr14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_or_b32_e32 v15, v1, v2
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v2
-; SI-NEXT:    v_or_b32_e32 v32, v2, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v36, v2, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_or_b32_e32 v13, v10, v1
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    ; implicit-def: $vgpr10
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; SI-NEXT:    v_or_b32_e32 v0, v1, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v15, v14, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v1
-; SI-NEXT:    ; implicit-def: $vgpr14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v36, v1, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v62
+; SI-NEXT:    v_or_b32_e32 v37, v22, v1
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v62
+; SI-NEXT:    v_or_b32_e32 v1, v1, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v1
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr22
+; SI-NEXT:    ; implicit-def: $vgpr4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v37, v26, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v24
-; SI-NEXT:    v_or_b32_e32 v0, v0, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v19, v18, v0
+; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v19, v0, v1
 ; SI-NEXT:    v_and_b32_e32 v0, 0xff, v44
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v28
 ; SI-NEXT:    v_or_b32_e32 v48, v1, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v57
-; SI-NEXT:    ; implicit-def: $vgpr24
-; SI-NEXT:    ; implicit-def: $vgpr17
-; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr26
-; SI-NEXT:    ; implicit-def: $vgpr8
-; SI-NEXT:    ; implicit-def: $vgpr18
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v24
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v21, v30, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v22
-; SI-NEXT:    v_or_b32_e32 v0, v0, v12
+; SI-NEXT:    v_or_b32_e32 v21, v18, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v47
+; SI-NEXT:    v_or_b32_e32 v0, v0, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v6
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v20
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v23, v25, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v58
+; SI-NEXT:    v_or_b32_e32 v23, v8, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v49
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v59
 ; SI-NEXT:    v_or_b32_e32 v52, v1, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v46
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v56
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v53, v16, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v4
-; SI-NEXT:    v_or_b32_e32 v0, v0, v47
+; SI-NEXT:    v_or_b32_e32 v53, v12, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v60
+; SI-NEXT:    v_or_b32_e32 v0, v0, v57
 ; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v9
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v26
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v27, v45, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v35
+; SI-NEXT:    v_or_b32_e32 v27, v46, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v30
 ; SI-NEXT:    v_or_b32_e32 v40, v1, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v59
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; SI-NEXT:    v_or_b32_e32 v29, v56, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v20
-; SI-NEXT:    v_or_b32_e32 v0, v0, v60
+; SI-NEXT:    v_or_b32_e32 v29, v58, v0
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v45
+; SI-NEXT:    v_or_b32_e32 v0, v0, v35
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xff, v28
+; SI-NEXT:    v_and_b32_e32 v0, 0xff, v16
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; SI-NEXT:    v_or_b32_e32 v31, v61, v0
 ; SI-NEXT:    ; implicit-def: $vgpr0
@@ -93955,177 +93954,183 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr0
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    ; implicit-def: $vgpr6
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr4
-; SI-NEXT:    ; implicit-def: $vgpr9
-; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr28
+; SI-NEXT:    ; implicit-def: $vgpr24
+; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr59
-; SI-NEXT:    ; implicit-def: $vgpr20
-; SI-NEXT:    ; implicit-def: $vgpr28
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr26
+; SI-NEXT:    ; implicit-def: $vgpr25
+; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; implicit-def: $vgpr9
+; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr16
 ; SI-NEXT:    ; kill: killed $vgpr0
 ; SI-NEXT:    ; implicit-def: $vgpr1
 ; SI-NEXT:    ; implicit-def: $vgpr0
-; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; implicit-def: $vgpr18
+; SI-NEXT:    ; implicit-def: $vgpr17
+; SI-NEXT:    ; implicit-def: $vgpr8
 ; SI-NEXT:    ; implicit-def: $vgpr12
-; SI-NEXT:    ; implicit-def: $vgpr25
-; SI-NEXT:    ; implicit-def: $vgpr16
-; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr56
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr46
+; SI-NEXT:    ; implicit-def: $vgpr58
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:  .LBB110_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB110_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v20
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v45
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v28
-; SI-NEXT:    v_or_b32_e32 v3, v60, v3
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v16
+; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x300, v3
 ; SI-NEXT:    v_or_b32_e32 v5, v61, v5
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v35
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v25
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v49
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 8, v30
 ; SI-NEXT:    v_or_b32_e32 v5, v7, v5
 ; SI-NEXT:    s_movk_i32 s6, 0x300
 ; SI-NEXT:    s_mov_b32 s7, 0x3000000
-; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v3
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v3
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v5
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v59
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v9
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v5, v56, v5
+; SI-NEXT:    v_or_b32_e32 v5, v58, v5
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v4
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v3
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v60
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v9
-; SI-NEXT:    v_or_b32_e32 v3, v47, v3
-; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v26
+; SI-NEXT:    v_or_b32_e32 v3, v57, v3
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v4, v45, v4
-; SI-NEXT:    v_or_b32_e32 v3, v4, v3
-; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v3
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v58
+; SI-NEXT:    v_or_b32_e32 v5, v46, v5
+; SI-NEXT:    v_or_b32_e32 v3, v5, v3
+; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v3
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v49
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v57
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v59
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v46
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v56
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v5, v16, v5
+; SI-NEXT:    v_or_b32_e32 v5, v12, v5
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v22
+; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v3
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v47
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v6
-; SI-NEXT:    v_or_b32_e32 v3, v12, v3
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v20
+; SI-NEXT:    v_or_b32_e32 v3, v17, v3
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_or_b32_e32 v5, v25, v5
+; SI-NEXT:    v_or_b32_e32 v5, v8, v5
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v3
+; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v3
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v44
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v62
-; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v6
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v6
-; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v9
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v4
-; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v28
-; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v20
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v20
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
-; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v30, v5
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v28
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v24
-; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v17
-; SI-NEXT:    v_or_b32_e32 v3, v8, v3
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v24
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_or_b32_e32 v5, v18, v5
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v8
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v12
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v12
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
+; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v3
+; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v62
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v17
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v8
+; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v12
+; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v25
+; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v9
+; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v16
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v0, v0, v4
+; SI-NEXT:    v_or_b32_e32 v0, v0, v3
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v0
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v26, v5
-; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    v_or_b32_e32 v1, v1, v5
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v16
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v0, v0, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
-; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
+; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v4, v22, v4
+; SI-NEXT:    v_or_b32_e32 v3, v4, v3
+; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v3
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v4
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v14, v3
-; SI-NEXT:    v_or_b32_e32 v0, v3, v0
+; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT:    v_or_b32_e32 v1, v1, v3
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v0
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v0
+; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v1
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
@@ -94138,50 +94143,50 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT:    v_or_b32_e32 v5, v10, v5
+; SI-NEXT:    v_or_b32_e32 v5, v6, v5
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v3
+; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v3
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_or_b32_e32 v1, v1, v3
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_or_b32_e32 v3, v10, v3
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT:    v_or_b32_e32 v2, v2, v5
 ; SI-NEXT:    v_or_b32_e32 v2, v2, v3
-; SI-NEXT:    v_or_b32_e32 v1, v2, v1
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v1
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
-; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
-; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
-; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
 ; SI-NEXT:    v_or_b32_e32 v3, v5, v3
-; SI-NEXT:    v_or_b32_e32 v2, v3, v2
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
+; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_or_b32_e32 v5, v7, v5
+; SI-NEXT:    v_or_b32_e32 v3, v5, v3
+; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v3
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
-; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v2
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v2
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v10
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -94196,7 +94201,7 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v3
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v11
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
@@ -94253,21 +94258,21 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v3
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v5
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v7
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v11
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v1
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v10
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v2
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v6
 ; SI-NEXT:  .LBB110_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mov_b32_e32 v2, v43
 ; SI-NEXT:    v_mov_b32_e32 v10, v41
-; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_mov_b32_e32 v28, v40
 ; SI-NEXT:    v_mov_b32_e32 v30, v42
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
@@ -94293,7 +94298,6 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v12, v32
 ; SI-NEXT:    v_mov_b32_e32 v14, v34
 ; SI-NEXT:    v_mov_b32_e32 v16, v36
-; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_mov_b32_e32 v17, v37
 ; SI-NEXT:    v_mov_b32_e32 v18, v38
 ; SI-NEXT:    v_mov_b32_e32 v20, v48
@@ -94339,78 +94343,78 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:132
-; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:80
-; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:88
+; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
+; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:56
+; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:64
+; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:72
+; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:80
+; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:88
 ; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:96
 ; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:104
 ; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:112
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
 ; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:68
-; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:36
-; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:4
-; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:116
-; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:124
-; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v1
-; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v3
-; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v5
-; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v7
-; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v9
-; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v11
-; VI-NEXT:    v_lshlrev_b16_e32 v41, 8, v13
-; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v15
+; VI-NEXT:    v_lshlrev_b16_e32 v40, 8, v27
+; VI-NEXT:    v_lshlrev_b16_e32 v42, 8, v29
+; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:116
+; VI-NEXT:    v_lshlrev_b16_e32 v39, 8, v1
+; VI-NEXT:    v_lshlrev_b16_e32 v48, 8, v3
+; VI-NEXT:    v_lshlrev_b16_e32 v49, 8, v5
+; VI-NEXT:    v_lshlrev_b16_e32 v50, 8, v7
+; VI-NEXT:    v_lshlrev_b16_e32 v51, 8, v9
+; VI-NEXT:    v_lshlrev_b16_e32 v52, 8, v11
+; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v13
+; VI-NEXT:    v_lshlrev_b16_e32 v54, 8, v15
 ; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
 ; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
 ; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
 ; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
 ; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
-; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
-; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
 ; VI-NEXT:    s_waitcnt vmcnt(14)
-; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v24
-; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v0
-; VI-NEXT:    v_lshlrev_b16_e32 v46, 8, v2
-; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v4
-; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v6
-; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v8
-; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v10
-; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v12
-; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v14
-; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v16
-; VI-NEXT:    v_lshlrev_b16_e32 v63, 8, v18
-; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v20
+; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; VI-NEXT:    v_lshlrev_b16_e32 v43, 8, v2
+; VI-NEXT:    v_lshlrev_b16_e32 v45, 8, v4
+; VI-NEXT:    v_lshlrev_b16_e32 v47, 8, v6
+; VI-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
+; VI-NEXT:    v_lshlrev_b16_e32 v57, 8, v10
+; VI-NEXT:    v_lshlrev_b16_e32 v58, 8, v12
+; VI-NEXT:    v_lshlrev_b16_e32 v59, 8, v14
+; VI-NEXT:    v_lshlrev_b16_e32 v60, 8, v16
+; VI-NEXT:    v_lshlrev_b16_e32 v61, 8, v18
+; VI-NEXT:    v_lshlrev_b16_e32 v62, 8, v20
 ; VI-NEXT:    s_waitcnt vmcnt(13)
-; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v22
+; VI-NEXT:    v_lshlrev_b16_e32 v38, 8, v22
 ; VI-NEXT:    s_waitcnt vmcnt(12)
-; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v26
+; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v24
 ; VI-NEXT:    s_waitcnt vmcnt(11)
-; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v28
+; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v26
+; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:124
 ; VI-NEXT:    s_waitcnt vmcnt(10)
-; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v30
+; VI-NEXT:    v_lshlrev_b16_e32 v36, 8, v30
 ; VI-NEXT:    s_waitcnt vmcnt(9)
-; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v31
+; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v31
 ; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:108
-; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:92
+; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:100
+; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:92
 ; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:84
 ; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:68
 ; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:52
+; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
 ; VI-NEXT:    s_waitcnt vmcnt(14)
 ; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
 ; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
@@ -94422,79 +94426,80 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; VI-NEXT:    v_or_b32_sdwa v9, v39, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v10, v20, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v28, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v22, v62 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v30, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v16, v38, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr39
-; VI-NEXT:    ; implicit-def: $vgpr20
-; VI-NEXT:    ; implicit-def: $vgpr28
+; VI-NEXT:    s_waitcnt vmcnt(14)
+; VI-NEXT:    v_or_b32_sdwa v9, v29, v47 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v41, v57 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v46, v59 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_or_b32_sdwa v12, v22, v61 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v30, v38 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v34, v24 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v31, v36 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v16, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr29
+; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr46
 ; VI-NEXT:    ; implicit-def: $vgpr22
 ; VI-NEXT:    ; implicit-def: $vgpr30
-; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr34
 ; VI-NEXT:    ; implicit-def: $vgpr31
-; VI-NEXT:    ; implicit-def: $vgpr38
 ; VI-NEXT:    ; implicit-def: $vgpr47
 ; VI-NEXT:    ; implicit-def: $vgpr57
 ; VI-NEXT:    ; implicit-def: $vgpr59
-; VI-NEXT:    ; implicit-def: $vgpr62
-; VI-NEXT:    ; implicit-def: $vgpr18
-; VI-NEXT:    ; implicit-def: $vgpr33
+; VI-NEXT:    ; implicit-def: $vgpr61
+; VI-NEXT:    ; implicit-def: $vgpr38
+; VI-NEXT:    ; implicit-def: $vgpr24
+; VI-NEXT:    ; implicit-def: $vgpr36
 ; VI-NEXT:    ; implicit-def: $vgpr32
 ; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(7)
-; VI-NEXT:    v_or_b32_sdwa v0, v0, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v0, v0, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_or_b32_sdwa v1, v1, v51 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v1, v1, v48 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v53 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v50 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; VI-NEXT:    ; implicit-def: $vgpr39
+; VI-NEXT:    ; implicit-def: $vgpr48
+; VI-NEXT:    ; implicit-def: $vgpr49
 ; VI-NEXT:    ; implicit-def: $vgpr50
-; VI-NEXT:    ; implicit-def: $vgpr51
-; VI-NEXT:    ; implicit-def: $vgpr52
-; VI-NEXT:    ; implicit-def: $vgpr53
 ; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_or_b32_sdwa v5, v5, v19 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_or_b32_sdwa v6, v6, v23 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v27 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr19
 ; VI-NEXT:    ; implicit-def: $vgpr23
-; VI-NEXT:    ; implicit-def: $vgpr27
+; VI-NEXT:    ; implicit-def: $vgpr40
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_or_b32_sdwa v8, v8, v45 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; VI-NEXT:    ; implicit-def: $vgpr45
+; VI-NEXT:    v_or_b32_sdwa v8, v8, v43 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr43
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_or_b32_sdwa v2, v2, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v2, v2, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v40 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v52 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_or_b32_sdwa v4, v4, v42 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v4, v4, v54 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    ; implicit-def: $vgpr51
+; VI-NEXT:    ; implicit-def: $vgpr52
 ; VI-NEXT:    ; implicit-def: $vgpr54
-; VI-NEXT:    ; implicit-def: $vgpr40
-; VI-NEXT:    ; implicit-def: $vgpr42
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v3, v3, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; VI-NEXT:    ; implicit-def: $vgpr41
+; VI-NEXT:    ; implicit-def: $vgpr53
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
@@ -94511,23 +94516,23 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; VI-NEXT:    ; implicit-def: $vgpr25
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v7, v7, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v8, v49, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v8, v27, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v9, v48, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v9, v55, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v10, v55, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v10, v44, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v10, v10, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v11, v36, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v11, v18, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v12, v43, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v12, v26, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v13, v37, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v13, v37, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v14, v44, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v14, v35, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
-; VI-NEXT:    v_or_b32_sdwa v15, v61, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v15, v63, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
@@ -94561,98 +94566,92 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    ; kill: killed $vgpr16
 ; VI-NEXT:    ; implicit-def: $vgpr16
 ; VI-NEXT:    ; kill: killed $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr49
-; VI-NEXT:    ; implicit-def: $vgpr48
+; VI-NEXT:    ; implicit-def: $vgpr27
 ; VI-NEXT:    ; implicit-def: $vgpr55
-; VI-NEXT:    ; implicit-def: $vgpr36
-; VI-NEXT:    ; implicit-def: $vgpr43
-; VI-NEXT:    ; implicit-def: $vgpr37
 ; VI-NEXT:    ; implicit-def: $vgpr44
-; VI-NEXT:    ; implicit-def: $vgpr61
-; VI-NEXT:    ; implicit-def: $vgpr29
-; VI-NEXT:    ; implicit-def: $vgpr46
+; VI-NEXT:    ; implicit-def: $vgpr18
+; VI-NEXT:    ; implicit-def: $vgpr26
+; VI-NEXT:    ; implicit-def: $vgpr37
+; VI-NEXT:    ; implicit-def: $vgpr35
+; VI-NEXT:    ; implicit-def: $vgpr63
+; VI-NEXT:    ; implicit-def: $vgpr16
+; VI-NEXT:    ; implicit-def: $vgpr42
+; VI-NEXT:    ; implicit-def: $vgpr45
 ; VI-NEXT:    ; implicit-def: $vgpr56
 ; VI-NEXT:    ; implicit-def: $vgpr58
 ; VI-NEXT:    ; implicit-def: $vgpr60
-; VI-NEXT:    ; implicit-def: $vgpr63
-; VI-NEXT:    ; implicit-def: $vgpr16
-; VI-NEXT:    ; implicit-def: $vgpr24
-; VI-NEXT:    ; implicit-def: $vgpr26
-; VI-NEXT:    ; implicit-def: $vgpr34
+; VI-NEXT:    ; implicit-def: $vgpr62
+; VI-NEXT:    ; implicit-def: $vgpr20
+; VI-NEXT:    ; implicit-def: $vgpr28
+; VI-NEXT:    ; implicit-def: $vgpr33
 ; VI-NEXT:  .LBB110_2: ; %Flow
 ; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; VI-NEXT:    s_cbranch_execz .LBB110_4
 ; VI-NEXT:  ; %bb.3: ; %cmp.true
 ; VI-NEXT:    s_waitcnt vmcnt(8)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v38
-; VI-NEXT:    v_add_u16_e32 v2, 3, v44
+; VI-NEXT:    v_add_u16_e32 v0, 3, v16
 ; VI-NEXT:    v_or_b32_sdwa v0, v32, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_or_b32_sdwa v14, v26, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_add_u16_e32 v2, 3, v35
 ; VI-NEXT:    v_mov_b32_e32 v3, 0x300
-; VI-NEXT:    v_or_b32_sdwa v2, v18, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_sdwa v18, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    s_waitcnt vmcnt(5)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v37
-; VI-NEXT:    v_or_b32_sdwa v24, v24, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v16, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(4)
+; VI-NEXT:    v_add_u16_e32 v0, 3, v37
+; VI-NEXT:    v_or_b32_sdwa v20, v20, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v30
-; VI-NEXT:    v_or_b32_sdwa v0, v16, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v38, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v12, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v43
-; VI-NEXT:    v_or_b32_sdwa v16, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v22
-; VI-NEXT:    v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v36
-; VI-NEXT:    v_or_b32_sdwa v22, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v26
+; VI-NEXT:    v_or_b32_sdwa v26, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v28
+; VI-NEXT:    v_add_u16_e32 v0, 3, v22
+; VI-NEXT:    v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v11, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_add_u16_e32 v0, 3, v18
+; VI-NEXT:    v_or_b32_sdwa v18, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v46
 ; VI-NEXT:    v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v10, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v55
-; VI-NEXT:    v_or_b32_sdwa v28, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_add_u16_e32 v0, 3, v20
+; VI-NEXT:    v_add_u16_e32 v0, 3, v44
+; VI-NEXT:    v_or_b32_sdwa v22, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v41
 ; VI-NEXT:    v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v2, 3, v35
 ; VI-NEXT:    v_add_u16_sdwa v9, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v48
-; VI-NEXT:    v_or_b32_sdwa v20, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_e32 v0, 3, v39
+; VI-NEXT:    v_add_u16_e32 v0, 3, v55
+; VI-NEXT:    v_or_b32_sdwa v14, v28, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v28, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v29
 ; VI-NEXT:    v_or_b32_sdwa v0, v47, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v8, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v0, 3, v49
-; VI-NEXT:    v_or_b32_sdwa v30, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v0, 3, v27
+; VI-NEXT:    v_or_b32_sdwa v27, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; VI-NEXT:    v_add_u16_e32 v2, 3, v34
+; VI-NEXT:    v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v13, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v1, 3, v61
-; VI-NEXT:    v_or_b32_sdwa v15, v34, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_e32 v1, 3, v63
+; VI-NEXT:    v_or_b32_sdwa v15, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v1, 3, v31
-; VI-NEXT:    v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; VI-NEXT:    v_add_u16_sdwa v26, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
-; VI-NEXT:    v_or_b32_e32 v12, v16, v12
-; VI-NEXT:    v_add_u16_e32 v16, 0x300, v24
+; VI-NEXT:    v_or_b32_sdwa v1, v36, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_add_u16_sdwa v24, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
 ; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
-; VI-NEXT:    v_or_b32_e32 v13, v16, v13
-; VI-NEXT:    v_or_b32_e32 v14, v14, v26
-; VI-NEXT:    v_or_b32_e32 v15, v15, v18
+; VI-NEXT:    v_or_b32_e32 v14, v14, v24
+; VI-NEXT:    v_or_b32_e32 v15, v15, v16
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v7, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v29, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v29, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v6, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -94685,54 +94684,58 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; VI-NEXT:    v_or_b32_e32 v6, v17, v6
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v29
 ; VI-NEXT:    v_or_b32_e32 v7, v17, v7
-; VI-NEXT:    v_add_u16_e32 v17, 0x300, v30
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v27
 ; VI-NEXT:    v_or_b32_e32 v8, v17, v8
-; VI-NEXT:    v_add_u16_e32 v17, 0x300, v20
-; VI-NEXT:    v_or_b32_e32 v9, v17, v9
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v28
-; VI-NEXT:    v_or_b32_e32 v10, v17, v10
+; VI-NEXT:    v_or_b32_e32 v9, v17, v9
 ; VI-NEXT:    v_add_u16_e32 v17, 0x300, v22
+; VI-NEXT:    v_or_b32_e32 v10, v17, v10
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v18
 ; VI-NEXT:    v_or_b32_e32 v11, v17, v11
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v26
+; VI-NEXT:    v_or_b32_e32 v12, v17, v12
+; VI-NEXT:    v_add_u16_e32 v17, 0x300, v20
+; VI-NEXT:    v_or_b32_e32 v13, v17, v13
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v42, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v19, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v23, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v23, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v2, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v27, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v30, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v1, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v31, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v31, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v0, 3, v0
-; VI-NEXT:    v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_sdwa v0, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_u16_e32 v3, 3, v3
-; VI-NEXT:    v_or_b32_sdwa v3, v50, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; VI-NEXT:    v_or_b32_sdwa v3, v39, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
 ; VI-NEXT:    v_or_b32_e32 v0, v3, v0
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v31
 ; VI-NEXT:    v_or_b32_e32 v1, v3, v1
-; VI-NEXT:    v_add_u16_e32 v3, 0x300, v27
+; VI-NEXT:    v_add_u16_e32 v3, 0x300, v30
 ; VI-NEXT:    v_or_b32_e32 v2, v3, v2
 ; VI-NEXT:    v_add_u16_e32 v3, 0x300, v23
 ; VI-NEXT:    v_or_b32_e32 v3, v3, v19
@@ -94792,99 +94795,100 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:132
+; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
 ; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32
-; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:24
-; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:32
-; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:40
-; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:48
-; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:56
-; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:64
-; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:72
-; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:80
+; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:40
+; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:56
+; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:64
+; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:72
+; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:80
 ; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:88
 ; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:96
 ; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:104
-; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:112
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:68
-; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:36
-; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v1
-; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v3
-; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v5
-; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v7
-; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v9
-; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v11
-; GFX9-NEXT:    v_lshlrev_b16_e32 v41, 8, v13
-; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v15
-; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v17
+; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:112
+; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    v_lshlrev_b16_e32 v55, 8, v17
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v19
-; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v23
-; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
+; GFX9-NEXT:    v_lshlrev_b16_e32 v40, 8, v25
 ; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v27
-; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
-; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    v_lshlrev_b16_e32 v44, 8, v29
+; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:116
+; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    v_lshlrev_b16_e32 v48, 8, v1
+; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v3
+; GFX9-NEXT:    v_lshlrev_b16_e32 v50, 8, v5
+; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v7
+; GFX9-NEXT:    v_lshlrev_b16_e32 v52, 8, v9
+; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v11
+; GFX9-NEXT:    v_lshlrev_b16_e32 v54, 8, v13
+; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v15
+; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(24)
-; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v22
+; GFX9-NEXT:    v_lshlrev_b16_e32 v43, 8, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(22)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v46, 8, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v6
 ; GFX9-NEXT:    s_waitcnt vmcnt(21)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v45, 8, v4
+; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v8
 ; GFX9-NEXT:    s_waitcnt vmcnt(20)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v57, 8, v6
+; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v10
 ; GFX9-NEXT:    s_waitcnt vmcnt(19)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v56, 8, v8
+; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v12
 ; GFX9-NEXT:    s_waitcnt vmcnt(18)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v59, 8, v10
+; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v14
 ; GFX9-NEXT:    s_waitcnt vmcnt(17)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v58, 8, v12
+; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v61, 8, v14
+; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v18
 ; GFX9-NEXT:    s_waitcnt vmcnt(15)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v60, 8, v16
+; GFX9-NEXT:    v_lshlrev_b16_e32 v63, 8, v20
 ; GFX9-NEXT:    s_waitcnt vmcnt(14)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v18
+; GFX9-NEXT:    v_lshlrev_b16_e32 v62, 8, v22
 ; GFX9-NEXT:    s_waitcnt vmcnt(13)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v36, 8, v20
+; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v24
 ; GFX9-NEXT:    s_waitcnt vmcnt(12)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v24
-; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:116
+; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v26
 ; GFX9-NEXT:    s_waitcnt vmcnt(11)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v28
+; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v28
 ; GFX9-NEXT:    s_waitcnt vmcnt(10)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v31
+; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
 ; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v32
-; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:108
-; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:92
-; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
-; GFX9-NEXT:    s_waitcnt vmcnt(16)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
+; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v31
+; GFX9-NEXT:    s_waitcnt vmcnt(8)
+; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v32
+; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:100
+; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:92
+; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:84
+; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB110_2
 ; GFX9-NEXT:  ; %bb.1: ; %cmp.false
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
@@ -94892,108 +94896,107 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_waitcnt vmcnt(23)
-; GFX9-NEXT:    v_or_b32_sdwa v9, v39, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    s_waitcnt vmcnt(9)
-; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v11, v63, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v12, v35, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v13, v34, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v14, v16, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v15, v32, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    v_or_b32_sdwa v16, v38, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr39
-; GFX9-NEXT:    ; implicit-def: $vgpr62
-; GFX9-NEXT:    ; implicit-def: $vgpr63
-; GFX9-NEXT:    ; implicit-def: $vgpr35
-; GFX9-NEXT:    ; implicit-def: $vgpr34
-; GFX9-NEXT:    ; implicit-def: $vgpr32
-; GFX9-NEXT:    ; implicit-def: $vgpr38
+; GFX9-NEXT:    s_waitcnt vmcnt(20)
+; GFX9-NEXT:    v_or_b32_sdwa v9, v25, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v27, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v41, v58 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    s_waitcnt vmcnt(10)
+; GFX9-NEXT:    v_or_b32_sdwa v12, v26, v60 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v37, v62 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v20, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr25
+; GFX9-NEXT:    ; implicit-def: $vgpr27
+; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr26
+; GFX9-NEXT:    ; implicit-def: $vgpr37
+; GFX9-NEXT:    ; implicit-def: $vgpr20
+; GFX9-NEXT:    ; implicit-def: $vgpr31
 ; GFX9-NEXT:    ; implicit-def: $vgpr45
 ; GFX9-NEXT:    ; implicit-def: $vgpr56
 ; GFX9-NEXT:    ; implicit-def: $vgpr58
 ; GFX9-NEXT:    ; implicit-def: $vgpr60
-; GFX9-NEXT:    ; implicit-def: $vgpr36
-; GFX9-NEXT:    ; implicit-def: $vgpr26
-; GFX9-NEXT:    ; implicit-def: $vgpr20
-; GFX9-NEXT:    ; implicit-def: $vgpr33
+; GFX9-NEXT:    ; implicit-def: $vgpr62
+; GFX9-NEXT:    ; implicit-def: $vgpr18
+; GFX9-NEXT:    ; implicit-def: $vgpr30
 ; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_or_b32_sdwa v1, v1, v48 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v50 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
 ; GFX9-NEXT:    v_perm_b32 v1, v3, v2, s6
 ; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(4)
 ; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(3)
 ; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr49
 ; GFX9-NEXT:    ; implicit-def: $vgpr48
-; GFX9-NEXT:    ; implicit-def: $vgpr53
+; GFX9-NEXT:    ; implicit-def: $vgpr39
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
-; GFX9-NEXT:    ; implicit-def: $vgpr40
+; GFX9-NEXT:    ; implicit-def: $vgpr49
+; GFX9-NEXT:    ; implicit-def: $vgpr53
 ; GFX9-NEXT:    ; implicit-def: $vgpr17
 ; GFX9-NEXT:    ; implicit-def: $vgpr19
 ; GFX9-NEXT:    ; implicit-def: $vgpr23
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
-; GFX9-NEXT:    ; implicit-def: $vgpr27
+; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    ; implicit-def: $vgpr43
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
 ; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr55
-; GFX9-NEXT:    ; implicit-def: $vgpr54
+; GFX9-NEXT:    ; implicit-def: $vgpr52
+; GFX9-NEXT:    ; implicit-def: $vgpr51
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
 ; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr41
+; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
 ; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr43
+; GFX9-NEXT:    ; implicit-def: $vgpr55
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
 ; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; GFX9-NEXT:    ; implicit-def: $vgpr21
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
 ; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT:    ; implicit-def: $vgpr25
+; GFX9-NEXT:    ; implicit-def: $vgpr40
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v44 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
-; GFX9-NEXT:    v_or_b32_sdwa v8, v51, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v8, v29, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
-; GFX9-NEXT:    v_or_b32_sdwa v9, v52, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v9, v42, v57 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
-; GFX9-NEXT:    v_or_b32_sdwa v10, v42, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v10, v47, v59 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
-; GFX9-NEXT:    v_or_b32_sdwa v11, v37, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v11, v35, v61 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
-; GFX9-NEXT:    v_or_b32_sdwa v12, v44, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v12, v16, v63 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
-; GFX9-NEXT:    v_or_b32_sdwa v13, v28, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v13, v24, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
-; GFX9-NEXT:    v_or_b32_sdwa v14, v47, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v14, v32, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
-; GFX9-NEXT:    v_or_b32_sdwa v15, v18, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v15, v38, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_or_b32_sdwa v16, v36, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
@@ -95027,111 +95030,110 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    ; kill: killed $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; kill: killed $vgpr16
-; GFX9-NEXT:    ; implicit-def: $vgpr51
-; GFX9-NEXT:    ; implicit-def: $vgpr52
+; GFX9-NEXT:    ; implicit-def: $vgpr29
 ; GFX9-NEXT:    ; implicit-def: $vgpr42
-; GFX9-NEXT:    ; implicit-def: $vgpr37
-; GFX9-NEXT:    ; implicit-def: $vgpr44
-; GFX9-NEXT:    ; implicit-def: $vgpr28
-; GFX9-NEXT:    ; implicit-def: $vgpr16
 ; GFX9-NEXT:    ; implicit-def: $vgpr47
-; GFX9-NEXT:    ; implicit-def: $vgpr18
-; GFX9-NEXT:    ; implicit-def: $vgpr29
+; GFX9-NEXT:    ; implicit-def: $vgpr35
+; GFX9-NEXT:    ; implicit-def: $vgpr16
+; GFX9-NEXT:    ; implicit-def: $vgpr24
+; GFX9-NEXT:    ; implicit-def: $vgpr32
+; GFX9-NEXT:    ; implicit-def: $vgpr38
+; GFX9-NEXT:    ; implicit-def: $vgpr36
+; GFX9-NEXT:    ; implicit-def: $vgpr44
 ; GFX9-NEXT:    ; implicit-def: $vgpr46
 ; GFX9-NEXT:    ; implicit-def: $vgpr57
 ; GFX9-NEXT:    ; implicit-def: $vgpr59
 ; GFX9-NEXT:    ; implicit-def: $vgpr61
+; GFX9-NEXT:    ; implicit-def: $vgpr63
 ; GFX9-NEXT:    ; implicit-def: $vgpr22
-; GFX9-NEXT:    ; implicit-def: $vgpr30
-; GFX9-NEXT:    ; implicit-def: $vgpr24
-; GFX9-NEXT:    ; implicit-def: $vgpr31
+; GFX9-NEXT:    ; implicit-def: $vgpr34
+; GFX9-NEXT:    ; implicit-def: $vgpr28
+; GFX9-NEXT:    ; implicit-def: $vgpr33
 ; GFX9-NEXT:  .LBB110_2: ; %Flow
 ; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; GFX9-NEXT:    s_cbranch_execz .LBB110_4
 ; GFX9-NEXT:  ; %bb.3: ; %cmp.true
-; GFX9-NEXT:    s_waitcnt vmcnt(8)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v18
-; GFX9-NEXT:    v_or_b32_sdwa v0, v31, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(9)
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v38
+; GFX9-NEXT:    v_or_b32_sdwa v0, v28, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(5)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v28
-; GFX9-NEXT:    v_or_b32_sdwa v0, v30, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    s_waitcnt vmcnt(4)
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v24
+; GFX9-NEXT:    v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v20
+; GFX9-NEXT:    v_or_b32_sdwa v0, v18, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v0
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v16
-; GFX9-NEXT:    v_or_b32_sdwa v0, v26, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v44
-; GFX9-NEXT:    v_or_b32_sdwa v0, v22, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v63, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(4)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v34
-; GFX9-NEXT:    v_or_b32_sdwa v0, v36, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v37
+; GFX9-NEXT:    v_or_b32_sdwa v0, v62, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v35
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v61, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v35
-; GFX9-NEXT:    v_add_u16_e32 v2, 3, v47
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v26
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v60, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_or_b32_sdwa v2, v24, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v42
+; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v47
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v59, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v63
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v41
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v58, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v52
+; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v42
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v57, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v62
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v27
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v56, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v51
+; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v29
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v46, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v0
-; GFX9-NEXT:    v_add_u16_e32 v0, 3, v39
+; GFX9-NEXT:    v_add_u16_e32 v0, 3, v25
 ; GFX9-NEXT:    v_or_b32_sdwa v0, v45, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v0
+; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; GFX9-NEXT:    v_add_u16_e32 v3, 3, v32
-; GFX9-NEXT:    v_or_b32_sdwa v3, v20, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v3
+; GFX9-NEXT:    v_add_u16_e32 v3, 3, v31
+; GFX9-NEXT:    v_or_b32_sdwa v3, v30, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v3
+; GFX9-NEXT:    v_add_u16_e32 v2, 3, v32
+; GFX9-NEXT:    v_or_b32_sdwa v2, v34, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v2
-; GFX9-NEXT:    v_add_u16_e32 v1, 3, v38
+; GFX9-NEXT:    v_add_u16_e32 v1, 3, v36
 ; GFX9-NEXT:    v_or_b32_sdwa v1, v33, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v1
+; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v1
 ; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
-; GFX9-NEXT:    v_perm_b32 v8, v30, v8, s6
-; GFX9-NEXT:    v_perm_b32 v9, v28, v9, s6
-; GFX9-NEXT:    v_perm_b32 v10, v26, v10, s6
-; GFX9-NEXT:    v_perm_b32 v11, v24, v11, s6
-; GFX9-NEXT:    v_perm_b32 v12, v22, v12, s6
-; GFX9-NEXT:    v_perm_b32 v13, v16, v13, s6
-; GFX9-NEXT:    v_perm_b32 v14, v20, v14, s6
-; GFX9-NEXT:    v_perm_b32 v15, v18, v15, s6
+; GFX9-NEXT:    v_perm_b32 v8, v25, v8, s6
+; GFX9-NEXT:    v_perm_b32 v9, v24, v9, s6
+; GFX9-NEXT:    v_perm_b32 v10, v22, v10, s6
+; GFX9-NEXT:    v_perm_b32 v11, v20, v11, s6
+; GFX9-NEXT:    v_perm_b32 v12, v16, v12, s6
+; GFX9-NEXT:    v_perm_b32 v13, v18, v13, s6
+; GFX9-NEXT:    v_perm_b32 v14, v30, v14, s6
+; GFX9-NEXT:    v_perm_b32 v15, v28, v15, s6
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v29, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v44, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
-; GFX9-NEXT:    v_or_b32_sdwa v31, v48, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v31, v39, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v27, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; GFX9-NEXT:    v_perm_b32 v7, v27, v7, s6
+; GFX9-NEXT:    v_perm_b32 v7, v26, v7, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v25, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -95153,7 +95155,7 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v5, v19, v5, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v43, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -95164,40 +95166,40 @@ define <32 x bfloat> @bitcast_v64i8_to_v32bf16(<64 x i8> %a, i32 %b) {
 ; GFX9-NEXT:    v_perm_b32 v4, v17, v4, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v41, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v40, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v3, v21, v3, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v55, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v52, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v54, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v51, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; GFX9-NEXT:    v_perm_b32 v2, v25, v2, s6
+; GFX9-NEXT:    v_perm_b32 v2, v27, v2, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v53, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v1, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v50, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v0
 ; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; GFX9-NEXT:    v_perm_b32 v1, v29, v1, s6
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
-; GFX9-NEXT:    v_or_b32_sdwa v0, v49, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v48, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
 ; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
 ; GFX9-NEXT:    v_perm_b32 v0, v31, v0, s6
 ; GFX9-NEXT:  .LBB110_4: ; %end
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll
index dda05a8897979..3a26a5c263d78 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll
@@ -3547,8 +3547,8 @@ define <18 x i32> @bitcast_v36i16_to_v18i32(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:4
 ; SI-NEXT:    v_mov_b32_e32 v36, v22
@@ -3580,9 +3580,9 @@ define <18 x i32> @bitcast_v36i16_to_v18i32(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v6
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -6864,6 +6864,7 @@ define <18 x i32> @bitcast_v36f16_to_v18i32(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v50
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v12
 ; SI-NEXT:    v_cvt_f32_f16_e32 v12, v51
@@ -6887,13 +6888,12 @@ define <18 x i32> @bitcast_v36f16_to_v18i32(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
 ; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
 ; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
@@ -6905,7 +6905,6 @@ define <18 x i32> @bitcast_v36f16_to_v18i32(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    v_or_b32_e32 v14, v16, v14
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
@@ -11192,8 +11191,8 @@ define <18 x float> @bitcast_v36i16_to_v18f32(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:4
 ; SI-NEXT:    v_mov_b32_e32 v36, v22
@@ -11225,9 +11224,9 @@ define <18 x float> @bitcast_v36i16_to_v18f32(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v6
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -14548,6 +14547,7 @@ define <18 x float> @bitcast_v36f16_to_v18f32(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v50
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v12
 ; SI-NEXT:    v_cvt_f32_f16_e32 v12, v51
@@ -14571,13 +14571,12 @@ define <18 x float> @bitcast_v36f16_to_v18f32(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
 ; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
 ; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
@@ -14589,7 +14588,6 @@ define <18 x float> @bitcast_v36f16_to_v18f32(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    v_or_b32_e32 v14, v16, v14
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
@@ -18055,8 +18053,8 @@ define <9 x i64> @bitcast_v36i16_to_v9i64(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:4
 ; SI-NEXT:    v_mov_b32_e32 v36, v22
@@ -18088,9 +18086,9 @@ define <9 x i64> @bitcast_v36i16_to_v9i64(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v6
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -21382,6 +21380,7 @@ define <9 x i64> @bitcast_v36f16_to_v9i64(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v50
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v12
 ; SI-NEXT:    v_cvt_f32_f16_e32 v12, v51
@@ -21405,13 +21404,12 @@ define <9 x i64> @bitcast_v36f16_to_v9i64(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
 ; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
 ; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
@@ -21423,7 +21421,6 @@ define <9 x i64> @bitcast_v36f16_to_v9i64(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    v_or_b32_e32 v14, v16, v14
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
@@ -24140,8 +24137,8 @@ define <9 x double> @bitcast_v36i16_to_v9f64(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:12
 ; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:4
 ; SI-NEXT:    v_mov_b32_e32 v36, v22
@@ -24173,9 +24170,9 @@ define <9 x double> @bitcast_v36i16_to_v9f64(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v6
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
@@ -27415,6 +27412,7 @@ define <9 x double> @bitcast_v36f16_to_v9f64(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v50
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v10, v11, v10
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v12
 ; SI-NEXT:    v_cvt_f32_f16_e32 v12, v51
@@ -27438,13 +27436,12 @@ define <9 x double> @bitcast_v36f16_to_v9f64(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
 ; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
 ; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
@@ -27456,7 +27453,6 @@ define <9 x double> @bitcast_v36f16_to_v9f64(<36 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    v_or_b32_e32 v14, v16, v14
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
@@ -28860,13 +28856,13 @@ define <36 x half> @bitcast_v36i16_to_v36f16(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:4
 ; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; kill: killed $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr48
@@ -28911,7 +28907,7 @@ define <36 x half> @bitcast_v36i16_to_v36f16(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; kill: killed $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
 ; SI-NEXT:    ; implicit-def: $vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
@@ -28939,37 +28935,37 @@ define <36 x half> @bitcast_v36i16_to_v36f16(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v60, v10
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v34
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v39
 ; SI-NEXT:    v_cvt_f32_f16_e32 v45, v11
 ; SI-NEXT:    v_cvt_f32_f16_e32 v58, v12
 ; SI-NEXT:    v_cvt_f32_f16_e32 v43, v13
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v34
 ; SI-NEXT:    v_cvt_f32_f16_e32 v56, v14
 ; SI-NEXT:    v_cvt_f32_f16_e32 v41, v15
 ; SI-NEXT:    v_cvt_f32_f16_e32 v46, v16
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v36
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v35
 ; SI-NEXT:    v_cvt_f32_f16_e32 v55, v17
 ; SI-NEXT:    v_cvt_f32_f16_e32 v44, v18
 ; SI-NEXT:    v_cvt_f32_f16_e32 v53, v19
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v37
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v36
 ; SI-NEXT:    v_cvt_f32_f16_e32 v42, v20
 ; SI-NEXT:    v_cvt_f32_f16_e32 v51, v21
 ; SI-NEXT:    v_cvt_f32_f16_e32 v40, v22
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v38
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v37
 ; SI-NEXT:    v_cvt_f32_f16_e32 v49, v23
 ; SI-NEXT:    v_cvt_f32_f16_e32 v54, v24
 ; SI-NEXT:    v_cvt_f32_f16_e32 v52, v26
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v39
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v38
 ; SI-NEXT:    v_cvt_f32_f16_e32 v50, v28
 ; SI-NEXT:    v_cvt_f32_f16_e32 v48, v29
 ; SI-NEXT:    ; implicit-def: $vgpr2
@@ -29001,12 +28997,12 @@ define <36 x half> @bitcast_v36i16_to_v36f16(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr29
 ; SI-NEXT:    ; implicit-def: $vgpr30
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr38
-; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr1
 ; SI-NEXT:  .LBB56_2: ; %Flow
@@ -29024,36 +29020,36 @@ define <36 x half> @bitcast_v36i16_to_v36f16(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v27
 ; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_i32_e32 v39, vcc, 3, v39
 ; SI-NEXT:    v_add_i32_e32 v34, vcc, 3, v34
-; SI-NEXT:    v_add_i32_e32 v35, vcc, 3, v35
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v30
+; SI-NEXT:    v_add_i32_e32 v35, vcc, 3, v35
 ; SI-NEXT:    v_add_i32_e32 v36, vcc, 3, v36
-; SI-NEXT:    v_add_i32_e32 v37, vcc, 3, v37
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v34
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v39
+; SI-NEXT:    v_add_i32_e32 v37, vcc, 3, v37
 ; SI-NEXT:    v_add_i32_e32 v38, vcc, 3, v38
-; SI-NEXT:    v_add_i32_e32 v39, vcc, 3, v39
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v34
 ; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
 ; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v36
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v35
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v37
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v36
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v38
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v37
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
@@ -29102,7 +29098,7 @@ define <36 x half> @bitcast_v36i16_to_v36f16(<36 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v48, v29
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v39
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v38
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
 ; SI-NEXT:  .LBB56_4: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll
index 4a8000ba52752..cc55ba1d84df6 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll
@@ -3698,26 +3698,24 @@ define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v50, v10
-; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    v_mov_b32_e32 v50, v10
+; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v37, v20
 ; SI-NEXT:    v_mov_b32_e32 v38, v18
 ; SI-NEXT:    v_mov_b32_e32 v39, v16
 ; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v11
@@ -3732,23 +3730,25 @@ define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v29
 ; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v0
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -3761,14 +3761,14 @@ define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB14_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr20
@@ -3795,8 +3795,8 @@ define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v42
-; SI-NEXT:    v_or_b32_e32 v1, v1, v36
-; SI-NEXT:    v_or_b32_e32 v2, v2, v41
+; SI-NEXT:    v_or_b32_e32 v1, v1, v41
+; SI-NEXT:    v_or_b32_e32 v2, v2, v36
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v35
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v40
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v34
@@ -3822,8 +3822,8 @@ define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr34
@@ -3836,19 +3836,19 @@ define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v60
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v59
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v58
@@ -3870,14 +3870,14 @@ define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB14_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -3904,8 +3904,8 @@ define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    v_or_b32_e32 v0, v42, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v36, v1
-; SI-NEXT:    v_or_b32_e32 v2, v41, v2
+; SI-NEXT:    v_or_b32_e32 v1, v41, v1
+; SI-NEXT:    v_or_b32_e32 v2, v36, v2
 ; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_or_b32_e32 v4, v40, v4
 ; SI-NEXT:    v_or_b32_e32 v5, v34, v5
@@ -3929,15 +3929,13 @@ define <20 x i32> @bitcast_v40i16_to_v20i32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
@@ -7274,6 +7272,8 @@ define <20 x i32> @bitcast_v40f16_to_v20i32(<40 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v39
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v37
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v38
 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v36
 ; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
@@ -7362,58 +7362,55 @@ define <20 x i32> @bitcast_v40f16_to_v20i32(<40 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
 ; SI-NEXT:    v_cvt_f32_f16_e32 v11, v41
 ; SI-NEXT:    v_or_b32_e32 v10, v12, v10
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
+; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
 ; SI-NEXT:    v_or_b32_e32 v11, v13, v11
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
-; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
 ; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
-; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
-; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
+; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v13, v15, v13
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -11800,26 +11797,24 @@ define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v50, v10
-; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    v_mov_b32_e32 v50, v10
+; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v37, v20
 ; SI-NEXT:    v_mov_b32_e32 v38, v18
 ; SI-NEXT:    v_mov_b32_e32 v39, v16
 ; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v11
@@ -11834,23 +11829,25 @@ define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v29
 ; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v0
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -11863,14 +11860,14 @@ define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB30_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr20
@@ -11897,8 +11894,8 @@ define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v42
-; SI-NEXT:    v_or_b32_e32 v1, v1, v36
-; SI-NEXT:    v_or_b32_e32 v2, v2, v41
+; SI-NEXT:    v_or_b32_e32 v1, v1, v41
+; SI-NEXT:    v_or_b32_e32 v2, v2, v36
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v35
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v40
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v34
@@ -11924,8 +11921,8 @@ define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr34
@@ -11938,19 +11935,19 @@ define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v60
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v59
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v58
@@ -11972,14 +11969,14 @@ define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB30_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -12006,8 +12003,8 @@ define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    v_or_b32_e32 v0, v42, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v36, v1
-; SI-NEXT:    v_or_b32_e32 v2, v41, v2
+; SI-NEXT:    v_or_b32_e32 v1, v41, v1
+; SI-NEXT:    v_or_b32_e32 v2, v36, v2
 ; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_or_b32_e32 v4, v40, v4
 ; SI-NEXT:    v_or_b32_e32 v5, v34, v5
@@ -12031,15 +12028,13 @@ define <20 x float> @bitcast_v40i16_to_v20f32(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
@@ -15364,6 +15359,8 @@ define <20 x float> @bitcast_v40f16_to_v20f32(<40 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v39
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v37
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v38
 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v36
 ; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
@@ -15452,58 +15449,55 @@ define <20 x float> @bitcast_v40f16_to_v20f32(<40 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
 ; SI-NEXT:    v_cvt_f32_f16_e32 v11, v41
 ; SI-NEXT:    v_or_b32_e32 v10, v12, v10
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
+; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
 ; SI-NEXT:    v_or_b32_e32 v11, v13, v11
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
-; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
 ; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
-; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
-; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
+; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v13, v15, v13
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -19214,26 +19208,24 @@ define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v50, v10
-; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    v_mov_b32_e32 v50, v10
+; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v37, v20
 ; SI-NEXT:    v_mov_b32_e32 v38, v18
 ; SI-NEXT:    v_mov_b32_e32 v39, v16
 ; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v11
@@ -19248,23 +19240,25 @@ define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v29
 ; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v0
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -19277,14 +19271,14 @@ define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB42_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr20
@@ -19311,8 +19305,8 @@ define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v42
-; SI-NEXT:    v_or_b32_e32 v1, v1, v36
-; SI-NEXT:    v_or_b32_e32 v2, v2, v41
+; SI-NEXT:    v_or_b32_e32 v1, v1, v41
+; SI-NEXT:    v_or_b32_e32 v2, v2, v36
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v35
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v40
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v34
@@ -19338,8 +19332,8 @@ define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr34
@@ -19352,19 +19346,19 @@ define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v60
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v59
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v58
@@ -19386,14 +19380,14 @@ define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB42_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -19420,8 +19414,8 @@ define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    v_or_b32_e32 v0, v42, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v36, v1
-; SI-NEXT:    v_or_b32_e32 v2, v41, v2
+; SI-NEXT:    v_or_b32_e32 v1, v41, v1
+; SI-NEXT:    v_or_b32_e32 v2, v36, v2
 ; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_or_b32_e32 v4, v40, v4
 ; SI-NEXT:    v_or_b32_e32 v5, v34, v5
@@ -19445,15 +19439,13 @@ define <10 x i64> @bitcast_v40i16_to_v10i64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
@@ -22800,6 +22792,8 @@ define <10 x i64> @bitcast_v40f16_to_v10i64(<40 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v39
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v37
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v38
 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v36
 ; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
@@ -22888,58 +22882,55 @@ define <10 x i64> @bitcast_v40f16_to_v10i64(<40 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
 ; SI-NEXT:    v_cvt_f32_f16_e32 v11, v41
 ; SI-NEXT:    v_or_b32_e32 v10, v12, v10
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
+; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
 ; SI-NEXT:    v_or_b32_e32 v11, v13, v11
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
-; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
 ; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
-; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
-; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
+; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v13, v15, v13
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -25855,26 +25846,24 @@ define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v50, v10
-; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    v_mov_b32_e32 v50, v10
+; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v37, v20
 ; SI-NEXT:    v_mov_b32_e32 v38, v18
 ; SI-NEXT:    v_mov_b32_e32 v39, v16
 ; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v5
 ; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v11
@@ -25889,23 +25878,25 @@ define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v29
 ; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v0
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
@@ -25918,14 +25909,14 @@ define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB50_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr20
@@ -25952,8 +25943,8 @@ define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v42
-; SI-NEXT:    v_or_b32_e32 v1, v1, v36
-; SI-NEXT:    v_or_b32_e32 v2, v2, v41
+; SI-NEXT:    v_or_b32_e32 v1, v1, v41
+; SI-NEXT:    v_or_b32_e32 v2, v2, v36
 ; SI-NEXT:    v_or_b32_e32 v3, v3, v35
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v40
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v34
@@ -25979,8 +25970,8 @@ define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr20
 ; SI-NEXT:    ; kill: killed $vgpr20
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr34
@@ -25993,19 +25984,19 @@ define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v60
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v59
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v58
@@ -26027,14 +26018,14 @@ define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_cbranch_execz .LBB50_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -26061,8 +26052,8 @@ define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    v_or_b32_e32 v0, v42, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v36, v1
-; SI-NEXT:    v_or_b32_e32 v2, v41, v2
+; SI-NEXT:    v_or_b32_e32 v1, v41, v1
+; SI-NEXT:    v_or_b32_e32 v2, v36, v2
 ; SI-NEXT:    v_or_b32_e32 v3, v35, v3
 ; SI-NEXT:    v_or_b32_e32 v4, v40, v4
 ; SI-NEXT:    v_or_b32_e32 v5, v34, v5
@@ -26086,15 +26077,13 @@ define <10 x double> @bitcast_v40i16_to_v10f64(<40 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
+; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
@@ -29351,6 +29340,8 @@ define <10 x double> @bitcast_v40f16_to_v10f64(<40 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v39
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v37
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v38
 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v36
 ; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
@@ -29439,58 +29430,55 @@ define <10 x double> @bitcast_v40f16_to_v10f64(<40 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
 ; SI-NEXT:    v_or_b32_e32 v9, v10, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
 ; SI-NEXT:    v_cvt_f32_f16_e32 v11, v41
 ; SI-NEXT:    v_or_b32_e32 v10, v12, v10
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
+; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
 ; SI-NEXT:    v_or_b32_e32 v11, v13, v11
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
-; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
 ; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
-; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
-; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
+; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    v_or_b32_e32 v12, v13, v12
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v13, v15, v13
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(1)
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll
index 4141c33eca786..2a96722ccce0b 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll
@@ -3960,93 +3960,98 @@ define <22 x i32> @bitcast_v44i16_to_v22i32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    v_mov_b32_e32 v54, v2
+; SI-NEXT:    v_mov_b32_e32 v55, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:12
+; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v50, v10
 ; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
-; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_mov_b32_e32 v54, v2
-; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:44
-; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_mov_b32_e32 v39, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
+; SI-NEXT:    v_mov_b32_e32 v48, v14
+; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    v_mov_b32_e32 v38, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
+; SI-NEXT:    s_waitcnt expcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v2
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v4
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v4
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v8
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:24
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v10
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
@@ -4075,20 +4080,19 @@ define <22 x i32> @bitcast_v44i16_to_v22i32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v49
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v48
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v39
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v45
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v38
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    v_or_b32_e32 v0, v0, v38
-; SI-NEXT:    v_or_b32_e32 v1, v1, v44
-; SI-NEXT:    v_or_b32_e32 v2, v2, v43
-; SI-NEXT:    v_or_b32_e32 v3, v3, v37
+; SI-NEXT:    v_or_b32_e32 v0, v0, v37
+; SI-NEXT:    v_or_b32_e32 v1, v1, v45
+; SI-NEXT:    v_or_b32_e32 v2, v2, v44
+; SI-NEXT:    v_or_b32_e32 v3, v3, v43
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v36
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v42
-; SI-NEXT:    v_or_b32_e32 v6, v6, v35
-; SI-NEXT:    v_or_b32_e32 v7, v7, v41
+; SI-NEXT:    v_or_b32_e32 v6, v6, v41
+; SI-NEXT:    v_or_b32_e32 v7, v7, v35
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v34
-; SI-NEXT:    v_or_b32_e32 v16, v16, v59
+; SI-NEXT:    v_or_b32_e32 v9, v9, v40
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr53
@@ -4098,63 +4102,63 @@ define <22 x i32> @bitcast_v44i16_to_v22i32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    ; kill: killed $vgpr22
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_or_b32_e32 v9, v9, v40
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v33
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v32
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v63
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v62
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v61
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v60
+; SI-NEXT:    v_or_b32_e32 v16, v16, v59
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v58
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v57
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v56
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v47
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v46
-; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr59
 ; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr56
@@ -4164,18 +4168,19 @@ define <22 x i32> @bitcast_v44i16_to_v22i32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -4185,8 +4190,7 @@ define <22 x i32> @bitcast_v44i16_to_v22i32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v49
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v48
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v39
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v45
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v38
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -4196,18 +4200,18 @@ define <22 x i32> @bitcast_v44i16_to_v22i32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v0, v38, v0
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_or_b32_e32 v0, v37, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v44, v1
-; SI-NEXT:    v_or_b32_e32 v2, v43, v2
-; SI-NEXT:    v_or_b32_e32 v3, v37, v3
+; SI-NEXT:    v_or_b32_e32 v1, v45, v1
+; SI-NEXT:    v_or_b32_e32 v2, v44, v2
+; SI-NEXT:    v_or_b32_e32 v3, v43, v3
 ; SI-NEXT:    v_or_b32_e32 v4, v36, v4
 ; SI-NEXT:    v_or_b32_e32 v5, v42, v5
-; SI-NEXT:    v_or_b32_e32 v6, v35, v6
-; SI-NEXT:    v_or_b32_e32 v7, v41, v7
+; SI-NEXT:    v_or_b32_e32 v6, v41, v6
+; SI-NEXT:    v_or_b32_e32 v7, v35, v7
 ; SI-NEXT:    v_or_b32_e32 v8, v34, v8
-; SI-NEXT:    v_or_b32_e32 v16, v59, v16
+; SI-NEXT:    v_or_b32_e32 v9, v40, v9
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
@@ -4217,58 +4221,57 @@ define <22 x i32> @bitcast_v44i16_to_v22i32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_or_b32_e32 v9, v40, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v33, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v32, v11
 ; SI-NEXT:    v_or_b32_e32 v12, v63, v12
 ; SI-NEXT:    v_or_b32_e32 v13, v62, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v61, v14
 ; SI-NEXT:    v_or_b32_e32 v15, v60, v15
+; SI-NEXT:    v_or_b32_e32 v16, v59, v16
 ; SI-NEXT:    v_or_b32_e32 v17, v58, v17
 ; SI-NEXT:    v_or_b32_e32 v18, v57, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v56, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v47, v20
 ; SI-NEXT:    v_or_b32_e32 v21, v46, v21
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
@@ -7996,73 +7999,72 @@ define <22 x i32> @bitcast_v44f16_to_v22i32(<44 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v6, v5
 ; SI-NEXT:    v_cvt_f32_f16_e32 v6, v63
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v40
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v53
-; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
-; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v61
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v40
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v53
 ; SI-NEXT:    v_cvt_f32_f16_e32 v23, v52
-; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
-; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
 ; SI-NEXT:    v_cvt_f32_f16_e32 v8, v59
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
+; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
+; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v10
 ; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
+; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_or_b32_e32 v9, v11, v9
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v12, v10
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
@@ -12848,93 +12850,98 @@ define <22 x float> @bitcast_v44i16_to_v22f32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    v_mov_b32_e32 v54, v2
+; SI-NEXT:    v_mov_b32_e32 v55, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:12
+; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v50, v10
 ; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
-; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_mov_b32_e32 v54, v2
-; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:44
-; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_mov_b32_e32 v39, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
+; SI-NEXT:    v_mov_b32_e32 v48, v14
+; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    v_mov_b32_e32 v38, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
+; SI-NEXT:    s_waitcnt expcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v2
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v4
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v0
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:4
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v4
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:24
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v10
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
@@ -12963,20 +12970,19 @@ define <22 x float> @bitcast_v44i16_to_v22f32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v49
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v48
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v39
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v45
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v38
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    v_or_b32_e32 v0, v0, v38
-; SI-NEXT:    v_or_b32_e32 v1, v1, v44
-; SI-NEXT:    v_or_b32_e32 v2, v2, v43
-; SI-NEXT:    v_or_b32_e32 v3, v3, v37
+; SI-NEXT:    v_or_b32_e32 v0, v0, v37
+; SI-NEXT:    v_or_b32_e32 v1, v1, v45
+; SI-NEXT:    v_or_b32_e32 v2, v2, v44
+; SI-NEXT:    v_or_b32_e32 v3, v3, v43
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v36
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v42
-; SI-NEXT:    v_or_b32_e32 v6, v6, v35
-; SI-NEXT:    v_or_b32_e32 v7, v7, v41
+; SI-NEXT:    v_or_b32_e32 v6, v6, v41
+; SI-NEXT:    v_or_b32_e32 v7, v7, v35
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v34
-; SI-NEXT:    v_or_b32_e32 v16, v16, v59
+; SI-NEXT:    v_or_b32_e32 v9, v9, v40
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr53
@@ -12986,63 +12992,63 @@ define <22 x float> @bitcast_v44i16_to_v22f32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    ; kill: killed $vgpr22
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_or_b32_e32 v9, v9, v40
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v33
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v32
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v63
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v62
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v61
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v60
+; SI-NEXT:    v_or_b32_e32 v16, v16, v59
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v58
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v57
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v56
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v47
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v46
-; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr59
 ; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr56
@@ -13052,18 +13058,19 @@ define <22 x float> @bitcast_v44i16_to_v22f32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -13073,8 +13080,7 @@ define <22 x float> @bitcast_v44i16_to_v22f32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v49
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v48
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v39
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v45
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v38
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -13084,18 +13090,18 @@ define <22 x float> @bitcast_v44i16_to_v22f32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v0, v38, v0
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_or_b32_e32 v0, v37, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v44, v1
-; SI-NEXT:    v_or_b32_e32 v2, v43, v2
-; SI-NEXT:    v_or_b32_e32 v3, v37, v3
+; SI-NEXT:    v_or_b32_e32 v1, v45, v1
+; SI-NEXT:    v_or_b32_e32 v2, v44, v2
+; SI-NEXT:    v_or_b32_e32 v3, v43, v3
 ; SI-NEXT:    v_or_b32_e32 v4, v36, v4
 ; SI-NEXT:    v_or_b32_e32 v5, v42, v5
-; SI-NEXT:    v_or_b32_e32 v6, v35, v6
-; SI-NEXT:    v_or_b32_e32 v7, v41, v7
+; SI-NEXT:    v_or_b32_e32 v6, v41, v6
+; SI-NEXT:    v_or_b32_e32 v7, v35, v7
 ; SI-NEXT:    v_or_b32_e32 v8, v34, v8
-; SI-NEXT:    v_or_b32_e32 v16, v59, v16
+; SI-NEXT:    v_or_b32_e32 v9, v40, v9
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
@@ -13105,58 +13111,57 @@ define <22 x float> @bitcast_v44i16_to_v22f32(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_or_b32_e32 v9, v40, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v33, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v32, v11
 ; SI-NEXT:    v_or_b32_e32 v12, v63, v12
 ; SI-NEXT:    v_or_b32_e32 v13, v62, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v61, v14
 ; SI-NEXT:    v_or_b32_e32 v15, v60, v15
+; SI-NEXT:    v_or_b32_e32 v16, v59, v16
 ; SI-NEXT:    v_or_b32_e32 v17, v58, v17
 ; SI-NEXT:    v_or_b32_e32 v18, v57, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v56, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v47, v20
 ; SI-NEXT:    v_or_b32_e32 v21, v46, v21
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
@@ -16884,73 +16889,72 @@ define <22 x float> @bitcast_v44f16_to_v22f32(<44 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v6, v5
 ; SI-NEXT:    v_cvt_f32_f16_e32 v6, v63
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v40
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v53
-; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
-; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v61
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v40
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v53
 ; SI-NEXT:    v_cvt_f32_f16_e32 v23, v52
-; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
-; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
 ; SI-NEXT:    v_cvt_f32_f16_e32 v8, v59
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
+; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
+; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v10
 ; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
+; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_or_b32_e32 v9, v11, v9
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v12, v10
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
@@ -21026,93 +21030,98 @@ define <11 x i64> @bitcast_v44i16_to_v11i64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    v_mov_b32_e32 v54, v2
+; SI-NEXT:    v_mov_b32_e32 v55, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:12
+; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v50, v10
 ; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
-; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_mov_b32_e32 v54, v2
-; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:44
-; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_mov_b32_e32 v39, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
+; SI-NEXT:    v_mov_b32_e32 v48, v14
+; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    v_mov_b32_e32 v38, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
+; SI-NEXT:    s_waitcnt expcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v2
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v4
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v0
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:4
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v4
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:24
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v10
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
@@ -21141,20 +21150,19 @@ define <11 x i64> @bitcast_v44i16_to_v11i64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v49
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v48
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v39
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v45
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v38
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    v_or_b32_e32 v0, v0, v38
-; SI-NEXT:    v_or_b32_e32 v1, v1, v44
-; SI-NEXT:    v_or_b32_e32 v2, v2, v43
-; SI-NEXT:    v_or_b32_e32 v3, v3, v37
+; SI-NEXT:    v_or_b32_e32 v0, v0, v37
+; SI-NEXT:    v_or_b32_e32 v1, v1, v45
+; SI-NEXT:    v_or_b32_e32 v2, v2, v44
+; SI-NEXT:    v_or_b32_e32 v3, v3, v43
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v36
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v42
-; SI-NEXT:    v_or_b32_e32 v6, v6, v35
-; SI-NEXT:    v_or_b32_e32 v7, v7, v41
+; SI-NEXT:    v_or_b32_e32 v6, v6, v41
+; SI-NEXT:    v_or_b32_e32 v7, v7, v35
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v34
-; SI-NEXT:    v_or_b32_e32 v16, v16, v59
+; SI-NEXT:    v_or_b32_e32 v9, v9, v40
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr53
@@ -21164,63 +21172,63 @@ define <11 x i64> @bitcast_v44i16_to_v11i64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    ; kill: killed $vgpr22
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_or_b32_e32 v9, v9, v40
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v33
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v32
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v63
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v62
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v61
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v60
+; SI-NEXT:    v_or_b32_e32 v16, v16, v59
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v58
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v57
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v56
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v47
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v46
-; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr59
 ; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr56
@@ -21230,18 +21238,19 @@ define <11 x i64> @bitcast_v44i16_to_v11i64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -21251,8 +21260,7 @@ define <11 x i64> @bitcast_v44i16_to_v11i64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v49
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v48
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v39
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v45
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v38
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -21262,18 +21270,18 @@ define <11 x i64> @bitcast_v44i16_to_v11i64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v0, v38, v0
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_or_b32_e32 v0, v37, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v44, v1
-; SI-NEXT:    v_or_b32_e32 v2, v43, v2
-; SI-NEXT:    v_or_b32_e32 v3, v37, v3
+; SI-NEXT:    v_or_b32_e32 v1, v45, v1
+; SI-NEXT:    v_or_b32_e32 v2, v44, v2
+; SI-NEXT:    v_or_b32_e32 v3, v43, v3
 ; SI-NEXT:    v_or_b32_e32 v4, v36, v4
 ; SI-NEXT:    v_or_b32_e32 v5, v42, v5
-; SI-NEXT:    v_or_b32_e32 v6, v35, v6
-; SI-NEXT:    v_or_b32_e32 v7, v41, v7
+; SI-NEXT:    v_or_b32_e32 v6, v41, v6
+; SI-NEXT:    v_or_b32_e32 v7, v35, v7
 ; SI-NEXT:    v_or_b32_e32 v8, v34, v8
-; SI-NEXT:    v_or_b32_e32 v16, v59, v16
+; SI-NEXT:    v_or_b32_e32 v9, v40, v9
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
@@ -21283,58 +21291,57 @@ define <11 x i64> @bitcast_v44i16_to_v11i64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_or_b32_e32 v9, v40, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v33, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v32, v11
 ; SI-NEXT:    v_or_b32_e32 v12, v63, v12
 ; SI-NEXT:    v_or_b32_e32 v13, v62, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v61, v14
 ; SI-NEXT:    v_or_b32_e32 v15, v60, v15
+; SI-NEXT:    v_or_b32_e32 v16, v59, v16
 ; SI-NEXT:    v_or_b32_e32 v17, v58, v17
 ; SI-NEXT:    v_or_b32_e32 v18, v57, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v56, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v47, v20
 ; SI-NEXT:    v_or_b32_e32 v21, v46, v21
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
@@ -25074,73 +25081,72 @@ define <11 x i64> @bitcast_v44f16_to_v11i64(<44 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v6, v5
 ; SI-NEXT:    v_cvt_f32_f16_e32 v6, v63
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v40
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v53
-; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
-; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v61
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v40
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v53
 ; SI-NEXT:    v_cvt_f32_f16_e32 v23, v52
-; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
-; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
 ; SI-NEXT:    v_cvt_f32_f16_e32 v8, v59
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
+; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
+; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v10
 ; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
+; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_or_b32_e32 v9, v11, v9
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v12, v10
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
@@ -28366,93 +28372,98 @@ define <11 x double> @bitcast_v44i16_to_v11f64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    v_mov_b32_e32 v54, v2
+; SI-NEXT:    v_mov_b32_e32 v55, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:12
+; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v50, v10
 ; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
-; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_mov_b32_e32 v54, v2
-; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:44
-; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_mov_b32_e32 v39, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
+; SI-NEXT:    v_mov_b32_e32 v48, v14
+; SI-NEXT:    v_mov_b32_e32 v49, v12
+; SI-NEXT:    v_mov_b32_e32 v38, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v7
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v9
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v15
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
+; SI-NEXT:    s_waitcnt expcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v2
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v4
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v0
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:4
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v4
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:24
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v10
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
@@ -28481,20 +28492,19 @@ define <11 x double> @bitcast_v44i16_to_v11f64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v49
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v48
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v39
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v45
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v38
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
-; SI-NEXT:    v_or_b32_e32 v0, v0, v38
-; SI-NEXT:    v_or_b32_e32 v1, v1, v44
-; SI-NEXT:    v_or_b32_e32 v2, v2, v43
-; SI-NEXT:    v_or_b32_e32 v3, v3, v37
+; SI-NEXT:    v_or_b32_e32 v0, v0, v37
+; SI-NEXT:    v_or_b32_e32 v1, v1, v45
+; SI-NEXT:    v_or_b32_e32 v2, v2, v44
+; SI-NEXT:    v_or_b32_e32 v3, v3, v43
 ; SI-NEXT:    v_or_b32_e32 v4, v4, v36
 ; SI-NEXT:    v_or_b32_e32 v5, v5, v42
-; SI-NEXT:    v_or_b32_e32 v6, v6, v35
-; SI-NEXT:    v_or_b32_e32 v7, v7, v41
+; SI-NEXT:    v_or_b32_e32 v6, v6, v41
+; SI-NEXT:    v_or_b32_e32 v7, v7, v35
 ; SI-NEXT:    v_or_b32_e32 v8, v8, v34
-; SI-NEXT:    v_or_b32_e32 v16, v16, v59
+; SI-NEXT:    v_or_b32_e32 v9, v9, v40
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr53
@@ -28504,63 +28514,63 @@ define <11 x double> @bitcast_v44i16_to_v11f64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; kill: killed $vgpr22
 ; SI-NEXT:    ; implicit-def: $vgpr22
 ; SI-NEXT:    ; kill: killed $vgpr22
-; SI-NEXT:    ; implicit-def: $vgpr38
+; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr45
 ; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; SI-NEXT:    v_or_b32_e32 v9, v9, v40
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v33
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v32
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v63
 ; SI-NEXT:    v_or_b32_e32 v13, v13, v62
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v61
 ; SI-NEXT:    v_or_b32_e32 v15, v15, v60
+; SI-NEXT:    v_or_b32_e32 v16, v16, v59
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v58
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v57
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v56
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v47
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v46
-; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr59
 ; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr56
@@ -28570,18 +28580,19 @@ define <11 x double> @bitcast_v44i16_to_v11f64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -28591,8 +28602,7 @@ define <11 x double> @bitcast_v44i16_to_v11f64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v49
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v48
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v39
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v45
+; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v38
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -28602,18 +28612,18 @@ define <11 x double> @bitcast_v44i16_to_v11f64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v0, v38, v0
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    v_or_b32_e32 v0, v37, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v44, v1
-; SI-NEXT:    v_or_b32_e32 v2, v43, v2
-; SI-NEXT:    v_or_b32_e32 v3, v37, v3
+; SI-NEXT:    v_or_b32_e32 v1, v45, v1
+; SI-NEXT:    v_or_b32_e32 v2, v44, v2
+; SI-NEXT:    v_or_b32_e32 v3, v43, v3
 ; SI-NEXT:    v_or_b32_e32 v4, v36, v4
 ; SI-NEXT:    v_or_b32_e32 v5, v42, v5
-; SI-NEXT:    v_or_b32_e32 v6, v35, v6
-; SI-NEXT:    v_or_b32_e32 v7, v41, v7
+; SI-NEXT:    v_or_b32_e32 v6, v41, v6
+; SI-NEXT:    v_or_b32_e32 v7, v35, v7
 ; SI-NEXT:    v_or_b32_e32 v8, v34, v8
-; SI-NEXT:    v_or_b32_e32 v16, v59, v16
+; SI-NEXT:    v_or_b32_e32 v9, v40, v9
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
@@ -28623,58 +28633,57 @@ define <11 x double> @bitcast_v44i16_to_v11f64(<44 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
-; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
+; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_or_b32_e32 v9, v40, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v33, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v32, v11
 ; SI-NEXT:    v_or_b32_e32 v12, v63, v12
 ; SI-NEXT:    v_or_b32_e32 v13, v62, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v61, v14
 ; SI-NEXT:    v_or_b32_e32 v15, v60, v15
+; SI-NEXT:    v_or_b32_e32 v16, v59, v16
 ; SI-NEXT:    v_or_b32_e32 v17, v58, v17
 ; SI-NEXT:    v_or_b32_e32 v18, v57, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v56, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v47, v20
 ; SI-NEXT:    v_or_b32_e32 v21, v46, v21
-; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
+; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
@@ -32325,73 +32334,72 @@ define <11 x double> @bitcast_v44f16_to_v11f64(<44 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v6, v5
 ; SI-NEXT:    v_cvt_f32_f16_e32 v6, v63
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v40
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v53
-; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
-; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; SI-NEXT:    v_cvt_f32_f16_e32 v7, v61
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v40
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v53
 ; SI-NEXT:    v_cvt_f32_f16_e32 v23, v52
-; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
-; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
 ; SI-NEXT:    v_cvt_f32_f16_e32 v8, v59
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
+; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
+; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
 ; SI-NEXT:    v_or_b32_e32 v8, v9, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v10
 ; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
+; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_or_b32_e32 v9, v11, v9
 ; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
-; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
 ; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
 ; SI-NEXT:    v_or_b32_e32 v10, v12, v10
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
@@ -36331,162 +36339,157 @@ define <44 x i16> @bitcast_v44f16_to_v44i16(<44 x half> %a, i32 %b) {
 ; SI-LABEL: bitcast_v44f16_to_v44i16:
 ; SI:       ; %bb.0:
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:4
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:16
 ; SI-NEXT:    s_waitcnt expcnt(6)
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32
 ; SI-NEXT:    s_waitcnt expcnt(5)
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(4)
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(3)
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:12
 ; SI-NEXT:    s_waitcnt expcnt(2)
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:16
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:40
-; SI-NEXT:    v_cvt_f16_f32_e32 v51, v1
-; SI-NEXT:    v_cvt_f16_f32_e32 v41, v2
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:24
 ; SI-NEXT:    v_cvt_f16_f32_e32 v48, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v49, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v55, v10
+; SI-NEXT:    v_cvt_f16_f32_e32 v34, v11
+; SI-NEXT:    v_cvt_f16_f32_e32 v35, v12
+; SI-NEXT:    v_cvt_f16_f32_e32 v33, v17
+; SI-NEXT:    v_cvt_f16_f32_e32 v53, v18
+; SI-NEXT:    v_cvt_f16_f32_e32 v17, v25
+; SI-NEXT:    v_cvt_f16_f32_e32 v25, v26
+; SI-NEXT:    v_cvt_f16_f32_e32 v12, v27
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:52
+; SI-NEXT:    v_cvt_f16_f32_e32 v51, v1
+; SI-NEXT:    v_cvt_f16_f32_e32 v41, v2
 ; SI-NEXT:    v_cvt_f16_f32_e32 v50, v5
 ; SI-NEXT:    v_cvt_f16_f32_e32 v40, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v37, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v38, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v39, v9
-; SI-NEXT:    v_cvt_f16_f32_e32 v55, v10
-; SI-NEXT:    v_cvt_f16_f32_e32 v34, v11
-; SI-NEXT:    v_cvt_f16_f32_e32 v35, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v36, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v54, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v31, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v32, v16
-; SI-NEXT:    v_cvt_f16_f32_e32 v33, v17
-; SI-NEXT:    v_cvt_f16_f32_e32 v53, v18
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v52, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v16, v23
-; SI-NEXT:    v_cvt_f16_f32_e32 v17, v24
-; SI-NEXT:    v_cvt_f16_f32_e32 v18, v25
-; SI-NEXT:    v_cvt_f16_f32_e32 v25, v26
-; SI-NEXT:    v_cvt_f16_f32_e32 v11, v27
-; SI-NEXT:    v_cvt_f16_f32_e32 v12, v28
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v23
+; SI-NEXT:    v_cvt_f16_f32_e32 v16, v24
+; SI-NEXT:    v_cvt_f16_f32_e32 v13, v28
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v29
 ; SI-NEXT:    v_cvt_f16_f32_e32 v24, v30
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v42
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f16_f32_e32 v8, v43
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v43
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f16_f32_e32 v9, v44
+; SI-NEXT:    v_cvt_f16_f32_e32 v7, v44
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_cvt_f16_f32_e32 v10, v45
+; SI-NEXT:    v_cvt_f16_f32_e32 v8, v45
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v23, v46
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v47
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v47
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v56
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v56
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v57
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v57
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cvt_f16_f32_e32 v13, v58
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v3
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v59
+; SI-NEXT:    v_cvt_f16_f32_e32 v9, v4
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v60
+; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v61
+; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v62
+; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v27
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB58_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.true
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
 ; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
 ; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
-; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
-; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
-; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
-; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
-; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
-; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
-; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
+; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
+; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
-; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
+; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
+; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
+; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v2
-; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
-; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
-; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
-; SI-NEXT:    v_or_b32_e32 v1, v1, v26
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v6
-; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
-; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
-; SI-NEXT:    v_or_b32_e32 v5, v5, v26
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v4
+; SI-NEXT:    v_or_b32_e32 v3, v3, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v10
+; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
+; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
+; SI-NEXT:    v_or_b32_e32 v9, v9, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v2
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
-; SI-NEXT:    v_or_b32_e32 v3, v3, v26
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v9
+; SI-NEXT:    v_or_b32_e32 v1, v1, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    v_or_b32_e32 v8, v8, v26
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v12
-; SI-NEXT:    v_or_b32_e32 v11, v11, v26
-; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v17
-; SI-NEXT:    v_or_b32_e32 v16, v16, v26
+; SI-NEXT:    v_or_b32_e32 v6, v6, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v13
+; SI-NEXT:    v_or_b32_e32 v12, v12, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v16
+; SI-NEXT:    v_or_b32_e32 v15, v15, v26
 ; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v20
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v26
 ; SI-NEXT:    v_cvt_f32_f16_e32 v26, v32
@@ -36533,30 +36536,30 @@ define <44 x i16> @bitcast_v44f16_to_v44i16(<44 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v49, v26
 ; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v27
 ; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
-; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
+; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
 ; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v49
 ; SI-NEXT:    v_cvt_f32_f16_e32 v33, v33
 ; SI-NEXT:    v_or_b32_e32 v48, v26, v27
 ; SI-NEXT:    v_cvt_f32_f16_e32 v26, v41
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v51
 ; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
+; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
 ; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
 ; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
 ; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    v_or_b32_e32 v51, v27, v26
 ; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
 ; SI-NEXT:    v_cvt_f32_f16_e32 v28, v55
 ; SI-NEXT:    v_or_b32_e32 v50, v29, v27
 ; SI-NEXT:    v_cvt_f32_f16_e32 v29, v54
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
 ; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
 ; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
 ; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
 ; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
@@ -36570,29 +36573,29 @@ define <44 x i16> @bitcast_v44f16_to_v44i16(<44 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
 ; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
+; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
 ; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
 ; SI-NEXT:    v_cvt_f16_f32_e32 v53, v36
 ; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
 ; SI-NEXT:    v_cvt_f16_f32_e32 v52, v52
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
+; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
 ; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
-; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
+; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
-; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
-; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
+; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
+; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
-; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
+; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
 ; SI-NEXT:    v_or_b32_e32 v36, v30, v29
 ; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v53
@@ -36601,25 +36604,25 @@ define <44 x i16> @bitcast_v44f16_to_v44i16(<44 x half> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
 ; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; SI-NEXT:    v_or_b32_e32 v33, v33, v30
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v52
-; SI-NEXT:    v_or_b32_e32 v18, v18, v25
+; SI-NEXT:    v_or_b32_e32 v17, v17, v25
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v24
-; SI-NEXT:    v_or_b32_e32 v10, v10, v23
-; SI-NEXT:    v_or_b32_e32 v13, v13, v22
-; SI-NEXT:    v_or_b32_e32 v7, v7, v15
+; SI-NEXT:    v_or_b32_e32 v8, v8, v23
+; SI-NEXT:    v_or_b32_e32 v5, v5, v22
+; SI-NEXT:    v_or_b32_e32 v11, v11, v18
 ; SI-NEXT:    v_alignbit_b32 v41, v48, v26, 16
 ; SI-NEXT:    v_alignbit_b32 v40, v37, v27, 16
 ; SI-NEXT:    v_alignbit_b32 v55, v34, v28, 16
 ; SI-NEXT:    v_alignbit_b32 v54, v31, v29, 16
 ; SI-NEXT:    v_alignbit_b32 v53, v19, v30, 16
-; SI-NEXT:    v_alignbit_b32 v52, v16, v52, 16
-; SI-NEXT:    v_alignbit_b32 v25, v11, v25, 16
-; SI-NEXT:    v_alignbit_b32 v24, v8, v24, 16
-; SI-NEXT:    v_alignbit_b32 v23, v3, v23, 16
-; SI-NEXT:    v_alignbit_b32 v22, v5, v22, 16
-; SI-NEXT:    v_alignbit_b32 v15, v1, v15, 16
+; SI-NEXT:    v_alignbit_b32 v52, v15, v52, 16
+; SI-NEXT:    v_alignbit_b32 v25, v12, v25, 16
+; SI-NEXT:    v_alignbit_b32 v24, v6, v24, 16
+; SI-NEXT:    v_alignbit_b32 v23, v1, v23, 16
+; SI-NEXT:    v_alignbit_b32 v22, v9, v22, 16
+; SI-NEXT:    v_alignbit_b32 v18, v3, v18, 16
 ; SI-NEXT:  .LBB58_2: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v51
@@ -36682,85 +36685,81 @@ define <44 x i16> @bitcast_v44f16_to_v44i16(<44 x half> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v52
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v20
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 40, v0
-; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 44, v0
+; SI-NEXT:    v_or_b32_e32 v15, v15, v16
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 44, v0
 ; SI-NEXT:    buffer_store_dword v19, v20, s[0:3], 0 offen
-; SI-NEXT:    buffer_store_dword v16, v17, s[0:3], 0 offen
+; SI-NEXT:    buffer_store_dword v15, v16, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v18
-; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v25
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; SI-NEXT:    v_or_b32_e32 v16, v16, v17
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 48, v0
-; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 52, v0
-; SI-NEXT:    buffer_store_dword v16, v17, s[0:3], 0 offen
-; SI-NEXT:    buffer_store_dword v11, v12, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v25
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
+; SI-NEXT:    v_or_b32_e32 v15, v15, v16
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 48, v0
+; SI-NEXT:    v_or_b32_e32 v12, v12, v13
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 52, v0
+; SI-NEXT:    buffer_store_dword v15, v16, s[0:3], 0 offen
+; SI-NEXT:    buffer_store_dword v12, v13, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v14
-; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v24
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; SI-NEXT:    v_or_b32_e32 v11, v11, v12
-; SI-NEXT:    v_add_i32_e32 v12, vcc, 56, v0
-; SI-NEXT:    v_or_b32_e32 v8, v8, v9
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 60, v0
-; SI-NEXT:    buffer_store_dword v11, v12, s[0:3], 0 offen
-; SI-NEXT:    buffer_store_dword v8, v9, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v24
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; SI-NEXT:    v_or_b32_e32 v12, v12, v13
+; SI-NEXT:    v_add_i32_e32 v13, vcc, 56, v0
+; SI-NEXT:    v_or_b32_e32 v6, v6, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 60, v0
+; SI-NEXT:    buffer_store_dword v12, v13, s[0:3], 0 offen
+; SI-NEXT:    buffer_store_dword v6, v7, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v10
-; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v23
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v8, v8, v9
-; SI-NEXT:    v_add_i32_e32 v9, vcc, 64, v0
-; SI-NEXT:    v_or_b32_e32 v3, v3, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x44, v0
-; SI-NEXT:    buffer_store_dword v8, v9, s[0:3], 0 offen
-; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v23
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_or_b32_e32 v6, v6, v7
+; SI-NEXT:    v_add_i32_e32 v7, vcc, 64, v0
+; SI-NEXT:    v_or_b32_e32 v1, v1, v2
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
+; SI-NEXT:    buffer_store_dword v6, v7, s[0:3], 0 offen
+; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v22
-; SI-NEXT:    v_or_b32_e32 v3, v3, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x48, v0
-; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v22
+; SI-NEXT:    v_or_b32_e32 v1, v1, v2
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
+; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v6
-; SI-NEXT:    v_or_b32_e32 v3, v3, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x4c, v0
-; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v10
+; SI-NEXT:    v_or_b32_e32 v1, v1, v2
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
+; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v15
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v3, v3, v4
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x50, v0
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v18
+; SI-NEXT:    v_or_b32_e32 v1, v1, v2
+; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
+; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x54, v0
-; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
 ; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
 ; SI-NEXT:    s_setpc_b64 s[30:31]
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll
index 7ec521c4ed8f7..a2bd1d30cc634 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll
@@ -4207,21 +4207,27 @@ define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v54, v2
-; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
 ; SI-NEXT:    v_mov_b32_e32 v50, v10
 ; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
+; SI-NEXT:    v_mov_b32_e32 v54, v2
+; SI-NEXT:    v_mov_b32_e32 v55, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:60
 ; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v5
@@ -4230,25 +4236,15 @@ define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(5)
@@ -4276,22 +4272,26 @@ define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:12
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -4301,23 +4301,23 @@ define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr24
 ; SI-NEXT:    ; kill: killed $vgpr24
 ; SI-NEXT:    ; implicit-def: $vgpr24
@@ -4384,37 +4384,37 @@ define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_or_b32_e32 v8, v8, v35
-; SI-NEXT:    v_or_b32_e32 v9, v9, v43
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_or_b32_e32 v8, v8, v43
+; SI-NEXT:    v_or_b32_e32 v9, v9, v35
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v42
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v34
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v41
@@ -4429,8 +4429,8 @@ define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v58
 ; SI-NEXT:    v_or_b32_e32 v22, v22, v57
 ; SI-NEXT:    v_or_b32_e32 v23, v23, v56
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr41
@@ -4449,23 +4449,23 @@ define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -4501,24 +4501,23 @@ define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
@@ -4540,8 +4539,8 @@ define <24 x i32> @bitcast_v48i16_to_v24i32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    v_or_b32_e32 v8, v35, v8
-; SI-NEXT:    v_or_b32_e32 v9, v43, v9
+; SI-NEXT:    v_or_b32_e32 v8, v43, v8
+; SI-NEXT:    v_or_b32_e32 v9, v35, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v42, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v34, v11
 ; SI-NEXT:    v_or_b32_e32 v12, v41, v12
@@ -8770,15 +8769,15 @@ define <24 x i32> @bitcast_v48f16_to_v24i32(<48 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
 ; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
@@ -8805,7 +8804,7 @@ define <24 x i32> @bitcast_v48f16_to_v24i32(<48 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
@@ -9733,8 +9732,8 @@ define inreg <24 x i32> @bitcast_v48f16_to_v24i32_scalar(<48 x half> inreg %a, i
 ; SI-NEXT:    s_cbranch_execnz .LBB19_3
 ; SI-NEXT:  .LBB19_2: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v54
@@ -9801,14 +9800,14 @@ define inreg <24 x i32> @bitcast_v48f16_to_v24i32_scalar(<48 x half> inreg %a, i
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
-; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
 ; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
@@ -13996,21 +13995,27 @@ define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v54, v2
-; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
 ; SI-NEXT:    v_mov_b32_e32 v50, v10
 ; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
+; SI-NEXT:    v_mov_b32_e32 v54, v2
+; SI-NEXT:    v_mov_b32_e32 v55, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:60
 ; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v5
@@ -14019,25 +14024,15 @@ define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(5)
@@ -14065,22 +14060,26 @@ define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:12
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -14090,23 +14089,23 @@ define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr24
 ; SI-NEXT:    ; kill: killed $vgpr24
 ; SI-NEXT:    ; implicit-def: $vgpr24
@@ -14173,37 +14172,37 @@ define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_or_b32_e32 v8, v8, v35
-; SI-NEXT:    v_or_b32_e32 v9, v9, v43
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_or_b32_e32 v8, v8, v43
+; SI-NEXT:    v_or_b32_e32 v9, v9, v35
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v42
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v34
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v41
@@ -14218,8 +14217,8 @@ define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v58
 ; SI-NEXT:    v_or_b32_e32 v22, v22, v57
 ; SI-NEXT:    v_or_b32_e32 v23, v23, v56
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr41
@@ -14238,23 +14237,23 @@ define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -14290,24 +14289,23 @@ define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
@@ -14329,8 +14327,8 @@ define <24 x float> @bitcast_v48i16_to_v24f32(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    v_or_b32_e32 v8, v35, v8
-; SI-NEXT:    v_or_b32_e32 v9, v43, v9
+; SI-NEXT:    v_or_b32_e32 v8, v43, v8
+; SI-NEXT:    v_or_b32_e32 v9, v35, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v42, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v34, v11
 ; SI-NEXT:    v_or_b32_e32 v12, v41, v12
@@ -18555,15 +18553,15 @@ define <24 x float> @bitcast_v48f16_to_v24f32(<48 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
 ; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
@@ -18590,7 +18588,7 @@ define <24 x float> @bitcast_v48f16_to_v24f32(<48 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
@@ -19518,8 +19516,8 @@ define inreg <24 x float> @bitcast_v48f16_to_v24f32_scalar(<48 x half> inreg %a,
 ; SI-NEXT:    s_cbranch_execnz .LBB35_3
 ; SI-NEXT:  .LBB35_2: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v54
@@ -19586,14 +19584,14 @@ define inreg <24 x float> @bitcast_v48f16_to_v24f32_scalar(<48 x half> inreg %a,
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
-; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
 ; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
@@ -23044,21 +23042,27 @@ define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v54, v2
-; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
 ; SI-NEXT:    v_mov_b32_e32 v50, v10
 ; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
+; SI-NEXT:    v_mov_b32_e32 v54, v2
+; SI-NEXT:    v_mov_b32_e32 v55, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:60
 ; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v5
@@ -23067,25 +23071,15 @@ define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(5)
@@ -23113,22 +23107,26 @@ define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:12
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -23138,23 +23136,23 @@ define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr24
 ; SI-NEXT:    ; kill: killed $vgpr24
 ; SI-NEXT:    ; implicit-def: $vgpr24
@@ -23221,37 +23219,37 @@ define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_or_b32_e32 v8, v8, v35
-; SI-NEXT:    v_or_b32_e32 v9, v9, v43
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_or_b32_e32 v8, v8, v43
+; SI-NEXT:    v_or_b32_e32 v9, v9, v35
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v42
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v34
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v41
@@ -23266,8 +23264,8 @@ define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v58
 ; SI-NEXT:    v_or_b32_e32 v22, v22, v57
 ; SI-NEXT:    v_or_b32_e32 v23, v23, v56
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr41
@@ -23286,23 +23284,23 @@ define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -23338,24 +23336,23 @@ define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
@@ -23377,8 +23374,8 @@ define <12 x i64> @bitcast_v48i16_to_v12i64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    v_or_b32_e32 v8, v35, v8
-; SI-NEXT:    v_or_b32_e32 v9, v43, v9
+; SI-NEXT:    v_or_b32_e32 v8, v43, v8
+; SI-NEXT:    v_or_b32_e32 v9, v35, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v42, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v34, v11
 ; SI-NEXT:    v_or_b32_e32 v12, v41, v12
@@ -27619,15 +27616,15 @@ define <12 x i64> @bitcast_v48f16_to_v12i64(<48 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
 ; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
@@ -27654,7 +27651,7 @@ define <12 x i64> @bitcast_v48f16_to_v12i64(<48 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
@@ -28582,8 +28579,8 @@ define inreg <12 x i64> @bitcast_v48f16_to_v12i64_scalar(<48 x half> inreg %a, i
 ; SI-NEXT:    s_cbranch_execnz .LBB47_3
 ; SI-NEXT:  .LBB47_2: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v54
@@ -28650,14 +28647,14 @@ define inreg <12 x i64> @bitcast_v48f16_to_v12i64_scalar(<48 x half> inreg %a, i
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
-; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
 ; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
@@ -31213,21 +31210,27 @@ define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    v_mov_b32_e32 v54, v2
-; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v48, v14
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
 ; SI-NEXT:    v_mov_b32_e32 v50, v10
 ; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_mov_b32_e32 v52, v6
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
+; SI-NEXT:    v_mov_b32_e32 v54, v2
+; SI-NEXT:    v_mov_b32_e32 v55, v0
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:60
 ; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v5
@@ -31236,25 +31239,15 @@ define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
 ; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v23
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v25
 ; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(5)
@@ -31282,22 +31275,26 @@ define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:12
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -31307,23 +31304,23 @@ define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr24
 ; SI-NEXT:    ; kill: killed $vgpr24
 ; SI-NEXT:    ; implicit-def: $vgpr24
@@ -31390,37 +31387,37 @@ define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
 ; SI-NEXT:    s_waitcnt vmcnt(8)
 ; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_or_b32_e32 v8, v8, v35
-; SI-NEXT:    v_or_b32_e32 v9, v9, v43
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_or_b32_e32 v8, v8, v43
+; SI-NEXT:    v_or_b32_e32 v9, v9, v35
 ; SI-NEXT:    v_or_b32_e32 v10, v10, v42
 ; SI-NEXT:    v_or_b32_e32 v11, v11, v34
 ; SI-NEXT:    v_or_b32_e32 v12, v12, v41
@@ -31435,8 +31432,8 @@ define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v21, v21, v58
 ; SI-NEXT:    v_or_b32_e32 v22, v22, v57
 ; SI-NEXT:    v_or_b32_e32 v23, v23, v56
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr41
@@ -31455,23 +31452,23 @@ define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -31507,24 +31504,23 @@ define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
@@ -31546,8 +31542,8 @@ define <12 x double> @bitcast_v48i16_to_v12f64(<48 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    v_or_b32_e32 v8, v35, v8
-; SI-NEXT:    v_or_b32_e32 v9, v43, v9
+; SI-NEXT:    v_or_b32_e32 v8, v43, v8
+; SI-NEXT:    v_or_b32_e32 v9, v35, v9
 ; SI-NEXT:    v_or_b32_e32 v10, v42, v10
 ; SI-NEXT:    v_or_b32_e32 v11, v34, v11
 ; SI-NEXT:    v_or_b32_e32 v12, v41, v12
@@ -35690,15 +35686,15 @@ define <12 x double> @bitcast_v48f16_to_v12f64(<48 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    v_or_b32_e32 v7, v8, v7
 ; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
 ; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
@@ -35725,7 +35721,7 @@ define <12 x double> @bitcast_v48f16_to_v12f64(<48 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
 ; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
@@ -36653,8 +36649,8 @@ define inreg <12 x double> @bitcast_v48f16_to_v12f64_scalar(<48 x half> inreg %a
 ; SI-NEXT:    s_cbranch_execnz .LBB55_3
 ; SI-NEXT:  .LBB55_2: ; %cmp.true
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v54
@@ -36721,14 +36717,14 @@ define inreg <12 x double> @bitcast_v48f16_to_v12f64_scalar(<48 x half> inreg %a
 ; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
-; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
 ; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll
index 028e61a1ef687..fbee320c82c7f 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll
@@ -4507,73 +4507,69 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    v_mov_b32_e32 v50, v10
+; SI-NEXT:    v_mov_b32_e32 v51, v8
+; SI-NEXT:    v_mov_b32_e32 v52, v6
+; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:44
-; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_mov_b32_e32 v52, v6
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:12
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
-; SI-NEXT:    v_mov_b32_e32 v50, v10
-; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v29
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v2
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:84
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v8
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v14
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v14
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -4584,60 +4580,64 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v55
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v54
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v53
@@ -4647,11 +4647,11 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v49
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v58
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v48
-; SI-NEXT:    v_or_b32_e32 v2, v2, v57
-; SI-NEXT:    v_or_b32_e32 v3, v3, v39
-; SI-NEXT:    v_or_b32_e32 v4, v4, v56
-; SI-NEXT:    v_or_b32_e32 v5, v5, v47
-; SI-NEXT:    v_or_b32_e32 v6, v6, v38
+; SI-NEXT:    v_or_b32_e32 v2, v2, v39
+; SI-NEXT:    v_or_b32_e32 v3, v3, v57
+; SI-NEXT:    v_or_b32_e32 v4, v4, v38
+; SI-NEXT:    v_or_b32_e32 v5, v5, v56
+; SI-NEXT:    v_or_b32_e32 v6, v6, v47
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr53
@@ -4661,27 +4661,25 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    ; kill: killed $vgpr26
@@ -4718,34 +4716,37 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr26
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v37
-; SI-NEXT:    v_or_b32_e32 v8, v8, v36
-; SI-NEXT:    v_or_b32_e32 v9, v9, v46
-; SI-NEXT:    v_or_b32_e32 v10, v10, v35
-; SI-NEXT:    v_or_b32_e32 v11, v11, v45
-; SI-NEXT:    v_or_b32_e32 v12, v12, v44
-; SI-NEXT:    v_or_b32_e32 v13, v13, v34
-; SI-NEXT:    v_or_b32_e32 v14, v14, v43
-; SI-NEXT:    v_or_b32_e32 v15, v15, v42
-; SI-NEXT:    v_or_b32_e32 v16, v16, v33
-; SI-NEXT:    v_or_b32_e32 v17, v17, v41
+; SI-NEXT:    v_or_b32_e32 v8, v8, v46
+; SI-NEXT:    v_or_b32_e32 v9, v9, v45
+; SI-NEXT:    v_or_b32_e32 v10, v10, v36
+; SI-NEXT:    v_or_b32_e32 v11, v11, v44
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v35
+; SI-NEXT:    v_or_b32_e32 v14, v14, v42
+; SI-NEXT:    v_or_b32_e32 v15, v15, v34
+; SI-NEXT:    v_or_b32_e32 v16, v16, v41
+; SI-NEXT:    v_or_b32_e32 v17, v17, v33
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v40
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v32
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v63
@@ -4755,16 +4756,16 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v59
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
@@ -4778,27 +4779,27 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -4816,11 +4817,11 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v58, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
 ; SI-NEXT:    v_or_b32_e32 v1, v48, v1
-; SI-NEXT:    v_or_b32_e32 v2, v57, v2
-; SI-NEXT:    v_or_b32_e32 v3, v39, v3
-; SI-NEXT:    v_or_b32_e32 v4, v56, v4
-; SI-NEXT:    v_or_b32_e32 v5, v47, v5
-; SI-NEXT:    v_or_b32_e32 v6, v38, v6
+; SI-NEXT:    v_or_b32_e32 v2, v39, v2
+; SI-NEXT:    v_or_b32_e32 v3, v57, v3
+; SI-NEXT:    v_or_b32_e32 v4, v38, v4
+; SI-NEXT:    v_or_b32_e32 v5, v56, v5
+; SI-NEXT:    v_or_b32_e32 v6, v47, v6
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
@@ -4830,30 +4831,33 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
@@ -4875,16 +4879,16 @@ define <26 x i32> @bitcast_v52i16_to_v26i32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
-; SI-NEXT:    v_or_b32_e32 v8, v36, v8
-; SI-NEXT:    v_or_b32_e32 v9, v46, v9
-; SI-NEXT:    v_or_b32_e32 v10, v35, v10
-; SI-NEXT:    v_or_b32_e32 v11, v45, v11
-; SI-NEXT:    v_or_b32_e32 v12, v44, v12
-; SI-NEXT:    v_or_b32_e32 v13, v34, v13
-; SI-NEXT:    v_or_b32_e32 v14, v43, v14
-; SI-NEXT:    v_or_b32_e32 v15, v42, v15
-; SI-NEXT:    v_or_b32_e32 v16, v33, v16
-; SI-NEXT:    v_or_b32_e32 v17, v41, v17
+; SI-NEXT:    v_or_b32_e32 v8, v46, v8
+; SI-NEXT:    v_or_b32_e32 v9, v45, v9
+; SI-NEXT:    v_or_b32_e32 v10, v36, v10
+; SI-NEXT:    v_or_b32_e32 v11, v44, v11
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v35, v13
+; SI-NEXT:    v_or_b32_e32 v14, v42, v14
+; SI-NEXT:    v_or_b32_e32 v15, v34, v15
+; SI-NEXT:    v_or_b32_e32 v16, v41, v16
+; SI-NEXT:    v_or_b32_e32 v17, v33, v17
 ; SI-NEXT:    v_or_b32_e32 v18, v40, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v32, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v63, v20
@@ -9474,50 +9478,50 @@ define <26 x i32> @bitcast_v52f16_to_v26i32(<52 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v58
-; SI-NEXT:    v_cvt_f32_f16_e32 v23, v47
-; SI-NEXT:    v_cvt_f32_f16_e32 v24, v46
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v4, v3
 ; SI-NEXT:    v_cvt_f32_f16_e32 v4, v51
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v4, v5, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v49
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v58
+; SI-NEXT:    v_cvt_f32_f16_e32 v23, v47
+; SI-NEXT:    v_cvt_f32_f16_e32 v24, v46
+; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
-; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v44
 ; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v5, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v49
 ; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
+; SI-NEXT:    v_cvt_f32_f16_e32 v26, v33
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    v_cvt_f32_f16_e32 v26, v33
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v32
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
 ; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
 ; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
@@ -9525,50 +9529,50 @@ define <26 x i32> @bitcast_v52f16_to_v26i32(<52 x half> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v5, v6, v5
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
+; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
 ; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
 ; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
+; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
-; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
@@ -10544,7 +10548,9 @@ define inreg <26 x i32> @bitcast_v52f16_to_v26i32_scalar(<52 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v53
@@ -10590,8 +10596,6 @@ define inreg <26 x i32> @bitcast_v52f16_to_v26i32_scalar(<52 x half> inreg %a, i
 ; SI-NEXT:    v_or_b32_e32 v18, v33, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v59, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v27, v20
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
 ; SI-NEXT:    s_waitcnt vmcnt(4)
@@ -15201,73 +15205,69 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    v_mov_b32_e32 v50, v10
+; SI-NEXT:    v_mov_b32_e32 v51, v8
+; SI-NEXT:    v_mov_b32_e32 v52, v6
+; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:44
-; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_mov_b32_e32 v52, v6
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:12
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
-; SI-NEXT:    v_mov_b32_e32 v50, v10
-; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v29
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v2
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:84
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v8
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v14
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v14
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -15278,60 +15278,64 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v55
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v54
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v53
@@ -15341,11 +15345,11 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v49
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v58
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v48
-; SI-NEXT:    v_or_b32_e32 v2, v2, v57
-; SI-NEXT:    v_or_b32_e32 v3, v3, v39
-; SI-NEXT:    v_or_b32_e32 v4, v4, v56
-; SI-NEXT:    v_or_b32_e32 v5, v5, v47
-; SI-NEXT:    v_or_b32_e32 v6, v6, v38
+; SI-NEXT:    v_or_b32_e32 v2, v2, v39
+; SI-NEXT:    v_or_b32_e32 v3, v3, v57
+; SI-NEXT:    v_or_b32_e32 v4, v4, v38
+; SI-NEXT:    v_or_b32_e32 v5, v5, v56
+; SI-NEXT:    v_or_b32_e32 v6, v6, v47
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr53
@@ -15355,27 +15359,25 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    ; kill: killed $vgpr26
@@ -15412,34 +15414,37 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr26
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v37
-; SI-NEXT:    v_or_b32_e32 v8, v8, v36
-; SI-NEXT:    v_or_b32_e32 v9, v9, v46
-; SI-NEXT:    v_or_b32_e32 v10, v10, v35
-; SI-NEXT:    v_or_b32_e32 v11, v11, v45
-; SI-NEXT:    v_or_b32_e32 v12, v12, v44
-; SI-NEXT:    v_or_b32_e32 v13, v13, v34
-; SI-NEXT:    v_or_b32_e32 v14, v14, v43
-; SI-NEXT:    v_or_b32_e32 v15, v15, v42
-; SI-NEXT:    v_or_b32_e32 v16, v16, v33
-; SI-NEXT:    v_or_b32_e32 v17, v17, v41
+; SI-NEXT:    v_or_b32_e32 v8, v8, v46
+; SI-NEXT:    v_or_b32_e32 v9, v9, v45
+; SI-NEXT:    v_or_b32_e32 v10, v10, v36
+; SI-NEXT:    v_or_b32_e32 v11, v11, v44
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v35
+; SI-NEXT:    v_or_b32_e32 v14, v14, v42
+; SI-NEXT:    v_or_b32_e32 v15, v15, v34
+; SI-NEXT:    v_or_b32_e32 v16, v16, v41
+; SI-NEXT:    v_or_b32_e32 v17, v17, v33
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v40
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v32
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v63
@@ -15449,16 +15454,16 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v59
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
@@ -15472,27 +15477,27 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -15510,11 +15515,11 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v58, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
 ; SI-NEXT:    v_or_b32_e32 v1, v48, v1
-; SI-NEXT:    v_or_b32_e32 v2, v57, v2
-; SI-NEXT:    v_or_b32_e32 v3, v39, v3
-; SI-NEXT:    v_or_b32_e32 v4, v56, v4
-; SI-NEXT:    v_or_b32_e32 v5, v47, v5
-; SI-NEXT:    v_or_b32_e32 v6, v38, v6
+; SI-NEXT:    v_or_b32_e32 v2, v39, v2
+; SI-NEXT:    v_or_b32_e32 v3, v57, v3
+; SI-NEXT:    v_or_b32_e32 v4, v38, v4
+; SI-NEXT:    v_or_b32_e32 v5, v56, v5
+; SI-NEXT:    v_or_b32_e32 v6, v47, v6
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
@@ -15524,30 +15529,33 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
@@ -15569,16 +15577,16 @@ define <26 x float> @bitcast_v52i16_to_v26f32(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
-; SI-NEXT:    v_or_b32_e32 v8, v36, v8
-; SI-NEXT:    v_or_b32_e32 v9, v46, v9
-; SI-NEXT:    v_or_b32_e32 v10, v35, v10
-; SI-NEXT:    v_or_b32_e32 v11, v45, v11
-; SI-NEXT:    v_or_b32_e32 v12, v44, v12
-; SI-NEXT:    v_or_b32_e32 v13, v34, v13
-; SI-NEXT:    v_or_b32_e32 v14, v43, v14
-; SI-NEXT:    v_or_b32_e32 v15, v42, v15
-; SI-NEXT:    v_or_b32_e32 v16, v33, v16
-; SI-NEXT:    v_or_b32_e32 v17, v41, v17
+; SI-NEXT:    v_or_b32_e32 v8, v46, v8
+; SI-NEXT:    v_or_b32_e32 v9, v45, v9
+; SI-NEXT:    v_or_b32_e32 v10, v36, v10
+; SI-NEXT:    v_or_b32_e32 v11, v44, v11
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v35, v13
+; SI-NEXT:    v_or_b32_e32 v14, v42, v14
+; SI-NEXT:    v_or_b32_e32 v15, v34, v15
+; SI-NEXT:    v_or_b32_e32 v16, v41, v16
+; SI-NEXT:    v_or_b32_e32 v17, v33, v17
 ; SI-NEXT:    v_or_b32_e32 v18, v40, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v32, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v63, v20
@@ -20169,50 +20177,50 @@ define <26 x float> @bitcast_v52f16_to_v26f32(<52 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v58
-; SI-NEXT:    v_cvt_f32_f16_e32 v23, v47
-; SI-NEXT:    v_cvt_f32_f16_e32 v24, v46
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v4, v3
 ; SI-NEXT:    v_cvt_f32_f16_e32 v4, v51
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v4, v5, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v49
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v58
+; SI-NEXT:    v_cvt_f32_f16_e32 v23, v47
+; SI-NEXT:    v_cvt_f32_f16_e32 v24, v46
+; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
-; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v44
 ; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v5, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v49
 ; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
+; SI-NEXT:    v_cvt_f32_f16_e32 v26, v33
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    v_cvt_f32_f16_e32 v26, v33
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v32
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
 ; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
 ; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
@@ -20220,50 +20228,50 @@ define <26 x float> @bitcast_v52f16_to_v26f32(<52 x half> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v5, v6, v5
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
+; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
 ; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
 ; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
+; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
-; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
@@ -21239,7 +21247,9 @@ define inreg <26 x float> @bitcast_v52f16_to_v26f32_scalar(<52 x half> inreg %a,
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v53
@@ -21285,8 +21295,6 @@ define inreg <26 x float> @bitcast_v52f16_to_v26f32_scalar(<52 x half> inreg %a,
 ; SI-NEXT:    v_or_b32_e32 v18, v33, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v59, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v27, v20
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
 ; SI-NEXT:    s_waitcnt vmcnt(4)
@@ -25096,73 +25104,69 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    v_mov_b32_e32 v50, v10
+; SI-NEXT:    v_mov_b32_e32 v51, v8
+; SI-NEXT:    v_mov_b32_e32 v52, v6
+; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:44
-; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_mov_b32_e32 v52, v6
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:12
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
-; SI-NEXT:    v_mov_b32_e32 v50, v10
-; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v29
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v2
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:84
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v8
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v14
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v14
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -25173,60 +25177,64 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v55
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v54
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v53
@@ -25236,11 +25244,11 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v49
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v58
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v48
-; SI-NEXT:    v_or_b32_e32 v2, v2, v57
-; SI-NEXT:    v_or_b32_e32 v3, v3, v39
-; SI-NEXT:    v_or_b32_e32 v4, v4, v56
-; SI-NEXT:    v_or_b32_e32 v5, v5, v47
-; SI-NEXT:    v_or_b32_e32 v6, v6, v38
+; SI-NEXT:    v_or_b32_e32 v2, v2, v39
+; SI-NEXT:    v_or_b32_e32 v3, v3, v57
+; SI-NEXT:    v_or_b32_e32 v4, v4, v38
+; SI-NEXT:    v_or_b32_e32 v5, v5, v56
+; SI-NEXT:    v_or_b32_e32 v6, v6, v47
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr53
@@ -25250,27 +25258,25 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    ; kill: killed $vgpr26
@@ -25307,34 +25313,37 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr26
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v37
-; SI-NEXT:    v_or_b32_e32 v8, v8, v36
-; SI-NEXT:    v_or_b32_e32 v9, v9, v46
-; SI-NEXT:    v_or_b32_e32 v10, v10, v35
-; SI-NEXT:    v_or_b32_e32 v11, v11, v45
-; SI-NEXT:    v_or_b32_e32 v12, v12, v44
-; SI-NEXT:    v_or_b32_e32 v13, v13, v34
-; SI-NEXT:    v_or_b32_e32 v14, v14, v43
-; SI-NEXT:    v_or_b32_e32 v15, v15, v42
-; SI-NEXT:    v_or_b32_e32 v16, v16, v33
-; SI-NEXT:    v_or_b32_e32 v17, v17, v41
+; SI-NEXT:    v_or_b32_e32 v8, v8, v46
+; SI-NEXT:    v_or_b32_e32 v9, v9, v45
+; SI-NEXT:    v_or_b32_e32 v10, v10, v36
+; SI-NEXT:    v_or_b32_e32 v11, v11, v44
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v35
+; SI-NEXT:    v_or_b32_e32 v14, v14, v42
+; SI-NEXT:    v_or_b32_e32 v15, v15, v34
+; SI-NEXT:    v_or_b32_e32 v16, v16, v41
+; SI-NEXT:    v_or_b32_e32 v17, v17, v33
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v40
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v32
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v63
@@ -25344,16 +25353,16 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v59
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
@@ -25367,27 +25376,27 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -25405,11 +25414,11 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v58, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
 ; SI-NEXT:    v_or_b32_e32 v1, v48, v1
-; SI-NEXT:    v_or_b32_e32 v2, v57, v2
-; SI-NEXT:    v_or_b32_e32 v3, v39, v3
-; SI-NEXT:    v_or_b32_e32 v4, v56, v4
-; SI-NEXT:    v_or_b32_e32 v5, v47, v5
-; SI-NEXT:    v_or_b32_e32 v6, v38, v6
+; SI-NEXT:    v_or_b32_e32 v2, v39, v2
+; SI-NEXT:    v_or_b32_e32 v3, v57, v3
+; SI-NEXT:    v_or_b32_e32 v4, v38, v4
+; SI-NEXT:    v_or_b32_e32 v5, v56, v5
+; SI-NEXT:    v_or_b32_e32 v6, v47, v6
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
@@ -25419,30 +25428,33 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
@@ -25464,16 +25476,16 @@ define <13 x i64> @bitcast_v52i16_to_v13i64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
-; SI-NEXT:    v_or_b32_e32 v8, v36, v8
-; SI-NEXT:    v_or_b32_e32 v9, v46, v9
-; SI-NEXT:    v_or_b32_e32 v10, v35, v10
-; SI-NEXT:    v_or_b32_e32 v11, v45, v11
-; SI-NEXT:    v_or_b32_e32 v12, v44, v12
-; SI-NEXT:    v_or_b32_e32 v13, v34, v13
-; SI-NEXT:    v_or_b32_e32 v14, v43, v14
-; SI-NEXT:    v_or_b32_e32 v15, v42, v15
-; SI-NEXT:    v_or_b32_e32 v16, v33, v16
-; SI-NEXT:    v_or_b32_e32 v17, v41, v17
+; SI-NEXT:    v_or_b32_e32 v8, v46, v8
+; SI-NEXT:    v_or_b32_e32 v9, v45, v9
+; SI-NEXT:    v_or_b32_e32 v10, v36, v10
+; SI-NEXT:    v_or_b32_e32 v11, v44, v11
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v35, v13
+; SI-NEXT:    v_or_b32_e32 v14, v42, v14
+; SI-NEXT:    v_or_b32_e32 v15, v34, v15
+; SI-NEXT:    v_or_b32_e32 v16, v41, v16
+; SI-NEXT:    v_or_b32_e32 v17, v33, v17
 ; SI-NEXT:    v_or_b32_e32 v18, v40, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v32, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v63, v20
@@ -30078,50 +30090,50 @@ define <13 x i64> @bitcast_v52f16_to_v13i64(<52 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v58
-; SI-NEXT:    v_cvt_f32_f16_e32 v23, v47
-; SI-NEXT:    v_cvt_f32_f16_e32 v24, v46
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v4, v3
 ; SI-NEXT:    v_cvt_f32_f16_e32 v4, v51
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v4, v5, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v49
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v58
+; SI-NEXT:    v_cvt_f32_f16_e32 v23, v47
+; SI-NEXT:    v_cvt_f32_f16_e32 v24, v46
+; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
-; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v44
 ; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v5, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v49
 ; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
+; SI-NEXT:    v_cvt_f32_f16_e32 v26, v33
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    v_cvt_f32_f16_e32 v26, v33
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v32
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
 ; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
 ; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
@@ -30129,50 +30141,50 @@ define <13 x i64> @bitcast_v52f16_to_v13i64(<52 x half> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v5, v6, v5
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
+; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
 ; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
 ; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
+; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
-; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
@@ -31148,7 +31160,9 @@ define inreg <13 x i64> @bitcast_v52f16_to_v13i64_scalar(<52 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v53
@@ -31194,8 +31208,6 @@ define inreg <13 x i64> @bitcast_v52f16_to_v13i64_scalar(<52 x half> inreg %a, i
 ; SI-NEXT:    v_or_b32_e32 v18, v33, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v59, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v27, v20
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
 ; SI-NEXT:    s_waitcnt vmcnt(4)
@@ -34085,73 +34097,69 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    v_mov_b32_e32 v50, v10
+; SI-NEXT:    v_mov_b32_e32 v51, v8
+; SI-NEXT:    v_mov_b32_e32 v52, v6
+; SI-NEXT:    v_mov_b32_e32 v53, v4
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:44
-; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_mov_b32_e32 v52, v6
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:12
 ; SI-NEXT:    v_mov_b32_e32 v49, v12
-; SI-NEXT:    v_mov_b32_e32 v50, v10
-; SI-NEXT:    v_mov_b32_e32 v51, v8
 ; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v13
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v29
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v29
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v2
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:84
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v8
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v14
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v14
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -34162,60 +34170,64 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:4
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v55
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v54
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v53
@@ -34225,11 +34237,11 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v49
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v58
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v48
-; SI-NEXT:    v_or_b32_e32 v2, v2, v57
-; SI-NEXT:    v_or_b32_e32 v3, v3, v39
-; SI-NEXT:    v_or_b32_e32 v4, v4, v56
-; SI-NEXT:    v_or_b32_e32 v5, v5, v47
-; SI-NEXT:    v_or_b32_e32 v6, v6, v38
+; SI-NEXT:    v_or_b32_e32 v2, v2, v39
+; SI-NEXT:    v_or_b32_e32 v3, v3, v57
+; SI-NEXT:    v_or_b32_e32 v4, v4, v38
+; SI-NEXT:    v_or_b32_e32 v5, v5, v56
+; SI-NEXT:    v_or_b32_e32 v6, v6, v47
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
 ; SI-NEXT:    ; implicit-def: $vgpr53
@@ -34239,27 +34251,25 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr57
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr47
-; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    s_waitcnt vmcnt(13)
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(12)
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_or_b32_e32 v25, v25, v26
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    ; kill: killed $vgpr26
@@ -34296,34 +34306,37 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr26
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
 ; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr26
 ; SI-NEXT:    v_or_b32_e32 v7, v7, v37
-; SI-NEXT:    v_or_b32_e32 v8, v8, v36
-; SI-NEXT:    v_or_b32_e32 v9, v9, v46
-; SI-NEXT:    v_or_b32_e32 v10, v10, v35
-; SI-NEXT:    v_or_b32_e32 v11, v11, v45
-; SI-NEXT:    v_or_b32_e32 v12, v12, v44
-; SI-NEXT:    v_or_b32_e32 v13, v13, v34
-; SI-NEXT:    v_or_b32_e32 v14, v14, v43
-; SI-NEXT:    v_or_b32_e32 v15, v15, v42
-; SI-NEXT:    v_or_b32_e32 v16, v16, v33
-; SI-NEXT:    v_or_b32_e32 v17, v17, v41
+; SI-NEXT:    v_or_b32_e32 v8, v8, v46
+; SI-NEXT:    v_or_b32_e32 v9, v9, v45
+; SI-NEXT:    v_or_b32_e32 v10, v10, v36
+; SI-NEXT:    v_or_b32_e32 v11, v11, v44
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v35
+; SI-NEXT:    v_or_b32_e32 v14, v14, v42
+; SI-NEXT:    v_or_b32_e32 v15, v15, v34
+; SI-NEXT:    v_or_b32_e32 v16, v16, v41
+; SI-NEXT:    v_or_b32_e32 v17, v17, v33
 ; SI-NEXT:    v_or_b32_e32 v18, v18, v40
 ; SI-NEXT:    v_or_b32_e32 v19, v19, v32
 ; SI-NEXT:    v_or_b32_e32 v20, v20, v63
@@ -34333,16 +34346,16 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v24, v24, v59
 ; SI-NEXT:    ; kill: killed $vgpr26
 ; SI-NEXT:    ; implicit-def: $vgpr37
-; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr35
 ; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr34
 ; SI-NEXT:    ; implicit-def: $vgpr41
+; SI-NEXT:    ; implicit-def: $vgpr33
 ; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
@@ -34356,27 +34369,27 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -34394,11 +34407,11 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v0, v58, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
 ; SI-NEXT:    v_or_b32_e32 v1, v48, v1
-; SI-NEXT:    v_or_b32_e32 v2, v57, v2
-; SI-NEXT:    v_or_b32_e32 v3, v39, v3
-; SI-NEXT:    v_or_b32_e32 v4, v56, v4
-; SI-NEXT:    v_or_b32_e32 v5, v47, v5
-; SI-NEXT:    v_or_b32_e32 v6, v38, v6
+; SI-NEXT:    v_or_b32_e32 v2, v39, v2
+; SI-NEXT:    v_or_b32_e32 v3, v57, v3
+; SI-NEXT:    v_or_b32_e32 v4, v38, v4
+; SI-NEXT:    v_or_b32_e32 v5, v56, v5
+; SI-NEXT:    v_or_b32_e32 v6, v47, v6
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
@@ -34408,30 +34421,33 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
@@ -34453,16 +34469,16 @@ define <13 x double> @bitcast_v52i16_to_v13f64(<52 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
-; SI-NEXT:    v_or_b32_e32 v8, v36, v8
-; SI-NEXT:    v_or_b32_e32 v9, v46, v9
-; SI-NEXT:    v_or_b32_e32 v10, v35, v10
-; SI-NEXT:    v_or_b32_e32 v11, v45, v11
-; SI-NEXT:    v_or_b32_e32 v12, v44, v12
-; SI-NEXT:    v_or_b32_e32 v13, v34, v13
-; SI-NEXT:    v_or_b32_e32 v14, v43, v14
-; SI-NEXT:    v_or_b32_e32 v15, v42, v15
-; SI-NEXT:    v_or_b32_e32 v16, v33, v16
-; SI-NEXT:    v_or_b32_e32 v17, v41, v17
+; SI-NEXT:    v_or_b32_e32 v8, v46, v8
+; SI-NEXT:    v_or_b32_e32 v9, v45, v9
+; SI-NEXT:    v_or_b32_e32 v10, v36, v10
+; SI-NEXT:    v_or_b32_e32 v11, v44, v11
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v35, v13
+; SI-NEXT:    v_or_b32_e32 v14, v42, v14
+; SI-NEXT:    v_or_b32_e32 v15, v34, v15
+; SI-NEXT:    v_or_b32_e32 v16, v41, v16
+; SI-NEXT:    v_or_b32_e32 v17, v33, v17
 ; SI-NEXT:    v_or_b32_e32 v18, v40, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v32, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v63, v20
@@ -38962,50 +38978,50 @@ define <13 x double> @bitcast_v52f16_to_v13f64(<52 x half> %a, i32 %b) {
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
 ; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v58
-; SI-NEXT:    v_cvt_f32_f16_e32 v23, v47
-; SI-NEXT:    v_cvt_f32_f16_e32 v24, v46
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v3, v4, v3
 ; SI-NEXT:    v_cvt_f32_f16_e32 v4, v51
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; SI-NEXT:    v_or_b32_e32 v4, v5, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v49
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v58
+; SI-NEXT:    v_cvt_f32_f16_e32 v23, v47
+; SI-NEXT:    v_cvt_f32_f16_e32 v24, v46
+; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
-; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v44
 ; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
-; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT:    v_or_b32_e32 v4, v5, v4
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v49
 ; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
+; SI-NEXT:    v_cvt_f32_f16_e32 v26, v33
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    v_cvt_f32_f16_e32 v26, v33
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v32
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
 ; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
 ; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
@@ -39013,50 +39029,50 @@ define <13 x double> @bitcast_v52f16_to_v13f64(<52 x half> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v5, v6, v5
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
-; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
+; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
 ; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
 ; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
 ; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
 ; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
 ; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
 ; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
+; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
 ; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
-; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
@@ -40032,7 +40048,9 @@ define inreg <13 x double> @bitcast_v52f16_to_v13f64_scalar(<52 x half> inreg %a
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v53
@@ -40078,8 +40096,6 @@ define inreg <13 x double> @bitcast_v52f16_to_v13f64_scalar(<52 x half> inreg %a
 ; SI-NEXT:    v_or_b32_e32 v18, v33, v18
 ; SI-NEXT:    v_or_b32_e32 v19, v59, v19
 ; SI-NEXT:    v_or_b32_e32 v20, v27, v20
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
 ; SI-NEXT:    s_waitcnt vmcnt(4)
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll
index f3ef202a22f31..462e50ac8412c 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll
@@ -4831,152 +4831,152 @@ define <28 x i32> @bitcast_v56i16_to_v28i32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v29
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:84
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:68
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v8
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:20
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:52
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v20
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:84
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v22
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v24
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:12
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v26
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
@@ -5026,9 +5026,9 @@ define <28 x i32> @bitcast_v56i16_to_v28i32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v59
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    v_or_b32_e32 v0, v0, v52
-; SI-NEXT:    v_or_b32_e32 v1, v1, v58
-; SI-NEXT:    v_or_b32_e32 v2, v2, v51
+; SI-NEXT:    v_or_b32_e32 v0, v0, v58
+; SI-NEXT:    v_or_b32_e32 v1, v1, v52
+; SI-NEXT:    v_or_b32_e32 v2, v2, v57
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v49
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
@@ -5037,125 +5037,125 @@ define <28 x i32> @bitcast_v56i16_to_v28i32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; kill: killed $vgpr28
-; SI-NEXT:    ; implicit-def: $vgpr52
 ; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; SI-NEXT:    v_or_b32_e32 v3, v3, v57
-; SI-NEXT:    v_or_b32_e32 v4, v4, v48
-; SI-NEXT:    v_or_b32_e32 v5, v5, v39
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_or_b32_e32 v3, v3, v51
+; SI-NEXT:    v_or_b32_e32 v4, v4, v56
+; SI-NEXT:    v_or_b32_e32 v5, v5, v50
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v46
-; SI-NEXT:    v_or_b32_e32 v7, v7, v37
-; SI-NEXT:    v_or_b32_e32 v8, v8, v45
-; SI-NEXT:    v_or_b32_e32 v9, v9, v44
-; SI-NEXT:    v_or_b32_e32 v10, v10, v34
-; SI-NEXT:    v_or_b32_e32 v11, v11, v33
-; SI-NEXT:    v_or_b32_e32 v12, v12, v32
-; SI-NEXT:    v_or_b32_e32 v13, v13, v63
-; SI-NEXT:    v_or_b32_e32 v14, v14, v61
-; SI-NEXT:    v_or_b32_e32 v15, v15, v50
-; SI-NEXT:    v_or_b32_e32 v17, v17, v56
-; SI-NEXT:    v_or_b32_e32 v18, v18, v47
-; SI-NEXT:    v_or_b32_e32 v19, v19, v38
-; SI-NEXT:    v_or_b32_e32 v20, v20, v36
-; SI-NEXT:    v_or_b32_e32 v21, v21, v35
-; SI-NEXT:    v_or_b32_e32 v22, v22, v43
-; SI-NEXT:    v_or_b32_e32 v23, v23, v42
-; SI-NEXT:    v_or_b32_e32 v24, v24, v41
-; SI-NEXT:    v_or_b32_e32 v25, v25, v40
-; SI-NEXT:    v_or_b32_e32 v26, v26, v62
+; SI-NEXT:    v_or_b32_e32 v7, v7, v45
+; SI-NEXT:    v_or_b32_e32 v8, v8, v39
+; SI-NEXT:    v_or_b32_e32 v9, v9, v37
+; SI-NEXT:    v_or_b32_e32 v10, v10, v36
+; SI-NEXT:    v_or_b32_e32 v11, v11, v34
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v42
+; SI-NEXT:    v_or_b32_e32 v14, v14, v40
+; SI-NEXT:    v_or_b32_e32 v15, v15, v47
+; SI-NEXT:    v_or_b32_e32 v17, v17, v48
+; SI-NEXT:    v_or_b32_e32 v18, v18, v38
+; SI-NEXT:    v_or_b32_e32 v19, v19, v44
+; SI-NEXT:    v_or_b32_e32 v20, v20, v35
+; SI-NEXT:    v_or_b32_e32 v21, v21, v33
+; SI-NEXT:    v_or_b32_e32 v22, v22, v32
+; SI-NEXT:    v_or_b32_e32 v23, v23, v41
+; SI-NEXT:    v_or_b32_e32 v24, v24, v63
+; SI-NEXT:    v_or_b32_e32 v25, v25, v62
+; SI-NEXT:    v_or_b32_e32 v26, v26, v61
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v60
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr38
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:  .LBB14_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -5164,10 +5164,10 @@ define <28 x i32> @bitcast_v56i16_to_v28i32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v0, v52, v0
+; SI-NEXT:    v_or_b32_e32 v0, v58, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v58, v1
-; SI-NEXT:    v_or_b32_e32 v2, v51, v2
+; SI-NEXT:    v_or_b32_e32 v1, v52, v1
+; SI-NEXT:    v_or_b32_e32 v2, v57, v2
 ; SI-NEXT:    v_or_b32_e32 v16, v49, v16
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
@@ -5177,35 +5177,37 @@ define <28 x i32> @bitcast_v56i16_to_v28i32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
 ; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
@@ -5231,29 +5233,29 @@ define <28 x i32> @bitcast_v56i16_to_v28i32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; SI-NEXT:    v_or_b32_e32 v3, v57, v3
-; SI-NEXT:    v_or_b32_e32 v4, v48, v4
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
+; SI-NEXT:    v_or_b32_e32 v3, v51, v3
+; SI-NEXT:    v_or_b32_e32 v4, v56, v4
+; SI-NEXT:    v_or_b32_e32 v5, v50, v5
 ; SI-NEXT:    v_or_b32_e32 v6, v46, v6
-; SI-NEXT:    v_or_b32_e32 v7, v37, v7
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
-; SI-NEXT:    v_or_b32_e32 v9, v44, v9
-; SI-NEXT:    v_or_b32_e32 v10, v34, v10
-; SI-NEXT:    v_or_b32_e32 v11, v33, v11
-; SI-NEXT:    v_or_b32_e32 v12, v32, v12
-; SI-NEXT:    v_or_b32_e32 v13, v63, v13
-; SI-NEXT:    v_or_b32_e32 v14, v61, v14
-; SI-NEXT:    v_or_b32_e32 v15, v50, v15
-; SI-NEXT:    v_or_b32_e32 v17, v56, v17
-; SI-NEXT:    v_or_b32_e32 v18, v47, v18
-; SI-NEXT:    v_or_b32_e32 v19, v38, v19
-; SI-NEXT:    v_or_b32_e32 v20, v36, v20
-; SI-NEXT:    v_or_b32_e32 v21, v35, v21
-; SI-NEXT:    v_or_b32_e32 v22, v43, v22
-; SI-NEXT:    v_or_b32_e32 v23, v42, v23
-; SI-NEXT:    v_or_b32_e32 v24, v41, v24
-; SI-NEXT:    v_or_b32_e32 v25, v40, v25
-; SI-NEXT:    v_or_b32_e32 v26, v62, v26
+; SI-NEXT:    v_or_b32_e32 v7, v45, v7
+; SI-NEXT:    v_or_b32_e32 v8, v39, v8
+; SI-NEXT:    v_or_b32_e32 v9, v37, v9
+; SI-NEXT:    v_or_b32_e32 v10, v36, v10
+; SI-NEXT:    v_or_b32_e32 v11, v34, v11
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v42, v13
+; SI-NEXT:    v_or_b32_e32 v14, v40, v14
+; SI-NEXT:    v_or_b32_e32 v15, v47, v15
+; SI-NEXT:    v_or_b32_e32 v17, v48, v17
+; SI-NEXT:    v_or_b32_e32 v18, v38, v18
+; SI-NEXT:    v_or_b32_e32 v19, v44, v19
+; SI-NEXT:    v_or_b32_e32 v20, v35, v20
+; SI-NEXT:    v_or_b32_e32 v21, v33, v21
+; SI-NEXT:    v_or_b32_e32 v22, v32, v22
+; SI-NEXT:    v_or_b32_e32 v23, v41, v23
+; SI-NEXT:    v_or_b32_e32 v24, v63, v24
+; SI-NEXT:    v_or_b32_e32 v25, v62, v25
+; SI-NEXT:    v_or_b32_e32 v26, v61, v26
 ; SI-NEXT:    v_or_b32_e32 v27, v60, v27
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
@@ -11430,12 +11432,12 @@ define inreg <28 x i32> @bitcast_v56f16_to_v28i32_scalar(<56 x half> inreg %a, i
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v18, v16
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mov_b32_e32 v49, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
@@ -11485,7 +11487,7 @@ define inreg <28 x i32> @bitcast_v56f16_to_v28i32_scalar(<56 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -11584,21 +11586,21 @@ define inreg <28 x i32> @bitcast_v56f16_to_v28i32_scalar(<56 x half> inreg %a, i
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v29
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v29
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
@@ -11611,37 +11613,37 @@ define inreg <28 x i32> @bitcast_v56f16_to_v28i32_scalar(<56 x half> inreg %a, i
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
 ; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
-; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
+; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
-; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
+; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
-; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
 ; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
 ; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
 ; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
@@ -16445,152 +16447,152 @@ define <28 x float> @bitcast_v56i16_to_v28f32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v29
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:84
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:68
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v8
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v14
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:20
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:52
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v20
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:84
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v22
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v24
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:12
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v26
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
@@ -16640,9 +16642,9 @@ define <28 x float> @bitcast_v56i16_to_v28f32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v59
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    v_or_b32_e32 v0, v0, v52
-; SI-NEXT:    v_or_b32_e32 v1, v1, v58
-; SI-NEXT:    v_or_b32_e32 v2, v2, v51
+; SI-NEXT:    v_or_b32_e32 v0, v0, v58
+; SI-NEXT:    v_or_b32_e32 v1, v1, v52
+; SI-NEXT:    v_or_b32_e32 v2, v2, v57
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v49
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
@@ -16651,125 +16653,125 @@ define <28 x float> @bitcast_v56i16_to_v28f32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; kill: killed $vgpr28
-; SI-NEXT:    ; implicit-def: $vgpr52
 ; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; SI-NEXT:    v_or_b32_e32 v3, v3, v57
-; SI-NEXT:    v_or_b32_e32 v4, v4, v48
-; SI-NEXT:    v_or_b32_e32 v5, v5, v39
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_or_b32_e32 v3, v3, v51
+; SI-NEXT:    v_or_b32_e32 v4, v4, v56
+; SI-NEXT:    v_or_b32_e32 v5, v5, v50
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v46
-; SI-NEXT:    v_or_b32_e32 v7, v7, v37
-; SI-NEXT:    v_or_b32_e32 v8, v8, v45
-; SI-NEXT:    v_or_b32_e32 v9, v9, v44
-; SI-NEXT:    v_or_b32_e32 v10, v10, v34
-; SI-NEXT:    v_or_b32_e32 v11, v11, v33
-; SI-NEXT:    v_or_b32_e32 v12, v12, v32
-; SI-NEXT:    v_or_b32_e32 v13, v13, v63
-; SI-NEXT:    v_or_b32_e32 v14, v14, v61
-; SI-NEXT:    v_or_b32_e32 v15, v15, v50
-; SI-NEXT:    v_or_b32_e32 v17, v17, v56
-; SI-NEXT:    v_or_b32_e32 v18, v18, v47
-; SI-NEXT:    v_or_b32_e32 v19, v19, v38
-; SI-NEXT:    v_or_b32_e32 v20, v20, v36
-; SI-NEXT:    v_or_b32_e32 v21, v21, v35
-; SI-NEXT:    v_or_b32_e32 v22, v22, v43
-; SI-NEXT:    v_or_b32_e32 v23, v23, v42
-; SI-NEXT:    v_or_b32_e32 v24, v24, v41
-; SI-NEXT:    v_or_b32_e32 v25, v25, v40
-; SI-NEXT:    v_or_b32_e32 v26, v26, v62
+; SI-NEXT:    v_or_b32_e32 v7, v7, v45
+; SI-NEXT:    v_or_b32_e32 v8, v8, v39
+; SI-NEXT:    v_or_b32_e32 v9, v9, v37
+; SI-NEXT:    v_or_b32_e32 v10, v10, v36
+; SI-NEXT:    v_or_b32_e32 v11, v11, v34
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v42
+; SI-NEXT:    v_or_b32_e32 v14, v14, v40
+; SI-NEXT:    v_or_b32_e32 v15, v15, v47
+; SI-NEXT:    v_or_b32_e32 v17, v17, v48
+; SI-NEXT:    v_or_b32_e32 v18, v18, v38
+; SI-NEXT:    v_or_b32_e32 v19, v19, v44
+; SI-NEXT:    v_or_b32_e32 v20, v20, v35
+; SI-NEXT:    v_or_b32_e32 v21, v21, v33
+; SI-NEXT:    v_or_b32_e32 v22, v22, v32
+; SI-NEXT:    v_or_b32_e32 v23, v23, v41
+; SI-NEXT:    v_or_b32_e32 v24, v24, v63
+; SI-NEXT:    v_or_b32_e32 v25, v25, v62
+; SI-NEXT:    v_or_b32_e32 v26, v26, v61
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v60
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr38
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:  .LBB30_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -16778,10 +16780,10 @@ define <28 x float> @bitcast_v56i16_to_v28f32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v0, v52, v0
+; SI-NEXT:    v_or_b32_e32 v0, v58, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v58, v1
-; SI-NEXT:    v_or_b32_e32 v2, v51, v2
+; SI-NEXT:    v_or_b32_e32 v1, v52, v1
+; SI-NEXT:    v_or_b32_e32 v2, v57, v2
 ; SI-NEXT:    v_or_b32_e32 v16, v49, v16
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
@@ -16791,35 +16793,37 @@ define <28 x float> @bitcast_v56i16_to_v28f32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
 ; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
@@ -16845,29 +16849,29 @@ define <28 x float> @bitcast_v56i16_to_v28f32(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; SI-NEXT:    v_or_b32_e32 v3, v57, v3
-; SI-NEXT:    v_or_b32_e32 v4, v48, v4
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
+; SI-NEXT:    v_or_b32_e32 v3, v51, v3
+; SI-NEXT:    v_or_b32_e32 v4, v56, v4
+; SI-NEXT:    v_or_b32_e32 v5, v50, v5
 ; SI-NEXT:    v_or_b32_e32 v6, v46, v6
-; SI-NEXT:    v_or_b32_e32 v7, v37, v7
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
-; SI-NEXT:    v_or_b32_e32 v9, v44, v9
-; SI-NEXT:    v_or_b32_e32 v10, v34, v10
-; SI-NEXT:    v_or_b32_e32 v11, v33, v11
-; SI-NEXT:    v_or_b32_e32 v12, v32, v12
-; SI-NEXT:    v_or_b32_e32 v13, v63, v13
-; SI-NEXT:    v_or_b32_e32 v14, v61, v14
-; SI-NEXT:    v_or_b32_e32 v15, v50, v15
-; SI-NEXT:    v_or_b32_e32 v17, v56, v17
-; SI-NEXT:    v_or_b32_e32 v18, v47, v18
-; SI-NEXT:    v_or_b32_e32 v19, v38, v19
-; SI-NEXT:    v_or_b32_e32 v20, v36, v20
-; SI-NEXT:    v_or_b32_e32 v21, v35, v21
-; SI-NEXT:    v_or_b32_e32 v22, v43, v22
-; SI-NEXT:    v_or_b32_e32 v23, v42, v23
-; SI-NEXT:    v_or_b32_e32 v24, v41, v24
-; SI-NEXT:    v_or_b32_e32 v25, v40, v25
-; SI-NEXT:    v_or_b32_e32 v26, v62, v26
+; SI-NEXT:    v_or_b32_e32 v7, v45, v7
+; SI-NEXT:    v_or_b32_e32 v8, v39, v8
+; SI-NEXT:    v_or_b32_e32 v9, v37, v9
+; SI-NEXT:    v_or_b32_e32 v10, v36, v10
+; SI-NEXT:    v_or_b32_e32 v11, v34, v11
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v42, v13
+; SI-NEXT:    v_or_b32_e32 v14, v40, v14
+; SI-NEXT:    v_or_b32_e32 v15, v47, v15
+; SI-NEXT:    v_or_b32_e32 v17, v48, v17
+; SI-NEXT:    v_or_b32_e32 v18, v38, v18
+; SI-NEXT:    v_or_b32_e32 v19, v44, v19
+; SI-NEXT:    v_or_b32_e32 v20, v35, v20
+; SI-NEXT:    v_or_b32_e32 v21, v33, v21
+; SI-NEXT:    v_or_b32_e32 v22, v32, v22
+; SI-NEXT:    v_or_b32_e32 v23, v41, v23
+; SI-NEXT:    v_or_b32_e32 v24, v63, v24
+; SI-NEXT:    v_or_b32_e32 v25, v62, v25
+; SI-NEXT:    v_or_b32_e32 v26, v61, v26
 ; SI-NEXT:    v_or_b32_e32 v27, v60, v27
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
@@ -23034,12 +23038,12 @@ define inreg <28 x float> @bitcast_v56f16_to_v28f32_scalar(<56 x half> inreg %a,
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v18, v16
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mov_b32_e32 v49, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
@@ -23089,7 +23093,7 @@ define inreg <28 x float> @bitcast_v56f16_to_v28f32_scalar(<56 x half> inreg %a,
 ; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -23188,21 +23192,21 @@ define inreg <28 x float> @bitcast_v56f16_to_v28f32_scalar(<56 x half> inreg %a,
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v29
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v29
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
@@ -23215,37 +23219,37 @@ define inreg <28 x float> @bitcast_v56f16_to_v28f32_scalar(<56 x half> inreg %a,
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
 ; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
-; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
+; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
-; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
+; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
-; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
 ; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
 ; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
 ; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
@@ -27200,152 +27204,152 @@ define <14 x i64> @bitcast_v56i16_to_v14i64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v29
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:84
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:68
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v8
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:20
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:52
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v20
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:84
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v22
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v24
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:12
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v26
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
@@ -27395,9 +27399,9 @@ define <14 x i64> @bitcast_v56i16_to_v14i64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v59
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    v_or_b32_e32 v0, v0, v52
-; SI-NEXT:    v_or_b32_e32 v1, v1, v58
-; SI-NEXT:    v_or_b32_e32 v2, v2, v51
+; SI-NEXT:    v_or_b32_e32 v0, v0, v58
+; SI-NEXT:    v_or_b32_e32 v1, v1, v52
+; SI-NEXT:    v_or_b32_e32 v2, v2, v57
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v49
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
@@ -27406,125 +27410,125 @@ define <14 x i64> @bitcast_v56i16_to_v14i64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; kill: killed $vgpr28
-; SI-NEXT:    ; implicit-def: $vgpr52
 ; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; SI-NEXT:    v_or_b32_e32 v3, v3, v57
-; SI-NEXT:    v_or_b32_e32 v4, v4, v48
-; SI-NEXT:    v_or_b32_e32 v5, v5, v39
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_or_b32_e32 v3, v3, v51
+; SI-NEXT:    v_or_b32_e32 v4, v4, v56
+; SI-NEXT:    v_or_b32_e32 v5, v5, v50
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v46
-; SI-NEXT:    v_or_b32_e32 v7, v7, v37
-; SI-NEXT:    v_or_b32_e32 v8, v8, v45
-; SI-NEXT:    v_or_b32_e32 v9, v9, v44
-; SI-NEXT:    v_or_b32_e32 v10, v10, v34
-; SI-NEXT:    v_or_b32_e32 v11, v11, v33
-; SI-NEXT:    v_or_b32_e32 v12, v12, v32
-; SI-NEXT:    v_or_b32_e32 v13, v13, v63
-; SI-NEXT:    v_or_b32_e32 v14, v14, v61
-; SI-NEXT:    v_or_b32_e32 v15, v15, v50
-; SI-NEXT:    v_or_b32_e32 v17, v17, v56
-; SI-NEXT:    v_or_b32_e32 v18, v18, v47
-; SI-NEXT:    v_or_b32_e32 v19, v19, v38
-; SI-NEXT:    v_or_b32_e32 v20, v20, v36
-; SI-NEXT:    v_or_b32_e32 v21, v21, v35
-; SI-NEXT:    v_or_b32_e32 v22, v22, v43
-; SI-NEXT:    v_or_b32_e32 v23, v23, v42
-; SI-NEXT:    v_or_b32_e32 v24, v24, v41
-; SI-NEXT:    v_or_b32_e32 v25, v25, v40
-; SI-NEXT:    v_or_b32_e32 v26, v26, v62
+; SI-NEXT:    v_or_b32_e32 v7, v7, v45
+; SI-NEXT:    v_or_b32_e32 v8, v8, v39
+; SI-NEXT:    v_or_b32_e32 v9, v9, v37
+; SI-NEXT:    v_or_b32_e32 v10, v10, v36
+; SI-NEXT:    v_or_b32_e32 v11, v11, v34
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v42
+; SI-NEXT:    v_or_b32_e32 v14, v14, v40
+; SI-NEXT:    v_or_b32_e32 v15, v15, v47
+; SI-NEXT:    v_or_b32_e32 v17, v17, v48
+; SI-NEXT:    v_or_b32_e32 v18, v18, v38
+; SI-NEXT:    v_or_b32_e32 v19, v19, v44
+; SI-NEXT:    v_or_b32_e32 v20, v20, v35
+; SI-NEXT:    v_or_b32_e32 v21, v21, v33
+; SI-NEXT:    v_or_b32_e32 v22, v22, v32
+; SI-NEXT:    v_or_b32_e32 v23, v23, v41
+; SI-NEXT:    v_or_b32_e32 v24, v24, v63
+; SI-NEXT:    v_or_b32_e32 v25, v25, v62
+; SI-NEXT:    v_or_b32_e32 v26, v26, v61
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v60
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr38
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:  .LBB42_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -27533,10 +27537,10 @@ define <14 x i64> @bitcast_v56i16_to_v14i64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v0, v52, v0
+; SI-NEXT:    v_or_b32_e32 v0, v58, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v58, v1
-; SI-NEXT:    v_or_b32_e32 v2, v51, v2
+; SI-NEXT:    v_or_b32_e32 v1, v52, v1
+; SI-NEXT:    v_or_b32_e32 v2, v57, v2
 ; SI-NEXT:    v_or_b32_e32 v16, v49, v16
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
@@ -27546,35 +27550,37 @@ define <14 x i64> @bitcast_v56i16_to_v14i64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
 ; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
@@ -27600,29 +27606,29 @@ define <14 x i64> @bitcast_v56i16_to_v14i64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; SI-NEXT:    v_or_b32_e32 v3, v57, v3
-; SI-NEXT:    v_or_b32_e32 v4, v48, v4
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
+; SI-NEXT:    v_or_b32_e32 v3, v51, v3
+; SI-NEXT:    v_or_b32_e32 v4, v56, v4
+; SI-NEXT:    v_or_b32_e32 v5, v50, v5
 ; SI-NEXT:    v_or_b32_e32 v6, v46, v6
-; SI-NEXT:    v_or_b32_e32 v7, v37, v7
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
-; SI-NEXT:    v_or_b32_e32 v9, v44, v9
-; SI-NEXT:    v_or_b32_e32 v10, v34, v10
-; SI-NEXT:    v_or_b32_e32 v11, v33, v11
-; SI-NEXT:    v_or_b32_e32 v12, v32, v12
-; SI-NEXT:    v_or_b32_e32 v13, v63, v13
-; SI-NEXT:    v_or_b32_e32 v14, v61, v14
-; SI-NEXT:    v_or_b32_e32 v15, v50, v15
-; SI-NEXT:    v_or_b32_e32 v17, v56, v17
-; SI-NEXT:    v_or_b32_e32 v18, v47, v18
-; SI-NEXT:    v_or_b32_e32 v19, v38, v19
-; SI-NEXT:    v_or_b32_e32 v20, v36, v20
-; SI-NEXT:    v_or_b32_e32 v21, v35, v21
-; SI-NEXT:    v_or_b32_e32 v22, v43, v22
-; SI-NEXT:    v_or_b32_e32 v23, v42, v23
-; SI-NEXT:    v_or_b32_e32 v24, v41, v24
-; SI-NEXT:    v_or_b32_e32 v25, v40, v25
-; SI-NEXT:    v_or_b32_e32 v26, v62, v26
+; SI-NEXT:    v_or_b32_e32 v7, v45, v7
+; SI-NEXT:    v_or_b32_e32 v8, v39, v8
+; SI-NEXT:    v_or_b32_e32 v9, v37, v9
+; SI-NEXT:    v_or_b32_e32 v10, v36, v10
+; SI-NEXT:    v_or_b32_e32 v11, v34, v11
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v42, v13
+; SI-NEXT:    v_or_b32_e32 v14, v40, v14
+; SI-NEXT:    v_or_b32_e32 v15, v47, v15
+; SI-NEXT:    v_or_b32_e32 v17, v48, v17
+; SI-NEXT:    v_or_b32_e32 v18, v38, v18
+; SI-NEXT:    v_or_b32_e32 v19, v44, v19
+; SI-NEXT:    v_or_b32_e32 v20, v35, v20
+; SI-NEXT:    v_or_b32_e32 v21, v33, v21
+; SI-NEXT:    v_or_b32_e32 v22, v32, v22
+; SI-NEXT:    v_or_b32_e32 v23, v41, v23
+; SI-NEXT:    v_or_b32_e32 v24, v63, v24
+; SI-NEXT:    v_or_b32_e32 v25, v62, v25
+; SI-NEXT:    v_or_b32_e32 v26, v61, v26
 ; SI-NEXT:    v_or_b32_e32 v27, v60, v27
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
@@ -33813,12 +33819,12 @@ define inreg <14 x i64> @bitcast_v56f16_to_v14i64_scalar(<56 x half> inreg %a, i
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v18, v16
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mov_b32_e32 v49, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
@@ -33868,7 +33874,7 @@ define inreg <14 x i64> @bitcast_v56f16_to_v14i64_scalar(<56 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -33967,21 +33973,21 @@ define inreg <14 x i64> @bitcast_v56f16_to_v14i64_scalar(<56 x half> inreg %a, i
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v29
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v29
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
@@ -33994,37 +34000,37 @@ define inreg <14 x i64> @bitcast_v56f16_to_v14i64_scalar(<56 x half> inreg %a, i
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
 ; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
-; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
+; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
-; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
+; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
-; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
 ; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
 ; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
 ; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
@@ -37005,152 +37011,152 @@ define <14 x double> @bitcast_v56i16_to_v14f64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v54, v2
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92
 ; SI-NEXT:    v_mov_b32_e32 v53, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v11
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v27
-; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v29
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v29
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v0
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:84
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:68
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v6
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v8
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v10
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v16
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v18
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:20
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:16
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:52
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v20
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:84
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v22
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:44
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v24
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:36
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:12
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v26
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:28
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
@@ -37200,9 +37206,9 @@ define <14 x double> @bitcast_v56i16_to_v14f64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v59
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
-; SI-NEXT:    v_or_b32_e32 v0, v0, v52
-; SI-NEXT:    v_or_b32_e32 v1, v1, v58
-; SI-NEXT:    v_or_b32_e32 v2, v2, v51
+; SI-NEXT:    v_or_b32_e32 v0, v0, v58
+; SI-NEXT:    v_or_b32_e32 v1, v1, v52
+; SI-NEXT:    v_or_b32_e32 v2, v2, v57
 ; SI-NEXT:    v_or_b32_e32 v16, v16, v49
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr54
@@ -37211,125 +37217,125 @@ define <14 x double> @bitcast_v56i16_to_v14f64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    ; kill: killed $vgpr28
 ; SI-NEXT:    ; implicit-def: $vgpr28
 ; SI-NEXT:    ; kill: killed $vgpr28
-; SI-NEXT:    ; implicit-def: $vgpr52
 ; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; SI-NEXT:    v_or_b32_e32 v3, v3, v57
-; SI-NEXT:    v_or_b32_e32 v4, v4, v48
-; SI-NEXT:    v_or_b32_e32 v5, v5, v39
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_or_b32_e32 v3, v3, v51
+; SI-NEXT:    v_or_b32_e32 v4, v4, v56
+; SI-NEXT:    v_or_b32_e32 v5, v5, v50
 ; SI-NEXT:    v_or_b32_e32 v6, v6, v46
-; SI-NEXT:    v_or_b32_e32 v7, v7, v37
-; SI-NEXT:    v_or_b32_e32 v8, v8, v45
-; SI-NEXT:    v_or_b32_e32 v9, v9, v44
-; SI-NEXT:    v_or_b32_e32 v10, v10, v34
-; SI-NEXT:    v_or_b32_e32 v11, v11, v33
-; SI-NEXT:    v_or_b32_e32 v12, v12, v32
-; SI-NEXT:    v_or_b32_e32 v13, v13, v63
-; SI-NEXT:    v_or_b32_e32 v14, v14, v61
-; SI-NEXT:    v_or_b32_e32 v15, v15, v50
-; SI-NEXT:    v_or_b32_e32 v17, v17, v56
-; SI-NEXT:    v_or_b32_e32 v18, v18, v47
-; SI-NEXT:    v_or_b32_e32 v19, v19, v38
-; SI-NEXT:    v_or_b32_e32 v20, v20, v36
-; SI-NEXT:    v_or_b32_e32 v21, v21, v35
-; SI-NEXT:    v_or_b32_e32 v22, v22, v43
-; SI-NEXT:    v_or_b32_e32 v23, v23, v42
-; SI-NEXT:    v_or_b32_e32 v24, v24, v41
-; SI-NEXT:    v_or_b32_e32 v25, v25, v40
-; SI-NEXT:    v_or_b32_e32 v26, v26, v62
+; SI-NEXT:    v_or_b32_e32 v7, v7, v45
+; SI-NEXT:    v_or_b32_e32 v8, v8, v39
+; SI-NEXT:    v_or_b32_e32 v9, v9, v37
+; SI-NEXT:    v_or_b32_e32 v10, v10, v36
+; SI-NEXT:    v_or_b32_e32 v11, v11, v34
+; SI-NEXT:    v_or_b32_e32 v12, v12, v43
+; SI-NEXT:    v_or_b32_e32 v13, v13, v42
+; SI-NEXT:    v_or_b32_e32 v14, v14, v40
+; SI-NEXT:    v_or_b32_e32 v15, v15, v47
+; SI-NEXT:    v_or_b32_e32 v17, v17, v48
+; SI-NEXT:    v_or_b32_e32 v18, v18, v38
+; SI-NEXT:    v_or_b32_e32 v19, v19, v44
+; SI-NEXT:    v_or_b32_e32 v20, v20, v35
+; SI-NEXT:    v_or_b32_e32 v21, v21, v33
+; SI-NEXT:    v_or_b32_e32 v22, v22, v32
+; SI-NEXT:    v_or_b32_e32 v23, v23, v41
+; SI-NEXT:    v_or_b32_e32 v24, v24, v63
+; SI-NEXT:    v_or_b32_e32 v25, v25, v62
+; SI-NEXT:    v_or_b32_e32 v26, v26, v61
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v60
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr51
+; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr45
-; SI-NEXT:    ; implicit-def: $vgpr44
+; SI-NEXT:    ; implicit-def: $vgpr39
+; SI-NEXT:    ; implicit-def: $vgpr37
+; SI-NEXT:    ; implicit-def: $vgpr36
 ; SI-NEXT:    ; implicit-def: $vgpr34
-; SI-NEXT:    ; implicit-def: $vgpr33
-; SI-NEXT:    ; implicit-def: $vgpr32
-; SI-NEXT:    ; implicit-def: $vgpr63
-; SI-NEXT:    ; implicit-def: $vgpr61
-; SI-NEXT:    ; implicit-def: $vgpr50
-; SI-NEXT:    ; implicit-def: $vgpr56
+; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr48
 ; SI-NEXT:    ; implicit-def: $vgpr38
-; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr44
 ; SI-NEXT:    ; implicit-def: $vgpr35
-; SI-NEXT:    ; implicit-def: $vgpr43
-; SI-NEXT:    ; implicit-def: $vgpr42
+; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
+; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
+; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; implicit-def: $vgpr60
 ; SI-NEXT:  .LBB50_2: ; %Flow
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v54
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v53
@@ -37338,10 +37344,10 @@ define <14 x double> @bitcast_v56i16_to_v14f64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; SI-NEXT:    v_or_b32_e32 v0, v52, v0
+; SI-NEXT:    v_or_b32_e32 v0, v58, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v1, v58, v1
-; SI-NEXT:    v_or_b32_e32 v2, v51, v2
+; SI-NEXT:    v_or_b32_e32 v1, v52, v1
+; SI-NEXT:    v_or_b32_e32 v2, v57, v2
 ; SI-NEXT:    v_or_b32_e32 v16, v49, v16
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
@@ -37351,35 +37357,37 @@ define <14 x double> @bitcast_v56i16_to_v14f64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
+; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
-; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
 ; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
 ; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
-; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
 ; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
-; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
@@ -37405,29 +37413,29 @@ define <14 x double> @bitcast_v56i16_to_v14f64(<56 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; SI-NEXT:    v_or_b32_e32 v3, v57, v3
-; SI-NEXT:    v_or_b32_e32 v4, v48, v4
-; SI-NEXT:    v_or_b32_e32 v5, v39, v5
+; SI-NEXT:    v_or_b32_e32 v3, v51, v3
+; SI-NEXT:    v_or_b32_e32 v4, v56, v4
+; SI-NEXT:    v_or_b32_e32 v5, v50, v5
 ; SI-NEXT:    v_or_b32_e32 v6, v46, v6
-; SI-NEXT:    v_or_b32_e32 v7, v37, v7
-; SI-NEXT:    v_or_b32_e32 v8, v45, v8
-; SI-NEXT:    v_or_b32_e32 v9, v44, v9
-; SI-NEXT:    v_or_b32_e32 v10, v34, v10
-; SI-NEXT:    v_or_b32_e32 v11, v33, v11
-; SI-NEXT:    v_or_b32_e32 v12, v32, v12
-; SI-NEXT:    v_or_b32_e32 v13, v63, v13
-; SI-NEXT:    v_or_b32_e32 v14, v61, v14
-; SI-NEXT:    v_or_b32_e32 v15, v50, v15
-; SI-NEXT:    v_or_b32_e32 v17, v56, v17
-; SI-NEXT:    v_or_b32_e32 v18, v47, v18
-; SI-NEXT:    v_or_b32_e32 v19, v38, v19
-; SI-NEXT:    v_or_b32_e32 v20, v36, v20
-; SI-NEXT:    v_or_b32_e32 v21, v35, v21
-; SI-NEXT:    v_or_b32_e32 v22, v43, v22
-; SI-NEXT:    v_or_b32_e32 v23, v42, v23
-; SI-NEXT:    v_or_b32_e32 v24, v41, v24
-; SI-NEXT:    v_or_b32_e32 v25, v40, v25
-; SI-NEXT:    v_or_b32_e32 v26, v62, v26
+; SI-NEXT:    v_or_b32_e32 v7, v45, v7
+; SI-NEXT:    v_or_b32_e32 v8, v39, v8
+; SI-NEXT:    v_or_b32_e32 v9, v37, v9
+; SI-NEXT:    v_or_b32_e32 v10, v36, v10
+; SI-NEXT:    v_or_b32_e32 v11, v34, v11
+; SI-NEXT:    v_or_b32_e32 v12, v43, v12
+; SI-NEXT:    v_or_b32_e32 v13, v42, v13
+; SI-NEXT:    v_or_b32_e32 v14, v40, v14
+; SI-NEXT:    v_or_b32_e32 v15, v47, v15
+; SI-NEXT:    v_or_b32_e32 v17, v48, v17
+; SI-NEXT:    v_or_b32_e32 v18, v38, v18
+; SI-NEXT:    v_or_b32_e32 v19, v44, v19
+; SI-NEXT:    v_or_b32_e32 v20, v35, v20
+; SI-NEXT:    v_or_b32_e32 v21, v33, v21
+; SI-NEXT:    v_or_b32_e32 v22, v32, v22
+; SI-NEXT:    v_or_b32_e32 v23, v41, v23
+; SI-NEXT:    v_or_b32_e32 v24, v63, v24
+; SI-NEXT:    v_or_b32_e32 v25, v62, v25
+; SI-NEXT:    v_or_b32_e32 v26, v61, v26
 ; SI-NEXT:    v_or_b32_e32 v27, v60, v27
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
 ; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
@@ -43505,12 +43513,12 @@ define inreg <14 x double> @bitcast_v56f16_to_v14f64_scalar(<56 x half> inreg %a
 ; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
 ; SI-NEXT:    v_or_b32_e32 v16, v18, v16
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    v_mov_b32_e32 v49, v2
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
@@ -43560,7 +43568,7 @@ define inreg <14 x double> @bitcast_v56f16_to_v14f64_scalar(<56 x half> inreg %a
 ; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(10)
 ; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_or_b32_e32 v18, v19, v18
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(0)
@@ -43659,21 +43667,21 @@ define inreg <14 x double> @bitcast_v56f16_to_v14f64_scalar(<56 x half> inreg %a
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    v_cvt_f32_f16_e32 v22, v29
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v22, v29
 ; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
@@ -43686,37 +43694,37 @@ define inreg <14 x double> @bitcast_v56f16_to_v14f64_scalar(<56 x half> inreg %a
 ; SI-NEXT:    v_or_b32_e32 v4, v5, v4
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
 ; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
-; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
 ; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
+; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
+; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
 ; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
-; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
+; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
 ; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
 ; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
-; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
 ; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
 ; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
 ; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
index 0a7790a27f5ae..94ed6276bd051 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
@@ -5148,167 +5148,167 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:12
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v29
 ; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:4
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:108
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v8
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:100
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v14
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:92
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:84
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:76
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v16
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v18
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v20
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:108
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:68
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:52
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v28
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v30
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:88
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v30
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
@@ -5365,111 +5365,111 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v61
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    v_or_b32_e32 v0, v0, v54
-; SI-NEXT:    v_or_b32_e32 v18, v18, v43
+; SI-NEXT:    v_or_b32_e32 v0, v0, v60
+; SI-NEXT:    v_or_b32_e32 v18, v18, v37
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    ; kill: killed $vgpr30
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; SI-NEXT:    v_or_b32_e32 v1, v1, v60
-; SI-NEXT:    v_or_b32_e32 v2, v2, v53
-; SI-NEXT:    v_or_b32_e32 v3, v3, v52
-; SI-NEXT:    v_or_b32_e32 v4, v4, v51
-; SI-NEXT:    v_or_b32_e32 v5, v5, v59
-; SI-NEXT:    v_or_b32_e32 v6, v6, v50
-; SI-NEXT:    v_or_b32_e32 v7, v7, v49
-; SI-NEXT:    v_or_b32_e32 v8, v8, v48
-; SI-NEXT:    v_or_b32_e32 v9, v9, v58
-; SI-NEXT:    v_or_b32_e32 v10, v10, v57
-; SI-NEXT:    v_or_b32_e32 v11, v11, v39
-; SI-NEXT:    v_or_b32_e32 v12, v12, v56
-; SI-NEXT:    v_or_b32_e32 v13, v13, v47
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_or_b32_e32 v1, v1, v54
+; SI-NEXT:    v_or_b32_e32 v2, v2, v59
+; SI-NEXT:    v_or_b32_e32 v3, v3, v58
+; SI-NEXT:    v_or_b32_e32 v4, v4, v53
+; SI-NEXT:    v_or_b32_e32 v5, v5, v57
+; SI-NEXT:    v_or_b32_e32 v6, v6, v52
+; SI-NEXT:    v_or_b32_e32 v7, v7, v56
+; SI-NEXT:    v_or_b32_e32 v8, v8, v51
+; SI-NEXT:    v_or_b32_e32 v9, v9, v50
+; SI-NEXT:    v_or_b32_e32 v10, v10, v49
+; SI-NEXT:    v_or_b32_e32 v11, v11, v48
+; SI-NEXT:    v_or_b32_e32 v12, v12, v47
+; SI-NEXT:    v_or_b32_e32 v13, v13, v39
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v46
-; SI-NEXT:    v_or_b32_e32 v15, v15, v38
-; SI-NEXT:    v_or_b32_e32 v16, v16, v45
+; SI-NEXT:    v_or_b32_e32 v15, v15, v45
+; SI-NEXT:    v_or_b32_e32 v16, v16, v38
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v44
-; SI-NEXT:    v_or_b32_e32 v19, v19, v42
-; SI-NEXT:    v_or_b32_e32 v20, v20, v41
-; SI-NEXT:    v_or_b32_e32 v21, v21, v40
-; SI-NEXT:    v_or_b32_e32 v22, v22, v37
-; SI-NEXT:    v_or_b32_e32 v23, v23, v36
-; SI-NEXT:    v_or_b32_e32 v24, v24, v35
-; SI-NEXT:    v_or_b32_e32 v25, v25, v34
-; SI-NEXT:    v_or_b32_e32 v26, v26, v33
+; SI-NEXT:    v_or_b32_e32 v19, v19, v36
+; SI-NEXT:    v_or_b32_e32 v20, v20, v43
+; SI-NEXT:    v_or_b32_e32 v21, v21, v35
+; SI-NEXT:    v_or_b32_e32 v22, v22, v42
+; SI-NEXT:    v_or_b32_e32 v23, v23, v34
+; SI-NEXT:    v_or_b32_e32 v24, v24, v41
+; SI-NEXT:    v_or_b32_e32 v25, v25, v33
+; SI-NEXT:    v_or_b32_e32 v26, v26, v40
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v32
 ; SI-NEXT:    v_or_b32_e32 v28, v28, v63
 ; SI-NEXT:    v_or_b32_e32 v29, v29, v62
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr59
 ; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
@@ -5477,79 +5477,75 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB14_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v0, v54, v0
+; SI-NEXT:    v_or_b32_e32 v0, v60, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v18, v43, v18
+; SI-NEXT:    v_or_b32_e32 v18, v37, v18
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
@@ -5583,31 +5579,31 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; SI-NEXT:    v_or_b32_e32 v1, v60, v1
-; SI-NEXT:    v_or_b32_e32 v2, v53, v2
-; SI-NEXT:    v_or_b32_e32 v3, v52, v3
-; SI-NEXT:    v_or_b32_e32 v4, v51, v4
-; SI-NEXT:    v_or_b32_e32 v5, v59, v5
-; SI-NEXT:    v_or_b32_e32 v6, v50, v6
-; SI-NEXT:    v_or_b32_e32 v7, v49, v7
-; SI-NEXT:    v_or_b32_e32 v8, v48, v8
-; SI-NEXT:    v_or_b32_e32 v9, v58, v9
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_or_b32_e32 v11, v39, v11
-; SI-NEXT:    v_or_b32_e32 v12, v56, v12
-; SI-NEXT:    v_or_b32_e32 v13, v47, v13
+; SI-NEXT:    v_or_b32_e32 v1, v54, v1
+; SI-NEXT:    v_or_b32_e32 v2, v59, v2
+; SI-NEXT:    v_or_b32_e32 v3, v58, v3
+; SI-NEXT:    v_or_b32_e32 v4, v53, v4
+; SI-NEXT:    v_or_b32_e32 v5, v57, v5
+; SI-NEXT:    v_or_b32_e32 v6, v52, v6
+; SI-NEXT:    v_or_b32_e32 v7, v56, v7
+; SI-NEXT:    v_or_b32_e32 v8, v51, v8
+; SI-NEXT:    v_or_b32_e32 v9, v50, v9
+; SI-NEXT:    v_or_b32_e32 v10, v49, v10
+; SI-NEXT:    v_or_b32_e32 v11, v48, v11
+; SI-NEXT:    v_or_b32_e32 v12, v47, v12
+; SI-NEXT:    v_or_b32_e32 v13, v39, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v46, v14
-; SI-NEXT:    v_or_b32_e32 v15, v38, v15
-; SI-NEXT:    v_or_b32_e32 v16, v45, v16
+; SI-NEXT:    v_or_b32_e32 v15, v45, v15
+; SI-NEXT:    v_or_b32_e32 v16, v38, v16
 ; SI-NEXT:    v_or_b32_e32 v17, v44, v17
-; SI-NEXT:    v_or_b32_e32 v19, v42, v19
-; SI-NEXT:    v_or_b32_e32 v20, v41, v20
-; SI-NEXT:    v_or_b32_e32 v21, v40, v21
-; SI-NEXT:    v_or_b32_e32 v22, v37, v22
-; SI-NEXT:    v_or_b32_e32 v23, v36, v23
-; SI-NEXT:    v_or_b32_e32 v24, v35, v24
-; SI-NEXT:    v_or_b32_e32 v25, v34, v25
-; SI-NEXT:    v_or_b32_e32 v26, v33, v26
+; SI-NEXT:    v_or_b32_e32 v19, v36, v19
+; SI-NEXT:    v_or_b32_e32 v20, v43, v20
+; SI-NEXT:    v_or_b32_e32 v21, v35, v21
+; SI-NEXT:    v_or_b32_e32 v22, v42, v22
+; SI-NEXT:    v_or_b32_e32 v23, v34, v23
+; SI-NEXT:    v_or_b32_e32 v24, v41, v24
+; SI-NEXT:    v_or_b32_e32 v25, v33, v25
+; SI-NEXT:    v_or_b32_e32 v26, v40, v26
 ; SI-NEXT:    v_or_b32_e32 v27, v32, v27
 ; SI-NEXT:    v_or_b32_e32 v28, v63, v28
 ; SI-NEXT:    v_or_b32_e32 v29, v62, v29
@@ -10599,17 +10595,17 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:28
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:40
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:36
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:48
 ; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:60
 ; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:72
 ; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:68
@@ -10714,35 +10710,35 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v61
-; SI-NEXT:    v_cvt_f16_f32_e32 v61, v51
-; SI-NEXT:    v_cvt_f16_f32_e32 v51, v53
+; SI-NEXT:    v_cvt_f16_f32_e32 v61, v49
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_cvt_f16_f32_e32 v49, v55
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v62
-; SI-NEXT:    v_cvt_f16_f32_e32 v62, v50
-; SI-NEXT:    v_cvt_f16_f32_e32 v50, v54
+; SI-NEXT:    v_cvt_f16_f32_e32 v62, v51
+; SI-NEXT:    v_cvt_f16_f32_e32 v51, v53
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v63
-; SI-NEXT:    v_cvt_f16_f32_e32 v63, v49
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v55
+; SI-NEXT:    v_cvt_f16_f32_e32 v63, v50
+; SI-NEXT:    v_cvt_f16_f32_e32 v50, v54
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v32
 ; SI-NEXT:    v_cvt_f16_f32_e32 v32, v47
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v33
-; SI-NEXT:    v_cvt_f16_f32_e32 v33, v46
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v34
 ; SI-NEXT:    v_cvt_f16_f32_e32 v34, v45
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v35
 ; SI-NEXT:    v_cvt_f16_f32_e32 v35, v44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v33
+; SI-NEXT:    v_cvt_f16_f32_e32 v33, v46
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v36
@@ -12312,21 +12308,26 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    v_mov_b32_e32 v33, v32
 ; SI-NEXT:    v_or_b32_e32 v10, v32, v10
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_mov_b32_e32 v44, v43
 ; SI-NEXT:    v_or_b32_e32 v13, v43, v13
 ; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v8, v5
 ; SI-NEXT:    v_mov_b32_e32 v57, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v39
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v49
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
@@ -12366,11 +12367,6 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    v_or_b32_e32 v22, v30, v22
 ; SI-NEXT:    v_or_b32_e32 v23, v31, v23
 ; SI-NEXT:    s_mov_b64 s[4:5], 0
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(10)
@@ -12503,9 +12499,9 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v31
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
-; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
@@ -12514,8 +12510,8 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
+; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 ; SI-NEXT:    s_waitcnt vmcnt(13)
@@ -12530,38 +12526,38 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
-; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
+; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
 ; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
-; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
 ; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -17703,167 +17699,167 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:12
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v29
 ; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:4
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:108
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:100
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:92
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v14
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:84
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:68
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v18
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:76
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v16
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v18
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v20
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:108
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:68
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:52
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v28
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v30
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:88
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v30
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
@@ -17920,111 +17916,111 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v61
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    v_or_b32_e32 v0, v0, v54
-; SI-NEXT:    v_or_b32_e32 v18, v18, v43
+; SI-NEXT:    v_or_b32_e32 v0, v0, v60
+; SI-NEXT:    v_or_b32_e32 v18, v18, v37
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    ; kill: killed $vgpr30
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; SI-NEXT:    v_or_b32_e32 v1, v1, v60
-; SI-NEXT:    v_or_b32_e32 v2, v2, v53
-; SI-NEXT:    v_or_b32_e32 v3, v3, v52
-; SI-NEXT:    v_or_b32_e32 v4, v4, v51
-; SI-NEXT:    v_or_b32_e32 v5, v5, v59
-; SI-NEXT:    v_or_b32_e32 v6, v6, v50
-; SI-NEXT:    v_or_b32_e32 v7, v7, v49
-; SI-NEXT:    v_or_b32_e32 v8, v8, v48
-; SI-NEXT:    v_or_b32_e32 v9, v9, v58
-; SI-NEXT:    v_or_b32_e32 v10, v10, v57
-; SI-NEXT:    v_or_b32_e32 v11, v11, v39
-; SI-NEXT:    v_or_b32_e32 v12, v12, v56
-; SI-NEXT:    v_or_b32_e32 v13, v13, v47
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_or_b32_e32 v1, v1, v54
+; SI-NEXT:    v_or_b32_e32 v2, v2, v59
+; SI-NEXT:    v_or_b32_e32 v3, v3, v58
+; SI-NEXT:    v_or_b32_e32 v4, v4, v53
+; SI-NEXT:    v_or_b32_e32 v5, v5, v57
+; SI-NEXT:    v_or_b32_e32 v6, v6, v52
+; SI-NEXT:    v_or_b32_e32 v7, v7, v56
+; SI-NEXT:    v_or_b32_e32 v8, v8, v51
+; SI-NEXT:    v_or_b32_e32 v9, v9, v50
+; SI-NEXT:    v_or_b32_e32 v10, v10, v49
+; SI-NEXT:    v_or_b32_e32 v11, v11, v48
+; SI-NEXT:    v_or_b32_e32 v12, v12, v47
+; SI-NEXT:    v_or_b32_e32 v13, v13, v39
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v46
-; SI-NEXT:    v_or_b32_e32 v15, v15, v38
-; SI-NEXT:    v_or_b32_e32 v16, v16, v45
+; SI-NEXT:    v_or_b32_e32 v15, v15, v45
+; SI-NEXT:    v_or_b32_e32 v16, v16, v38
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v44
-; SI-NEXT:    v_or_b32_e32 v19, v19, v42
-; SI-NEXT:    v_or_b32_e32 v20, v20, v41
-; SI-NEXT:    v_or_b32_e32 v21, v21, v40
-; SI-NEXT:    v_or_b32_e32 v22, v22, v37
-; SI-NEXT:    v_or_b32_e32 v23, v23, v36
-; SI-NEXT:    v_or_b32_e32 v24, v24, v35
-; SI-NEXT:    v_or_b32_e32 v25, v25, v34
-; SI-NEXT:    v_or_b32_e32 v26, v26, v33
+; SI-NEXT:    v_or_b32_e32 v19, v19, v36
+; SI-NEXT:    v_or_b32_e32 v20, v20, v43
+; SI-NEXT:    v_or_b32_e32 v21, v21, v35
+; SI-NEXT:    v_or_b32_e32 v22, v22, v42
+; SI-NEXT:    v_or_b32_e32 v23, v23, v34
+; SI-NEXT:    v_or_b32_e32 v24, v24, v41
+; SI-NEXT:    v_or_b32_e32 v25, v25, v33
+; SI-NEXT:    v_or_b32_e32 v26, v26, v40
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v32
 ; SI-NEXT:    v_or_b32_e32 v28, v28, v63
 ; SI-NEXT:    v_or_b32_e32 v29, v29, v62
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr59
 ; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
@@ -18032,79 +18028,75 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB30_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v0, v54, v0
+; SI-NEXT:    v_or_b32_e32 v0, v60, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v18, v43, v18
+; SI-NEXT:    v_or_b32_e32 v18, v37, v18
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
@@ -18138,31 +18130,31 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; SI-NEXT:    v_or_b32_e32 v1, v60, v1
-; SI-NEXT:    v_or_b32_e32 v2, v53, v2
-; SI-NEXT:    v_or_b32_e32 v3, v52, v3
-; SI-NEXT:    v_or_b32_e32 v4, v51, v4
-; SI-NEXT:    v_or_b32_e32 v5, v59, v5
-; SI-NEXT:    v_or_b32_e32 v6, v50, v6
-; SI-NEXT:    v_or_b32_e32 v7, v49, v7
-; SI-NEXT:    v_or_b32_e32 v8, v48, v8
-; SI-NEXT:    v_or_b32_e32 v9, v58, v9
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_or_b32_e32 v11, v39, v11
-; SI-NEXT:    v_or_b32_e32 v12, v56, v12
-; SI-NEXT:    v_or_b32_e32 v13, v47, v13
+; SI-NEXT:    v_or_b32_e32 v1, v54, v1
+; SI-NEXT:    v_or_b32_e32 v2, v59, v2
+; SI-NEXT:    v_or_b32_e32 v3, v58, v3
+; SI-NEXT:    v_or_b32_e32 v4, v53, v4
+; SI-NEXT:    v_or_b32_e32 v5, v57, v5
+; SI-NEXT:    v_or_b32_e32 v6, v52, v6
+; SI-NEXT:    v_or_b32_e32 v7, v56, v7
+; SI-NEXT:    v_or_b32_e32 v8, v51, v8
+; SI-NEXT:    v_or_b32_e32 v9, v50, v9
+; SI-NEXT:    v_or_b32_e32 v10, v49, v10
+; SI-NEXT:    v_or_b32_e32 v11, v48, v11
+; SI-NEXT:    v_or_b32_e32 v12, v47, v12
+; SI-NEXT:    v_or_b32_e32 v13, v39, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v46, v14
-; SI-NEXT:    v_or_b32_e32 v15, v38, v15
-; SI-NEXT:    v_or_b32_e32 v16, v45, v16
+; SI-NEXT:    v_or_b32_e32 v15, v45, v15
+; SI-NEXT:    v_or_b32_e32 v16, v38, v16
 ; SI-NEXT:    v_or_b32_e32 v17, v44, v17
-; SI-NEXT:    v_or_b32_e32 v19, v42, v19
-; SI-NEXT:    v_or_b32_e32 v20, v41, v20
-; SI-NEXT:    v_or_b32_e32 v21, v40, v21
-; SI-NEXT:    v_or_b32_e32 v22, v37, v22
-; SI-NEXT:    v_or_b32_e32 v23, v36, v23
-; SI-NEXT:    v_or_b32_e32 v24, v35, v24
-; SI-NEXT:    v_or_b32_e32 v25, v34, v25
-; SI-NEXT:    v_or_b32_e32 v26, v33, v26
+; SI-NEXT:    v_or_b32_e32 v19, v36, v19
+; SI-NEXT:    v_or_b32_e32 v20, v43, v20
+; SI-NEXT:    v_or_b32_e32 v21, v35, v21
+; SI-NEXT:    v_or_b32_e32 v22, v42, v22
+; SI-NEXT:    v_or_b32_e32 v23, v34, v23
+; SI-NEXT:    v_or_b32_e32 v24, v41, v24
+; SI-NEXT:    v_or_b32_e32 v25, v33, v25
+; SI-NEXT:    v_or_b32_e32 v26, v40, v26
 ; SI-NEXT:    v_or_b32_e32 v27, v32, v27
 ; SI-NEXT:    v_or_b32_e32 v28, v63, v28
 ; SI-NEXT:    v_or_b32_e32 v29, v62, v29
@@ -23136,17 +23128,17 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:28
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:40
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:36
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:48
 ; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:60
 ; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:72
 ; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:68
@@ -23251,35 +23243,35 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v61
-; SI-NEXT:    v_cvt_f16_f32_e32 v61, v51
-; SI-NEXT:    v_cvt_f16_f32_e32 v51, v53
+; SI-NEXT:    v_cvt_f16_f32_e32 v61, v49
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_cvt_f16_f32_e32 v49, v55
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v62
-; SI-NEXT:    v_cvt_f16_f32_e32 v62, v50
-; SI-NEXT:    v_cvt_f16_f32_e32 v50, v54
+; SI-NEXT:    v_cvt_f16_f32_e32 v62, v51
+; SI-NEXT:    v_cvt_f16_f32_e32 v51, v53
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v63
-; SI-NEXT:    v_cvt_f16_f32_e32 v63, v49
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v55
+; SI-NEXT:    v_cvt_f16_f32_e32 v63, v50
+; SI-NEXT:    v_cvt_f16_f32_e32 v50, v54
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v32
 ; SI-NEXT:    v_cvt_f16_f32_e32 v32, v47
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v33
-; SI-NEXT:    v_cvt_f16_f32_e32 v33, v46
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v34
 ; SI-NEXT:    v_cvt_f16_f32_e32 v34, v45
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v35
 ; SI-NEXT:    v_cvt_f16_f32_e32 v35, v44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v33
+; SI-NEXT:    v_cvt_f16_f32_e32 v33, v46
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v36
@@ -24849,21 +24841,26 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
 ; SI-NEXT:    v_mov_b32_e32 v33, v32
 ; SI-NEXT:    v_or_b32_e32 v10, v32, v10
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_mov_b32_e32 v44, v43
 ; SI-NEXT:    v_or_b32_e32 v13, v43, v13
 ; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v8, v5
 ; SI-NEXT:    v_mov_b32_e32 v57, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v39
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v49
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
@@ -24903,11 +24900,6 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
 ; SI-NEXT:    v_or_b32_e32 v22, v30, v22
 ; SI-NEXT:    v_or_b32_e32 v23, v31, v23
 ; SI-NEXT:    s_mov_b64 s[4:5], 0
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(10)
@@ -25040,9 +25032,9 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v31
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
-; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
@@ -25051,8 +25043,8 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
+; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 ; SI-NEXT:    s_waitcnt vmcnt(13)
@@ -25067,38 +25059,38 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
-; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
+; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
 ; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
-; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
 ; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -29360,167 +29352,167 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:12
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v29
 ; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:4
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:108
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v6
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v8
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:36
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:100
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v14
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:92
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:84
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:76
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v16
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v18
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v20
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:108
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:68
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:52
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v28
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v30
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:88
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v30
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
@@ -29577,111 +29569,111 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v61
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    v_or_b32_e32 v0, v0, v54
-; SI-NEXT:    v_or_b32_e32 v18, v18, v43
+; SI-NEXT:    v_or_b32_e32 v0, v0, v60
+; SI-NEXT:    v_or_b32_e32 v18, v18, v37
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    ; kill: killed $vgpr30
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; SI-NEXT:    v_or_b32_e32 v1, v1, v60
-; SI-NEXT:    v_or_b32_e32 v2, v2, v53
-; SI-NEXT:    v_or_b32_e32 v3, v3, v52
-; SI-NEXT:    v_or_b32_e32 v4, v4, v51
-; SI-NEXT:    v_or_b32_e32 v5, v5, v59
-; SI-NEXT:    v_or_b32_e32 v6, v6, v50
-; SI-NEXT:    v_or_b32_e32 v7, v7, v49
-; SI-NEXT:    v_or_b32_e32 v8, v8, v48
-; SI-NEXT:    v_or_b32_e32 v9, v9, v58
-; SI-NEXT:    v_or_b32_e32 v10, v10, v57
-; SI-NEXT:    v_or_b32_e32 v11, v11, v39
-; SI-NEXT:    v_or_b32_e32 v12, v12, v56
-; SI-NEXT:    v_or_b32_e32 v13, v13, v47
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_or_b32_e32 v1, v1, v54
+; SI-NEXT:    v_or_b32_e32 v2, v2, v59
+; SI-NEXT:    v_or_b32_e32 v3, v3, v58
+; SI-NEXT:    v_or_b32_e32 v4, v4, v53
+; SI-NEXT:    v_or_b32_e32 v5, v5, v57
+; SI-NEXT:    v_or_b32_e32 v6, v6, v52
+; SI-NEXT:    v_or_b32_e32 v7, v7, v56
+; SI-NEXT:    v_or_b32_e32 v8, v8, v51
+; SI-NEXT:    v_or_b32_e32 v9, v9, v50
+; SI-NEXT:    v_or_b32_e32 v10, v10, v49
+; SI-NEXT:    v_or_b32_e32 v11, v11, v48
+; SI-NEXT:    v_or_b32_e32 v12, v12, v47
+; SI-NEXT:    v_or_b32_e32 v13, v13, v39
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v46
-; SI-NEXT:    v_or_b32_e32 v15, v15, v38
-; SI-NEXT:    v_or_b32_e32 v16, v16, v45
+; SI-NEXT:    v_or_b32_e32 v15, v15, v45
+; SI-NEXT:    v_or_b32_e32 v16, v16, v38
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v44
-; SI-NEXT:    v_or_b32_e32 v19, v19, v42
-; SI-NEXT:    v_or_b32_e32 v20, v20, v41
-; SI-NEXT:    v_or_b32_e32 v21, v21, v40
-; SI-NEXT:    v_or_b32_e32 v22, v22, v37
-; SI-NEXT:    v_or_b32_e32 v23, v23, v36
-; SI-NEXT:    v_or_b32_e32 v24, v24, v35
-; SI-NEXT:    v_or_b32_e32 v25, v25, v34
-; SI-NEXT:    v_or_b32_e32 v26, v26, v33
+; SI-NEXT:    v_or_b32_e32 v19, v19, v36
+; SI-NEXT:    v_or_b32_e32 v20, v20, v43
+; SI-NEXT:    v_or_b32_e32 v21, v21, v35
+; SI-NEXT:    v_or_b32_e32 v22, v22, v42
+; SI-NEXT:    v_or_b32_e32 v23, v23, v34
+; SI-NEXT:    v_or_b32_e32 v24, v24, v41
+; SI-NEXT:    v_or_b32_e32 v25, v25, v33
+; SI-NEXT:    v_or_b32_e32 v26, v26, v40
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v32
 ; SI-NEXT:    v_or_b32_e32 v28, v28, v63
 ; SI-NEXT:    v_or_b32_e32 v29, v29, v62
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr59
 ; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
@@ -29689,79 +29681,75 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB42_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v0, v54, v0
+; SI-NEXT:    v_or_b32_e32 v0, v60, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v18, v43, v18
+; SI-NEXT:    v_or_b32_e32 v18, v37, v18
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
@@ -29795,31 +29783,31 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; SI-NEXT:    v_or_b32_e32 v1, v60, v1
-; SI-NEXT:    v_or_b32_e32 v2, v53, v2
-; SI-NEXT:    v_or_b32_e32 v3, v52, v3
-; SI-NEXT:    v_or_b32_e32 v4, v51, v4
-; SI-NEXT:    v_or_b32_e32 v5, v59, v5
-; SI-NEXT:    v_or_b32_e32 v6, v50, v6
-; SI-NEXT:    v_or_b32_e32 v7, v49, v7
-; SI-NEXT:    v_or_b32_e32 v8, v48, v8
-; SI-NEXT:    v_or_b32_e32 v9, v58, v9
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_or_b32_e32 v11, v39, v11
-; SI-NEXT:    v_or_b32_e32 v12, v56, v12
-; SI-NEXT:    v_or_b32_e32 v13, v47, v13
+; SI-NEXT:    v_or_b32_e32 v1, v54, v1
+; SI-NEXT:    v_or_b32_e32 v2, v59, v2
+; SI-NEXT:    v_or_b32_e32 v3, v58, v3
+; SI-NEXT:    v_or_b32_e32 v4, v53, v4
+; SI-NEXT:    v_or_b32_e32 v5, v57, v5
+; SI-NEXT:    v_or_b32_e32 v6, v52, v6
+; SI-NEXT:    v_or_b32_e32 v7, v56, v7
+; SI-NEXT:    v_or_b32_e32 v8, v51, v8
+; SI-NEXT:    v_or_b32_e32 v9, v50, v9
+; SI-NEXT:    v_or_b32_e32 v10, v49, v10
+; SI-NEXT:    v_or_b32_e32 v11, v48, v11
+; SI-NEXT:    v_or_b32_e32 v12, v47, v12
+; SI-NEXT:    v_or_b32_e32 v13, v39, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v46, v14
-; SI-NEXT:    v_or_b32_e32 v15, v38, v15
-; SI-NEXT:    v_or_b32_e32 v16, v45, v16
+; SI-NEXT:    v_or_b32_e32 v15, v45, v15
+; SI-NEXT:    v_or_b32_e32 v16, v38, v16
 ; SI-NEXT:    v_or_b32_e32 v17, v44, v17
-; SI-NEXT:    v_or_b32_e32 v19, v42, v19
-; SI-NEXT:    v_or_b32_e32 v20, v41, v20
-; SI-NEXT:    v_or_b32_e32 v21, v40, v21
-; SI-NEXT:    v_or_b32_e32 v22, v37, v22
-; SI-NEXT:    v_or_b32_e32 v23, v36, v23
-; SI-NEXT:    v_or_b32_e32 v24, v35, v24
-; SI-NEXT:    v_or_b32_e32 v25, v34, v25
-; SI-NEXT:    v_or_b32_e32 v26, v33, v26
+; SI-NEXT:    v_or_b32_e32 v19, v36, v19
+; SI-NEXT:    v_or_b32_e32 v20, v43, v20
+; SI-NEXT:    v_or_b32_e32 v21, v35, v21
+; SI-NEXT:    v_or_b32_e32 v22, v42, v22
+; SI-NEXT:    v_or_b32_e32 v23, v34, v23
+; SI-NEXT:    v_or_b32_e32 v24, v41, v24
+; SI-NEXT:    v_or_b32_e32 v25, v33, v25
+; SI-NEXT:    v_or_b32_e32 v26, v40, v26
 ; SI-NEXT:    v_or_b32_e32 v27, v32, v27
 ; SI-NEXT:    v_or_b32_e32 v28, v63, v28
 ; SI-NEXT:    v_or_b32_e32 v29, v62, v29
@@ -34828,17 +34816,17 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:28
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:40
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:36
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:48
 ; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:60
 ; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:72
 ; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:68
@@ -34943,35 +34931,35 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v61
-; SI-NEXT:    v_cvt_f16_f32_e32 v61, v51
-; SI-NEXT:    v_cvt_f16_f32_e32 v51, v53
+; SI-NEXT:    v_cvt_f16_f32_e32 v61, v49
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_cvt_f16_f32_e32 v49, v55
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v62
-; SI-NEXT:    v_cvt_f16_f32_e32 v62, v50
-; SI-NEXT:    v_cvt_f16_f32_e32 v50, v54
+; SI-NEXT:    v_cvt_f16_f32_e32 v62, v51
+; SI-NEXT:    v_cvt_f16_f32_e32 v51, v53
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v63
-; SI-NEXT:    v_cvt_f16_f32_e32 v63, v49
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v55
+; SI-NEXT:    v_cvt_f16_f32_e32 v63, v50
+; SI-NEXT:    v_cvt_f16_f32_e32 v50, v54
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v32
 ; SI-NEXT:    v_cvt_f16_f32_e32 v32, v47
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v33
-; SI-NEXT:    v_cvt_f16_f32_e32 v33, v46
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v34
 ; SI-NEXT:    v_cvt_f16_f32_e32 v34, v45
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v35
 ; SI-NEXT:    v_cvt_f16_f32_e32 v35, v44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v33
+; SI-NEXT:    v_cvt_f16_f32_e32 v33, v46
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v36
@@ -36541,21 +36529,26 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    v_mov_b32_e32 v33, v32
 ; SI-NEXT:    v_or_b32_e32 v10, v32, v10
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_mov_b32_e32 v44, v43
 ; SI-NEXT:    v_or_b32_e32 v13, v43, v13
 ; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v8, v5
 ; SI-NEXT:    v_mov_b32_e32 v57, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v39
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v49
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
@@ -36595,11 +36588,6 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    v_or_b32_e32 v22, v30, v22
 ; SI-NEXT:    v_or_b32_e32 v23, v31, v23
 ; SI-NEXT:    s_mov_b64 s[4:5], 0
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(10)
@@ -36732,9 +36720,9 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v31
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
-; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
@@ -36743,8 +36731,8 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
+; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 ; SI-NEXT:    s_waitcnt vmcnt(13)
@@ -36759,38 +36747,38 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
-; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
+; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
 ; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
-; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
 ; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -40020,167 +40008,167 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
 ; SI-NEXT:    v_mov_b32_e32 v55, v0
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:12
-; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v5
-; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v9
-; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v13
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v15
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v17
-; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v21
-; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v23
-; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v27
+; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v7
+; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v13
+; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v15
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v17
+; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v19
+; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v21
+; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v23
+; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v27
 ; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v29
 ; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20
 ; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
-; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:84
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:60
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:8
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:4
 ; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v4
+; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v4
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:108
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v6
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v8
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:36
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v10
+; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:100
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v10
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:76
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v14
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:92
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v16
+; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:88
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:84
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v14
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:60
-; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:76
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v16
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v18
+; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v20
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:108
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v22
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:68
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v26
+; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v26
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:52
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v28
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v30
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:88
-; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:44
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v30
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
 ; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.false
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
@@ -40237,111 +40225,111 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v61
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
-; SI-NEXT:    v_or_b32_e32 v0, v0, v54
-; SI-NEXT:    v_or_b32_e32 v18, v18, v43
+; SI-NEXT:    v_or_b32_e32 v0, v0, v60
+; SI-NEXT:    v_or_b32_e32 v18, v18, v37
 ; SI-NEXT:    ; implicit-def: $vgpr55
 ; SI-NEXT:    ; implicit-def: $vgpr61
 ; SI-NEXT:    ; kill: killed $vgpr30
 ; SI-NEXT:    ; implicit-def: $vgpr30
 ; SI-NEXT:    ; kill: killed $vgpr30
-; SI-NEXT:    ; implicit-def: $vgpr54
-; SI-NEXT:    ; implicit-def: $vgpr43
+; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
 ; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
 ; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
 ; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
 ; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
 ; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
 ; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
 ; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
 ; SI-NEXT:    s_waitcnt vmcnt(3)
-; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; SI-NEXT:    v_or_b32_e32 v1, v1, v60
-; SI-NEXT:    v_or_b32_e32 v2, v2, v53
-; SI-NEXT:    v_or_b32_e32 v3, v3, v52
-; SI-NEXT:    v_or_b32_e32 v4, v4, v51
-; SI-NEXT:    v_or_b32_e32 v5, v5, v59
-; SI-NEXT:    v_or_b32_e32 v6, v6, v50
-; SI-NEXT:    v_or_b32_e32 v7, v7, v49
-; SI-NEXT:    v_or_b32_e32 v8, v8, v48
-; SI-NEXT:    v_or_b32_e32 v9, v9, v58
-; SI-NEXT:    v_or_b32_e32 v10, v10, v57
-; SI-NEXT:    v_or_b32_e32 v11, v11, v39
-; SI-NEXT:    v_or_b32_e32 v12, v12, v56
-; SI-NEXT:    v_or_b32_e32 v13, v13, v47
+; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; SI-NEXT:    v_or_b32_e32 v1, v1, v54
+; SI-NEXT:    v_or_b32_e32 v2, v2, v59
+; SI-NEXT:    v_or_b32_e32 v3, v3, v58
+; SI-NEXT:    v_or_b32_e32 v4, v4, v53
+; SI-NEXT:    v_or_b32_e32 v5, v5, v57
+; SI-NEXT:    v_or_b32_e32 v6, v6, v52
+; SI-NEXT:    v_or_b32_e32 v7, v7, v56
+; SI-NEXT:    v_or_b32_e32 v8, v8, v51
+; SI-NEXT:    v_or_b32_e32 v9, v9, v50
+; SI-NEXT:    v_or_b32_e32 v10, v10, v49
+; SI-NEXT:    v_or_b32_e32 v11, v11, v48
+; SI-NEXT:    v_or_b32_e32 v12, v12, v47
+; SI-NEXT:    v_or_b32_e32 v13, v13, v39
 ; SI-NEXT:    v_or_b32_e32 v14, v14, v46
-; SI-NEXT:    v_or_b32_e32 v15, v15, v38
-; SI-NEXT:    v_or_b32_e32 v16, v16, v45
+; SI-NEXT:    v_or_b32_e32 v15, v15, v45
+; SI-NEXT:    v_or_b32_e32 v16, v16, v38
 ; SI-NEXT:    v_or_b32_e32 v17, v17, v44
-; SI-NEXT:    v_or_b32_e32 v19, v19, v42
-; SI-NEXT:    v_or_b32_e32 v20, v20, v41
-; SI-NEXT:    v_or_b32_e32 v21, v21, v40
-; SI-NEXT:    v_or_b32_e32 v22, v22, v37
-; SI-NEXT:    v_or_b32_e32 v23, v23, v36
-; SI-NEXT:    v_or_b32_e32 v24, v24, v35
-; SI-NEXT:    v_or_b32_e32 v25, v25, v34
-; SI-NEXT:    v_or_b32_e32 v26, v26, v33
+; SI-NEXT:    v_or_b32_e32 v19, v19, v36
+; SI-NEXT:    v_or_b32_e32 v20, v20, v43
+; SI-NEXT:    v_or_b32_e32 v21, v21, v35
+; SI-NEXT:    v_or_b32_e32 v22, v22, v42
+; SI-NEXT:    v_or_b32_e32 v23, v23, v34
+; SI-NEXT:    v_or_b32_e32 v24, v24, v41
+; SI-NEXT:    v_or_b32_e32 v25, v25, v33
+; SI-NEXT:    v_or_b32_e32 v26, v26, v40
 ; SI-NEXT:    v_or_b32_e32 v27, v27, v32
 ; SI-NEXT:    v_or_b32_e32 v28, v28, v63
 ; SI-NEXT:    v_or_b32_e32 v29, v29, v62
-; SI-NEXT:    ; implicit-def: $vgpr60
+; SI-NEXT:    ; implicit-def: $vgpr54
+; SI-NEXT:    ; implicit-def: $vgpr59
+; SI-NEXT:    ; implicit-def: $vgpr58
 ; SI-NEXT:    ; implicit-def: $vgpr53
+; SI-NEXT:    ; implicit-def: $vgpr57
 ; SI-NEXT:    ; implicit-def: $vgpr52
+; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr51
-; SI-NEXT:    ; implicit-def: $vgpr59
 ; SI-NEXT:    ; implicit-def: $vgpr50
 ; SI-NEXT:    ; implicit-def: $vgpr49
 ; SI-NEXT:    ; implicit-def: $vgpr48
-; SI-NEXT:    ; implicit-def: $vgpr58
-; SI-NEXT:    ; implicit-def: $vgpr57
-; SI-NEXT:    ; implicit-def: $vgpr39
-; SI-NEXT:    ; implicit-def: $vgpr56
 ; SI-NEXT:    ; implicit-def: $vgpr47
+; SI-NEXT:    ; implicit-def: $vgpr39
 ; SI-NEXT:    ; implicit-def: $vgpr46
-; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr45
+; SI-NEXT:    ; implicit-def: $vgpr38
 ; SI-NEXT:    ; implicit-def: $vgpr44
-; SI-NEXT:    ; implicit-def: $vgpr42
-; SI-NEXT:    ; implicit-def: $vgpr41
-; SI-NEXT:    ; implicit-def: $vgpr40
-; SI-NEXT:    ; implicit-def: $vgpr37
 ; SI-NEXT:    ; implicit-def: $vgpr36
+; SI-NEXT:    ; implicit-def: $vgpr43
 ; SI-NEXT:    ; implicit-def: $vgpr35
+; SI-NEXT:    ; implicit-def: $vgpr42
 ; SI-NEXT:    ; implicit-def: $vgpr34
+; SI-NEXT:    ; implicit-def: $vgpr41
 ; SI-NEXT:    ; implicit-def: $vgpr33
+; SI-NEXT:    ; implicit-def: $vgpr40
 ; SI-NEXT:    ; implicit-def: $vgpr32
 ; SI-NEXT:    ; implicit-def: $vgpr63
 ; SI-NEXT:    ; implicit-def: $vgpr62
@@ -40349,79 +40337,75 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB50_4
 ; SI-NEXT:  ; %bb.3: ; %cmp.true
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v55
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v61
 ; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; SI-NEXT:    v_or_b32_e32 v0, v54, v0
+; SI-NEXT:    v_or_b32_e32 v0, v60, v0
 ; SI-NEXT:    s_mov_b32 s6, 0x30000
-; SI-NEXT:    v_or_b32_e32 v18, v43, v18
+; SI-NEXT:    v_or_b32_e32 v18, v37, v18
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
 ; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
+; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
 ; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
-; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
 ; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
+; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
 ; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
 ; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
+; SI-NEXT:    s_waitcnt vmcnt(13)
 ; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
-; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
-; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    s_waitcnt vmcnt(7)
 ; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
-; SI-NEXT:    s_waitcnt vmcnt(9)
+; SI-NEXT:    s_waitcnt vmcnt(6)
 ; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
-; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(4)
 ; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
-; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
 ; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
-; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
-; SI-NEXT:    s_waitcnt vmcnt(3)
 ; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
-; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
 ; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
 ; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
 ; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
-; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
 ; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
 ; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
@@ -40455,31 +40439,31 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
 ; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
 ; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
 ; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; SI-NEXT:    v_or_b32_e32 v1, v60, v1
-; SI-NEXT:    v_or_b32_e32 v2, v53, v2
-; SI-NEXT:    v_or_b32_e32 v3, v52, v3
-; SI-NEXT:    v_or_b32_e32 v4, v51, v4
-; SI-NEXT:    v_or_b32_e32 v5, v59, v5
-; SI-NEXT:    v_or_b32_e32 v6, v50, v6
-; SI-NEXT:    v_or_b32_e32 v7, v49, v7
-; SI-NEXT:    v_or_b32_e32 v8, v48, v8
-; SI-NEXT:    v_or_b32_e32 v9, v58, v9
-; SI-NEXT:    v_or_b32_e32 v10, v57, v10
-; SI-NEXT:    v_or_b32_e32 v11, v39, v11
-; SI-NEXT:    v_or_b32_e32 v12, v56, v12
-; SI-NEXT:    v_or_b32_e32 v13, v47, v13
+; SI-NEXT:    v_or_b32_e32 v1, v54, v1
+; SI-NEXT:    v_or_b32_e32 v2, v59, v2
+; SI-NEXT:    v_or_b32_e32 v3, v58, v3
+; SI-NEXT:    v_or_b32_e32 v4, v53, v4
+; SI-NEXT:    v_or_b32_e32 v5, v57, v5
+; SI-NEXT:    v_or_b32_e32 v6, v52, v6
+; SI-NEXT:    v_or_b32_e32 v7, v56, v7
+; SI-NEXT:    v_or_b32_e32 v8, v51, v8
+; SI-NEXT:    v_or_b32_e32 v9, v50, v9
+; SI-NEXT:    v_or_b32_e32 v10, v49, v10
+; SI-NEXT:    v_or_b32_e32 v11, v48, v11
+; SI-NEXT:    v_or_b32_e32 v12, v47, v12
+; SI-NEXT:    v_or_b32_e32 v13, v39, v13
 ; SI-NEXT:    v_or_b32_e32 v14, v46, v14
-; SI-NEXT:    v_or_b32_e32 v15, v38, v15
-; SI-NEXT:    v_or_b32_e32 v16, v45, v16
+; SI-NEXT:    v_or_b32_e32 v15, v45, v15
+; SI-NEXT:    v_or_b32_e32 v16, v38, v16
 ; SI-NEXT:    v_or_b32_e32 v17, v44, v17
-; SI-NEXT:    v_or_b32_e32 v19, v42, v19
-; SI-NEXT:    v_or_b32_e32 v20, v41, v20
-; SI-NEXT:    v_or_b32_e32 v21, v40, v21
-; SI-NEXT:    v_or_b32_e32 v22, v37, v22
-; SI-NEXT:    v_or_b32_e32 v23, v36, v23
-; SI-NEXT:    v_or_b32_e32 v24, v35, v24
-; SI-NEXT:    v_or_b32_e32 v25, v34, v25
-; SI-NEXT:    v_or_b32_e32 v26, v33, v26
+; SI-NEXT:    v_or_b32_e32 v19, v36, v19
+; SI-NEXT:    v_or_b32_e32 v20, v43, v20
+; SI-NEXT:    v_or_b32_e32 v21, v35, v21
+; SI-NEXT:    v_or_b32_e32 v22, v42, v22
+; SI-NEXT:    v_or_b32_e32 v23, v34, v23
+; SI-NEXT:    v_or_b32_e32 v24, v41, v24
+; SI-NEXT:    v_or_b32_e32 v25, v33, v25
+; SI-NEXT:    v_or_b32_e32 v26, v40, v26
 ; SI-NEXT:    v_or_b32_e32 v27, v32, v27
 ; SI-NEXT:    v_or_b32_e32 v28, v63, v28
 ; SI-NEXT:    v_or_b32_e32 v29, v62, v29
@@ -45357,17 +45341,17 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:16
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:12
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:20
 ; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:28
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:40
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:36
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:48
 ; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:52
 ; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:60
 ; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:72
 ; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:68
@@ -45472,35 +45456,35 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v61
-; SI-NEXT:    v_cvt_f16_f32_e32 v61, v51
-; SI-NEXT:    v_cvt_f16_f32_e32 v51, v53
+; SI-NEXT:    v_cvt_f16_f32_e32 v61, v49
+; SI-NEXT:    s_waitcnt vmcnt(14)
+; SI-NEXT:    v_cvt_f16_f32_e32 v49, v55
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v62
-; SI-NEXT:    v_cvt_f16_f32_e32 v62, v50
-; SI-NEXT:    v_cvt_f16_f32_e32 v50, v54
+; SI-NEXT:    v_cvt_f16_f32_e32 v62, v51
+; SI-NEXT:    v_cvt_f16_f32_e32 v51, v53
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v63
-; SI-NEXT:    v_cvt_f16_f32_e32 v63, v49
-; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v55
+; SI-NEXT:    v_cvt_f16_f32_e32 v63, v50
+; SI-NEXT:    v_cvt_f16_f32_e32 v50, v54
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v32
 ; SI-NEXT:    v_cvt_f16_f32_e32 v32, v47
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v0, v33
-; SI-NEXT:    v_cvt_f16_f32_e32 v33, v46
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v34
 ; SI-NEXT:    v_cvt_f16_f32_e32 v34, v45
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v35
 ; SI-NEXT:    v_cvt_f16_f32_e32 v35, v44
+; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v0, v33
+; SI-NEXT:    v_cvt_f16_f32_e32 v33, v46
 ; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_cvt_f16_f32_e32 v0, v36
@@ -47070,21 +47054,26 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
 ; SI-NEXT:    v_mov_b32_e32 v33, v32
 ; SI-NEXT:    v_or_b32_e32 v10, v32, v10
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; SI-NEXT:    v_mov_b32_e32 v44, v43
 ; SI-NEXT:    v_or_b32_e32 v13, v43, v13
 ; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; SI-NEXT:    v_or_b32_e32 v5, v8, v5
 ; SI-NEXT:    v_mov_b32_e32 v57, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v39
 ; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; SI-NEXT:    v_or_b32_e32 v6, v7, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v49
 ; SI-NEXT:    v_or_b32_e32 v7, v37, v7
 ; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
@@ -47124,11 +47113,6 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
 ; SI-NEXT:    v_or_b32_e32 v22, v30, v22
 ; SI-NEXT:    v_or_b32_e32 v23, v31, v23
 ; SI-NEXT:    s_mov_b64 s[4:5], 0
-; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(11)
 ; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
 ; SI-NEXT:    s_waitcnt vmcnt(10)
@@ -47261,9 +47245,9 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
 ; SI-NEXT:    v_cvt_f32_f16_e32 v25, v31
 ; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
-; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
@@ -47272,8 +47256,8 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
 ; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
+; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
 ; SI-NEXT:    s_waitcnt vmcnt(13)
@@ -47288,38 +47272,38 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
 ; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt vmcnt(14)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT:    s_waitcnt vmcnt(12)
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
-; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    s_waitcnt vmcnt(11)
+; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    s_waitcnt vmcnt(9)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
 ; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
-; SI-NEXT:    s_waitcnt vmcnt(10)
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
 ; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
 ; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
 ; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
-; SI-NEXT:    s_waitcnt vmcnt(8)
-; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
 ; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
 ; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
-; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT:    s_waitcnt vmcnt(5)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
+; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
-; SI-NEXT:    s_waitcnt vmcnt(5)
-; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
 ; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
 ; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
-; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
-; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
 ; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
 ; SI-NEXT:    s_waitcnt vmcnt(1)
@@ -51934,502 +51918,496 @@ define <60 x i16> @bitcast_v60f16_to_v60i16(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
 ; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
-; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32
-; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4
-; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:8
+; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:120
+; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32
+; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:4
 ; SI-NEXT:    s_waitcnt expcnt(2)
-; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:12
+; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:8
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:20
-; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:24
-; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:28
-; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:32
-; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:36
-; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:40
-; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:44
-; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48
-; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52
-; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:56
-; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:60
-; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:64
-; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:68
-; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:72
+; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:16
+; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:20
+; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:24
+; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
+; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:32
+; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:36
+; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:40
+; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44
+; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:52
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:56
+; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:60
+; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:64
+; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:68
+; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:72
 ; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:76
 ; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:80
 ; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:84
-; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:92
-; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:88
-; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:100
-; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:96
-; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:108
-; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:104
-; SI-NEXT:    v_cvt_f16_f32_e32 v53, v8
-; SI-NEXT:    v_cvt_f16_f32_e32 v8, v9
-; SI-NEXT:    v_cvt_f16_f32_e32 v9, v22
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:88
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT:    v_cvt_f16_f32_e32 v52, v7
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v18
-; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v9, v26
+; SI-NEXT:    v_cvt_f16_f32_e32 v55, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v22
+; SI-NEXT:    v_cvt_f16_f32_e32 v60, v2
 ; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v1, v2
-; SI-NEXT:    v_cvt_f16_f32_e32 v2, v5
-; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v9, v30
-; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v12
-; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; SI-NEXT:    v_cvt_f16_f32_e32 v55, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v18
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:    v_cvt_f16_f32_e32 v40, v4
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v6
-; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v11
-; SI-NEXT:    v_cvt_f16_f32_e32 v6, v13
-; SI-NEXT:    v_cvt_f16_f32_e32 v13, v14
-; SI-NEXT:    v_cvt_f16_f32_e32 v14, v15
+; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(1)
-; SI-NEXT:    v_cvt_f16_f32_e32 v2, v16
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v19
-; SI-NEXT:    v_cvt_f16_f32_e32 v18, v20
-; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
+; SI-NEXT:    v_cvt_f16_f32_e32 v1, v6
+; SI-NEXT:    v_cvt_f16_f32_e32 v53, v8
+; SI-NEXT:    v_cvt_f16_f32_e32 v8, v10
+; SI-NEXT:    v_cvt_f16_f32_e32 v49, v12
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v13
+; SI-NEXT:    v_cvt_f16_f32_e32 v37, v15
+; SI-NEXT:    v_cvt_f16_f32_e32 v52, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v7, v9
+; SI-NEXT:    v_cvt_f16_f32_e32 v48, v11
+; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
+; SI-NEXT:    v_cvt_f16_f32_e32 v38, v16
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
+; SI-NEXT:    v_cvt_f16_f32_e32 v11, v23
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v24
+; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
+; SI-NEXT:    v_cvt_f16_f32_e32 v24, v26
 ; SI-NEXT:    v_cvt_f16_f32_e32 v26, v27
 ; SI-NEXT:    v_cvt_f16_f32_e32 v27, v28
-; SI-NEXT:    v_cvt_f16_f32_e32 v28, v29
+; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
+; SI-NEXT:    v_cvt_f16_f32_e32 v28, v30
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
-; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    s_waitcnt vmcnt(14)
-; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
-; SI-NEXT:    v_cvt_f16_f32_e32 v31, v23
-; SI-NEXT:    v_cvt_f16_f32_e32 v22, v51
-; SI-NEXT:    v_cvt_f16_f32_e32 v23, v43
-; SI-NEXT:    v_cvt_f16_f32_e32 v29, v44
-; SI-NEXT:    v_cvt_f16_f32_e32 v44, v61
-; SI-NEXT:    v_cvt_f16_f32_e32 v19, v62
-; SI-NEXT:    v_cvt_f16_f32_e32 v20, v63
-; SI-NEXT:    v_cvt_f16_f32_e32 v61, v33
-; SI-NEXT:    v_cvt_f16_f32_e32 v43, v36
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v39
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v9, v54
-; SI-NEXT:    v_cvt_f16_f32_e32 v16, v50
-; SI-NEXT:    v_cvt_f16_f32_e32 v51, v41
+; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v51
+; SI-NEXT:    v_cvt_f16_f32_e32 v22, v43
+; SI-NEXT:    v_cvt_f16_f32_e32 v23, v44
+; SI-NEXT:    v_cvt_f16_f32_e32 v61, v61
+; SI-NEXT:    v_cvt_f16_f32_e32 v44, v62
+; SI-NEXT:    v_cvt_f16_f32_e32 v18, v63
+; SI-NEXT:    v_cvt_f16_f32_e32 v19, v33
+; SI-NEXT:    v_cvt_f16_f32_e32 v33, v36
+; SI-NEXT:    v_cvt_f16_f32_e32 v43, v39
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v50
+; SI-NEXT:    v_cvt_f16_f32_e32 v16, v54
+; SI-NEXT:    v_cvt_f16_f32_e32 v54, v41
+; SI-NEXT:    v_cvt_f16_f32_e32 v51, v42
 ; SI-NEXT:    v_cvt_f16_f32_e32 v12, v45
-; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; SI-NEXT:    v_cvt_f16_f32_e32 v30, v47
-; SI-NEXT:    v_cvt_f16_f32_e32 v11, v42
-; SI-NEXT:    v_cvt_f16_f32_e32 v50, v46
-; SI-NEXT:    v_cvt_f16_f32_e32 v54, v56
-; SI-NEXT:    v_cvt_f16_f32_e32 v45, v32
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v9, v57
 ; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f16_f32_e32 v47, v34
-; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:112
-; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:116
-; SI-NEXT:    v_cvt_f16_f32_e32 v36, v58
-; SI-NEXT:    v_cvt_f16_f32_e32 v46, v59
-; SI-NEXT:    s_waitcnt vmcnt(13)
-; SI-NEXT:    v_cvt_f16_f32_e32 v39, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v42, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v13, v46
+; SI-NEXT:    s_waitcnt vmcnt(12)
+; SI-NEXT:    v_cvt_f16_f32_e32 v50, v47
 ; SI-NEXT:    s_waitcnt vmcnt(11)
-; SI-NEXT:    v_cvt_f16_f32_e32 v35, v60
-; SI-NEXT:    v_cvt_f16_f32_e32 v62, v38
+; SI-NEXT:    v_cvt_f16_f32_e32 v30, v56
+; SI-NEXT:    s_waitcnt vmcnt(10)
+; SI-NEXT:    v_cvt_f16_f32_e32 v63, v57
 ; SI-NEXT:    s_waitcnt vmcnt(9)
-; SI-NEXT:    v_cvt_f16_f32_e32 v33, v49
-; SI-NEXT:    v_cvt_f16_f32_e32 v41, v48
+; SI-NEXT:    v_cvt_f16_f32_e32 v62, v58
+; SI-NEXT:    s_waitcnt vmcnt(8)
+; SI-NEXT:    v_cvt_f16_f32_e32 v36, v31
+; SI-NEXT:    s_waitcnt vmcnt(7)
+; SI-NEXT:    v_cvt_f16_f32_e32 v45, v32
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f16_f32_e32 v31, v59
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_cvt_f16_f32_e32 v32, v34
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v35
+; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92
+; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96
+; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:100
+; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:104
+; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:108
+; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:112
+; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:116
+; SI-NEXT:    s_waitcnt vmcnt(6)
+; SI-NEXT:    v_cvt_f16_f32_e32 v42, v3
+; SI-NEXT:    s_waitcnt vmcnt(5)
+; SI-NEXT:    v_cvt_f16_f32_e32 v9, v4
+; SI-NEXT:    s_waitcnt vmcnt(4)
+; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_cvt_f16_f32_e32 v39, v34
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f16_f32_e32 v41, v35
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v46
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v47
 ; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
 ; SI-NEXT:    s_or_saveexec_b64 s[4:5], s[4:5]
-; SI-NEXT:    v_mov_b32_e32 v56, v29
-; SI-NEXT:    v_mov_b32_e32 v29, v28
-; SI-NEXT:    v_mov_b32_e32 v57, v25
-; SI-NEXT:    v_mov_b32_e32 v58, v21
-; SI-NEXT:    v_mov_b32_e32 v59, v17
-; SI-NEXT:    v_mov_b32_e32 v60, v6
-; SI-NEXT:    v_mov_b32_e32 v63, v8
+; SI-NEXT:    v_mov_b32_e32 v47, v21
+; SI-NEXT:    v_mov_b32_e32 v56, v17
+; SI-NEXT:    v_mov_b32_e32 v57, v6
+; SI-NEXT:    v_mov_b32_e32 v58, v7
+; SI-NEXT:    v_mov_b32_e32 v59, v33
 ; SI-NEXT:    s_xor_b64 exec, exec, s[4:5]
 ; SI-NEXT:    s_cbranch_execz .LBB58_2
 ; SI-NEXT:  ; %bb.1: ; %cmp.true
-; SI-NEXT:    v_cvt_f32_f16_e32 v6, v62
-; SI-NEXT:    v_cvt_f32_f16_e32 v34, v34
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
 ; SI-NEXT:    v_cvt_f32_f16_e32 v32, v32
-; SI-NEXT:    v_cvt_f32_f16_e32 v35, v35
-; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
-; SI-NEXT:    v_cvt_f16_f32_e32 v62, v6
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
+; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT:    v_cvt_f32_f16_e32 v7, v62
+; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v6, v63
+; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
+; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
 ; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
-; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
+; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
 ; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
-; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
-; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v34
-; SI-NEXT:    v_or_b32_e32 v32, v32, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v62
-; SI-NEXT:    v_or_b32_e32 v35, v35, v37
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v47
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v46
-; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
-; SI-NEXT:    v_cvt_f32_f16_e32 v8, v54
+; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
+; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
+; SI-NEXT:    v_cvt_f16_f32_e32 v62, v7
+; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; SI-NEXT:    v_or_b32_e32 v3, v3, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v10
+; SI-NEXT:    v_or_b32_e32 v9, v9, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v32
+; SI-NEXT:    v_or_b32_e32 v31, v31, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v62
+; SI-NEXT:    v_or_b32_e32 v63, v6, v34
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
 ; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
 ; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
 ; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
-; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
-; SI-NEXT:    v_cvt_f16_f32_e32 v47, v37
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v38
-; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
 ; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
+; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
 ; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
-; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
-; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
+; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
 ; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
 ; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
-; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
-; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
+; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
+; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
 ; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
 ; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
-; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
-; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
 ; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
 ; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
-; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
-; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
-; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v47
-; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
-; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
+; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
+; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
+; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
 ; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
-; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
 ; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT:    v_or_b32_e32 v46, v37, v38
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v9
-; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
 ; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
 ; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
 ; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
-; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
-; SI-NEXT:    v_or_b32_e32 v54, v8, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v13
 ; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
 ; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
 ; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
-; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
-; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
-; SI-NEXT:    v_or_b32_e32 v11, v11, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v16
+; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
+; SI-NEXT:    v_or_b32_e32 v12, v12, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v16
 ; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
-; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
-; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
-; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
-; SI-NEXT:    v_or_b32_e32 v15, v15, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v20
-; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
-; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
-; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
+; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
+; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
+; SI-NEXT:    v_or_b32_e32 v15, v15, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v19
+; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
+; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
 ; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
-; SI-NEXT:    v_or_b32_e32 v19, v19, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v23
-; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
+; SI-NEXT:    v_or_b32_e32 v18, v18, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v23
 ; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
-; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
-; SI-NEXT:    v_or_b32_e32 v22, v22, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v27
-; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
-; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
-; SI-NEXT:    v_or_b32_e32 v26, v26, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v24
-; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
-; SI-NEXT:    v_or_b32_e32 v31, v31, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v18
-; SI-NEXT:    v_or_b32_e32 v7, v7, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v2
-; SI-NEXT:    v_or_b32_e32 v14, v14, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v5
-; SI-NEXT:    v_or_b32_e32 v4, v4, v37
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v53
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v52
+; SI-NEXT:    v_or_b32_e32 v22, v22, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v27
+; SI-NEXT:    v_or_b32_e32 v26, v26, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v5
+; SI-NEXT:    v_or_b32_e32 v11, v11, v34
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v20
+; SI-NEXT:    v_or_b32_e32 v2, v2, v34
+; SI-NEXT:    v_cvt_f32_f16_e32 v34, v38
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v37
 ; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
-; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v53, v37
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v38
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
+; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
+; SI-NEXT:    v_cvt_f16_f32_e32 v38, v34
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
 ; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v53
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v38
 ; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT:    v_or_b32_e32 v52, v37, v38
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v40
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v55
+; SI-NEXT:    v_or_b32_e32 v37, v34, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v34, v49
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v48
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v40, v37
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v38
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v40
-; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
-; SI-NEXT:    v_or_b32_e32 v55, v37, v38
+; SI-NEXT:    v_cvt_f32_f16_e32 v46, v58
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
+; SI-NEXT:    v_cvt_f16_f32_e32 v49, v34
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
+; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v49
+; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
+; SI-NEXT:    v_or_b32_e32 v48, v34, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v34, v53
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v52
+; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
+; SI-NEXT:    v_cvt_f16_f32_e32 v53, v34
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
+; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v53
+; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
+; SI-NEXT:    v_or_b32_e32 v52, v34, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v34, v40
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v55
+; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
+; SI-NEXT:    v_cvt_f32_f16_e32 v42, v42
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
+; SI-NEXT:    v_cvt_f16_f32_e32 v40, v34
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
+; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v40
+; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
+; SI-NEXT:    v_or_b32_e32 v55, v34, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v34, v60
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v6
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v63
-; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
-; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
-; SI-NEXT:    v_cvt_f32_f16_e32 v48, v43
-; SI-NEXT:    v_or_b32_e32 v6, v37, v1
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v6
+; SI-NEXT:    v_cvt_f32_f16_e32 v43, v43
+; SI-NEXT:    v_cvt_f32_f16_e32 v51, v51
+; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
+; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
+; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
+; SI-NEXT:    v_cvt_f32_f16_e32 v41, v41
+; SI-NEXT:    v_or_b32_e32 v6, v35, v34
 ; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
-; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
-; SI-NEXT:    v_cvt_f32_f16_e32 v49, v61
-; SI-NEXT:    v_add_f32_e32 v48, 0x38000000, v48
-; SI-NEXT:    v_cvt_f16_f32_e32 v48, v48
-; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v49
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
-; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
-; SI-NEXT:    v_cvt_f32_f16_e32 v41, v41
-; SI-NEXT:    v_or_b32_e32 v61, v49, v48
+; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
 ; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
+; SI-NEXT:    v_cvt_f32_f16_e32 v54, v54
 ; SI-NEXT:    v_cvt_f32_f16_e32 v50, v50
 ; SI-NEXT:    v_cvt_f32_f16_e32 v36, v36
 ; SI-NEXT:    v_cvt_f32_f16_e32 v39, v39
-; SI-NEXT:    v_cvt_f32_f16_e32 v33, v33
+; SI-NEXT:    v_add_f32_e32 v42, 0x38000000, v42
+; SI-NEXT:    v_cvt_f16_f32_e32 v42, v42
+; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
+; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
+; SI-NEXT:    v_add_f32_e32 v43, 0x38000000, v43
+; SI-NEXT:    v_add_f32_e32 v51, 0x38000000, v51
 ; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
 ; SI-NEXT:    v_add_f32_e32 v41, 0x38000000, v41
+; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
+; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
+; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
 ; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
+; SI-NEXT:    v_cvt_f16_f32_e32 v43, v43
+; SI-NEXT:    v_cvt_f16_f32_e32 v51, v51
+; SI-NEXT:    v_add_f32_e32 v54, 0x38000000, v54
 ; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
 ; SI-NEXT:    v_add_f32_e32 v50, 0x38000000, v50
 ; SI-NEXT:    v_add_f32_e32 v36, 0x38000000, v36
-; SI-NEXT:    v_add_f32_e32 v39, 0x38000000, v39
 ; SI-NEXT:    v_cvt_f16_f32_e32 v41, v41
-; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
+; SI-NEXT:    v_add_f32_e32 v39, 0x38000000, v39
+; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
 ; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
+; SI-NEXT:    v_cvt_f16_f32_e32 v54, v54
 ; SI-NEXT:    v_cvt_f16_f32_e32 v50, v50
 ; SI-NEXT:    v_cvt_f16_f32_e32 v36, v36
 ; SI-NEXT:    v_cvt_f16_f32_e32 v39, v39
-; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
+; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v42
+; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
+; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v43
+; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
 ; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
 ; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v41
+; SI-NEXT:    v_or_b32_e32 v25, v25, v24
+; SI-NEXT:    v_or_b32_e32 v29, v29, v28
+; SI-NEXT:    v_or_b32_e32 v54, v54, v51
 ; SI-NEXT:    v_or_b32_e32 v50, v50, v30
-; SI-NEXT:    v_or_b32_e32 v33, v33, v41
-; SI-NEXT:    v_alignbit_b32 v1, v55, v1, 16
-; SI-NEXT:    v_alignbit_b32 v43, v15, v48, 16
-; SI-NEXT:    v_alignbit_b32 v30, v54, v30, 16
-; SI-NEXT:    v_alignbit_b32 v41, v32, v41, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v6
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_or_b32_e32 v6, v37, v3
+; SI-NEXT:    v_or_b32_e32 v39, v39, v41
+; SI-NEXT:    v_alignbit_b32 v60, v55, v34, 16
+; SI-NEXT:    v_alignbit_b32 v24, v26, v24, 16
+; SI-NEXT:    v_alignbit_b32 v28, v22, v28, 16
+; SI-NEXT:    v_alignbit_b32 v51, v12, v51, 16
+; SI-NEXT:    v_alignbit_b32 v30, v63, v30, 16
+; SI-NEXT:    v_alignbit_b32 v41, v3, v41, 16
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v6
+; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
+; SI-NEXT:    v_or_b32_e32 v6, v35, v1
 ; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v38
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v60
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_alignbit_b32 v3, v52, v3, 16
-; SI-NEXT:    v_add_f32_e32 v38, 0x38000000, v38
-; SI-NEXT:    v_or_b32_e32 v63, v37, v10
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v59
-; SI-NEXT:    v_cvt_f16_f32_e32 v38, v38
-; SI-NEXT:    v_alignbit_b32 v10, v4, v10, 16
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
-; SI-NEXT:    v_or_b32_e32 v60, v38, v13
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v58
-; SI-NEXT:    v_alignbit_b32 v13, v14, v13, 16
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v46
+; SI-NEXT:    v_cvt_f32_f16_e32 v46, v57
+; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
+; SI-NEXT:    v_alignbit_b32 v1, v52, v1, 16
+; SI-NEXT:    v_add_f32_e32 v46, 0x38000000, v46
+; SI-NEXT:    v_or_b32_e32 v58, v35, v8
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v56
+; SI-NEXT:    v_cvt_f16_f32_e32 v46, v46
+; SI-NEXT:    v_alignbit_b32 v8, v48, v8, 16
+; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
+; SI-NEXT:    v_or_b32_e32 v57, v46, v14
+; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v46, v47
+; SI-NEXT:    v_alignbit_b32 v14, v37, v14, 16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v17, v6
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
 ; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
 ; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; SI-NEXT:    v_or_b32_e32 v59, v37, v17
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v38
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v57
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_add_f32_e32 v38, 0x38000000, v38
-; SI-NEXT:    v_cvt_f16_f32_e32 v38, v38
+; SI-NEXT:    v_or_b32_e32 v56, v35, v17
+; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v46
+; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v46, v59
+; SI-NEXT:    v_add_f32_e32 v46, 0x38000000, v46
+; SI-NEXT:    v_cvt_f16_f32_e32 v46, v46
+; SI-NEXT:    v_or_b32_e32 v59, v46, v43
+; SI-NEXT:    v_alignbit_b32 v43, v15, v43, 16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_cvt_f32_f16_e32 v21, v6
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
 ; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
 ; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; SI-NEXT:    v_or_b32_e32 v58, v37, v21
-; SI-NEXT:    v_cvt_f32_f16_e32 v37, v44
-; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
-; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
-; SI-NEXT:    v_alignbit_b32 v44, v19, v37, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v25, v6
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
-; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
-; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
-; SI-NEXT:    v_or_b32_e32 v57, v38, v25
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v56
-; SI-NEXT:    v_add_f32_e32 v38, 0x38000000, v38
-; SI-NEXT:    v_cvt_f16_f32_e32 v38, v38
-; SI-NEXT:    v_or_b32_e32 v56, v38, v37
-; SI-NEXT:    v_cvt_f32_f16_e32 v38, v51
-; SI-NEXT:    v_cvt_f32_f16_e32 v51, v42
-; SI-NEXT:    v_add_f32_e32 v38, 0x38000000, v38
-; SI-NEXT:    v_cvt_f16_f32_e32 v38, v38
-; SI-NEXT:    v_add_f32_e32 v51, 0x38000000, v51
-; SI-NEXT:    v_cvt_f16_f32_e32 v51, v51
-; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
-; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v51
-; SI-NEXT:    v_or_b32_e32 v39, v39, v42
-; SI-NEXT:    v_alignbit_b32 v51, v11, v38, 16
-; SI-NEXT:    v_alignbit_b32 v42, v35, v42, 16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v28, v6
-; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
-; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
-; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
-; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; SI-NEXT:    v_or_b32_e32 v29, v29, v28
+; SI-NEXT:    v_or_b32_e32 v47, v35, v21
+; SI-NEXT:    v_cvt_f32_f16_e32 v35, v44
+; SI-NEXT:    v_cvt_f32_f16_e32 v44, v61
+; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
+; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
+; SI-NEXT:    v_add_f32_e32 v44, 0x38000000, v44
+; SI-NEXT:    v_cvt_f16_f32_e32 v44, v44
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
+; SI-NEXT:    v_or_b32_e32 v61, v44, v35
+; SI-NEXT:    v_cvt_f32_f16_e32 v44, v45
+; SI-NEXT:    v_add_f32_e32 v44, 0x38000000, v44
+; SI-NEXT:    v_cvt_f16_f32_e32 v44, v44
+; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v44
+; SI-NEXT:    v_or_b32_e32 v36, v36, v45
+; SI-NEXT:    v_alignbit_b32 v44, v18, v35, 16
+; SI-NEXT:    v_alignbit_b32 v45, v31, v45, 16
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_cvt_f32_f16_e32 v49, v6
-; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v49
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
-; SI-NEXT:    v_or_b32_e32 v6, v49, v38
-; SI-NEXT:    v_cvt_f32_f16_e32 v49, v45
+; SI-NEXT:    v_cvt_f32_f16_e32 v33, v6
+; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
+; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
+; SI-NEXT:    v_or_b32_e32 v6, v33, v42
 ; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v6, v7, v17, 16
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v49
-; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v6, v31, v21, 16
-; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v6, v26, v25, 16
-; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
-; SI-NEXT:    v_or_b32_e32 v36, v36, v49
+; SI-NEXT:    v_alignbit_b32 v6, v2, v17, 16
 ; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_alignbit_b32 v6, v22, v28, 16
-; SI-NEXT:    v_alignbit_b32 v45, v46, v49, 16
+; SI-NEXT:    v_alignbit_b32 v6, v11, v21, 16
+; SI-NEXT:    v_alignbit_b32 v42, v9, v42, 16
 ; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
 ; SI-NEXT:  .LBB58_2: ; %end
 ; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v60
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v37, 0xffff, v6
-; SI-NEXT:    v_or_b32_e32 v1, v37, v1
-; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v34, 0xffff, v6
+; SI-NEXT:    v_or_b32_e32 v34, v34, v35
+; SI-NEXT:    buffer_store_dword v34, v0, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v55
-; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v40
-; SI-NEXT:    v_or_b32_e32 v1, v1, v37
-; SI-NEXT:    v_add_i32_e32 v37, vcc, 4, v0
-; SI-NEXT:    buffer_store_dword v1, v37, s[0:3], 0 offen
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; SI-NEXT:    v_or_b32_e32 v1, v1, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 8, v0
-; SI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v34, 0xffff, v55
+; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v40
+; SI-NEXT:    v_or_b32_e32 v34, v34, v35
+; SI-NEXT:    v_add_i32_e32 v35, vcc, 4, v0
+; SI-NEXT:    buffer_store_dword v34, v35, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v34, 0xffff, v6
+; SI-NEXT:    v_or_b32_e32 v1, v34, v1
+; SI-NEXT:    v_add_i32_e32 v34, vcc, 8, v0
+; SI-NEXT:    buffer_store_dword v1, v34, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v52
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v53
-; SI-NEXT:    v_or_b32_e32 v1, v1, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 12, v0
-; SI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v53
+; SI-NEXT:    v_or_b32_e32 v1, v1, v34
+; SI-NEXT:    v_add_i32_e32 v34, vcc, 12, v0
+; SI-NEXT:    buffer_store_dword v1, v34, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v63
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v10
-; SI-NEXT:    v_or_b32_e32 v1, v1, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 16, v0
-; SI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
-; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
-; SI-NEXT:    v_or_b32_e32 v1, v1, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 20, v0
-; SI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v58
+; SI-NEXT:    v_or_b32_e32 v1, v1, v8
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 16, v0
+; SI-NEXT:    buffer_store_dword v1, v8, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v60
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v13
-; SI-NEXT:    v_or_b32_e32 v1, v1, v3
-; SI-NEXT:    v_add_i32_e32 v3, vcc, 24, v0
-; SI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v48
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v49
+; SI-NEXT:    v_or_b32_e32 v1, v1, v8
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 20, v0
+; SI-NEXT:    buffer_store_dword v1, v8, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14
-; SI-NEXT:    v_or_b32_e32 v1, v1, v2
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
-; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v57
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v14
+; SI-NEXT:    v_or_b32_e32 v1, v1, v8
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 24, v0
+; SI-NEXT:    buffer_store_dword v1, v8, s[0:3], 0 offen
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v37
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v38
+; SI-NEXT:    v_or_b32_e32 v1, v1, v8
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 28, v0
+; SI-NEXT:    buffer_store_dword v1, v8, s[0:3], 0 offen
+; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v59
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v56
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT:    v_or_b32_e32 v1, v1, v2
-; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
-; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
+; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
+; SI-NEXT:    v_or_b32_e32 v1, v1, v8
+; SI-NEXT:    v_add_i32_e32 v8, vcc, 32, v0
+; SI-NEXT:    buffer_store_dword v1, v8, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v7
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v18
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v20
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
+; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v58
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v47
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v31
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v11
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v57
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v25
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v24
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
@@ -52439,11 +52417,9 @@ define <60 x i16> @bitcast_v60f16_to_v60i16(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v29
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v28
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
@@ -52454,19 +52430,19 @@ define <60 x i16> @bitcast_v60f16_to_v60i16(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v56
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v44
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v19
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v20
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v18
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v19
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v59
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v43
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
@@ -52478,16 +52454,14 @@ define <60 x i16> @bitcast_v60f16_to_v60i16(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v54
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v51
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v11
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v12
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v12
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v13
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
@@ -52498,8 +52472,8 @@ define <60 x i16> @bitcast_v60f16_to_v60i16(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v54
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v9
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v63
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v62
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
@@ -52510,32 +52484,34 @@ define <60 x i16> @bitcast_v60f16_to_v60i16(<60 x half> %a, i32 %b) {
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v46
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v47
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v31
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v32
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v39
+; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v42
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v62
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v9
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v10
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v33
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v39
 ; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v41
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
 ; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
 ; SI-NEXT:    s_waitcnt expcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v32
-; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v34
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x74, v0
 ; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
diff --git a/llvm/test/CodeGen/AMDGPU/bf16.ll b/llvm/test/CodeGen/AMDGPU/bf16.ll
index 516a14edbc260..2bdf994496421 100644
--- a/llvm/test/CodeGen/AMDGPU/bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/bf16.ll
@@ -5633,12 +5633,11 @@ define { <32 x i32>, bfloat } @test_overflow_stack(bfloat %a, <32 x i32> %b) {
 ; GFX9-NEXT:    buffer_store_dword v28, v0, s[0:3], 0 offen offset:104
 ; GFX9-NEXT:    buffer_store_dword v27, v0, s[0:3], 0 offen offset:100
 ; GFX9-NEXT:    buffer_store_dword v26, v0, s[0:3], 0 offen offset:96
-; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:92
-; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    s_nop 0
+; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dword v24, v0, s[0:3], 0 offen offset:88
 ; GFX9-NEXT:    buffer_store_dword v23, v0, s[0:3], 0 offen offset:84
@@ -5664,10 +5663,11 @@ define { <32 x i32>, bfloat } @test_overflow_stack(bfloat %a, <32 x i32> %b) {
 ; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:4
 ; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen
 ; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    buffer_store_dword v27, v0, s[0:3], 0 offen offset:124
+; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:124
+; GFX9-NEXT:    s_waitcnt vmcnt(25)
 ; GFX9-NEXT:    buffer_store_dword v26, v0, s[0:3], 0 offen offset:120
 ; GFX9-NEXT:    s_waitcnt vmcnt(25)
-; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:116
+; GFX9-NEXT:    buffer_store_dword v27, v0, s[0:3], 0 offen offset:116
 ; GFX9-NEXT:    buffer_store_short v1, v0, s[0:3], 0 offen offset:128
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
@@ -8479,30 +8479,30 @@ define <32 x double> @global_extload_v32bf16_to_v32f64(ptr addrspace(1) %ptr) {
 ; GFX9-LABEL: global_extload_v32bf16_to_v32f64:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_load_ushort v8, v[1:2], off offset:62
-; GFX9-NEXT:    global_load_ushort v10, v[1:2], off offset:60
-; GFX9-NEXT:    global_load_ushort v11, v[1:2], off offset:58
-; GFX9-NEXT:    global_load_ushort v12, v[1:2], off offset:56
-; GFX9-NEXT:    global_load_ushort v13, v[1:2], off offset:54
-; GFX9-NEXT:    global_load_ushort v14, v[1:2], off offset:52
-; GFX9-NEXT:    global_load_ushort v15, v[1:2], off offset:50
-; GFX9-NEXT:    global_load_ushort v16, v[1:2], off offset:48
-; GFX9-NEXT:    global_load_ushort v17, v[1:2], off offset:46
-; GFX9-NEXT:    global_load_ushort v18, v[1:2], off offset:44
-; GFX9-NEXT:    global_load_ushort v19, v[1:2], off offset:42
-; GFX9-NEXT:    global_load_ushort v20, v[1:2], off offset:40
-; GFX9-NEXT:    global_load_ushort v21, v[1:2], off offset:38
-; GFX9-NEXT:    global_load_ushort v22, v[1:2], off offset:36
-; GFX9-NEXT:    global_load_ushort v23, v[1:2], off offset:34
-; GFX9-NEXT:    global_load_ushort v24, v[1:2], off offset:32
-; GFX9-NEXT:    global_load_ushort v25, v[1:2], off
-; GFX9-NEXT:    global_load_ushort v26, v[1:2], off offset:2
-; GFX9-NEXT:    global_load_ushort v27, v[1:2], off offset:30
+; GFX9-NEXT:    global_load_ushort v9, v[1:2], off offset:62
+; GFX9-NEXT:    global_load_ushort v11, v[1:2], off offset:60
+; GFX9-NEXT:    global_load_ushort v12, v[1:2], off offset:58
+; GFX9-NEXT:    global_load_ushort v13, v[1:2], off offset:56
+; GFX9-NEXT:    global_load_ushort v14, v[1:2], off offset:54
+; GFX9-NEXT:    global_load_ushort v15, v[1:2], off offset:52
+; GFX9-NEXT:    global_load_ushort v16, v[1:2], off offset:50
+; GFX9-NEXT:    global_load_ushort v17, v[1:2], off offset:48
+; GFX9-NEXT:    global_load_ushort v18, v[1:2], off offset:46
+; GFX9-NEXT:    global_load_ushort v19, v[1:2], off offset:44
+; GFX9-NEXT:    global_load_ushort v20, v[1:2], off offset:42
+; GFX9-NEXT:    global_load_ushort v21, v[1:2], off offset:40
+; GFX9-NEXT:    global_load_ushort v22, v[1:2], off offset:38
+; GFX9-NEXT:    global_load_ushort v23, v[1:2], off offset:36
+; GFX9-NEXT:    global_load_ushort v24, v[1:2], off offset:34
+; GFX9-NEXT:    global_load_ushort v25, v[1:2], off offset:32
+; GFX9-NEXT:    global_load_ushort v26, v[1:2], off
+; GFX9-NEXT:    global_load_ushort v27, v[1:2], off offset:2
 ; GFX9-NEXT:    global_load_ushort v3, v[1:2], off offset:16
 ; GFX9-NEXT:    global_load_ushort v4, v[1:2], off offset:18
 ; GFX9-NEXT:    global_load_ushort v5, v[1:2], off offset:20
 ; GFX9-NEXT:    global_load_ushort v6, v[1:2], off offset:22
-; GFX9-NEXT:    global_load_ushort v28, v[1:2], off offset:24
+; GFX9-NEXT:    global_load_ushort v8, v[1:2], off offset:24
+; GFX9-NEXT:    global_load_ushort v28, v[1:2], off offset:30
 ; GFX9-NEXT:    global_load_ushort v29, v[1:2], off offset:26
 ; GFX9-NEXT:    global_load_ushort v30, v[1:2], off offset:28
 ; GFX9-NEXT:    global_load_ushort v31, v[1:2], off offset:4
@@ -8513,122 +8513,122 @@ define <32 x double> @global_extload_v32bf16_to_v32f64(ptr addrspace(1) %ptr) {
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    global_load_ushort v1, v[1:2], off offset:14
 ; GFX9-NEXT:    s_waitcnt vmcnt(31)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v8
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v2
+; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v9
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(30)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v10
-; GFX9-NEXT:    s_waitcnt vmcnt(28)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v12
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:252
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:248
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v2
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v11
-; GFX9-NEXT:    s_waitcnt vmcnt(29)
+; GFX9-NEXT:    s_waitcnt vmcnt(28)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v11, 16, v13
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:244
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:240
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v2
-; GFX9-NEXT:    s_waitcnt vmcnt(30)
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:252
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:248
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v2
+; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v12
+; GFX9-NEXT:    s_waitcnt vmcnt(29)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v14
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:236
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:232
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v10
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[10:11], v11
-; GFX9-NEXT:    s_waitcnt vmcnt(31)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v15
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:244
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:240
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(30)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v16
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:228
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:224
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v12
+; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v15
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:236
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:232
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v11
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[11:12], v12
 ; GFX9-NEXT:    s_waitcnt vmcnt(31)
+; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v16
+; GFX9-NEXT:    s_waitcnt vmcnt(30)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v17
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[12:13], v13
-; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:220
-; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:216
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[10:11], v14
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[14:15], v15
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:228
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:224
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v13
+; GFX9-NEXT:    s_waitcnt vmcnt(31)
+; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v18
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[13:14], v14
+; GFX9-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:220
+; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:216
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[11:12], v15
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[15:16], v16
 ; GFX9-NEXT:    s_waitcnt vmcnt(32)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v18
+; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v19
 ; GFX9-NEXT:    s_waitcnt vmcnt(30)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v20
+; GFX9-NEXT:    v_lshlrev_b32_e32 v19, 16, v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(28)
+; GFX9-NEXT:    v_lshlrev_b32_e32 v21, 16, v23
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:212
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:208
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v2
+; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v20
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v20, 16, v22
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:212
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:208
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v2
-; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v19
-; GFX9-NEXT:    v_lshlrev_b32_e32 v19, 16, v21
-; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:204
-; GFX9-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:200
-; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:196
-; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:192
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[10:11], v20
+; GFX9-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:204
+; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:200
+; GFX9-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:196
+; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:192
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[11:12], v21
 ; GFX9-NEXT:    s_waitcnt vmcnt(33)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v23
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[16:17], v16
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[12:13], v18
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[18:19], v19
-; GFX9-NEXT:    buffer_store_dword v15, v0, s[0:3], 0 offen offset:188
-; GFX9-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:184
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:180
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:176
-; GFX9-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:172
-; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:168
-; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:164
-; GFX9-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:160
-; GFX9-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:156
-; GFX9-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:152
-; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:148
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v2
-; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:144
+; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v24
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[17:18], v17
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[13:14], v19
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[19:20], v20
+; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:188
+; GFX9-NEXT:    buffer_store_dword v15, v0, s[0:3], 0 offen offset:184
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:180
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:176
+; GFX9-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:172
+; GFX9-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:168
+; GFX9-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:164
+; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:160
+; GFX9-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen offset:156
+; GFX9-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:152
+; GFX9-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:148
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v2
+; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:144
 ; GFX9-NEXT:    s_waitcnt vmcnt(44)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v24
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:140
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:136
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v10
-; GFX9-NEXT:    s_waitcnt vmcnt(43)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v27
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:132
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:128
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v12
+; GFX9-NEXT:    v_lshlrev_b32_e32 v11, 16, v25
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:140
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:136
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v11
 ; GFX9-NEXT:    s_waitcnt vmcnt(38)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v30
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:124
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:120
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v14
-; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v29
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:116
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:112
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v16
-; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v25
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[10:11], v2
+; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v28
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:132
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:128
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v13
+; GFX9-NEXT:    s_waitcnt vmcnt(38)
+; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v30
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:124
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:120
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v15
+; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v29
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:116
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:112
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[9:10], v17
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v26
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[12:13], v2
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[11:12], v2
+; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v27
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[13:14], v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(41)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v31
-; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v28
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[14:15], v2
+; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[15:16], v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(40)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v32
-; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:108
-; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:104
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v18
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[16:17], v2
+; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:108
+; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:104
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v8
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[17:18], v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(41)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v33
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[18:19], v2
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[19:20], v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(40)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v34
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[20:21], v2
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[21:22], v2
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
 ; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:100
 ; GFX9-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:96
 ; GFX9-NEXT:    v_cvt_f64_f32_e32 v[8:9], v6
 ; GFX9-NEXT:    v_cvt_f64_f32_e32 v[5:6], v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(41)
-; GFX9-NEXT:    v_lshlrev_b32_e32 v22, 16, v7
+; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v7
 ; GFX9-NEXT:    s_waitcnt vmcnt(40)
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
@@ -8642,25 +8642,25 @@ define <32 x double> @global_extload_v32bf16_to_v32f64(ptr addrspace(1) %ptr) {
 ; GFX9-NEXT:    v_cvt_f64_f32_e32 v[6:7], v3
 ; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:76
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
-; GFX9-NEXT:    v_cvt_f64_f32_e32 v[1:2], v22
+; GFX9-NEXT:    v_cvt_f64_f32_e32 v[1:2], v10
 ; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:68
 ; GFX9-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:64
 ; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:60
 ; GFX9-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:56
 ; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:52
 ; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
-; GFX9-NEXT:    buffer_store_dword v21, v0, s[0:3], 0 offen offset:44
-; GFX9-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen offset:40
-; GFX9-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:36
-; GFX9-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:32
-; GFX9-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:28
-; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:24
-; GFX9-NEXT:    buffer_store_dword v15, v0, s[0:3], 0 offen offset:20
-; GFX9-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:16
-; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:12
-; GFX9-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:8
-; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:4
-; GFX9-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen
+; GFX9-NEXT:    buffer_store_dword v22, v0, s[0:3], 0 offen offset:44
+; GFX9-NEXT:    buffer_store_dword v21, v0, s[0:3], 0 offen offset:40
+; GFX9-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen offset:36
+; GFX9-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:32
+; GFX9-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:28
+; GFX9-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:24
+; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:20
+; GFX9-NEXT:    buffer_store_dword v15, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:12
+; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:8
+; GFX9-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:4
+; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -8861,72 +8861,72 @@ define <32 x double> @global_extload_v32bf16_to_v32f64(ptr addrspace(1) %ptr) {
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_clause 0x1f
-; GFX11-NEXT:    global_load_u16 v3, v[1:2], off offset:12
-; GFX11-NEXT:    global_load_u16 v4, v[1:2], off offset:8
-; GFX11-NEXT:    global_load_u16 v5, v[1:2], off offset:4
-; GFX11-NEXT:    global_load_u16 v6, v[1:2], off offset:2
+; GFX11-NEXT:    global_load_u16 v3, v[1:2], off offset:2
+; GFX11-NEXT:    global_load_u16 v4, v[1:2], off offset:12
+; GFX11-NEXT:    global_load_u16 v5, v[1:2], off offset:8
+; GFX11-NEXT:    global_load_u16 v6, v[1:2], off offset:4
 ; GFX11-NEXT:    global_load_u16 v7, v[1:2], off
 ; GFX11-NEXT:    global_load_u16 v8, v[1:2], off offset:6
 ; GFX11-NEXT:    global_load_u16 v9, v[1:2], off offset:10
 ; GFX11-NEXT:    global_load_u16 v10, v[1:2], off offset:14
-; GFX11-NEXT:    global_load_u16 v11, v[1:2], off offset:28
-; GFX11-NEXT:    global_load_u16 v12, v[1:2], off offset:24
-; GFX11-NEXT:    global_load_u16 v13, v[1:2], off offset:20
-; GFX11-NEXT:    global_load_u16 v14, v[1:2], off offset:18
+; GFX11-NEXT:    global_load_u16 v11, v[1:2], off offset:18
+; GFX11-NEXT:    global_load_u16 v12, v[1:2], off offset:28
+; GFX11-NEXT:    global_load_u16 v13, v[1:2], off offset:24
+; GFX11-NEXT:    global_load_u16 v14, v[1:2], off offset:20
 ; GFX11-NEXT:    global_load_u16 v15, v[1:2], off offset:16
 ; GFX11-NEXT:    global_load_u16 v16, v[1:2], off offset:22
 ; GFX11-NEXT:    global_load_u16 v17, v[1:2], off offset:26
 ; GFX11-NEXT:    global_load_u16 v18, v[1:2], off offset:30
-; GFX11-NEXT:    global_load_u16 v19, v[1:2], off offset:44
-; GFX11-NEXT:    global_load_u16 v20, v[1:2], off offset:40
-; GFX11-NEXT:    global_load_u16 v21, v[1:2], off offset:36
-; GFX11-NEXT:    global_load_u16 v22, v[1:2], off offset:34
+; GFX11-NEXT:    global_load_u16 v19, v[1:2], off offset:34
+; GFX11-NEXT:    global_load_u16 v20, v[1:2], off offset:44
+; GFX11-NEXT:    global_load_u16 v21, v[1:2], off offset:40
+; GFX11-NEXT:    global_load_u16 v22, v[1:2], off offset:36
 ; GFX11-NEXT:    global_load_u16 v23, v[1:2], off offset:32
 ; GFX11-NEXT:    global_load_u16 v24, v[1:2], off offset:38
 ; GFX11-NEXT:    global_load_u16 v25, v[1:2], off offset:42
 ; GFX11-NEXT:    global_load_u16 v26, v[1:2], off offset:46
-; GFX11-NEXT:    global_load_u16 v27, v[1:2], off offset:60
-; GFX11-NEXT:    global_load_u16 v28, v[1:2], off offset:56
-; GFX11-NEXT:    global_load_u16 v29, v[1:2], off offset:52
-; GFX11-NEXT:    global_load_u16 v30, v[1:2], off offset:50
+; GFX11-NEXT:    global_load_u16 v27, v[1:2], off offset:50
+; GFX11-NEXT:    global_load_u16 v28, v[1:2], off offset:60
+; GFX11-NEXT:    global_load_u16 v29, v[1:2], off offset:56
+; GFX11-NEXT:    global_load_u16 v30, v[1:2], off offset:52
 ; GFX11-NEXT:    global_load_u16 v31, v[1:2], off offset:48
 ; GFX11-NEXT:    global_load_u16 v32, v[1:2], off offset:54
 ; GFX11-NEXT:    global_load_u16 v33, v[1:2], off offset:58
 ; GFX11-NEXT:    global_load_u16 v1, v[1:2], off offset:62
 ; GFX11-NEXT:    s_waitcnt vmcnt(31)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
 ; GFX11-NEXT:    s_waitcnt vmcnt(30)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v4
 ; GFX11-NEXT:    s_waitcnt vmcnt(29)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; GFX11-NEXT:    s_waitcnt vmcnt(28)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
 ; GFX11-NEXT:    s_waitcnt vmcnt(27)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
 ; GFX11-NEXT:    s_waitcnt vmcnt(26)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v8
 ; GFX11-NEXT:    s_waitcnt vmcnt(25)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
 ; GFX11-NEXT:    s_waitcnt vmcnt(24)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
 ; GFX11-NEXT:    s_waitcnt vmcnt(23)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v102, 16, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 16, v11
 ; GFX11-NEXT:    s_waitcnt vmcnt(22)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v101, 16, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v100, 16, v12
 ; GFX11-NEXT:    s_waitcnt vmcnt(21)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
 ; GFX11-NEXT:    s_waitcnt vmcnt(20)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; GFX11-NEXT:    s_waitcnt vmcnt(19)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v100, 16, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v15
 ; GFX11-NEXT:    s_waitcnt vmcnt(18)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 16, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v16
 ; GFX11-NEXT:    s_waitcnt vmcnt(17)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; GFX11-NEXT:    s_waitcnt vmcnt(16)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
 ; GFX11-NEXT:    s_waitcnt vmcnt(15)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 16, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v19
 ; GFX11-NEXT:    s_waitcnt vmcnt(14)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v20
 ; GFX11-NEXT:    s_waitcnt vmcnt(13)
@@ -8934,7 +8934,7 @@ define <32 x double> @global_extload_v32bf16_to_v32f64(ptr addrspace(1) %ptr) {
 ; GFX11-NEXT:    s_waitcnt vmcnt(12)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
 ; GFX11-NEXT:    s_waitcnt vmcnt(11)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v103, 16, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v101, 16, v23
 ; GFX11-NEXT:    s_waitcnt vmcnt(10)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v24
 ; GFX11-NEXT:    s_waitcnt vmcnt(9)
@@ -8942,7 +8942,7 @@ define <32 x double> @global_extload_v32bf16_to_v32f64(ptr addrspace(1) %ptr) {
 ; GFX11-NEXT:    s_waitcnt vmcnt(8)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
 ; GFX11-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 16, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 16, v27
 ; GFX11-NEXT:    s_waitcnt vmcnt(6)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v65, 16, v28
 ; GFX11-NEXT:    s_waitcnt vmcnt(5)
@@ -8957,36 +8957,36 @@ define <32 x double> @global_extload_v32bf16_to_v32f64(ptr addrspace(1) %ptr) {
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[96:97], v68
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[84:85], v65
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[96:97], v65
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[84:85], v29
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[82:83], v64
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[86:87], v33
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[98:99], v1
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[80:81], v29
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[70:71], v30
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[80:81], v30
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[70:71], v52
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[68:69], v53
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[66:67], v26
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[64:65], v52
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[64:65], v49
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[54:55], v25
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[52:53], v49
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[52:53], v21
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[50:51], v48
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[48:49], v21
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[23:24], v34
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[35:36], v22
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[33:34], v103
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[48:49], v22
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[19:20], v34
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[23:24], v35
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[35:36], v36
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[33:34], v101
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[31:32], v18
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[29:30], v102
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[29:30], v100
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[27:28], v17
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[25:26], v101
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[21:22], v13
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[19:20], v14
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[17:18], v100
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[25:26], v13
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[21:22], v14
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[17:18], v39
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[15:16], v10
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[13:14], v39
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[13:14], v38
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[11:12], v9
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[9:10], v38
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[7:8], v6
-; GFX11-NEXT:    v_cvt_f64_f32_e32 v[5:6], v5
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[9:10], v5
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[7:8], v7
+; GFX11-NEXT:    v_cvt_f64_f32_e32 v[5:6], v6
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[3:4], v2
 ; GFX11-NEXT:    v_cvt_f64_f32_e32 v[1:2], v37
 ; GFX11-NEXT:    s_clause 0xf
@@ -44505,15 +44505,14 @@ define <4 x bfloat> @v_fmuladd_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b, <4 x bfl
 ; GFX11TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
 ; GFX11TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
 ; GFX11TRUE16-NEXT:    v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11TRUE16-NEXT:    v_dual_cndmask_b32 v0, v9, v11 :: v_dual_and_b32 v1, 0xffff0000, v1
 ; GFX11TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
-; GFX11TRUE16-NEXT:    v_cndmask_b32_e32 v0, v9, v11, vcc_lo
 ; GFX11TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11TRUE16-NEXT:    v_dual_add_f32 v1, v1, v5 :: v_dual_add_f32 v2, v6, v8
+; GFX11TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11TRUE16-NEXT:    v_dual_add_f32 v1, v1, v5 :: v_dual_and_b32 v0, 0xffff0000, v0
+; GFX11TRUE16-NEXT:    v_add_f32_e32 v2, v6, v8
 ; GFX11TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v3
-; GFX11TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX11TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
 ; GFX11TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
 ; GFX11TRUE16-NEXT:    v_dual_cndmask_b32 v3, v7, v6 :: v_dual_lshlrev_b32 v6, 16, v4
diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll
index 198bf839cb1cb..3eba106b861c6 100644
--- a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll
+++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll
@@ -9090,11 +9090,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__amdgpu
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s4, s16, 0x400
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_lshlrev_b32 v3, 16, v1
 ; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
@@ -9146,11 +9145,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__amdgpu
 ; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-FAKE16-NEXT:    s_add_i32 s4, s16, 0x400
 ; GFX11-FAKE16-NEXT:    s_mov_b32 s5, 0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v1
 ; GFX11-FAKE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
 ; GFX11-FAKE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-FAKE16-NEXT:    .p2align 6
 ; GFX11-FAKE16-NEXT:  .LBB26_1: ; %atomicrmw.start
@@ -10698,11 +10696,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset(ptr add
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s4, s16, 0x400
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_lshlrev_b32 v3, 16, v1
 ; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
@@ -10754,11 +10751,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset(ptr add
 ; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-FAKE16-NEXT:    s_add_i32 s4, s16, 0x400
 ; GFX11-FAKE16-NEXT:    s_mov_b32 s5, 0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v1
 ; GFX11-FAKE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
 ; GFX11-FAKE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-FAKE16-NEXT:    .p2align 6
 ; GFX11-FAKE16-NEXT:  .LBB29_1: ; %atomicrmw.start
@@ -11563,11 +11559,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__amdgpu
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s4, s16, 0x400
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_lshlrev_b32 v3, 16, v1
 ; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
@@ -11619,11 +11614,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__amdgpu
 ; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-FAKE16-NEXT:    s_add_i32 s4, s16, 0x400
 ; GFX11-FAKE16-NEXT:    s_mov_b32 s5, 0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v1
 ; GFX11-FAKE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
 ; GFX11-FAKE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-FAKE16-NEXT:    .p2align 6
 ; GFX11-FAKE16-NEXT:  .LBB31_1: ; %atomicrmw.start
diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll
index bee2813ca30f0..c17225594164f 100644
--- a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll
+++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll
@@ -7385,11 +7385,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__amdgpu
 ; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX12-TRUE16-NEXT:    s_add_co_i32 s4, s16, 0x400
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_lshlrev_b32 v3, 16, v1
 ; GFX12-TRUE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
 ; GFX12-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX12-TRUE16-NEXT:  .LBB19_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -7445,10 +7444,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__amdgpu
 ; GFX12-FAKE16-NEXT:    s_add_co_i32 s4, s16, 0x400
 ; GFX12-FAKE16-NEXT:    s_mov_b32 s5, 0
 ; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-FAKE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX12-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v1
 ; GFX12-FAKE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
-; GFX12-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
+; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
 ; GFX12-FAKE16-NEXT:  .LBB19_1: ; %atomicrmw.start
 ; GFX12-FAKE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x0
@@ -7540,11 +7539,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__amdgpu
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s4, s16, 0x400
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_lshlrev_b32 v3, 16, v1
 ; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
@@ -7596,11 +7594,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__amdgpu
 ; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-FAKE16-NEXT:    s_add_i32 s4, s16, 0x400
 ; GFX11-FAKE16-NEXT:    s_mov_b32 s5, 0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v1
 ; GFX11-FAKE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
 ; GFX11-FAKE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-FAKE16-NEXT:    .p2align 6
 ; GFX11-FAKE16-NEXT:  .LBB19_1: ; %atomicrmw.start
diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll
index 1826743ed017d..56719dccbd08a 100644
--- a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll
+++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll
@@ -7385,11 +7385,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__amdgpu
 ; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX12-TRUE16-NEXT:    s_add_co_i32 s4, s16, 0x400
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_lshlrev_b32 v3, 16, v1
 ; GFX12-TRUE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
 ; GFX12-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX12-TRUE16-NEXT:  .LBB19_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -7445,10 +7444,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__amdgpu
 ; GFX12-FAKE16-NEXT:    s_add_co_i32 s4, s16, 0x400
 ; GFX12-FAKE16-NEXT:    s_mov_b32 s5, 0
 ; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-FAKE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX12-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v1
 ; GFX12-FAKE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
-; GFX12-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
+; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
 ; GFX12-FAKE16-NEXT:  .LBB19_1: ; %atomicrmw.start
 ; GFX12-FAKE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x0
@@ -7540,11 +7539,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__amdgpu
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s4, s16, 0x400
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_lshlrev_b32 v3, 16, v1
 ; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
@@ -7596,11 +7594,10 @@ define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__amdgpu
 ; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s16
 ; GFX11-FAKE16-NEXT:    s_add_i32 s4, s16, 0x400
 ; GFX11-FAKE16-NEXT:    s_mov_b32 s5, 0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v1
 ; GFX11-FAKE16-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
 ; GFX11-FAKE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-FAKE16-NEXT:    .p2align 6
 ; GFX11-FAKE16-NEXT:  .LBB19_1: ; %atomicrmw.start
diff --git a/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll b/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll
index 9503ffbdb4104..99b7c7737f4ae 100644
--- a/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll
+++ b/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll
@@ -910,38 +910,36 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i64_with_select(ptr addrspace(1) no
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    s_mov_b32 s8, s6
 ; SI-NEXT:    s_mov_b32 s9, s7
-; SI-NEXT:    buffer_load_ubyte v0, off, s[8:11], 0 offset:5
-; SI-NEXT:    buffer_load_ubyte v1, off, s[8:11], 0 offset:7
-; SI-NEXT:    buffer_load_ubyte v2, off, s[8:11], 0
-; SI-NEXT:    buffer_load_ubyte v3, off, s[8:11], 0 offset:1
-; SI-NEXT:    buffer_load_ubyte v4, off, s[8:11], 0 offset:2
-; SI-NEXT:    buffer_load_ubyte v5, off, s[8:11], 0 offset:3
+; SI-NEXT:    buffer_load_ubyte v0, off, s[8:11], 0
+; SI-NEXT:    buffer_load_ubyte v1, off, s[8:11], 0 offset:1
+; SI-NEXT:    buffer_load_ubyte v2, off, s[8:11], 0 offset:2
+; SI-NEXT:    buffer_load_ubyte v3, off, s[8:11], 0 offset:3
+; SI-NEXT:    buffer_load_ubyte v4, off, s[8:11], 0 offset:5
+; SI-NEXT:    buffer_load_ubyte v5, off, s[8:11], 0 offset:7
 ; SI-NEXT:    buffer_load_ubyte v6, off, s[8:11], 0 offset:4
 ; SI-NEXT:    buffer_load_ubyte v7, off, s[8:11], 0 offset:6
 ; SI-NEXT:    s_mov_b32 s0, s4
 ; SI-NEXT:    s_mov_b32 s1, s5
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v0
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v6
+; SI-NEXT:    v_or_b32_e32 v4, v4, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v1, v1, v7
-; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    v_or_b32_e32 v3, v5, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v5, v5, v7
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v2
-; SI-NEXT:    v_ffbh_u32_e32 v1, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v2, v2, v4
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_ffbh_u32_e32 v0, v0
-; SI-NEXT:    v_add_i32_e32 v1, vcc, 32, v1
-; SI-NEXT:    v_min_u32_e32 v0, v1, v0
+; SI-NEXT:    v_ffbh_u32_e32 v1, v2
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 32, v0
+; SI-NEXT:    v_min_u32_e32 v0, v0, v1
 ; SI-NEXT:    v_min_u32_e32 v0, 64, v0
 ; SI-NEXT:    v_mov_b32_e32 v1, 0
 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
diff --git a/llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll b/llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll
index 97bcd8b5ee68a..73fddb53d1dcc 100644
--- a/llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll
+++ b/llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll
@@ -874,38 +874,36 @@ define amdgpu_kernel void @v_cttz_zero_undef_i64_with_select(ptr addrspace(1) no
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    s_mov_b32 s8, s6
 ; SI-NEXT:    s_mov_b32 s9, s7
-; SI-NEXT:    buffer_load_ubyte v0, off, s[8:11], 0 offset:5
-; SI-NEXT:    buffer_load_ubyte v1, off, s[8:11], 0 offset:7
-; SI-NEXT:    buffer_load_ubyte v2, off, s[8:11], 0
-; SI-NEXT:    buffer_load_ubyte v3, off, s[8:11], 0 offset:1
-; SI-NEXT:    buffer_load_ubyte v4, off, s[8:11], 0 offset:2
-; SI-NEXT:    buffer_load_ubyte v5, off, s[8:11], 0 offset:3
+; SI-NEXT:    buffer_load_ubyte v0, off, s[8:11], 0
+; SI-NEXT:    buffer_load_ubyte v1, off, s[8:11], 0 offset:1
+; SI-NEXT:    buffer_load_ubyte v2, off, s[8:11], 0 offset:2
+; SI-NEXT:    buffer_load_ubyte v3, off, s[8:11], 0 offset:3
+; SI-NEXT:    buffer_load_ubyte v4, off, s[8:11], 0 offset:5
+; SI-NEXT:    buffer_load_ubyte v5, off, s[8:11], 0 offset:7
 ; SI-NEXT:    buffer_load_ubyte v6, off, s[8:11], 0 offset:4
 ; SI-NEXT:    buffer_load_ubyte v7, off, s[8:11], 0 offset:6
 ; SI-NEXT:    s_mov_b32 s0, s4
 ; SI-NEXT:    s_mov_b32 s1, s5
-; SI-NEXT:    s_waitcnt vmcnt(7)
-; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v0
-; SI-NEXT:    s_waitcnt vmcnt(6)
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; SI-NEXT:    s_waitcnt vmcnt(4)
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
+; SI-NEXT:    s_waitcnt vmcnt(3)
+; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
 ; SI-NEXT:    s_waitcnt vmcnt(1)
-; SI-NEXT:    v_or_b32_e32 v0, v0, v6
+; SI-NEXT:    v_or_b32_e32 v4, v4, v6
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_or_b32_e32 v1, v1, v7
-; SI-NEXT:    v_or_b32_e32 v2, v3, v2
-; SI-NEXT:    v_or_b32_e32 v3, v5, v4
-; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_or_b32_e32 v5, v5, v7
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v2
-; SI-NEXT:    v_ffbl_b32_e32 v1, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT:    v_or_b32_e32 v2, v2, v4
+; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_ffbl_b32_e32 v0, v0
-; SI-NEXT:    v_add_i32_e32 v0, vcc, 32, v0
-; SI-NEXT:    v_min_u32_e32 v0, v0, v1
+; SI-NEXT:    v_ffbl_b32_e32 v1, v2
+; SI-NEXT:    v_add_i32_e32 v1, vcc, 32, v1
+; SI-NEXT:    v_min_u32_e32 v0, v1, v0
 ; SI-NEXT:    v_min_u32_e32 v0, 64, v0
 ; SI-NEXT:    v_mov_b32_e32 v1, 0
 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
diff --git a/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll b/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll
index d1090738e24a6..745e047348626 100644
--- a/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll
+++ b/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll
@@ -1568,10 +1568,10 @@ define amdgpu_kernel void @load_v4i8_to_v4f32_unaligned_multiuse(ptr addrspace(1
 ; GFX9-NEXT:    s_mov_b32 s0, 0x4000405
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_load_ubyte v1, v0, s[12:13] offset:2
-; GFX9-NEXT:    global_load_ubyte v2, v0, s[14:15] offset:3
 ; GFX9-NEXT:    global_load_ubyte v3, v0, s[12:13] offset:3
+; GFX9-NEXT:    global_load_ubyte v2, v0, s[14:15] offset:3
 ; GFX9-NEXT:    global_load_ubyte v4, v0, s[14:15] offset:2
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_lshl_or_b32 v6, v3, 8, v1
 ; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v1, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -1904,37 +1904,37 @@ define amdgpu_kernel void @load_v7i8_to_v7f32(ptr addrspace(1) noalias %out, ptr
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 5, v0
 ; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_ubyte v10, v[2:3]
-; VI-NEXT:    v_add_u32_e32 v2, vcc, 6, v0
-; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v1, vcc
-; VI-NEXT:    v_add_u32_e32 v4, vcc, 1, v0
+; VI-NEXT:    v_add_u32_e32 v4, vcc, 6, v0
 ; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v1, vcc
-; VI-NEXT:    v_add_u32_e32 v6, vcc, 2, v0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, 1, v0
 ; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v1, vcc
-; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v0
+; VI-NEXT:    v_add_u32_e32 v8, vcc, 2, v0
 ; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_ubyte v6, v[6:7]
-; VI-NEXT:    flat_load_ubyte v7, v[8:9]
-; VI-NEXT:    flat_load_ubyte v8, v[2:3]
-; VI-NEXT:    flat_load_ubyte v2, v[0:1]
+; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v0
+; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v1, vcc
+; VI-NEXT:    flat_load_ubyte v12, v[2:3]
+; VI-NEXT:    flat_load_ubyte v2, v[8:9]
+; VI-NEXT:    flat_load_ubyte v3, v[10:11]
 ; VI-NEXT:    flat_load_ubyte v4, v[4:5]
+; VI-NEXT:    flat_load_ubyte v5, v[0:1]
+; VI-NEXT:    flat_load_ubyte v6, v[6:7]
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, 4, v0
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_ubyte v9, v[0:1]
+; VI-NEXT:    flat_load_ubyte v7, v[0:1]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
-; VI-NEXT:    s_waitcnt vmcnt(6)
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v5, v10
+; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v2, v2
 ; VI-NEXT:    s_waitcnt vmcnt(4)
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v3, v7
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v3, v3
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v0, v2
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v2, v6
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v0, v5
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v1, v4
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v6, v8
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v1, v6
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v6, v4
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v5, v12
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v4, v9
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v4, v7
 ; VI-NEXT:    buffer_store_dwordx3 v[4:6], off, s[0:3], 0 offset:16
 ; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/ds-alignment.ll b/llvm/test/CodeGen/AMDGPU/ds-alignment.ll
index b1664c59a7e4c..93422e259b827 100644
--- a/llvm/test/CodeGen/AMDGPU/ds-alignment.ll
+++ b/llvm/test/CodeGen/AMDGPU/ds-alignment.ll
@@ -209,27 +209,28 @@ define amdgpu_kernel void @ds8align1(ptr addrspace(3) %in, ptr addrspace(3) %out
 ; ALIGNED-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; ALIGNED-SDAG-NEXT:    v_mov_b32_e32 v0, s0
-; ALIGNED-SDAG-NEXT:    ds_read_u8 v1, v0
-; ALIGNED-SDAG-NEXT:    ds_read_u8 v2, v0 offset:1
-; ALIGNED-SDAG-NEXT:    ds_read_u8 v3, v0 offset:2
-; ALIGNED-SDAG-NEXT:    ds_read_u8 v4, v0 offset:3
+; ALIGNED-SDAG-NEXT:    ds_read_u8 v2, v0
+; ALIGNED-SDAG-NEXT:    ds_read_u8 v3, v0 offset:1
+; ALIGNED-SDAG-NEXT:    ds_read_u8 v4, v0 offset:2
 ; ALIGNED-SDAG-NEXT:    ds_read_u8 v5, v0 offset:4
 ; ALIGNED-SDAG-NEXT:    ds_read_u8 v6, v0 offset:5
+; ALIGNED-SDAG-NEXT:    ds_read_u8 v7, v0 offset:3
 ; ALIGNED-SDAG-NEXT:    ds_read_u8 v8, v0 offset:6
 ; ALIGNED-SDAG-NEXT:    ds_read_u8 v0, v0 offset:7
-; ALIGNED-SDAG-NEXT:    v_mov_b32_e32 v7, s1
-; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(3)
-; ALIGNED-SDAG-NEXT:    ds_write_b8 v7, v5 offset:4
-; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(3)
-; ALIGNED-SDAG-NEXT:    ds_write_b8 v7, v6 offset:5
-; ALIGNED-SDAG-NEXT:    ds_write_b8 v7, v1
-; ALIGNED-SDAG-NEXT:    ds_write_b8 v7, v2 offset:1
-; ALIGNED-SDAG-NEXT:    ds_write_b8 v7, v3 offset:2
-; ALIGNED-SDAG-NEXT:    ds_write_b8 v7, v4 offset:3
+; ALIGNED-SDAG-NEXT:    v_mov_b32_e32 v1, s1
+; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(4)
+; ALIGNED-SDAG-NEXT:    ds_write_b8 v1, v5 offset:4
+; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(4)
+; ALIGNED-SDAG-NEXT:    ds_write_b8 v1, v6 offset:5
+; ALIGNED-SDAG-NEXT:    ds_write_b8 v1, v2
+; ALIGNED-SDAG-NEXT:    ds_write_b8 v1, v3 offset:1
+; ALIGNED-SDAG-NEXT:    ds_write_b8 v1, v4 offset:2
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(7)
-; ALIGNED-SDAG-NEXT:    ds_write_b8 v7, v8 offset:6
+; ALIGNED-SDAG-NEXT:    ds_write_b8 v1, v7 offset:3
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(7)
-; ALIGNED-SDAG-NEXT:    ds_write_b8 v7, v0 offset:7
+; ALIGNED-SDAG-NEXT:    ds_write_b8 v1, v8 offset:6
+; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(7)
+; ALIGNED-SDAG-NEXT:    ds_write_b8 v1, v0 offset:7
 ; ALIGNED-SDAG-NEXT:    s_endpgm
 ;
 ; ALIGNED-GISEL-LABEL: ds8align1:
@@ -492,23 +493,24 @@ define amdgpu_kernel void @ds12align2(ptr addrspace(3) %in, ptr addrspace(3) %ou
 ; ALIGNED-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; ALIGNED-SDAG-NEXT:    v_mov_b32_e32 v0, s0
-; ALIGNED-SDAG-NEXT:    ds_read_u16 v1, v0 offset:8
 ; ALIGNED-SDAG-NEXT:    ds_read_u16 v2, v0
-; ALIGNED-SDAG-NEXT:    ds_read_u16 v3, v0 offset:2
+; ALIGNED-SDAG-NEXT:    ds_read_u16 v3, v0 offset:8
 ; ALIGNED-SDAG-NEXT:    ds_read_u16 v4, v0 offset:4
-; ALIGNED-SDAG-NEXT:    ds_read_u16 v5, v0 offset:6
-; ALIGNED-SDAG-NEXT:    v_mov_b32_e32 v6, s1
+; ALIGNED-SDAG-NEXT:    ds_read_u16 v5, v0 offset:2
+; ALIGNED-SDAG-NEXT:    ds_read_u16 v6, v0 offset:6
 ; ALIGNED-SDAG-NEXT:    ds_read_u16 v0, v0 offset:10
+; ALIGNED-SDAG-NEXT:    v_mov_b32_e32 v1, s1
+; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(4)
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v3 offset:8
+; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(4)
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v4 offset:4
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v2
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(5)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v6, v1 offset:8
-; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(3)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v6, v4 offset:4
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v6, v2
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v6, v3 offset:2
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v5 offset:2
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(5)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v6, v5 offset:6
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v6 offset:6
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(5)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v6, v0 offset:10
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v0 offset:10
 ; ALIGNED-SDAG-NEXT:    s_endpgm
 ;
 ; ALIGNED-GISEL-LABEL: ds12align2:
@@ -808,29 +810,27 @@ define amdgpu_kernel void @ds16align2(ptr addrspace(3) %in, ptr addrspace(3) %ou
 ; ALIGNED-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; ALIGNED-SDAG-NEXT:    v_mov_b32_e32 v0, s0
-; ALIGNED-SDAG-NEXT:    ds_read_u16 v1, v0 offset:12
 ; ALIGNED-SDAG-NEXT:    ds_read_u16 v2, v0
 ; ALIGNED-SDAG-NEXT:    ds_read_u16 v3, v0 offset:2
 ; ALIGNED-SDAG-NEXT:    ds_read_u16 v4, v0 offset:4
-; ALIGNED-SDAG-NEXT:    ds_read_u16 v5, v0 offset:6
-; ALIGNED-SDAG-NEXT:    ds_read_u16 v6, v0 offset:8
-; ALIGNED-SDAG-NEXT:    ds_read_u16 v7, v0 offset:10
-; ALIGNED-SDAG-NEXT:    v_mov_b32_e32 v8, s1
+; ALIGNED-SDAG-NEXT:    ds_read_u16 v5, v0 offset:12
+; ALIGNED-SDAG-NEXT:    ds_read_u16 v6, v0 offset:6
+; ALIGNED-SDAG-NEXT:    ds_read_u16 v7, v0 offset:8
+; ALIGNED-SDAG-NEXT:    ds_read_u16 v8, v0 offset:10
 ; ALIGNED-SDAG-NEXT:    ds_read_u16 v0, v0 offset:14
-; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(7)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v8, v1 offset:12
-; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(7)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v8, v2
-; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(6)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v8, v4 offset:4
+; ALIGNED-SDAG-NEXT:    v_mov_b32_e32 v1, s1
+; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(4)
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v5 offset:12
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v2
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v4 offset:4
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(5)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v8, v6 offset:8
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v8, v3 offset:2
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v8, v5 offset:6
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v7 offset:8
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v3 offset:2
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v6 offset:6
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(7)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v8, v7 offset:10
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v8 offset:10
 ; ALIGNED-SDAG-NEXT:    s_waitcnt lgkmcnt(7)
-; ALIGNED-SDAG-NEXT:    ds_write_b16 v8, v0 offset:14
+; ALIGNED-SDAG-NEXT:    ds_write_b16 v1, v0 offset:14
 ; ALIGNED-SDAG-NEXT:    s_endpgm
 ;
 ; ALIGNED-GISEL-LABEL: ds16align2:
diff --git a/llvm/test/CodeGen/AMDGPU/ds_read2.ll b/llvm/test/CodeGen/AMDGPU/ds_read2.ll
index 06c30dfd36033..d95f528442efd 100644
--- a/llvm/test/CodeGen/AMDGPU/ds_read2.ll
+++ b/llvm/test/CodeGen/AMDGPU/ds_read2.ll
@@ -522,36 +522,39 @@ define amdgpu_kernel void @simple_read2_f32_volatile_1(ptr addrspace(1) %out) #0
 define amdgpu_kernel void @unaligned_read2_f32(ptr addrspace(1) %out, ptr addrspace(3) %lds) #0 {
 ; CI-LABEL: unaligned_read2_f32:
 ; CI:       ; %bb.0:
-; CI-NEXT:    s_load_dword s0, s[4:5], 0x2
+; CI-NEXT:    s_load_dword s2, s[4:5], 0x2
+; CI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
 ; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; CI-NEXT:    s_mov_b32 m0, -1
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
-; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_add_i32_e32 v1, vcc, s0, v0
-; CI-NEXT:    ds_read_u8 v2, v1 offset:34
-; CI-NEXT:    ds_read_u8 v3, v1 offset:32
-; CI-NEXT:    ds_read_u8 v4, v1 offset:3
+; CI-NEXT:    v_add_i32_e32 v1, vcc, s2, v0
+; CI-NEXT:    ds_read_u8 v2, v1 offset:1
+; CI-NEXT:    ds_read_u8 v3, v1 offset:34
+; CI-NEXT:    ds_read_u8 v4, v1 offset:32
 ; CI-NEXT:    ds_read_u8 v5, v1 offset:2
-; CI-NEXT:    ds_read_u8 v6, v1 offset:1
-; CI-NEXT:    ds_read_u8 v7, v1
+; CI-NEXT:    ds_read_u8 v6, v1
+; CI-NEXT:    ds_read_u8 v7, v1 offset:3
 ; CI-NEXT:    ds_read_u8 v8, v1 offset:33
 ; CI-NEXT:    ds_read_u8 v1, v1 offset:35
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; CI-NEXT:    s_waitcnt lgkmcnt(7)
+; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; CI-NEXT:    s_waitcnt lgkmcnt(3)
+; CI-NEXT:    v_or_b32_e32 v2, v2, v6
+; CI-NEXT:    s_waitcnt lgkmcnt(2)
+; CI-NEXT:    v_lshlrev_b32_e32 v6, 8, v7
+; CI-NEXT:    v_or_b32_e32 v5, v6, v5
+; CI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
-; CI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; CI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; CI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
+; CI-NEXT:    v_or_b32_e32 v2, v5, v2
 ; CI-NEXT:    v_lshlrev_b32_e32 v5, 8, v8
-; CI-NEXT:    v_or_b32_e32 v1, v1, v2
-; CI-NEXT:    v_or_b32_e32 v6, v6, v7
-; CI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; CI-NEXT:    v_or_b32_e32 v3, v5, v3
-; CI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; CI-NEXT:    v_or_b32_e32 v4, v4, v6
 ; CI-NEXT:    v_or_b32_e32 v1, v1, v3
-; CI-NEXT:    v_add_f32_e32 v2, v4, v1
+; CI-NEXT:    v_or_b32_e32 v4, v5, v4
+; CI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; CI-NEXT:    v_or_b32_e32 v1, v1, v4
+; CI-NEXT:    v_add_f32_e32 v2, v2, v1
+; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    v_mov_b32_e32 v1, 0
 ; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
 ; CI-NEXT:    s_endpgm
@@ -612,36 +615,39 @@ define amdgpu_kernel void @unaligned_read2_f32(ptr addrspace(1) %out, ptr addrsp
 define amdgpu_kernel void @unaligned_offset_read2_f32(ptr addrspace(1) %out, ptr addrspace(3) %lds) #0 {
 ; CI-LABEL: unaligned_offset_read2_f32:
 ; CI:       ; %bb.0:
-; CI-NEXT:    s_load_dword s0, s[4:5], 0x2
+; CI-NEXT:    s_load_dword s2, s[4:5], 0x2
+; CI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
 ; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; CI-NEXT:    s_mov_b32 m0, -1
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
-; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_add_i32_e32 v1, vcc, s0, v0
-; CI-NEXT:    ds_read_u8 v2, v1 offset:11
-; CI-NEXT:    ds_read_u8 v3, v1 offset:9
-; CI-NEXT:    ds_read_u8 v4, v1 offset:8
+; CI-NEXT:    v_add_i32_e32 v1, vcc, s2, v0
+; CI-NEXT:    ds_read_u8 v2, v1 offset:6
+; CI-NEXT:    ds_read_u8 v3, v1 offset:11
+; CI-NEXT:    ds_read_u8 v4, v1 offset:9
 ; CI-NEXT:    ds_read_u8 v5, v1 offset:7
-; CI-NEXT:    ds_read_u8 v6, v1 offset:6
-; CI-NEXT:    ds_read_u8 v7, v1 offset:5
+; CI-NEXT:    ds_read_u8 v6, v1 offset:5
+; CI-NEXT:    ds_read_u8 v7, v1 offset:8
 ; CI-NEXT:    ds_read_u8 v8, v1 offset:10
 ; CI-NEXT:    ds_read_u8 v1, v1 offset:12
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; CI-NEXT:    s_waitcnt lgkmcnt(7)
+; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; CI-NEXT:    s_waitcnt lgkmcnt(3)
+; CI-NEXT:    v_or_b32_e32 v2, v2, v6
+; CI-NEXT:    s_waitcnt lgkmcnt(2)
+; CI-NEXT:    v_lshlrev_b32_e32 v6, 8, v7
+; CI-NEXT:    v_or_b32_e32 v5, v6, v5
+; CI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
-; CI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; CI-NEXT:    v_or_b32_e32 v4, v4, v5
 ; CI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
+; CI-NEXT:    v_or_b32_e32 v2, v5, v2
 ; CI-NEXT:    v_lshlrev_b32_e32 v5, 8, v8
-; CI-NEXT:    v_or_b32_e32 v1, v1, v2
-; CI-NEXT:    v_or_b32_e32 v6, v6, v7
-; CI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; CI-NEXT:    v_or_b32_e32 v3, v5, v3
-; CI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; CI-NEXT:    v_or_b32_e32 v4, v4, v6
 ; CI-NEXT:    v_or_b32_e32 v1, v1, v3
-; CI-NEXT:    v_add_f32_e32 v2, v4, v1
+; CI-NEXT:    v_or_b32_e32 v4, v5, v4
+; CI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; CI-NEXT:    v_or_b32_e32 v1, v1, v4
+; CI-NEXT:    v_add_f32_e32 v2, v2, v1
+; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    v_mov_b32_e32 v1, 0
 ; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
 ; CI-NEXT:    s_endpgm
@@ -709,17 +715,17 @@ define amdgpu_kernel void @misaligned_2_simple_read2_f32(ptr addrspace(1) %out,
 ; CI-NEXT:    s_mov_b32 s2, 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    v_add_i32_e32 v1, vcc, s0, v0
-; CI-NEXT:    ds_read_u16 v2, v1 offset:32
-; CI-NEXT:    ds_read_u16 v3, v1 offset:2
+; CI-NEXT:    ds_read_u16 v2, v1 offset:2
+; CI-NEXT:    ds_read_u16 v3, v1 offset:32
 ; CI-NEXT:    ds_read_u16 v4, v1
 ; CI-NEXT:    ds_read_u16 v1, v1 offset:34
 ; CI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; CI-NEXT:    v_or_b32_e32 v3, v3, v4
+; CI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; CI-NEXT:    v_or_b32_e32 v2, v2, v4
 ; CI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; CI-NEXT:    v_or_b32_e32 v1, v1, v2
-; CI-NEXT:    v_add_f32_e32 v2, v3, v1
+; CI-NEXT:    v_or_b32_e32 v1, v1, v3
+; CI-NEXT:    v_add_f32_e32 v2, v2, v1
 ; CI-NEXT:    v_mov_b32_e32 v1, 0
 ; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
 ; CI-NEXT:    s_endpgm
@@ -1460,17 +1466,17 @@ define amdgpu_kernel void @read2_v2i32_align1_odd_offset(ptr addrspace(1) %out)
 ; CI-NEXT:    v_or_b32_e32 v2, v2, v3
 ; CI-NEXT:    s_waitcnt lgkmcnt(1)
 ; CI-NEXT:    v_or_b32_e32 v1, v1, v4
-; CI-NEXT:    ds_read_u8 v4, v0 offset:67
-; CI-NEXT:    ds_read_u8 v6, v0 offset:66
+; CI-NEXT:    ds_read_u8 v4, v0 offset:66
+; CI-NEXT:    ds_read_u8 v6, v0 offset:67
 ; CI-NEXT:    ds_read_u8 v0, v0 offset:65
 ; CI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; CI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
 ; CI-NEXT:    v_or_b32_e32 v1, v2, v1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v6
+; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v4
 ; CI-NEXT:    v_or_b32_e32 v0, v2, v0
 ; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v5
-; CI-NEXT:    v_or_b32_e32 v2, v2, v4
+; CI-NEXT:    v_or_b32_e32 v2, v2, v6
 ; CI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
@@ -1481,26 +1487,25 @@ define amdgpu_kernel void @read2_v2i32_align1_odd_offset(ptr addrspace(1) %out)
 ; GFX9-ALIGNED-LABEL: read2_v2i32_align1_odd_offset:
 ; GFX9-ALIGNED:       ; %bb.0: ; %entry
 ; GFX9-ALIGNED-NEXT:    v_mov_b32_e32 v2, 0
-; GFX9-ALIGNED-NEXT:    ds_read_u8 v0, v2 offset:70
-; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v2 offset:65
-; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v2 offset:66
-; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v2 offset:67
-; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v2 offset:68
-; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v2 offset:69
+; GFX9-ALIGNED-NEXT:    ds_read_u8 v0, v2 offset:65
+; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v2 offset:66
+; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v2 offset:67
+; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v2 offset:68
+; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v2 offset:70
+; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v2 offset:69
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v7, v2 offset:72
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v8, v2 offset:71
-; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(7)
-; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 8, v0
 ; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v1, 8, v7
-; GFX9-ALIGNED-NEXT:    v_or_b32_sdwa v1, v1, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v1, v1, v0
-; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 8, v4
-; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v0, v0, v3
-; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 8, v6
-; GFX9-ALIGNED-NEXT:    v_or_b32_sdwa v3, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
+; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
+; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v1, v1, v6
+; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v6, 8, v7
+; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v0, v3, v0
+; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 8, v5
+; GFX9-ALIGNED-NEXT:    v_or_b32_sdwa v6, v6, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX9-ALIGNED-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v1, v6, v1
 ; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v0, v3, v0
 ; GFX9-ALIGNED-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
 ; GFX9-ALIGNED-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.global.ll b/llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.global.ll
index 60334e46a4454..52bcaed7ec75a 100644
--- a/llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.global.ll
+++ b/llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.global.ll
@@ -153,14 +153,14 @@ define i32 @global_load_2xi16_align1(ptr addrspace(1) %p) #0 {
 ; GFX7-ALIGNED-NEXT:    v_addc_u32_e32 v5, vcc, 0, v1, vcc
 ; GFX7-ALIGNED-NEXT:    v_add_i32_e32 v6, vcc, 3, v0
 ; GFX7-ALIGNED-NEXT:    v_addc_u32_e32 v7, vcc, 0, v1, vcc
-; GFX7-ALIGNED-NEXT:    flat_load_ubyte v6, v[6:7]
 ; GFX7-ALIGNED-NEXT:    flat_load_ubyte v4, v[4:5]
+; GFX7-ALIGNED-NEXT:    flat_load_ubyte v5, v[6:7]
 ; GFX7-ALIGNED-NEXT:    flat_load_ubyte v2, v[2:3]
 ; GFX7-ALIGNED-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; GFX7-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 24, v6
-; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(2)
 ; GFX7-ALIGNED-NEXT:    v_lshlrev_b32_e32 v1, 8, v4
+; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 24, v5
 ; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(1)
 ; GFX7-ALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.private.ll b/llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.private.ll
index f9694dcd89abf..6f8da57e223e5 100644
--- a/llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.private.ll
+++ b/llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.private.ll
@@ -204,14 +204,14 @@ define i32 @private_load_2xi16_align1(ptr addrspace(5) %p) #0 {
 ; GFX7-ALIGNED-NEXT:    v_add_i32_e32 v1, vcc, 2, v0
 ; GFX7-ALIGNED-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
 ; GFX7-ALIGNED-NEXT:    v_add_i32_e32 v3, vcc, 3, v0
-; GFX7-ALIGNED-NEXT:    buffer_load_ubyte v3, v3, s[0:3], 0 offen
 ; GFX7-ALIGNED-NEXT:    buffer_load_ubyte v2, v2, s[0:3], 0 offen
+; GFX7-ALIGNED-NEXT:    buffer_load_ubyte v3, v3, s[0:3], 0 offen
 ; GFX7-ALIGNED-NEXT:    buffer_load_ubyte v1, v1, s[0:3], 0 offen
 ; GFX7-ALIGNED-NEXT:    buffer_load_ubyte v0, v0, s[0:3], 0 offen
 ; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; GFX7-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
-; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(2)
 ; GFX7-ALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
 ; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(1)
 ; GFX7-ALIGNED-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/fcopysign.f32.ll b/llvm/test/CodeGen/AMDGPU/fcopysign.f32.ll
index fab45c9dc3bc3..61f5b73033f5e 100644
--- a/llvm/test/CodeGen/AMDGPU/fcopysign.f32.ll
+++ b/llvm/test/CodeGen/AMDGPU/fcopysign.f32.ll
@@ -569,14 +569,13 @@ define amdgpu_kernel void @s_test_copysign_v3f32(ptr addrspace(1) %out, <3 x flo
 ; GFX11-NEXT:    s_clause 0x1
 ; GFX11-NEXT:    s_load_b256 s[8:15], s[4:5], 0x34
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v3, s12
 ; GFX11-NEXT:    v_dual_mov_b32 v0, s14 :: v_dual_mov_b32 v1, s13
-; GFX11-NEXT:    v_mov_b32_e32 v3, s12
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_bfi_b32 v2, 0x7fffffff, s10, v0
 ; GFX11-NEXT:    v_bfi_b32 v1, 0x7fffffff, s9, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
 ; GFX11-NEXT:    v_bfi_b32 v0, 0x7fffffff, s8, v3
 ; GFX11-NEXT:    global_store_b96 v4, v[0:2], s[0:1]
 ; GFX11-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/fp-classify.ll b/llvm/test/CodeGen/AMDGPU/fp-classify.ll
index 498df8a65feda..200f74beec385 100644
--- a/llvm/test/CodeGen/AMDGPU/fp-classify.ll
+++ b/llvm/test/CodeGen/AMDGPU/fp-classify.ll
@@ -536,8 +536,8 @@ define amdgpu_kernel void @test_isfinite_pattern_4_commute_and(ptr addrspace(1)
 define amdgpu_kernel void @test_not_isfinite_pattern_4_wrong_ord_test(ptr addrspace(1) nocapture %out, float %x, [8 x i32], float %y) #0 {
 ; SI-LABEL: test_not_isfinite_pattern_4_wrong_ord_test:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dword s0, s[4:5], 0x14
 ; SI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x9
+; SI-NEXT:    s_load_dword s0, s[4:5], 0x14
 ; SI-NEXT:    s_load_dword s1, s[4:5], 0xb
 ; SI-NEXT:    s_mov_b32 s11, 0xf000
 ; SI-NEXT:    s_mov_b32 s10, -1
diff --git a/llvm/test/CodeGen/AMDGPU/freeze.ll b/llvm/test/CodeGen/AMDGPU/freeze.ll
index 2040aedc250e6..ac438062ae208 100644
--- a/llvm/test/CodeGen/AMDGPU/freeze.ll
+++ b/llvm/test/CodeGen/AMDGPU/freeze.ll
@@ -10546,14 +10546,14 @@ define void @freeze_v16p3(ptr addrspace(3) %ptra, ptr addrspace(3) %ptrb) {
 ; GFX6-SDAG-LABEL: freeze_v16p3:
 ; GFX6-SDAG:       ; %bb.0:
 ; GFX6-SDAG-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v6, vcc, 8, v0
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v2, vcc, 24, v0
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v4, vcc, 16, v0
-; GFX6-SDAG-NEXT:    s_mov_b32 m0, -1
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v12, vcc, 40, v0
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v14, vcc, 32, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v4, vcc, 24, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v6, vcc, 16, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v10, vcc, 40, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v12, vcc, 32, v0
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v16, vcc, 56, v0
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v10, vcc, 48, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v14, vcc, 48, v0
+; GFX6-SDAG-NEXT:    s_mov_b32 m0, -1
 ; GFX6-SDAG-NEXT:    ds_read_b64 v[2:3], v2
 ; GFX6-SDAG-NEXT:    ds_read_b64 v[4:5], v4
 ; GFX6-SDAG-NEXT:    ds_read_b64 v[6:7], v6
@@ -10563,22 +10563,23 @@ define void @freeze_v16p3(ptr addrspace(3) %ptra, ptr addrspace(3) %ptrb) {
 ; GFX6-SDAG-NEXT:    ds_read_b64 v[14:15], v14
 ; GFX6-SDAG-NEXT:    ds_read_b64 v[16:17], v16
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 48, v1
-; GFX6-SDAG-NEXT:    s_waitcnt lgkmcnt(3)
-; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[10:11]
+; GFX6-SDAG-NEXT:    s_waitcnt lgkmcnt(4)
+; GFX6-SDAG-NEXT:    ds_write_b64 v1, v[8:9]
+; GFX6-SDAG-NEXT:    s_waitcnt lgkmcnt(2)
+; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[14:15]
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 56, v1
-; GFX6-SDAG-NEXT:    s_waitcnt lgkmcnt(1)
+; GFX6-SDAG-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[16:17]
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 32, v1
-; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[14:15]
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 40, v1
 ; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[12:13]
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 40, v1
+; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[10:11]
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 16, v1
-; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[4:5]
+; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[6:7]
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 24, v1
-; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[2:3]
+; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[4:5]
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 8, v1
-; GFX6-SDAG-NEXT:    ds_write_b64 v1, v[8:9]
-; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[6:7]
+; GFX6-SDAG-NEXT:    ds_write_b64 v0, v[2:3]
 ; GFX6-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX6-SDAG-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -11565,22 +11566,22 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v9, vcc, 44, v0
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v10, vcc, 40, v0
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v11, vcc, 36, v0
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v12, vcc, 28, v0
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v13, vcc, 24, v0
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v14, vcc, 20, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v12, vcc, 32, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v13, vcc, 28, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v14, vcc, 24, v0
+; GFX6-SDAG-NEXT:    v_add_i32_e32 v15, vcc, 20, v0
 ; GFX6-SDAG-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_load_dword v3, v3, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_load_dword v4, v4, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_load_dword v9, v9, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_load_dword v10, v10, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_load_dword v11, v11, s[0:3], 0 offen
-; GFX6-SDAG-NEXT:    buffer_load_dword v15, v0, s[0:3], 0 offen
-; GFX6-SDAG-NEXT:    buffer_load_dword v12, v12, s[0:3], 0 offen
-; GFX6-SDAG-NEXT:    buffer_load_dword v13, v13, s[0:3], 0 offen
+; GFX6-SDAG-NEXT:    buffer_load_dword v16, v0, s[0:3], 0 offen
+; GFX6-SDAG-NEXT:    buffer_load_dword v15, v15, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_load_dword v14, v14, s[0:3], 0 offen
-; GFX6-SDAG-NEXT:    v_add_i32_e32 v16, vcc, 32, v0
+; GFX6-SDAG-NEXT:    buffer_load_dword v13, v13, s[0:3], 0 offen
+; GFX6-SDAG-NEXT:    buffer_load_dword v12, v12, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 60, v0
-; GFX6-SDAG-NEXT:    buffer_load_dword v16, v16, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_load_dword v0, v0, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v17, vcc, 4, v1
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v18, vcc, 8, v1
@@ -11603,13 +11604,15 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v19, vcc, 36, v1
 ; GFX6-SDAG-NEXT:    v_add_i32_e32 v8, vcc, 44, v1
 ; GFX6-SDAG-NEXT:    s_waitcnt vmcnt(9)
-; GFX6-SDAG-NEXT:    buffer_store_dword v15, v1, s[0:3], 0 offen
-; GFX6-SDAG-NEXT:    s_waitcnt vmcnt(7)
-; GFX6-SDAG-NEXT:    buffer_store_dword v14, v17, s[0:3], 0 offen
-; GFX6-SDAG-NEXT:    buffer_store_dword v13, v7, s[0:3], 0 offen
-; GFX6-SDAG-NEXT:    buffer_store_dword v12, v18, s[0:3], 0 offen
+; GFX6-SDAG-NEXT:    buffer_store_dword v16, v1, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    s_waitcnt vmcnt(9)
-; GFX6-SDAG-NEXT:    buffer_store_dword v16, v6, s[0:3], 0 offen
+; GFX6-SDAG-NEXT:    buffer_store_dword v15, v17, s[0:3], 0 offen
+; GFX6-SDAG-NEXT:    s_waitcnt vmcnt(9)
+; GFX6-SDAG-NEXT:    buffer_store_dword v14, v7, s[0:3], 0 offen
+; GFX6-SDAG-NEXT:    s_waitcnt vmcnt(9)
+; GFX6-SDAG-NEXT:    buffer_store_dword v13, v18, s[0:3], 0 offen
+; GFX6-SDAG-NEXT:    s_waitcnt vmcnt(9)
+; GFX6-SDAG-NEXT:    buffer_store_dword v12, v6, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_store_dword v11, v19, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_store_dword v10, v5, s[0:3], 0 offen
 ; GFX6-SDAG-NEXT:    buffer_store_dword v9, v8, s[0:3], 0 offen
@@ -11631,24 +11634,24 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX6-GISEL:       ; %bb.0:
 ; GFX6-GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
-; GFX6-GISEL-NEXT:    buffer_load_dword v3, v0, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v3, vcc, 8, v0
 ; GFX6-GISEL-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v4, vcc, 8, v0
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v5, vcc, 12, v0
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v6, vcc, 16, v0
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v7, vcc, 20, v0
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v8, vcc, 24, v0
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v9, vcc, 28, v0
+; GFX6-GISEL-NEXT:    buffer_load_dword v3, v3, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v4, vcc, 12, v0
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v5, vcc, 16, v0
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v6, vcc, 20, v0
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v7, vcc, 24, v0
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v8, vcc, 28, v0
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v10, vcc, 32, v0
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v11, vcc, 36, v0
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v12, vcc, 40, v0
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v13, vcc, 44, v0
 ; GFX6-GISEL-NEXT:    buffer_load_dword v4, v4, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    buffer_load_dword v9, v0, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    buffer_load_dword v5, v5, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    buffer_load_dword v6, v6, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    buffer_load_dword v7, v7, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    buffer_load_dword v8, v8, s[0:3], 0 offen
-; GFX6-GISEL-NEXT:    buffer_load_dword v9, v9, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    buffer_load_dword v10, v10, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    buffer_load_dword v11, v11, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    buffer_load_dword v12, v12, s[0:3], 0 offen
@@ -11669,30 +11672,32 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX6-GISEL-NEXT:    s_waitcnt expcnt(0)
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 16, v1
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v17, vcc, 20, v1
-; GFX6-GISEL-NEXT:    s_waitcnt vmcnt(12)
-; GFX6-GISEL-NEXT:    buffer_store_dword v6, v2, s[0:3], 0 offen
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 40, v1
-; GFX6-GISEL-NEXT:    buffer_store_dword v4, v18, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    buffer_store_dword v3, v18, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    s_waitcnt expcnt(0)
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v4, vcc, 24, v1
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v3, vcc, 24, v1
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v18, vcc, 28, v1
-; GFX6-GISEL-NEXT:    buffer_store_dword v5, v19, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    s_waitcnt vmcnt(14)
+; GFX6-GISEL-NEXT:    buffer_store_dword v4, v19, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    s_waitcnt expcnt(0)
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v5, vcc, 32, v1
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v4, vcc, 32, v1
+; GFX6-GISEL-NEXT:    s_waitcnt vmcnt(14)
+; GFX6-GISEL-NEXT:    buffer_store_dword v5, v2, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 40, v1
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v19, vcc, 36, v1
-; GFX6-GISEL-NEXT:    v_add_i32_e32 v6, vcc, 44, v1
-; GFX6-GISEL-NEXT:    buffer_store_dword v3, v1, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    s_waitcnt expcnt(0)
+; GFX6-GISEL-NEXT:    v_add_i32_e32 v5, vcc, 44, v1
+; GFX6-GISEL-NEXT:    buffer_store_dword v9, v1, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    s_waitcnt vmcnt(14)
-; GFX6-GISEL-NEXT:    buffer_store_dword v7, v17, s[0:3], 0 offen
-; GFX6-GISEL-NEXT:    buffer_store_dword v8, v4, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    buffer_store_dword v6, v17, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    buffer_store_dword v7, v3, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    s_waitcnt vmcnt(14)
-; GFX6-GISEL-NEXT:    buffer_store_dword v9, v18, s[0:3], 0 offen
-; GFX6-GISEL-NEXT:    buffer_store_dword v10, v5, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    buffer_store_dword v8, v18, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    buffer_store_dword v10, v4, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    s_waitcnt vmcnt(14)
 ; GFX6-GISEL-NEXT:    buffer_store_dword v11, v19, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    buffer_store_dword v12, v2, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    s_waitcnt vmcnt(14)
-; GFX6-GISEL-NEXT:    buffer_store_dword v13, v6, s[0:3], 0 offen
+; GFX6-GISEL-NEXT:    buffer_store_dword v13, v5, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 48, v1
 ; GFX6-GISEL-NEXT:    buffer_store_dword v14, v2, s[0:3], 0 offen
 ; GFX6-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 52, v1
@@ -11723,22 +11728,22 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX7-SDAG-NEXT:    v_add_i32_e32 v9, vcc, 44, v0
 ; GFX7-SDAG-NEXT:    v_add_i32_e32 v10, vcc, 40, v0
 ; GFX7-SDAG-NEXT:    v_add_i32_e32 v11, vcc, 36, v0
-; GFX7-SDAG-NEXT:    v_add_i32_e32 v12, vcc, 28, v0
-; GFX7-SDAG-NEXT:    v_add_i32_e32 v13, vcc, 24, v0
-; GFX7-SDAG-NEXT:    v_add_i32_e32 v14, vcc, 20, v0
+; GFX7-SDAG-NEXT:    v_add_i32_e32 v12, vcc, 32, v0
+; GFX7-SDAG-NEXT:    v_add_i32_e32 v13, vcc, 28, v0
+; GFX7-SDAG-NEXT:    v_add_i32_e32 v14, vcc, 24, v0
+; GFX7-SDAG-NEXT:    v_add_i32_e32 v15, vcc, 20, v0
 ; GFX7-SDAG-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_load_dword v3, v3, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_load_dword v4, v4, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_load_dword v9, v9, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_load_dword v10, v10, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_load_dword v11, v11, s[0:3], 0 offen
-; GFX7-SDAG-NEXT:    buffer_load_dword v15, v0, s[0:3], 0 offen
-; GFX7-SDAG-NEXT:    buffer_load_dword v12, v12, s[0:3], 0 offen
-; GFX7-SDAG-NEXT:    buffer_load_dword v13, v13, s[0:3], 0 offen
+; GFX7-SDAG-NEXT:    buffer_load_dword v16, v0, s[0:3], 0 offen
+; GFX7-SDAG-NEXT:    buffer_load_dword v15, v15, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_load_dword v14, v14, s[0:3], 0 offen
-; GFX7-SDAG-NEXT:    v_add_i32_e32 v16, vcc, 32, v0
+; GFX7-SDAG-NEXT:    buffer_load_dword v13, v13, s[0:3], 0 offen
+; GFX7-SDAG-NEXT:    buffer_load_dword v12, v12, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    v_add_i32_e32 v0, vcc, 60, v0
-; GFX7-SDAG-NEXT:    buffer_load_dword v16, v16, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_load_dword v0, v0, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    v_add_i32_e32 v17, vcc, 4, v1
 ; GFX7-SDAG-NEXT:    v_add_i32_e32 v18, vcc, 8, v1
@@ -11759,13 +11764,15 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX7-SDAG-NEXT:    v_add_i32_e32 v19, vcc, 36, v1
 ; GFX7-SDAG-NEXT:    v_add_i32_e32 v8, vcc, 44, v1
 ; GFX7-SDAG-NEXT:    s_waitcnt vmcnt(9)
-; GFX7-SDAG-NEXT:    buffer_store_dword v15, v1, s[0:3], 0 offen
-; GFX7-SDAG-NEXT:    s_waitcnt vmcnt(7)
-; GFX7-SDAG-NEXT:    buffer_store_dword v14, v17, s[0:3], 0 offen
-; GFX7-SDAG-NEXT:    buffer_store_dword v13, v7, s[0:3], 0 offen
-; GFX7-SDAG-NEXT:    buffer_store_dword v12, v18, s[0:3], 0 offen
+; GFX7-SDAG-NEXT:    buffer_store_dword v16, v1, s[0:3], 0 offen
+; GFX7-SDAG-NEXT:    s_waitcnt vmcnt(9)
+; GFX7-SDAG-NEXT:    buffer_store_dword v15, v17, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    s_waitcnt vmcnt(9)
-; GFX7-SDAG-NEXT:    buffer_store_dword v16, v6, s[0:3], 0 offen
+; GFX7-SDAG-NEXT:    buffer_store_dword v14, v7, s[0:3], 0 offen
+; GFX7-SDAG-NEXT:    s_waitcnt vmcnt(9)
+; GFX7-SDAG-NEXT:    buffer_store_dword v13, v18, s[0:3], 0 offen
+; GFX7-SDAG-NEXT:    s_waitcnt vmcnt(9)
+; GFX7-SDAG-NEXT:    buffer_store_dword v12, v6, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_store_dword v11, v19, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_store_dword v10, v5, s[0:3], 0 offen
 ; GFX7-SDAG-NEXT:    buffer_store_dword v9, v8, s[0:3], 0 offen
@@ -11785,24 +11792,24 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX7-GISEL:       ; %bb.0:
 ; GFX7-GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
-; GFX7-GISEL-NEXT:    buffer_load_dword v3, v0, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v3, vcc, 8, v0
 ; GFX7-GISEL-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v4, vcc, 8, v0
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v5, vcc, 12, v0
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v6, vcc, 16, v0
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v7, vcc, 20, v0
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v8, vcc, 24, v0
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v9, vcc, 28, v0
+; GFX7-GISEL-NEXT:    buffer_load_dword v3, v3, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v4, vcc, 12, v0
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v5, vcc, 16, v0
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v6, vcc, 20, v0
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v7, vcc, 24, v0
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v8, vcc, 28, v0
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v10, vcc, 32, v0
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v11, vcc, 36, v0
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v12, vcc, 40, v0
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v13, vcc, 44, v0
 ; GFX7-GISEL-NEXT:    buffer_load_dword v4, v4, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    buffer_load_dword v9, v0, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    buffer_load_dword v5, v5, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    buffer_load_dword v6, v6, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    buffer_load_dword v7, v7, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    buffer_load_dword v8, v8, s[0:3], 0 offen
-; GFX7-GISEL-NEXT:    buffer_load_dword v9, v9, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    buffer_load_dword v10, v10, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    buffer_load_dword v11, v11, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    buffer_load_dword v12, v12, s[0:3], 0 offen
@@ -11822,28 +11829,29 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX7-GISEL-NEXT:    buffer_store_dword v2, v17, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 16, v1
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v17, vcc, 20, v1
-; GFX7-GISEL-NEXT:    s_waitcnt vmcnt(12)
-; GFX7-GISEL-NEXT:    buffer_store_dword v6, v2, s[0:3], 0 offen
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 40, v1
-; GFX7-GISEL-NEXT:    buffer_store_dword v4, v18, s[0:3], 0 offen
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v4, vcc, 24, v1
+; GFX7-GISEL-NEXT:    buffer_store_dword v3, v18, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v3, vcc, 24, v1
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v18, vcc, 28, v1
-; GFX7-GISEL-NEXT:    buffer_store_dword v5, v19, s[0:3], 0 offen
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v5, vcc, 32, v1
+; GFX7-GISEL-NEXT:    s_waitcnt vmcnt(14)
+; GFX7-GISEL-NEXT:    buffer_store_dword v4, v19, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v4, vcc, 32, v1
+; GFX7-GISEL-NEXT:    s_waitcnt vmcnt(14)
+; GFX7-GISEL-NEXT:    buffer_store_dword v5, v2, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 40, v1
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v19, vcc, 36, v1
-; GFX7-GISEL-NEXT:    v_add_i32_e32 v6, vcc, 44, v1
-; GFX7-GISEL-NEXT:    buffer_store_dword v3, v1, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    v_add_i32_e32 v5, vcc, 44, v1
+; GFX7-GISEL-NEXT:    buffer_store_dword v9, v1, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    s_waitcnt vmcnt(14)
-; GFX7-GISEL-NEXT:    buffer_store_dword v7, v17, s[0:3], 0 offen
-; GFX7-GISEL-NEXT:    buffer_store_dword v8, v4, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    buffer_store_dword v6, v17, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    buffer_store_dword v7, v3, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    s_waitcnt vmcnt(14)
-; GFX7-GISEL-NEXT:    buffer_store_dword v9, v18, s[0:3], 0 offen
-; GFX7-GISEL-NEXT:    buffer_store_dword v10, v5, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    buffer_store_dword v8, v18, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    buffer_store_dword v10, v4, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    s_waitcnt vmcnt(14)
 ; GFX7-GISEL-NEXT:    buffer_store_dword v11, v19, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    buffer_store_dword v12, v2, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    s_waitcnt vmcnt(14)
-; GFX7-GISEL-NEXT:    buffer_store_dword v13, v6, s[0:3], 0 offen
+; GFX7-GISEL-NEXT:    buffer_store_dword v13, v5, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 48, v1
 ; GFX7-GISEL-NEXT:    buffer_store_dword v14, v2, s[0:3], 0 offen
 ; GFX7-GISEL-NEXT:    v_add_i32_e32 v2, vcc, 52, v1
@@ -11861,24 +11869,24 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX8-GISEL:       ; %bb.0:
 ; GFX8-GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v2, vcc, 4, v0
-; GFX8-GISEL-NEXT:    buffer_load_dword v3, v0, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v3, vcc, 8, v0
 ; GFX8-GISEL-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v4, vcc, 8, v0
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v5, vcc, 12, v0
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v6, vcc, 16, v0
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v7, vcc, 20, v0
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v8, vcc, 24, v0
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v9, vcc, 28, v0
+; GFX8-GISEL-NEXT:    buffer_load_dword v3, v3, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v4, vcc, 12, v0
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v5, vcc, 16, v0
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v6, vcc, 20, v0
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v7, vcc, 24, v0
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v8, vcc, 28, v0
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v10, vcc, 32, v0
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v11, vcc, 36, v0
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v12, vcc, 40, v0
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v13, vcc, 44, v0
 ; GFX8-GISEL-NEXT:    buffer_load_dword v4, v4, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    buffer_load_dword v9, v0, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    buffer_load_dword v5, v5, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    buffer_load_dword v6, v6, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    buffer_load_dword v7, v7, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    buffer_load_dword v8, v8, s[0:3], 0 offen
-; GFX8-GISEL-NEXT:    buffer_load_dword v9, v9, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    buffer_load_dword v10, v10, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    buffer_load_dword v11, v11, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    buffer_load_dword v12, v12, s[0:3], 0 offen
@@ -11898,28 +11906,29 @@ define void @freeze_v16p5(ptr addrspace(5) %ptra, ptr addrspace(5) %ptrb) {
 ; GFX8-GISEL-NEXT:    buffer_store_dword v2, v17, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v2, vcc, 16, v1
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v17, vcc, 20, v1
-; GFX8-GISEL-NEXT:    s_waitcnt vmcnt(12)
-; GFX8-GISEL-NEXT:    buffer_store_dword v6, v2, s[0:3], 0 offen
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v2, vcc, 40, v1
-; GFX8-GISEL-NEXT:    buffer_store_dword v4, v18, s[0:3], 0 offen
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v4, vcc, 24, v1
+; GFX8-GISEL-NEXT:    buffer_store_dword v3, v18, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v3, vcc, 24, v1
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v18, vcc, 28, v1
-; GFX8-GISEL-NEXT:    buffer_store_dword v5, v19, s[0:3], 0 offen
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v5, vcc, 32, v1
+; GFX8-GISEL-NEXT:    s_waitcnt vmcnt(14)
+; GFX8-GISEL-NEXT:    buffer_store_dword v4, v19, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v4, vcc, 32, v1
+; GFX8-GISEL-NEXT:    s_waitcnt vmcnt(14)
+; GFX8-GISEL-NEXT:    buffer_store_dword v5, v2, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v2, vcc, 40, v1
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v19, vcc, 36, v1
-; GFX8-GISEL-NEXT:    v_add_u32_e32 v6, vcc, 44, v1
-; GFX8-GISEL-NEXT:    buffer_store_dword v3, v1, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    v_add_u32_e32 v5, vcc, 44, v1
+; GFX8-GISEL-NEXT:    buffer_store_dword v9, v1, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    s_waitcnt vmcnt(14)
-; GFX8-GISEL-NEXT:    buffer_store_dword v7, v17, s[0:3], 0 offen
-; GFX8-GISEL-NEXT:    buffer_store_dword v8, v4, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    buffer_store_dword v6, v17, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    buffer_store_dword v7, v3, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    s_waitcnt vmcnt(14)
-; GFX8-GISEL-NEXT:    buffer_store_dword v9, v18, s[0:3], 0 offen
-; GFX8-GISEL-NEXT:    buffer_store_dword v10, v5, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    buffer_store_dword v8, v18, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    buffer_store_dword v10, v4, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    s_waitcnt vmcnt(14)
 ; GFX8-GISEL-NEXT:    buffer_store_dword v11, v19, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    buffer_store_dword v12, v2, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    s_waitcnt vmcnt(14)
-; GFX8-GISEL-NEXT:    buffer_store_dword v13, v6, s[0:3], 0 offen
+; GFX8-GISEL-NEXT:    buffer_store_dword v13, v5, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v2, vcc, 48, v1
 ; GFX8-GISEL-NEXT:    buffer_store_dword v14, v2, s[0:3], 0 offen
 ; GFX8-GISEL-NEXT:    v_add_u32_e32 v2, vcc, 52, v1
diff --git a/llvm/test/CodeGen/AMDGPU/function-args.ll b/llvm/test/CodeGen/AMDGPU/function-args.ll
index 81b8b36180746..a901d7f97eb37 100644
--- a/llvm/test/CodeGen/AMDGPU/function-args.ll
+++ b/llvm/test/CodeGen/AMDGPU/function-args.ll
@@ -3380,42 +3380,117 @@ define void @void_func_v32i32_v2i16_v2f16_v2bf16_v4bf16(<32 x i32> %arg0, <2 x i
 }
 
 define void @void_func_v32i32_v2i64_v2f64(<32 x i32> %arg0, <2 x i64> %arg1, <2 x double> %arg2) #0 {
-; CIGFX89-LABEL: void_func_v32i32_v2i64_v2f64:
-; CIGFX89:       ; %bb.0:
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CIGFX89-NEXT:    buffer_load_dword v31, off, s[0:3], s32
-; CIGFX89-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:20
-; CIGFX89-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:16
-; CIGFX89-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:12
-; CIGFX89-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
-; CIGFX89-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
-; CIGFX89-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:32
-; CIGFX89-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:28
-; CIGFX89-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:24
-; CIGFX89-NEXT:    s_mov_b32 s7, 0xf000
-; CIGFX89-NEXT:    s_mov_b32 s6, -1
-; CIGFX89-NEXT:    s_waitcnt vmcnt(8)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[28:31], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[24:27], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[20:23], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[32:35], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    buffer_store_dwordx4 v[36:39], off, s[4:7], 0
-; CIGFX89-NEXT:    s_waitcnt vmcnt(0)
-; CIGFX89-NEXT:    s_setpc_b64 s[30:31]
+; CI-LABEL: void_func_v32i32_v2i64_v2f64:
+; CI:       ; %bb.0:
+; CI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
+; CI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:32
+; CI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:28
+; CI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:24
+; CI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:16
+; CI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:12
+; CI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:8
+; CI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt vmcnt(7)
+; CI-NEXT:    buffer_store_dwordx4 v[28:31], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[24:27], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[20:23], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:20
+; CI-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[35:38], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[31:34], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    s_setpc_b64 s[30:31]
+;
+; VI-LABEL: void_func_v32i32_v2i64_v2f64:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
+; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_waitcnt vmcnt(7)
+; VI-NEXT:    buffer_store_dwordx4 v[28:31], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[24:27], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[20:23], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[35:38], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[31:34], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: void_func_v32i32_v2i64_v2f64:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
+; GFX9-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    s_mov_b32 s7, 0xf000
+; GFX9-NEXT:    s_mov_b32 s6, -1
+; GFX9-NEXT:    s_waitcnt vmcnt(7)
+; GFX9-NEXT:    buffer_store_dwordx4 v[28:31], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[24:27], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[20:23], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    s_nop 0
+; GFX9-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[35:38], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[31:34], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: void_func_v32i32_v2i64_v2f64:
 ; GFX11:       ; %bb.0:
@@ -3552,13 +3627,13 @@ define void @void_func_v32i32_v8i32_v8f32(<32 x i32> %arg0, <8 x i32> %arg1, <8
 ; CI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, -1
-; CI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:64
-; CI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:60
-; CI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:56
-; CI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:52
-; CI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:16
-; CI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:12
-; CI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:8
+; CI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:16
+; CI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:12
+; CI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
+; CI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; CI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:32
+; CI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:28
+; CI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:24
 ; CI-NEXT:    s_waitcnt vmcnt(7)
 ; CI-NEXT:    buffer_store_dwordx4 v[28:31], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
@@ -3570,29 +3645,29 @@ define void @void_func_v32i32_v8i32_v8f32(<32 x i32> %arg0, <8 x i32> %arg1, <8
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:4
-; CI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:32
-; CI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:28
-; CI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:24
-; CI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
-; CI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:48
-; CI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:44
-; CI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:40
-; CI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; CI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:20
+; CI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:48
+; CI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:44
+; CI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:40
+; CI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:64
+; CI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:60
+; CI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:56
+; CI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:52
+; CI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:36
 ; CI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[36:39], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[32:35], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: void_func_v32i32_v8i32_v8f32:
@@ -3601,13 +3676,13 @@ define void @void_func_v32i32_v8i32_v8f32(<32 x i32> %arg0, <8 x i32> %arg1, <8
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
-; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:64
-; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:56
-; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:52
-; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:24
 ; VI-NEXT:    s_waitcnt vmcnt(7)
 ; VI-NEXT:    buffer_store_dwordx4 v[28:31], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -3619,29 +3694,29 @@ define void @void_func_v32i32_v8i32_v8f32(<32 x i32> %arg0, <8 x i32> %arg1, <8
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:4
-; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:32
-; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:28
-; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:24
-; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:48
-; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:44
-; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:40
-; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; VI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:44
+; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:40
+; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:64
+; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:56
+; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:52
+; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:36
 ; VI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[36:39], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[32:35], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: void_func_v32i32_v8i32_v8f32:
@@ -3650,13 +3725,13 @@ define void @void_func_v32i32_v8i32_v8f32(<32 x i32> %arg0, <8 x i32> %arg1, <8
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX9-NEXT:    s_mov_b32 s7, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s6, -1
-; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:64
-; GFX9-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:56
-; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:52
-; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:24
 ; GFX9-NEXT:    s_waitcnt vmcnt(7)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[28:31], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -3668,15 +3743,15 @@ define void @void_func_v32i32_v8i32_v8f32(<32 x i32> %arg0, <8 x i32> %arg1, <8
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:32
-; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:24
-; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:48
-; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:40
-; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:36
+; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:44
+; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:40
+; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:64
+; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:56
+; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:52
+; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:36
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -3684,14 +3759,14 @@ define void @void_func_v32i32_v8i32_v8f32(<32 x i32> %arg0, <8 x i32> %arg1, <8
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[36:39], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[32:35], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: void_func_v32i32_v8i32_v8f32:
@@ -3791,40 +3866,40 @@ define void @void_func_v32i32_v16i32_v16f32(<32 x i32> %arg0, <16 x i32> %arg1,
 ; CI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:4
-; CI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:96
-; CI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:92
-; CI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:88
-; CI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:84
-; CI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:112
-; CI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:108
-; CI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:104
+; CI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:112
+; CI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:108
+; CI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:104
+; CI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:100
+; CI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:128
+; CI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:124
+; CI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:120
 ; CI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[32:35], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:100
-; CI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128
-; CI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124
-; CI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:120
-; CI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; CI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:80
-; CI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:76
-; CI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:72
-; CI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:68
+; CI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:116
+; CI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80
+; CI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76
+; CI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72
+; CI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:96
+; CI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:92
+; CI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:88
+; CI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:84
+; CI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:68
 ; CI-NEXT:    buffer_store_dwordx4 v[36:39], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[20:23], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: void_func_v32i32_v16i32_v16f32:
@@ -3864,40 +3939,40 @@ define void @void_func_v32i32_v16i32_v16f32(<32 x i32> %arg0, <16 x i32> %arg1,
 ; VI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:4
-; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:96
-; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:92
-; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:88
-; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:84
-; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:112
-; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:108
-; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:104
+; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:112
+; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:108
+; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:104
+; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:100
+; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:128
+; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:124
+; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:120
 ; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[32:35], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128
-; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124
-; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:120
-; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:80
-; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:76
-; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:72
-; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:68
+; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:116
+; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80
+; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76
+; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72
+; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:96
+; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:92
+; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:88
+; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:84
+; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:68
 ; VI-NEXT:    buffer_store_dwordx4 v[36:39], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[20:23], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: void_func_v32i32_v16i32_v16f32:
@@ -3938,27 +4013,27 @@ define void @void_func_v32i32_v16i32_v16f32(<32 x i32> %arg0, <16 x i32> %arg1,
 ; GFX9-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:96
-; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:92
-; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:88
-; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:84
-; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:112
-; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:108
-; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:104
+; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:112
+; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:108
+; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:104
+; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:100
+; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:128
+; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:124
+; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:120
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[32:35], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:100
-; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128
-; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124
-; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:120
-; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:80
-; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:76
-; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:72
-; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:68
+; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:116
+; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80
+; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76
+; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72
+; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:96
+; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:92
+; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:88
+; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:84
+; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:68
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dwordx4 v[36:39], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -3966,14 +4041,14 @@ define void @void_func_v32i32_v16i32_v16f32(<32 x i32> %arg0, <16 x i32> %arg1,
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[4:7], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[20:23], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: void_func_v32i32_v16i32_v16f32:
@@ -4259,9 +4334,9 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; CI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
 ; CI-NEXT:    s_mov_b32 s6, -1
-; CI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:60
-; CI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:64
-; CI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:48
+; CI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:48
+; CI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60
+; CI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64
 ; CI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:52
 ; CI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:56
 ; CI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:36
@@ -4275,16 +4350,16 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:28
-; CI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:32
-; CI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:20
-; CI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:24
 ; CI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:16
-; CI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:12
-; CI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:8
-; CI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:4
+; CI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:32
+; CI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:28
+; CI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:24
+; CI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:20
+; CI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:16
+; CI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:12
+; CI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:8
+; CI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:4
 ; CI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:44
 ; CI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
@@ -4292,15 +4367,15 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_store_byte v33, off, s[4:7], 0
+; CI-NEXT:    buffer_store_byte v34, off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_store_byte v32, off, s[4:7], 0
+; CI-NEXT:    buffer_store_byte v33, off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_byte v36, off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_byte v35, off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_store_byte v34, off, s[4:7], 0
+; CI-NEXT:    buffer_store_byte v32, off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_byte v20, off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
@@ -4308,14 +4383,6 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_byte v37, off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_store_byte v17, off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_store_byte v16, off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_store_byte v19, off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_store_byte v18, off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_byte v12, off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_byte v13, off, s[4:7], 0
@@ -4324,6 +4391,14 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    buffer_store_byte v15, off, s[4:7], 0
 ; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_byte v16, off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_byte v17, off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_byte v18, off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_store_byte v19, off, s[4:7], 0
+; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: void_func_v32i32_v16i8:
@@ -4332,9 +4407,9 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
-; VI-NEXT:    buffer_load_ubyte v32, off, s[0:3], s32 offset:60
-; VI-NEXT:    buffer_load_ubyte v33, off, s[0:3], s32 offset:64
-; VI-NEXT:    buffer_load_ubyte v34, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_ubyte v32, off, s[0:3], s32 offset:48
+; VI-NEXT:    buffer_load_ubyte v33, off, s[0:3], s32 offset:60
+; VI-NEXT:    buffer_load_ubyte v34, off, s[0:3], s32 offset:64
 ; VI-NEXT:    buffer_load_ubyte v35, off, s[0:3], s32 offset:52
 ; VI-NEXT:    buffer_load_ubyte v36, off, s[0:3], s32 offset:56
 ; VI-NEXT:    buffer_load_ubyte v37, off, s[0:3], s32 offset:36
@@ -4348,16 +4423,16 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_load_ubyte v16, off, s[0:3], s32 offset:28
-; VI-NEXT:    buffer_load_ubyte v17, off, s[0:3], s32 offset:32
-; VI-NEXT:    buffer_load_ubyte v18, off, s[0:3], s32 offset:20
-; VI-NEXT:    buffer_load_ubyte v19, off, s[0:3], s32 offset:24
 ; VI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_load_ubyte v12, off, s[0:3], s32 offset:16
-; VI-NEXT:    buffer_load_ubyte v13, off, s[0:3], s32 offset:12
-; VI-NEXT:    buffer_load_ubyte v14, off, s[0:3], s32 offset:8
-; VI-NEXT:    buffer_load_ubyte v15, off, s[0:3], s32 offset:4
+; VI-NEXT:    buffer_load_ubyte v12, off, s[0:3], s32 offset:32
+; VI-NEXT:    buffer_load_ubyte v13, off, s[0:3], s32 offset:28
+; VI-NEXT:    buffer_load_ubyte v14, off, s[0:3], s32 offset:24
+; VI-NEXT:    buffer_load_ubyte v15, off, s[0:3], s32 offset:20
+; VI-NEXT:    buffer_load_ubyte v16, off, s[0:3], s32 offset:16
+; VI-NEXT:    buffer_load_ubyte v17, off, s[0:3], s32 offset:12
+; VI-NEXT:    buffer_load_ubyte v18, off, s[0:3], s32 offset:8
+; VI-NEXT:    buffer_load_ubyte v19, off, s[0:3], s32 offset:4
 ; VI-NEXT:    buffer_load_ubyte v20, off, s[0:3], s32 offset:44
 ; VI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -4365,15 +4440,15 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_byte v33, off, s[4:7], 0
+; VI-NEXT:    buffer_store_byte v34, off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_byte v32, off, s[4:7], 0
+; VI-NEXT:    buffer_store_byte v33, off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_byte v36, off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_byte v35, off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_byte v34, off, s[4:7], 0
+; VI-NEXT:    buffer_store_byte v32, off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_byte v20, off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
@@ -4381,14 +4456,6 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_byte v37, off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_byte v17, off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_byte v16, off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_byte v19, off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_store_byte v18, off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_byte v12, off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_byte v13, off, s[4:7], 0
@@ -4397,6 +4464,14 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_store_byte v15, off, s[4:7], 0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_byte v16, off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_byte v17, off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_byte v18, off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_store_byte v19, off, s[4:7], 0
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: void_func_v32i32_v16i8:
@@ -4405,9 +4480,9 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX9-NEXT:    s_mov_b32 s7, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s6, -1
-; GFX9-NEXT:    buffer_load_ubyte v32, off, s[0:3], s32 offset:60
-; GFX9-NEXT:    buffer_load_ubyte v33, off, s[0:3], s32 offset:64
-; GFX9-NEXT:    buffer_load_ubyte v34, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_ubyte v32, off, s[0:3], s32 offset:48
+; GFX9-NEXT:    buffer_load_ubyte v33, off, s[0:3], s32 offset:60
+; GFX9-NEXT:    buffer_load_ubyte v34, off, s[0:3], s32 offset:64
 ; GFX9-NEXT:    buffer_load_ubyte v35, off, s[0:3], s32 offset:52
 ; GFX9-NEXT:    buffer_load_ubyte v36, off, s[0:3], s32 offset:56
 ; GFX9-NEXT:    buffer_load_ubyte v37, off, s[0:3], s32 offset:36
@@ -4421,18 +4496,17 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[16:19], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_load_ubyte v16, off, s[0:3], s32 offset:28
-; GFX9-NEXT:    buffer_load_ubyte v17, off, s[0:3], s32 offset:32
-; GFX9-NEXT:    buffer_load_ubyte v18, off, s[0:3], s32 offset:20
-; GFX9-NEXT:    buffer_load_ubyte v19, off, s[0:3], s32 offset:24
-; GFX9-NEXT:    buffer_load_ubyte v20, off, s[0:3], s32 offset:44
-; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_load_ubyte v12, off, s[0:3], s32 offset:16
-; GFX9-NEXT:    buffer_load_ubyte v13, off, s[0:3], s32 offset:12
-; GFX9-NEXT:    buffer_load_ubyte v14, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    buffer_load_ubyte v15, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ubyte v12, off, s[0:3], s32 offset:32
+; GFX9-NEXT:    buffer_load_ubyte v13, off, s[0:3], s32 offset:28
+; GFX9-NEXT:    buffer_load_ubyte v14, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    buffer_load_ubyte v15, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    buffer_load_ubyte v16, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    buffer_load_ubyte v17, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    buffer_load_ubyte v18, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_load_ubyte v19, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_load_ubyte v20, off, s[0:3], s32 offset:44
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -4440,15 +4514,15 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_byte v33, off, s[4:7], 0
+; GFX9-NEXT:    buffer_store_byte v34, off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_byte v32, off, s[4:7], 0
+; GFX9-NEXT:    buffer_store_byte v33, off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_byte v36, off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_byte v35, off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_byte v34, off, s[4:7], 0
+; GFX9-NEXT:    buffer_store_byte v32, off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_byte v20, off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
@@ -4456,14 +4530,6 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_byte v37, off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_byte v17, off, s[4:7], 0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_byte v16, off, s[4:7], 0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_byte v19, off, s[4:7], 0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_store_byte v18, off, s[4:7], 0
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_byte v12, off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_byte v13, off, s[4:7], 0
@@ -4472,6 +4538,14 @@ define void @void_func_v32i32_v16i8(<32 x i32> %arg0, <16 x i8> %arg1) #0 {
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_store_byte v15, off, s[4:7], 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_byte v16, off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_byte v17, off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_byte v18, off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_store_byte v19, off, s[4:7], 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-TRUE16-LABEL: void_func_v32i32_v16i8:
diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-return-types.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-return-types.ll
index 75619532a2e37..668219875db72 100644
--- a/llvm/test/CodeGen/AMDGPU/gfx-callable-return-types.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-return-types.ll
@@ -2450,22 +2450,21 @@ define amdgpu_gfx <72 x i32> @return_72xi32(<72 x i32> %val) #1 {
 ; GFX10-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:28
 ; GFX10-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:24
 ; GFX10-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:20
-; GFX10-NEXT:    s_clause 0x3
-; GFX10-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:12
-; GFX10-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:16
-; GFX10-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:20
-; GFX10-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:24
 ; GFX10-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:16
 ; GFX10-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:12
-; GFX10-NEXT:    s_clause 0x3
-; GFX10-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:4
-; GFX10-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:8
-; GFX10-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:160
-; GFX10-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:28
 ; GFX10-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:8
-; GFX10-NEXT:    buffer_load_dword v3, off, s[0:3], s32
+; GFX10-NEXT:    s_clause 0x8
+; GFX10-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:28
+; GFX10-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:24
+; GFX10-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:20
+; GFX10-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:16
+; GFX10-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:12
+; GFX10-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:8
+; GFX10-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:4
+; GFX10-NEXT:    buffer_load_dword v10, off, s[0:3], s32
+; GFX10-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:160
 ; GFX10-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
-; GFX10-NEXT:    s_waitcnt vmcnt(2)
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_store_dword v27, v0, s[0:3], 0 offen offset:284
 ; GFX10-NEXT:    buffer_store_dword v51, v0, s[0:3], 0 offen offset:280
 ; GFX10-NEXT:    buffer_store_dword v50, v0, s[0:3], 0 offen offset:276
@@ -2499,16 +2498,14 @@ define amdgpu_gfx <72 x i32> @return_72xi32(<72 x i32> %val) #1 {
 ; GFX10-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:164
 ; GFX10-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:160
 ; GFX10-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:156
-; GFX10-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:152
-; GFX10-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:148
-; GFX10-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:144
-; GFX10-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:140
-; GFX10-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:136
-; GFX10-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:132
-; GFX10-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:128
-; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:124
+; GFX10-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:152
+; GFX10-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:148
+; GFX10-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:144
+; GFX10-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:140
+; GFX10-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:136
+; GFX10-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:132
+; GFX10-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:128
+; GFX10-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:124
 ; GFX10-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/global_atomics.ll b/llvm/test/CodeGen/AMDGPU/global_atomics.ll
index 3e15b135eeab9..b7b69ed9f53ba 100644
--- a/llvm/test/CodeGen/AMDGPU/global_atomics.ll
+++ b/llvm/test/CodeGen/AMDGPU/global_atomics.ll
@@ -4276,14 +4276,14 @@ define amdgpu_kernel void @atomic_cmpxchg_i32_addr64_offset(ptr addrspace(1) %ou
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
 ; GFX9-NEXT:    s_load_dword s6, s[4:5], 0x2c
+; GFX9-NEXT:    s_load_dword s7, s[4:5], 0x3c
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
-; GFX9-NEXT:    s_load_dword s4, s[4:5], 0x3c
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
 ; GFX9-NEXT:    s_add_u32 s0, s2, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s3, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s6
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    global_atomic_cmpswap v2, v[0:1], s[0:1] offset:16
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
@@ -4523,14 +4523,14 @@ define amdgpu_kernel void @atomic_cmpxchg_i32_addr64(ptr addrspace(1) %out, i32
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
 ; GFX9-NEXT:    s_load_dword s6, s[4:5], 0x2c
+; GFX9-NEXT:    s_load_dword s7, s[4:5], 0x3c
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
-; GFX9-NEXT:    s_load_dword s4, s[4:5], 0x3c
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
 ; GFX9-NEXT:    s_add_u32 s0, s2, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s3, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s6
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    global_atomic_cmpswap v2, v[0:1], s[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
diff --git a/llvm/test/CodeGen/AMDGPU/idot4s.ll b/llvm/test/CodeGen/AMDGPU/idot4s.ll
index d28f0a190e117..9e7968f1acb84 100644
--- a/llvm/test/CodeGen/AMDGPU/idot4s.ll
+++ b/llvm/test/CodeGen/AMDGPU/idot4s.ll
@@ -2903,13 +2903,13 @@ define amdgpu_kernel void @idot4_acc32_3src_3ele_src0(ptr addrspace(1) %src1,
 ; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc0c0c01
 ; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc020101
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    global_load_dword v1, v0, s[12:13]
-; GFX9-DL-NEXT:    global_load_dword v2, v0, s[8:9]
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[8:9]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[12:13]
 ; GFX9-DL-NEXT:    global_load_dword v3, v0, s[10:11]
 ; GFX9-DL-NEXT:    s_load_dword s3, s[14:15], 0x0
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v2, s0
+; GFX9-DL-NEXT:    v_perm_b32 v1, v2, v1, s0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s1
 ; GFX9-DL-NEXT:    v_or_b32_e32 v1, v1, v2
@@ -2925,12 +2925,12 @@ define amdgpu_kernel void @idot4_acc32_3src_3ele_src0(ptr addrspace(1) %src1,
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x2
-; GFX10-DL-NEXT:    global_load_dword v1, v0, s[12:13]
-; GFX10-DL-NEXT:    global_load_dword v2, v0, s[8:9]
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[8:9]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[12:13]
 ; GFX10-DL-NEXT:    global_load_dword v3, v0, s[10:11]
 ; GFX10-DL-NEXT:    s_load_dword s0, s[14:15], 0x0
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v2, 0xc06010c
+; GFX10-DL-NEXT:    v_perm_b32 v0, v2, v1, 0xc06010c
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc0c0c01
 ; GFX10-DL-NEXT:    v_perm_b32 v2, v3, v3, 0xc020101
@@ -2950,12 +2950,12 @@ define amdgpu_kernel void @idot4_acc32_3src_3ele_src0(ptr addrspace(1) %src1,
 ; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-DL-NEXT:    s_clause 0x2
-; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
-; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
 ; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[2:3]
 ; GFX11-DL-NEXT:    s_load_b32 s0, s[6:7], 0x0
 ; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v2, 0xc06010c
+; GFX11-DL-NEXT:    v_perm_b32 v1, v2, v1, 0xc06010c
 ; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-DL-NEXT:    v_perm_b32 v2, v0, v0, 0xc0c0c01
 ; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020101
diff --git a/llvm/test/CodeGen/AMDGPU/idot4u.ll b/llvm/test/CodeGen/AMDGPU/idot4u.ll
index df77757443391..f995f426c6372 100644
--- a/llvm/test/CodeGen/AMDGPU/idot4u.ll
+++ b/llvm/test/CodeGen/AMDGPU/idot4u.ll
@@ -4454,13 +4454,13 @@ define amdgpu_kernel void @udot4_acc32_3src_3ele_src0(ptr addrspace(1) %src1,
 ; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc0c0c01
 ; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc020101
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    global_load_dword v1, v0, s[12:13]
-; GFX9-DL-NEXT:    global_load_dword v2, v0, s[8:9]
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[8:9]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[12:13]
 ; GFX9-DL-NEXT:    global_load_dword v3, v0, s[10:11]
 ; GFX9-DL-NEXT:    s_load_dword s3, s[14:15], 0x0
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v2, s0
+; GFX9-DL-NEXT:    v_perm_b32 v1, v2, v1, s0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s1
 ; GFX9-DL-NEXT:    v_or_b32_e32 v1, v1, v2
@@ -4476,12 +4476,12 @@ define amdgpu_kernel void @udot4_acc32_3src_3ele_src0(ptr addrspace(1) %src1,
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x2
-; GFX10-DL-NEXT:    global_load_dword v1, v0, s[12:13]
-; GFX10-DL-NEXT:    global_load_dword v2, v0, s[8:9]
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[8:9]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[12:13]
 ; GFX10-DL-NEXT:    global_load_dword v3, v0, s[10:11]
 ; GFX10-DL-NEXT:    s_load_dword s0, s[14:15], 0x0
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v2, 0xc06010c
+; GFX10-DL-NEXT:    v_perm_b32 v0, v2, v1, 0xc06010c
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc0c0c01
 ; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
@@ -4500,12 +4500,12 @@ define amdgpu_kernel void @udot4_acc32_3src_3ele_src0(ptr addrspace(1) %src1,
 ; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-DL-NEXT:    s_clause 0x2
-; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
-; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
 ; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[2:3]
 ; GFX11-DL-NEXT:    s_load_b32 s0, s[6:7], 0x0
 ; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v2, 0xc06010c
+; GFX11-DL-NEXT:    v_perm_b32 v1, v2, v1, 0xc06010c
 ; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-DL-NEXT:    v_perm_b32 v2, v0, v0, 0xc0c0c01
 ; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020101
diff --git a/llvm/test/CodeGen/AMDGPU/insert_vector_elt.ll b/llvm/test/CodeGen/AMDGPU/insert_vector_elt.ll
index f6330f4eb8216..be16fac4c53f7 100644
--- a/llvm/test/CodeGen/AMDGPU/insert_vector_elt.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert_vector_elt.ll
@@ -1101,14 +1101,14 @@ define amdgpu_kernel void @dynamic_insertelement_v3i32(ptr addrspace(1) %out, <3
 define amdgpu_kernel void @dynamic_insertelement_v4i32(ptr addrspace(1) %out, <4 x i32> %a, i32 %b, [8 x i32], i32 %val) nounwind {
 ; SI-LABEL: dynamic_insertelement_v4i32:
 ; SI:       ; %bb.0:
-; SI-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x4
 ; SI-NEXT:    s_load_dword s10, s[8:9], 0x8
 ; SI-NEXT:    s_load_dword s11, s[8:9], 0x11
+; SI-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x4
 ; SI-NEXT:    s_load_dwordx2 s[4:5], s[8:9], 0x0
 ; SI-NEXT:    s_mov_b32 s7, 0x100f000
-; SI-NEXT:    s_mov_b32 s6, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    s_cmp_eq_u32 s10, 3
+; SI-NEXT:    s_mov_b32 s6, -1
 ; SI-NEXT:    s_cselect_b32 s3, s11, s3
 ; SI-NEXT:    s_cmp_eq_u32 s10, 2
 ; SI-NEXT:    s_cselect_b32 s2, s11, s2
@@ -1125,14 +1125,14 @@ define amdgpu_kernel void @dynamic_insertelement_v4i32(ptr addrspace(1) %out, <4
 ;
 ; VI-LABEL: dynamic_insertelement_v4i32:
 ; VI:       ; %bb.0:
-; VI-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x10
 ; VI-NEXT:    s_load_dword s10, s[8:9], 0x20
 ; VI-NEXT:    s_load_dword s11, s[8:9], 0x44
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x10
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[8:9], 0x0
 ; VI-NEXT:    s_mov_b32 s7, 0x1100f000
-; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_cmp_eq_u32 s10, 3
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_cselect_b32 s3, s11, s3
 ; VI-NEXT:    s_cmp_eq_u32 s10, 2
 ; VI-NEXT:    s_cselect_b32 s2, s11, s2
diff --git a/llvm/test/CodeGen/AMDGPU/kernel-args.ll b/llvm/test/CodeGen/AMDGPU/kernel-args.ll
index 9df995b5a7066..a18b5b5396f63 100644
--- a/llvm/test/CodeGen/AMDGPU/kernel-args.ll
+++ b/llvm/test/CodeGen/AMDGPU/kernel-args.ll
@@ -1664,8 +1664,8 @@ entry:
 define amdgpu_kernel void @v5i16_arg(ptr addrspace(1) nocapture %out, <5 x i16> %in) nounwind {
 ; SI-LABEL: v5i16_arg:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s6, s[4:5], 0xf
 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
+; SI-NEXT:    s_load_dword s6, s[4:5], 0xf
 ; SI-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_mov_b32 s2, -1
@@ -5191,16 +5191,16 @@ define amdgpu_kernel void @array_3xi16(i8 %arg0, [3 x i16] %arg1) {
 ; VI-NEXT:    s_addc_u32 s1, s5, 0
 ; VI-NEXT:    s_add_u32 s2, s0, 2
 ; VI-NEXT:    s_addc_u32 s3, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    s_add_u32 s0, s4, 42
-; VI-NEXT:    s_addc_u32 s1, s5, 0
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    flat_load_ushort v4, v[0:1]
-; VI-NEXT:    flat_load_ushort v2, v[2:3]
+; VI-NEXT:    s_add_u32 s0, s4, 42
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    v_mov_b32_e32 v5, s1
 ; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v4, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    flat_load_ushort v4, v[4:5]
+; VI-NEXT:    flat_load_ushort v2, v[2:3]
 ; VI-NEXT:    flat_load_ushort v0, v[0:1]
 ; VI-NEXT:    s_load_dword s0, s[4:5], 0x24
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5208,10 +5208,10 @@ define amdgpu_kernel void @array_3xi16(i8 %arg0, [3 x i16] %arg1) {
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    flat_store_byte v[0:1], v1
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    flat_store_short v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    flat_store_short v[0:1], v4
 ; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    flat_store_short v[0:1], v2
+; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    flat_store_short v[0:1], v0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane64.ptr.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane64.ptr.ll
index f7c37caf41eab..393d8c1a1bf2f 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane64.ptr.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane64.ptr.ll
@@ -69,14 +69,13 @@ define amdgpu_kernel void @test_v3p3(ptr addrspace(1) %out, <3 x ptr addrspace(3
 ; GFX11-SDAG-NEXT:    s_clause 0x1
 ; GFX11-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x34
 ; GFX11-SDAG-NEXT:    s_load_b64 s[4:5], s[4:5], 0x24
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v3, s0
 ; GFX11-SDAG-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s1
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, s0
-; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-SDAG-NEXT:    v_permlane64_b32 v2, v0
 ; GFX11-SDAG-NEXT:    v_permlane64_b32 v1, v1
-; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_4)
 ; GFX11-SDAG-NEXT:    v_permlane64_b32 v0, v3
 ; GFX11-SDAG-NEXT:    global_store_b96 v4, v[0:2], s[4:5]
 ; GFX11-SDAG-NEXT:    s_endpgm
@@ -108,14 +107,13 @@ define amdgpu_kernel void @test_v3p5(ptr addrspace(1) %out, <3 x ptr addrspace(5
 ; GFX11-SDAG-NEXT:    s_clause 0x1
 ; GFX11-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x34
 ; GFX11-SDAG-NEXT:    s_load_b64 s[4:5], s[4:5], 0x24
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v3, s0
 ; GFX11-SDAG-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s1
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, s0
-; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-SDAG-NEXT:    v_permlane64_b32 v2, v0
 ; GFX11-SDAG-NEXT:    v_permlane64_b32 v1, v1
-; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_4)
 ; GFX11-SDAG-NEXT:    v_permlane64_b32 v0, v3
 ; GFX11-SDAG-NEXT:    global_store_b96 v4, v[0:2], s[4:5]
 ; GFX11-SDAG-NEXT:    s_endpgm
@@ -147,14 +145,13 @@ define amdgpu_kernel void @test_v3p6(ptr addrspace(1) %out, <3 x ptr addrspace(6
 ; GFX11-SDAG-NEXT:    s_clause 0x1
 ; GFX11-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x34
 ; GFX11-SDAG-NEXT:    s_load_b64 s[4:5], s[4:5], 0x24
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v3, s0
 ; GFX11-SDAG-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s1
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, s0
-; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-SDAG-NEXT:    v_permlane64_b32 v2, v0
 ; GFX11-SDAG-NEXT:    v_permlane64_b32 v1, v1
-; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_4)
 ; GFX11-SDAG-NEXT:    v_permlane64_b32 v0, v3
 ; GFX11-SDAG-NEXT:    global_store_b96 v4, v[0:2], s[4:5]
 ; GFX11-SDAG-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.readfirstlane.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.readfirstlane.ll
index c0afc0a443955..49a334b8b6c52 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.readfirstlane.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.readfirstlane.ll
@@ -984,9 +984,9 @@ define void @test_readfirstlane_v32f32(ptr addrspace(1) %out, <32 x float> %src)
 ; CHECK-SDAG-NEXT:    s_xor_saveexec_b64 s[4:5], -1
 ; CHECK-SDAG-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
 ; CHECK-SDAG-NEXT:    s_mov_b64 exec, s[4:5]
-; CHECK-SDAG-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
-; CHECK-SDAG-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
 ; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s61, v27
+; CHECK-SDAG-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:8
+; CHECK-SDAG-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4
 ; CHECK-SDAG-NEXT:    buffer_load_dword v27, off, s[0:3], s32
 ; CHECK-SDAG-NEXT:    v_writelane_b32 v31, s36, 0
 ; CHECK-SDAG-NEXT:    v_writelane_b32 v31, s37, 1
@@ -1033,9 +1033,9 @@ define void @test_readfirstlane_v32f32(ptr addrspace(1) %out, <32 x float> %src)
 ; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s41, v7
 ; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s40, v6
 ; CHECK-SDAG-NEXT:    s_waitcnt vmcnt(2)
-; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s66, v0
+; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s67, v0
 ; CHECK-SDAG-NEXT:    s_waitcnt vmcnt(1)
-; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s67, v1
+; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s66, v1
 ; CHECK-SDAG-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s65, v27
 ; CHECK-SDAG-NEXT:    ;;#ASMSTART
@@ -1429,9 +1429,9 @@ define void @test_readfirstlane_v32i32(ptr addrspace(1) %out, <32 x i32> %src) {
 ; CHECK-SDAG-NEXT:    s_xor_saveexec_b64 s[4:5], -1
 ; CHECK-SDAG-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
 ; CHECK-SDAG-NEXT:    s_mov_b64 exec, s[4:5]
-; CHECK-SDAG-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
-; CHECK-SDAG-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
 ; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s61, v27
+; CHECK-SDAG-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:8
+; CHECK-SDAG-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4
 ; CHECK-SDAG-NEXT:    buffer_load_dword v27, off, s[0:3], s32
 ; CHECK-SDAG-NEXT:    v_writelane_b32 v31, s36, 0
 ; CHECK-SDAG-NEXT:    v_writelane_b32 v31, s37, 1
@@ -1478,9 +1478,9 @@ define void @test_readfirstlane_v32i32(ptr addrspace(1) %out, <32 x i32> %src) {
 ; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s41, v7
 ; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s40, v6
 ; CHECK-SDAG-NEXT:    s_waitcnt vmcnt(2)
-; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s66, v0
+; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s67, v0
 ; CHECK-SDAG-NEXT:    s_waitcnt vmcnt(1)
-; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s67, v1
+; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s66, v1
 ; CHECK-SDAG-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-SDAG-NEXT:    v_readfirstlane_b32 s65, v27
 ; CHECK-SDAG-NEXT:    ;;#ASMSTART
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.maximum.f64.ll b/llvm/test/CodeGen/AMDGPU/llvm.maximum.f64.ll
index a18e5ace18704..f971080e02c5b 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.maximum.f64.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.maximum.f64.ll
@@ -2522,9 +2522,9 @@ define <16 x double> @v_maximum_v16f64(<16 x double> %src0, <16 x double> %src1)
 ; GFX10-LABEL: v_maximum_v16f64:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_clause 0x19
-; GFX10-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:8
+; GFX10-NEXT:    s_clause 0x18
 ; GFX10-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
+; GFX10-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:8
 ; GFX10-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:16
 ; GFX10-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:12
 ; GFX10-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:24
@@ -2548,22 +2548,21 @@ define <16 x double> @v_maximum_v16f64(<16 x double> %src0, <16 x double> %src1)
 ; GFX10-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:80
 ; GFX10-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:76
 ; GFX10-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:72
-; GFX10-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:104
-; GFX10-NEXT:    s_waitcnt vmcnt(24)
+; GFX10-NEXT:    s_waitcnt vmcnt(23)
 ; GFX10-NEXT:    v_max_f64 v[82:83], v[0:1], v[31:32]
 ; GFX10-NEXT:    v_cmp_u_f64_e32 vcc_lo, v[0:1], v[31:32]
-; GFX10-NEXT:    s_waitcnt vmcnt(22)
+; GFX10-NEXT:    s_waitcnt vmcnt(21)
 ; GFX10-NEXT:    v_max_f64 v[84:85], v[2:3], v[33:34]
 ; GFX10-NEXT:    v_cmp_u_f64_e64 s4, v[2:3], v[33:34]
-; GFX10-NEXT:    s_clause 0x3
-; GFX10-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:120
-; GFX10-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; GFX10-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:112
-; GFX10-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108
-; GFX10-NEXT:    s_waitcnt vmcnt(24)
+; GFX10-NEXT:    s_waitcnt vmcnt(19)
 ; GFX10-NEXT:    v_max_f64 v[32:33], v[4:5], v[35:36]
 ; GFX10-NEXT:    v_cmp_u_f64_e64 s5, v[4:5], v[35:36]
-; GFX10-NEXT:    s_clause 0x2
+; GFX10-NEXT:    s_clause 0x7
+; GFX10-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112
+; GFX10-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:104
+; GFX10-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
+; GFX10-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120
+; GFX10-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116
 ; GFX10-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX10-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:128
 ; GFX10-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:124
@@ -2593,9 +2592,6 @@ define <16 x double> @v_maximum_v16f64(<16 x double> %src0, <16 x double> %src1)
 ; GFX10-NEXT:    v_max_f64 v[50:51], v[18:19], v[80:81]
 ; GFX10-NEXT:    v_max_f64 v[70:71], v[22:23], v[68:69]
 ; GFX10-NEXT:    v_cmp_u_f64_e64 s14, v[22:23], v[68:69]
-; GFX10-NEXT:    s_waitcnt vmcnt(7)
-; GFX10-NEXT:    v_max_f64 v[68:69], v[24:25], v[66:67]
-; GFX10-NEXT:    v_cmp_u_f64_e64 s15, v[24:25], v[66:67]
 ; GFX10-NEXT:    v_cndmask_b32_e64 v6, v34, 0, s6
 ; GFX10-NEXT:    v_cndmask_b32_e64 v7, v35, 0x7ff80000, s6
 ; GFX10-NEXT:    v_cndmask_b32_e64 v8, v48, 0, s7
@@ -2614,27 +2610,30 @@ define <16 x double> @v_maximum_v16f64(<16 x double> %src0, <16 x double> %src1)
 ; GFX10-NEXT:    v_cndmask_b32_e64 v21, v65, 0x7ff80000, s13
 ; GFX10-NEXT:    v_cndmask_b32_e64 v22, v70, 0, s14
 ; GFX10-NEXT:    v_cndmask_b32_e64 v23, v71, 0x7ff80000, s14
-; GFX10-NEXT:    v_cndmask_b32_e64 v24, v68, 0, s15
-; GFX10-NEXT:    v_cndmask_b32_e64 v25, v69, 0x7ff80000, s15
+; GFX10-NEXT:    s_waitcnt vmcnt(6)
+; GFX10-NEXT:    v_max_f64 v[68:69], v[24:25], v[66:67]
+; GFX10-NEXT:    v_cmp_u_f64_e64 s15, v[24:25], v[66:67]
 ; GFX10-NEXT:    s_waitcnt vmcnt(5)
-; GFX10-NEXT:    v_max_f64 v[80:81], v[28:29], v[0:1]
-; GFX10-NEXT:    v_cmp_u_f64_e64 s17, v[28:29], v[0:1]
+; GFX10-NEXT:    v_max_f64 v[66:67], v[26:27], v[0:1]
+; GFX10-NEXT:    v_cmp_u_f64_e64 s16, v[26:27], v[0:1]
 ; GFX10-NEXT:    s_waitcnt vmcnt(3)
-; GFX10-NEXT:    v_max_f64 v[66:67], v[26:27], v[2:3]
-; GFX10-NEXT:    v_cmp_u_f64_e64 s16, v[26:27], v[2:3]
-; GFX10-NEXT:    v_cndmask_b32_e64 v0, v82, 0, vcc_lo
+; GFX10-NEXT:    v_max_f64 v[80:81], v[28:29], v[2:3]
+; GFX10-NEXT:    v_cmp_u_f64_e64 s17, v[28:29], v[2:3]
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    v_max_f64 v[86:87], v[30:31], v[4:5]
 ; GFX10-NEXT:    v_cmp_u_f64_e64 s18, v[30:31], v[4:5]
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v82, 0, vcc_lo
 ; GFX10-NEXT:    v_cndmask_b32_e64 v1, v83, 0x7ff80000, vcc_lo
 ; GFX10-NEXT:    v_cndmask_b32_e64 v2, v84, 0, s4
 ; GFX10-NEXT:    v_cndmask_b32_e64 v3, v85, 0x7ff80000, s4
 ; GFX10-NEXT:    v_cndmask_b32_e64 v4, v32, 0, s5
 ; GFX10-NEXT:    v_cndmask_b32_e64 v5, v33, 0x7ff80000, s5
-; GFX10-NEXT:    v_cndmask_b32_e64 v28, v80, 0, s17
-; GFX10-NEXT:    v_cndmask_b32_e64 v29, v81, 0x7ff80000, s17
+; GFX10-NEXT:    v_cndmask_b32_e64 v24, v68, 0, s15
+; GFX10-NEXT:    v_cndmask_b32_e64 v25, v69, 0x7ff80000, s15
 ; GFX10-NEXT:    v_cndmask_b32_e64 v26, v66, 0, s16
 ; GFX10-NEXT:    v_cndmask_b32_e64 v27, v67, 0x7ff80000, s16
+; GFX10-NEXT:    v_cndmask_b32_e64 v28, v80, 0, s17
+; GFX10-NEXT:    v_cndmask_b32_e64 v29, v81, 0x7ff80000, s17
 ; GFX10-NEXT:    v_cndmask_b32_e64 v30, v86, 0, s18
 ; GFX10-NEXT:    v_cndmask_b32_e64 v31, v87, 0x7ff80000, s18
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.minimum.f64.ll b/llvm/test/CodeGen/AMDGPU/llvm.minimum.f64.ll
index 4c413af878462..dfd67873c3b86 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.minimum.f64.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.minimum.f64.ll
@@ -2522,9 +2522,9 @@ define <16 x double> @v_minimum_v16f64(<16 x double> %src0, <16 x double> %src1)
 ; GFX10-LABEL: v_minimum_v16f64:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_clause 0x19
-; GFX10-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:8
+; GFX10-NEXT:    s_clause 0x18
 ; GFX10-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
+; GFX10-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:8
 ; GFX10-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:16
 ; GFX10-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:12
 ; GFX10-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:24
@@ -2548,22 +2548,21 @@ define <16 x double> @v_minimum_v16f64(<16 x double> %src0, <16 x double> %src1)
 ; GFX10-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:80
 ; GFX10-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:76
 ; GFX10-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:72
-; GFX10-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:104
-; GFX10-NEXT:    s_waitcnt vmcnt(24)
+; GFX10-NEXT:    s_waitcnt vmcnt(23)
 ; GFX10-NEXT:    v_min_f64 v[82:83], v[0:1], v[31:32]
 ; GFX10-NEXT:    v_cmp_u_f64_e32 vcc_lo, v[0:1], v[31:32]
-; GFX10-NEXT:    s_waitcnt vmcnt(22)
+; GFX10-NEXT:    s_waitcnt vmcnt(21)
 ; GFX10-NEXT:    v_min_f64 v[84:85], v[2:3], v[33:34]
 ; GFX10-NEXT:    v_cmp_u_f64_e64 s4, v[2:3], v[33:34]
-; GFX10-NEXT:    s_clause 0x3
-; GFX10-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:120
-; GFX10-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
-; GFX10-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:112
-; GFX10-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108
-; GFX10-NEXT:    s_waitcnt vmcnt(24)
+; GFX10-NEXT:    s_waitcnt vmcnt(19)
 ; GFX10-NEXT:    v_min_f64 v[32:33], v[4:5], v[35:36]
 ; GFX10-NEXT:    v_cmp_u_f64_e64 s5, v[4:5], v[35:36]
-; GFX10-NEXT:    s_clause 0x2
+; GFX10-NEXT:    s_clause 0x7
+; GFX10-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112
+; GFX10-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:104
+; GFX10-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108
+; GFX10-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120
+; GFX10-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116
 ; GFX10-NEXT:    buffer_load_dword v31, off, s[0:3], s32
 ; GFX10-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:128
 ; GFX10-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:124
@@ -2593,9 +2592,6 @@ define <16 x double> @v_minimum_v16f64(<16 x double> %src0, <16 x double> %src1)
 ; GFX10-NEXT:    v_min_f64 v[50:51], v[18:19], v[80:81]
 ; GFX10-NEXT:    v_min_f64 v[70:71], v[22:23], v[68:69]
 ; GFX10-NEXT:    v_cmp_u_f64_e64 s14, v[22:23], v[68:69]
-; GFX10-NEXT:    s_waitcnt vmcnt(7)
-; GFX10-NEXT:    v_min_f64 v[68:69], v[24:25], v[66:67]
-; GFX10-NEXT:    v_cmp_u_f64_e64 s15, v[24:25], v[66:67]
 ; GFX10-NEXT:    v_cndmask_b32_e64 v6, v34, 0, s6
 ; GFX10-NEXT:    v_cndmask_b32_e64 v7, v35, 0x7ff80000, s6
 ; GFX10-NEXT:    v_cndmask_b32_e64 v8, v48, 0, s7
@@ -2614,27 +2610,30 @@ define <16 x double> @v_minimum_v16f64(<16 x double> %src0, <16 x double> %src1)
 ; GFX10-NEXT:    v_cndmask_b32_e64 v21, v65, 0x7ff80000, s13
 ; GFX10-NEXT:    v_cndmask_b32_e64 v22, v70, 0, s14
 ; GFX10-NEXT:    v_cndmask_b32_e64 v23, v71, 0x7ff80000, s14
-; GFX10-NEXT:    v_cndmask_b32_e64 v24, v68, 0, s15
-; GFX10-NEXT:    v_cndmask_b32_e64 v25, v69, 0x7ff80000, s15
+; GFX10-NEXT:    s_waitcnt vmcnt(6)
+; GFX10-NEXT:    v_min_f64 v[68:69], v[24:25], v[66:67]
+; GFX10-NEXT:    v_cmp_u_f64_e64 s15, v[24:25], v[66:67]
 ; GFX10-NEXT:    s_waitcnt vmcnt(5)
-; GFX10-NEXT:    v_min_f64 v[80:81], v[28:29], v[0:1]
-; GFX10-NEXT:    v_cmp_u_f64_e64 s17, v[28:29], v[0:1]
+; GFX10-NEXT:    v_min_f64 v[66:67], v[26:27], v[0:1]
+; GFX10-NEXT:    v_cmp_u_f64_e64 s16, v[26:27], v[0:1]
 ; GFX10-NEXT:    s_waitcnt vmcnt(3)
-; GFX10-NEXT:    v_min_f64 v[66:67], v[26:27], v[2:3]
-; GFX10-NEXT:    v_cmp_u_f64_e64 s16, v[26:27], v[2:3]
-; GFX10-NEXT:    v_cndmask_b32_e64 v0, v82, 0, vcc_lo
+; GFX10-NEXT:    v_min_f64 v[80:81], v[28:29], v[2:3]
+; GFX10-NEXT:    v_cmp_u_f64_e64 s17, v[28:29], v[2:3]
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    v_min_f64 v[86:87], v[30:31], v[4:5]
 ; GFX10-NEXT:    v_cmp_u_f64_e64 s18, v[30:31], v[4:5]
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v82, 0, vcc_lo
 ; GFX10-NEXT:    v_cndmask_b32_e64 v1, v83, 0x7ff80000, vcc_lo
 ; GFX10-NEXT:    v_cndmask_b32_e64 v2, v84, 0, s4
 ; GFX10-NEXT:    v_cndmask_b32_e64 v3, v85, 0x7ff80000, s4
 ; GFX10-NEXT:    v_cndmask_b32_e64 v4, v32, 0, s5
 ; GFX10-NEXT:    v_cndmask_b32_e64 v5, v33, 0x7ff80000, s5
-; GFX10-NEXT:    v_cndmask_b32_e64 v28, v80, 0, s17
-; GFX10-NEXT:    v_cndmask_b32_e64 v29, v81, 0x7ff80000, s17
+; GFX10-NEXT:    v_cndmask_b32_e64 v24, v68, 0, s15
+; GFX10-NEXT:    v_cndmask_b32_e64 v25, v69, 0x7ff80000, s15
 ; GFX10-NEXT:    v_cndmask_b32_e64 v26, v66, 0, s16
 ; GFX10-NEXT:    v_cndmask_b32_e64 v27, v67, 0x7ff80000, s16
+; GFX10-NEXT:    v_cndmask_b32_e64 v28, v80, 0, s17
+; GFX10-NEXT:    v_cndmask_b32_e64 v29, v81, 0x7ff80000, s17
 ; GFX10-NEXT:    v_cndmask_b32_e64 v30, v86, 0, s18
 ; GFX10-NEXT:    v_cndmask_b32_e64 v31, v87, 0x7ff80000, s18
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/load-constant-i16.ll b/llvm/test/CodeGen/AMDGPU/load-constant-i16.ll
index cce5022213291..8e312a0e195ff 100644
--- a/llvm/test/CodeGen/AMDGPU/load-constant-i16.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-constant-i16.ll
@@ -760,34 +760,34 @@ define amdgpu_kernel void @constant_load_v16i16_align2(ptr addrspace(4) %ptr0) #
 ; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v8, 0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_clause 0x7
-; GFX12-TRUE16-NEXT:    global_load_d16_b16 v3, v8, s[0:1] offset:28
-; GFX12-TRUE16-NEXT:    global_load_d16_b16 v2, v8, s[0:1] offset:24
-; GFX12-TRUE16-NEXT:    global_load_d16_b16 v1, v8, s[0:1] offset:20
-; GFX12-TRUE16-NEXT:    global_load_d16_b16 v0, v8, s[0:1] offset:16
-; GFX12-TRUE16-NEXT:    global_load_d16_b16 v7, v8, s[0:1] offset:12
-; GFX12-TRUE16-NEXT:    global_load_d16_b16 v6, v8, s[0:1] offset:8
-; GFX12-TRUE16-NEXT:    global_load_d16_b16 v5, v8, s[0:1] offset:4
-; GFX12-TRUE16-NEXT:    global_load_d16_b16 v4, v8, s[0:1]
+; GFX12-TRUE16-NEXT:    global_load_d16_b16 v3, v8, s[0:1] offset:12
+; GFX12-TRUE16-NEXT:    global_load_d16_b16 v7, v8, s[0:1] offset:28
+; GFX12-TRUE16-NEXT:    global_load_d16_b16 v6, v8, s[0:1] offset:24
+; GFX12-TRUE16-NEXT:    global_load_d16_b16 v5, v8, s[0:1] offset:20
+; GFX12-TRUE16-NEXT:    global_load_d16_b16 v4, v8, s[0:1] offset:16
+; GFX12-TRUE16-NEXT:    global_load_d16_b16 v2, v8, s[0:1] offset:8
+; GFX12-TRUE16-NEXT:    global_load_d16_b16 v1, v8, s[0:1] offset:4
+; GFX12-TRUE16-NEXT:    global_load_d16_b16 v0, v8, s[0:1]
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v3, v8, s[0:1] offset:30
+; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v3, v8, s[0:1] offset:14
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v2, v8, s[0:1] offset:26
+; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v7, v8, s[0:1] offset:30
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v1, v8, s[0:1] offset:22
+; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v6, v8, s[0:1] offset:26
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v0, v8, s[0:1] offset:18
+; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v5, v8, s[0:1] offset:22
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v7, v8, s[0:1] offset:14
+; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v4, v8, s[0:1] offset:18
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v6, v8, s[0:1] offset:10
+; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v2, v8, s[0:1] offset:10
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v5, v8, s[0:1] offset:6
+; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v1, v8, s[0:1] offset:6
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v4, v8, s[0:1] offset:2
-; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x4
-; GFX12-TRUE16-NEXT:    global_store_b128 v[0:1], v[0:3], off
+; GFX12-TRUE16-NEXT:    global_load_d16_hi_b16 v0, v8, s[0:1] offset:2
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
+; GFX12-TRUE16-NEXT:    s_clause 0x1
 ; GFX12-TRUE16-NEXT:    global_store_b128 v[0:1], v[4:7], off
+; GFX12-TRUE16-NEXT:    global_store_b128 v[0:1], v[0:3], off
 ; GFX12-TRUE16-NEXT:    s_endpgm
 ;
 ; GFX12-FAKE16-LABEL: constant_load_v16i16_align2:
@@ -796,34 +796,34 @@ define amdgpu_kernel void @constant_load_v16i16_align2(ptr addrspace(4) %ptr0) #
 ; GFX12-FAKE16-NEXT:    v_mov_b32_e32 v8, 0
 ; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX12-FAKE16-NEXT:    s_clause 0x7
-; GFX12-FAKE16-NEXT:    global_load_u16 v3, v8, s[0:1] offset:28
-; GFX12-FAKE16-NEXT:    global_load_u16 v2, v8, s[0:1] offset:24
-; GFX12-FAKE16-NEXT:    global_load_u16 v1, v8, s[0:1] offset:20
-; GFX12-FAKE16-NEXT:    global_load_u16 v0, v8, s[0:1] offset:16
-; GFX12-FAKE16-NEXT:    global_load_u16 v7, v8, s[0:1] offset:12
-; GFX12-FAKE16-NEXT:    global_load_u16 v6, v8, s[0:1] offset:8
-; GFX12-FAKE16-NEXT:    global_load_u16 v5, v8, s[0:1] offset:4
-; GFX12-FAKE16-NEXT:    global_load_u16 v4, v8, s[0:1]
+; GFX12-FAKE16-NEXT:    global_load_u16 v3, v8, s[0:1] offset:12
+; GFX12-FAKE16-NEXT:    global_load_u16 v7, v8, s[0:1] offset:28
+; GFX12-FAKE16-NEXT:    global_load_u16 v6, v8, s[0:1] offset:24
+; GFX12-FAKE16-NEXT:    global_load_u16 v5, v8, s[0:1] offset:20
+; GFX12-FAKE16-NEXT:    global_load_u16 v4, v8, s[0:1] offset:16
+; GFX12-FAKE16-NEXT:    global_load_u16 v2, v8, s[0:1] offset:8
+; GFX12-FAKE16-NEXT:    global_load_u16 v1, v8, s[0:1] offset:4
+; GFX12-FAKE16-NEXT:    global_load_u16 v0, v8, s[0:1]
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v3, v8, s[0:1] offset:30
+; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v3, v8, s[0:1] offset:14
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v2, v8, s[0:1] offset:26
+; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v7, v8, s[0:1] offset:30
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v1, v8, s[0:1] offset:22
+; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v6, v8, s[0:1] offset:26
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v0, v8, s[0:1] offset:18
+; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v5, v8, s[0:1] offset:22
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v7, v8, s[0:1] offset:14
+; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v4, v8, s[0:1] offset:18
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v6, v8, s[0:1] offset:10
+; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v2, v8, s[0:1] offset:10
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v5, v8, s[0:1] offset:6
+; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v1, v8, s[0:1] offset:6
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x7
-; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v4, v8, s[0:1] offset:2
-; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x4
-; GFX12-FAKE16-NEXT:    global_store_b128 v[0:1], v[0:3], off
+; GFX12-FAKE16-NEXT:    global_load_d16_hi_b16 v0, v8, s[0:1] offset:2
 ; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x0
+; GFX12-FAKE16-NEXT:    s_clause 0x1
 ; GFX12-FAKE16-NEXT:    global_store_b128 v[0:1], v[4:7], off
+; GFX12-FAKE16-NEXT:    global_store_b128 v[0:1], v[0:3], off
 ; GFX12-FAKE16-NEXT:    s_endpgm
 entry:
   %ld =  load <16 x i16>, ptr addrspace(4) %ptr0, align 2
diff --git a/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll b/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll
index a6ce512164b89..8a3cc57e08579 100644
--- a/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll
@@ -7,13 +7,13 @@
 define amdgpu_vs void @test(ptr addrspace(8) inreg %arg1, ptr addrspace(3) %arg2) {
 ; CHECK-LABEL: test:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    v_add_i32_e32 v3, vcc, 12, v0
-; CHECK-NEXT:    v_add_i32_e32 v1, vcc, 8, v0
+; CHECK-NEXT:    v_add_i32_e32 v1, vcc, 12, v0
+; CHECK-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
 ; CHECK-NEXT:    v_add_i32_e32 v4, vcc, 4, v0
 ; CHECK-NEXT:    s_mov_b32 m0, -1
-; CHECK-NEXT:    ds_read_b32 v2, v1
+; CHECK-NEXT:    ds_read_b32 v3, v1
+; CHECK-NEXT:    ds_read_b32 v2, v2
 ; CHECK-NEXT:    ds_read_b32 v1, v4
-; CHECK-NEXT:    ds_read_b32 v3, v3
 ; CHECK-NEXT:    ds_read_b32 v0, v0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    exp mrt0 off, off, off, off
@@ -69,36 +69,36 @@ define amdgpu_vs void @test_3(i32 inreg %arg1, i32 inreg %arg2, ptr addrspace(8)
 ; CHECK-NEXT:    v_add_i32_e32 v0, vcc, 12, v1
 ; CHECK-NEXT:    v_add_i32_e32 v3, vcc, 8, v1
 ; CHECK-NEXT:    v_add_i32_e32 v4, vcc, 4, v1
-; CHECK-NEXT:    v_add_i32_e32 v6, vcc, 20, v1
-; CHECK-NEXT:    v_add_i32_e32 v7, vcc, 16, v1
-; CHECK-NEXT:    v_mov_b32_e32 v9, s0
-; CHECK-NEXT:    v_add_i32_e32 v10, vcc, 12, v2
-; CHECK-NEXT:    v_add_i32_e32 v11, vcc, 8, v2
+; CHECK-NEXT:    v_add_i32_e32 v7, vcc, 20, v1
+; CHECK-NEXT:    v_add_i32_e32 v9, vcc, 16, v1
+; CHECK-NEXT:    v_mov_b32_e32 v10, s0
+; CHECK-NEXT:    v_add_i32_e32 v11, vcc, 12, v2
+; CHECK-NEXT:    v_add_i32_e32 v12, vcc, 8, v2
 ; CHECK-NEXT:    s_mov_b32 m0, -1
+; CHECK-NEXT:    ds_read_b32 v6, v0
 ; CHECK-NEXT:    ds_read_b32 v5, v3
 ; CHECK-NEXT:    ds_read_b32 v4, v4
-; CHECK-NEXT:    ds_read_b32 v8, v6
-; CHECK-NEXT:    ds_read_b32 v7, v7
-; CHECK-NEXT:    ds_read_b32 v6, v0
+; CHECK-NEXT:    ds_read_b32 v8, v7
+; CHECK-NEXT:    ds_read_b32 v7, v9
 ; CHECK-NEXT:    ds_read_b32 v3, v1
 ; CHECK-NEXT:    v_add_i32_e32 v0, vcc, 4, v2
 ; CHECK-NEXT:    v_add_i32_e32 v1, vcc, 20, v2
-; CHECK-NEXT:    v_add_i32_e32 v12, vcc, 16, v2
+; CHECK-NEXT:    v_add_i32_e32 v9, vcc, 16, v2
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    tbuffer_store_format_xyzw v[3:6], v9, s[4:7], s1 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen offset:264 glc slc
-; CHECK-NEXT:    tbuffer_store_format_xy v[7:8], v9, s[4:7], s1 format:[BUF_DATA_FORMAT_INVALID,BUF_NUM_FORMAT_UINT] idxen offset:280 glc slc
+; CHECK-NEXT:    tbuffer_store_format_xyzw v[3:6], v10, s[4:7], s1 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen offset:264 glc slc
+; CHECK-NEXT:    tbuffer_store_format_xy v[7:8], v10, s[4:7], s1 format:[BUF_DATA_FORMAT_INVALID,BUF_NUM_FORMAT_UINT] idxen offset:280 glc slc
 ; CHECK-NEXT:    s_waitcnt expcnt(1)
-; CHECK-NEXT:    ds_read_b32 v4, v11
+; CHECK-NEXT:    ds_read_b32 v5, v11
+; CHECK-NEXT:    ds_read_b32 v4, v12
 ; CHECK-NEXT:    ds_read_b32 v3, v0
 ; CHECK-NEXT:    ds_read_b32 v1, v1
-; CHECK-NEXT:    ds_read_b32 v0, v12
-; CHECK-NEXT:    ds_read_b32 v5, v10
+; CHECK-NEXT:    ds_read_b32 v0, v9
 ; CHECK-NEXT:    ds_read_b32 v2, v2
-; CHECK-NEXT:    s_waitcnt lgkmcnt(2)
+; CHECK-NEXT:    s_waitcnt lgkmcnt(1)
 ; CHECK-NEXT:    exp mrt0 off, off, off, off
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    tbuffer_store_format_xyzw v[2:5], v9, s[4:7], s1 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen offset:240 glc slc
-; CHECK-NEXT:    tbuffer_store_format_xy v[0:1], v9, s[4:7], s1 format:[BUF_DATA_FORMAT_INVALID,BUF_NUM_FORMAT_UINT] idxen offset:256 glc slc
+; CHECK-NEXT:    tbuffer_store_format_xyzw v[2:5], v10, s[4:7], s1 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen offset:240 glc slc
+; CHECK-NEXT:    tbuffer_store_format_xy v[0:1], v10, s[4:7], s1 format:[BUF_DATA_FORMAT_INVALID,BUF_NUM_FORMAT_UINT] idxen offset:256 glc slc
 ; CHECK-NEXT:    s_endpgm
   %load1 = load <6 x float>, ptr addrspace(3) %arg5, align 4
   %vec11 = shufflevector <6 x float> %load1, <6 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
diff --git a/llvm/test/CodeGen/AMDGPU/load-local.128.ll b/llvm/test/CodeGen/AMDGPU/load-local.128.ll
index 10dca76cc389a..d634e40f1d79b 100644
--- a/llvm/test/CodeGen/AMDGPU/load-local.128.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-local.128.ll
@@ -95,51 +95,51 @@ define <4 x i32> @load_lds_v4i32_align1(ptr addrspace(3) %ptr) {
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    s_mov_b32 m0, -1
-; GFX7-NEXT:    ds_read_u8 v1, v0 offset:6
-; GFX7-NEXT:    ds_read_u8 v2, v0 offset:4
-; GFX7-NEXT:    ds_read_u8 v3, v0 offset:2
-; GFX7-NEXT:    ds_read_u8 v4, v0 offset:1
+; GFX7-NEXT:    ds_read_u8 v1, v0 offset:1
+; GFX7-NEXT:    ds_read_u8 v2, v0 offset:6
+; GFX7-NEXT:    ds_read_u8 v3, v0 offset:4
+; GFX7-NEXT:    ds_read_u8 v4, v0 offset:2
 ; GFX7-NEXT:    ds_read_u8 v5, v0
 ; GFX7-NEXT:    ds_read_u8 v6, v0 offset:3
 ; GFX7-NEXT:    ds_read_u8 v7, v0 offset:5
 ; GFX7-NEXT:    ds_read_u8 v8, v0 offset:7
-; GFX7-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(7)
+; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(3)
-; GFX7-NEXT:    v_or_b32_e32 v4, v4, v5
+; GFX7-NEXT:    v_or_b32_e32 v1, v1, v5
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v5, 8, v6
-; GFX7-NEXT:    v_or_b32_e32 v3, v5, v3
-; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX7-NEXT:    v_or_b32_e32 v4, v3, v4
+; GFX7-NEXT:    v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT:    v_or_b32_e32 v4, v4, v1
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(1)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 8, v7
-; GFX7-NEXT:    v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 8, v7
+; GFX7-NEXT:    v_or_b32_e32 v1, v1, v3
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 8, v8
-; GFX7-NEXT:    ds_read_u8 v5, v0 offset:15
-; GFX7-NEXT:    ds_read_u8 v6, v0 offset:14
+; GFX7-NEXT:    ds_read_u8 v5, v0 offset:9
+; GFX7-NEXT:    ds_read_u8 v6, v0 offset:11
 ; GFX7-NEXT:    ds_read_u8 v7, v0 offset:13
-; GFX7-NEXT:    ds_read_u8 v8, v0 offset:12
-; GFX7-NEXT:    ds_read_u8 v9, v0 offset:11
-; GFX7-NEXT:    ds_read_u8 v10, v0 offset:10
-; GFX7-NEXT:    ds_read_u8 v11, v0 offset:9
+; GFX7-NEXT:    ds_read_u8 v8, v0 offset:15
+; GFX7-NEXT:    ds_read_u8 v9, v0 offset:14
+; GFX7-NEXT:    ds_read_u8 v10, v0 offset:12
+; GFX7-NEXT:    ds_read_u8 v11, v0 offset:10
 ; GFX7-NEXT:    ds_read_u8 v0, v0 offset:8
-; GFX7-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX7-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX7-NEXT:    s_waitcnt lgkmcnt(1)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 8, v11
+; GFX7-NEXT:    v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT:    v_or_b32_e32 v1, v2, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(7)
+; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 8, v5
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 8, v9
-; GFX7-NEXT:    v_or_b32_e32 v2, v2, v10
+; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 8, v6
+; GFX7-NEXT:    v_or_b32_e32 v2, v2, v11
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 8, v5
+; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 8, v8
 ; GFX7-NEXT:    v_or_b32_e32 v2, v2, v0
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 8, v7
-; GFX7-NEXT:    v_or_b32_e32 v3, v3, v6
-; GFX7-NEXT:    v_or_b32_e32 v0, v0, v8
+; GFX7-NEXT:    v_or_b32_e32 v3, v3, v9
+; GFX7-NEXT:    v_or_b32_e32 v0, v0, v10
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX7-NEXT:    v_or_b32_e32 v3, v3, v0
 ; GFX7-NEXT:    v_mov_b32_e32 v0, v4
@@ -331,21 +331,21 @@ define <4 x i32> @load_lds_v4i32_align2(ptr addrspace(3) %ptr) {
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    ds_read_u16 v1, v0 offset:2
 ; GFX7-NEXT:    ds_read_u16 v3, v0 offset:12
 ; GFX7-NEXT:    ds_read_u16 v2, v0 offset:8
-; GFX7-NEXT:    ds_read_u16 v1, v0 offset:4
-; GFX7-NEXT:    ds_read_u16 v4, v0 offset:2
+; GFX7-NEXT:    ds_read_u16 v4, v0 offset:4
 ; GFX7-NEXT:    ds_read_u16 v5, v0
 ; GFX7-NEXT:    ds_read_u16 v6, v0 offset:6
 ; GFX7-NEXT:    ds_read_u16 v7, v0 offset:10
 ; GFX7-NEXT:    ds_read_u16 v8, v0 offset:14
-; GFX7-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(7)
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(3)
 ; GFX7-NEXT:    v_or_b32_e32 v0, v0, v5
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v6
-; GFX7-NEXT:    v_or_b32_e32 v1, v4, v1
+; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v6
+; GFX7-NEXT:    v_or_b32_e32 v1, v1, v4
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(1)
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v7
 ; GFX7-NEXT:    v_or_b32_e32 v2, v4, v2
diff --git a/llvm/test/CodeGen/AMDGPU/load-local.96.ll b/llvm/test/CodeGen/AMDGPU/load-local.96.ll
index 2da3fce72072e..b917b48b90e6a 100644
--- a/llvm/test/CodeGen/AMDGPU/load-local.96.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-local.96.ll
@@ -86,41 +86,41 @@ define <3 x i32> @load_lds_v3i32_align1(ptr addrspace(3) %ptr) {
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    s_mov_b32 m0, -1
-; GFX7-NEXT:    ds_read_u8 v1, v0 offset:6
-; GFX7-NEXT:    ds_read_u8 v2, v0 offset:4
+; GFX7-NEXT:    ds_read_u8 v1, v0 offset:1
+; GFX7-NEXT:    ds_read_u8 v2, v0 offset:6
+; GFX7-NEXT:    ds_read_u8 v4, v0 offset:4
 ; GFX7-NEXT:    ds_read_u8 v3, v0 offset:2
-; GFX7-NEXT:    ds_read_u8 v4, v0 offset:1
 ; GFX7-NEXT:    ds_read_u8 v5, v0
 ; GFX7-NEXT:    ds_read_u8 v6, v0 offset:3
 ; GFX7-NEXT:    ds_read_u8 v7, v0 offset:5
 ; GFX7-NEXT:    ds_read_u8 v8, v0 offset:7
-; GFX7-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(7)
+; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(3)
-; GFX7-NEXT:    v_or_b32_e32 v4, v4, v5
+; GFX7-NEXT:    v_or_b32_e32 v1, v1, v5
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v5, 8, v6
 ; GFX7-NEXT:    v_or_b32_e32 v3, v5, v3
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX7-NEXT:    v_or_b32_e32 v3, v3, v4
+; GFX7-NEXT:    v_or_b32_e32 v3, v3, v1
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(1)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 8, v7
-; GFX7-NEXT:    ds_read_u8 v5, v0 offset:11
-; GFX7-NEXT:    ds_read_u8 v6, v0 offset:10
-; GFX7-NEXT:    ds_read_u8 v7, v0 offset:9
+; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 8, v7
+; GFX7-NEXT:    ds_read_u8 v5, v0 offset:9
+; GFX7-NEXT:    ds_read_u8 v6, v0 offset:11
+; GFX7-NEXT:    ds_read_u8 v7, v0 offset:10
 ; GFX7-NEXT:    ds_read_u8 v0, v0 offset:8
-; GFX7-NEXT:    v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT:    v_or_b32_e32 v1, v1, v4
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(4)
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 8, v8
-; GFX7-NEXT:    v_or_b32_e32 v1, v4, v1
-; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX7-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX7-NEXT:    s_waitcnt lgkmcnt(1)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 8, v7
+; GFX7-NEXT:    v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT:    v_or_b32_e32 v1, v2, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(3)
+; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 8, v5
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 8, v5
-; GFX7-NEXT:    v_or_b32_e32 v2, v2, v6
+; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 8, v6
+; GFX7-NEXT:    v_or_b32_e32 v2, v2, v7
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; GFX7-NEXT:    v_or_b32_e32 v2, v2, v0
 ; GFX7-NEXT:    v_mov_b32_e32 v0, v3
@@ -274,19 +274,19 @@ define <3 x i32> @load_lds_v3i32_align2(ptr addrspace(3) %ptr) {
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    ds_read_u16 v1, v0 offset:2
 ; GFX7-NEXT:    ds_read_u16 v2, v0 offset:8
-; GFX7-NEXT:    ds_read_u16 v1, v0 offset:4
-; GFX7-NEXT:    ds_read_u16 v3, v0 offset:2
+; GFX7-NEXT:    ds_read_u16 v3, v0 offset:4
 ; GFX7-NEXT:    ds_read_u16 v4, v0
 ; GFX7-NEXT:    ds_read_u16 v5, v0 offset:6
 ; GFX7-NEXT:    ds_read_u16 v6, v0 offset:10
-; GFX7-NEXT:    s_waitcnt lgkmcnt(3)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(5)
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX7-NEXT:    v_or_b32_e32 v0, v0, v4
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(1)
-; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
-; GFX7-NEXT:    v_or_b32_e32 v1, v3, v1
+; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
+; GFX7-NEXT:    v_or_b32_e32 v1, v1, v3
 ; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
 ; GFX7-NEXT:    v_or_b32_e32 v2, v3, v2
diff --git a/llvm/test/CodeGen/AMDGPU/max.i16.ll b/llvm/test/CodeGen/AMDGPU/max.i16.ll
index 1857eaba0a2a9..1e246465ab1e3 100644
--- a/llvm/test/CodeGen/AMDGPU/max.i16.ll
+++ b/llvm/test/CodeGen/AMDGPU/max.i16.ll
@@ -139,17 +139,17 @@ define amdgpu_kernel void @v_test_imax_sge_v3i16(ptr addrspace(1) %out, ptr addr
 ;
 ; GFX9-LABEL: v_test_imax_sge_v3i16:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
-; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    global_load_short_d16 v2, v0, s[2:3] offset:4
 ; GFX9-NEXT:    s_nop 0
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
 ; GFX9-NEXT:    global_load_dword v4, v0, s[2:3]
+; GFX9-NEXT:    ; kill: killed $sgpr2_sgpr3
 ; GFX9-NEXT:    s_nop 0
 ; GFX9-NEXT:    global_load_short_d16 v1, v0, s[6:7] offset:4
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
diff --git a/llvm/test/CodeGen/AMDGPU/memcpy-libcall.ll b/llvm/test/CodeGen/AMDGPU/memcpy-libcall.ll
index 8d020b9e1a603..0003366f3a3ea 100644
--- a/llvm/test/CodeGen/AMDGPU/memcpy-libcall.ll
+++ b/llvm/test/CodeGen/AMDGPU/memcpy-libcall.ll
@@ -173,54 +173,53 @@ define amdgpu_kernel void @memcpy_p0_p5_minsize(ptr %generic, ptr addrspace(5) %
 ; CHECK-NEXT:    v_mov_b32_e32 v26, s0
 ; CHECK-NEXT:    buffer_load_dword v3, v26, s[20:23], 0 offen offset:124
 ; CHECK-NEXT:    buffer_load_dword v2, v26, s[20:23], 0 offen offset:120
+; CHECK-NEXT:    buffer_load_dword v5, v26, s[20:23], 0 offen offset:100
+; CHECK-NEXT:    buffer_load_dword v7, v26, s[20:23], 0 offen offset:108
 ; CHECK-NEXT:    buffer_load_dword v1, v26, s[20:23], 0 offen offset:116
 ; CHECK-NEXT:    buffer_load_dword v0, v26, s[20:23], 0 offen offset:112
-; CHECK-NEXT:    buffer_load_dword v7, v26, s[20:23], 0 offen offset:108
 ; CHECK-NEXT:    buffer_load_dword v6, v26, s[20:23], 0 offen offset:104
-; CHECK-NEXT:    buffer_load_dword v5, v26, s[20:23], 0 offen offset:100
 ; CHECK-NEXT:    buffer_load_dword v4, v26, s[20:23], 0 offen offset:96
 ; CHECK-NEXT:    s_load_dwordx2 s[0:1], s[8:9], 0x0
-; CHECK-NEXT:    buffer_load_dword v8, v26, s[20:23], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v9, v26, s[20:23], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v10, v26, s[20:23], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v11, v26, s[20:23], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v12, v26, s[20:23], 0 offen offset:32
-; CHECK-NEXT:    buffer_load_dword v13, v26, s[20:23], 0 offen offset:36
-; CHECK-NEXT:    buffer_load_dword v14, v26, s[20:23], 0 offen offset:40
-; CHECK-NEXT:    buffer_load_dword v15, v26, s[20:23], 0 offen offset:44
-; CHECK-NEXT:    buffer_load_dword v16, v26, s[20:23], 0 offen offset:48
-; CHECK-NEXT:    buffer_load_dword v17, v26, s[20:23], 0 offen offset:52
-; CHECK-NEXT:    buffer_load_dword v18, v26, s[20:23], 0 offen offset:56
-; CHECK-NEXT:    buffer_load_dword v19, v26, s[20:23], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v8, v26, s[20:23], 0 offen offset:32
+; CHECK-NEXT:    buffer_load_dword v9, v26, s[20:23], 0 offen offset:36
+; CHECK-NEXT:    buffer_load_dword v10, v26, s[20:23], 0 offen offset:40
+; CHECK-NEXT:    buffer_load_dword v11, v26, s[20:23], 0 offen offset:44
+; CHECK-NEXT:    buffer_load_dword v12, v26, s[20:23], 0 offen offset:48
+; CHECK-NEXT:    buffer_load_dword v13, v26, s[20:23], 0 offen offset:52
+; CHECK-NEXT:    buffer_load_dword v14, v26, s[20:23], 0 offen offset:56
+; CHECK-NEXT:    buffer_load_dword v15, v26, s[20:23], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v17, v26, s[20:23], 0 offen offset:68
+; CHECK-NEXT:    buffer_load_dword v19, v26, s[20:23], 0 offen offset:76
+; CHECK-NEXT:    buffer_load_dword v21, v26, s[20:23], 0 offen offset:84
 ; CHECK-NEXT:    buffer_load_dword v23, v26, s[20:23], 0 offen offset:92
 ; CHECK-NEXT:    buffer_load_dword v22, v26, s[20:23], 0 offen offset:88
-; CHECK-NEXT:    buffer_load_dword v21, v26, s[20:23], 0 offen offset:84
 ; CHECK-NEXT:    buffer_load_dword v20, v26, s[20:23], 0 offen offset:80
+; CHECK-NEXT:    buffer_load_dword v18, v26, s[20:23], 0 offen offset:72
+; CHECK-NEXT:    buffer_load_dword v16, v26, s[20:23], 0 offen offset:64
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    v_mov_b32_e32 v25, s1
 ; CHECK-NEXT:    v_mov_b32_e32 v24, s0
-; CHECK-NEXT:    s_waitcnt vmcnt(20)
+; CHECK-NEXT:    s_waitcnt vmcnt(18)
 ; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[0:3] offset:112
-; CHECK-NEXT:    buffer_load_dword v3, v26, s[20:23], 0 offen offset:76
-; CHECK-NEXT:    s_nop 0
-; CHECK-NEXT:    buffer_load_dword v2, v26, s[20:23], 0 offen offset:72
-; CHECK-NEXT:    buffer_load_dword v1, v26, s[20:23], 0 offen offset:68
-; CHECK-NEXT:    buffer_load_dword v0, v26, s[20:23], 0 offen offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[4:7] offset:96
-; CHECK-NEXT:    buffer_load_dword v4, v26, s[20:23], 0 offen
+; CHECK-NEXT:    buffer_load_dword v0, v26, s[20:23], 0 offen
+; CHECK-NEXT:    buffer_load_dword v1, v26, s[20:23], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v2, v26, s[20:23], 0 offen offset:8
 ; CHECK-NEXT:    s_nop 0
-; CHECK-NEXT:    buffer_load_dword v5, v26, s[20:23], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v6, v26, s[20:23], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v7, v26, s[20:23], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_dword v4, v26, s[20:23], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v5, v26, s[20:23], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v6, v26, s[20:23], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v7, v26, s[20:23], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v3, v26, s[20:23], 0 offen offset:12
 ; CHECK-NEXT:    s_nop 0
 ; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[20:23] offset:80
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[0:3] offset:64
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[16:19] offset:48
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[12:15] offset:32
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[8:11] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[16:19] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[12:15] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[8:11] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[4:7]
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[4:7] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[0:3]
 ; CHECK-NEXT:    s_endpgm
 entry:
   tail call void @llvm.memcpy.p0.p5.i64(ptr %generic, ptr addrspace(5) %src, i64 128, i1 false)
@@ -464,54 +463,53 @@ define amdgpu_kernel void @memcpy_p0_p5_optsize(ptr %generic, ptr addrspace(5) %
 ; CHECK-NEXT:    v_mov_b32_e32 v26, s0
 ; CHECK-NEXT:    buffer_load_dword v3, v26, s[20:23], 0 offen offset:124
 ; CHECK-NEXT:    buffer_load_dword v2, v26, s[20:23], 0 offen offset:120
+; CHECK-NEXT:    buffer_load_dword v5, v26, s[20:23], 0 offen offset:100
+; CHECK-NEXT:    buffer_load_dword v7, v26, s[20:23], 0 offen offset:108
 ; CHECK-NEXT:    buffer_load_dword v1, v26, s[20:23], 0 offen offset:116
 ; CHECK-NEXT:    buffer_load_dword v0, v26, s[20:23], 0 offen offset:112
-; CHECK-NEXT:    buffer_load_dword v7, v26, s[20:23], 0 offen offset:108
 ; CHECK-NEXT:    buffer_load_dword v6, v26, s[20:23], 0 offen offset:104
-; CHECK-NEXT:    buffer_load_dword v5, v26, s[20:23], 0 offen offset:100
 ; CHECK-NEXT:    buffer_load_dword v4, v26, s[20:23], 0 offen offset:96
 ; CHECK-NEXT:    s_load_dwordx2 s[0:1], s[8:9], 0x0
-; CHECK-NEXT:    buffer_load_dword v8, v26, s[20:23], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v9, v26, s[20:23], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v10, v26, s[20:23], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v11, v26, s[20:23], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v12, v26, s[20:23], 0 offen offset:32
-; CHECK-NEXT:    buffer_load_dword v13, v26, s[20:23], 0 offen offset:36
-; CHECK-NEXT:    buffer_load_dword v14, v26, s[20:23], 0 offen offset:40
-; CHECK-NEXT:    buffer_load_dword v15, v26, s[20:23], 0 offen offset:44
-; CHECK-NEXT:    buffer_load_dword v16, v26, s[20:23], 0 offen offset:48
-; CHECK-NEXT:    buffer_load_dword v17, v26, s[20:23], 0 offen offset:52
-; CHECK-NEXT:    buffer_load_dword v18, v26, s[20:23], 0 offen offset:56
-; CHECK-NEXT:    buffer_load_dword v19, v26, s[20:23], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v8, v26, s[20:23], 0 offen offset:32
+; CHECK-NEXT:    buffer_load_dword v9, v26, s[20:23], 0 offen offset:36
+; CHECK-NEXT:    buffer_load_dword v10, v26, s[20:23], 0 offen offset:40
+; CHECK-NEXT:    buffer_load_dword v11, v26, s[20:23], 0 offen offset:44
+; CHECK-NEXT:    buffer_load_dword v12, v26, s[20:23], 0 offen offset:48
+; CHECK-NEXT:    buffer_load_dword v13, v26, s[20:23], 0 offen offset:52
+; CHECK-NEXT:    buffer_load_dword v14, v26, s[20:23], 0 offen offset:56
+; CHECK-NEXT:    buffer_load_dword v15, v26, s[20:23], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v17, v26, s[20:23], 0 offen offset:68
+; CHECK-NEXT:    buffer_load_dword v19, v26, s[20:23], 0 offen offset:76
+; CHECK-NEXT:    buffer_load_dword v21, v26, s[20:23], 0 offen offset:84
 ; CHECK-NEXT:    buffer_load_dword v23, v26, s[20:23], 0 offen offset:92
 ; CHECK-NEXT:    buffer_load_dword v22, v26, s[20:23], 0 offen offset:88
-; CHECK-NEXT:    buffer_load_dword v21, v26, s[20:23], 0 offen offset:84
 ; CHECK-NEXT:    buffer_load_dword v20, v26, s[20:23], 0 offen offset:80
+; CHECK-NEXT:    buffer_load_dword v18, v26, s[20:23], 0 offen offset:72
+; CHECK-NEXT:    buffer_load_dword v16, v26, s[20:23], 0 offen offset:64
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    v_mov_b32_e32 v25, s1
 ; CHECK-NEXT:    v_mov_b32_e32 v24, s0
-; CHECK-NEXT:    s_waitcnt vmcnt(20)
+; CHECK-NEXT:    s_waitcnt vmcnt(18)
 ; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[0:3] offset:112
-; CHECK-NEXT:    buffer_load_dword v3, v26, s[20:23], 0 offen offset:76
-; CHECK-NEXT:    s_nop 0
-; CHECK-NEXT:    buffer_load_dword v2, v26, s[20:23], 0 offen offset:72
-; CHECK-NEXT:    buffer_load_dword v1, v26, s[20:23], 0 offen offset:68
-; CHECK-NEXT:    buffer_load_dword v0, v26, s[20:23], 0 offen offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[4:7] offset:96
-; CHECK-NEXT:    buffer_load_dword v4, v26, s[20:23], 0 offen
+; CHECK-NEXT:    buffer_load_dword v0, v26, s[20:23], 0 offen
+; CHECK-NEXT:    buffer_load_dword v1, v26, s[20:23], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v2, v26, s[20:23], 0 offen offset:8
 ; CHECK-NEXT:    s_nop 0
-; CHECK-NEXT:    buffer_load_dword v5, v26, s[20:23], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v6, v26, s[20:23], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v7, v26, s[20:23], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_dword v4, v26, s[20:23], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v5, v26, s[20:23], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v6, v26, s[20:23], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v7, v26, s[20:23], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v3, v26, s[20:23], 0 offen offset:12
 ; CHECK-NEXT:    s_nop 0
 ; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[20:23] offset:80
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[0:3] offset:64
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[16:19] offset:48
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[12:15] offset:32
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[8:11] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[16:19] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[12:15] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[8:11] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[4:7]
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[4:7] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[24:25], v[0:3]
 ; CHECK-NEXT:    s_endpgm
 entry:
   tail call void @llvm.memcpy.p0.p5.i64(ptr %generic, ptr addrspace(5) %src, i64 128, i1 false)
diff --git a/llvm/test/CodeGen/AMDGPU/memcpy-param-combinations.ll b/llvm/test/CodeGen/AMDGPU/memcpy-param-combinations.ll
index cadc3dadb0a1e..b43ccc551ca95 100644
--- a/llvm/test/CodeGen/AMDGPU/memcpy-param-combinations.ll
+++ b/llvm/test/CodeGen/AMDGPU/memcpy-param-combinations.ll
@@ -451,12 +451,12 @@ define void @memcpy_p0_p3_sz31_align_1_1(ptr addrspace(0) align 1 %dst, ptr addr
 ; CHECK-LABEL: memcpy_p0_p3_sz31_align_1_1:
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT:    ds_read_u8 v9, v2 offset:30
 ; CHECK-NEXT:    ds_read_b32 v8, v2 offset:24
+; CHECK-NEXT:    ds_read_u8 v9, v2 offset:30
 ; CHECK-NEXT:    ds_read_u16 v10, v2 offset:28
 ; CHECK-NEXT:    ds_read_b64 v[6:7], v2 offset:16
 ; CHECK-NEXT:    ds_read2_b64 v[2:5], v2 offset1:1
-; CHECK-NEXT:    s_waitcnt lgkmcnt(4)
+; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
 ; CHECK-NEXT:    flat_store_byte v[0:1], v9 offset:30
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
 ; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
@@ -506,12 +506,12 @@ define void @memcpy_p0_p3_sz31_align_2_2(ptr addrspace(0) align 2 %dst, ptr addr
 ; CHECK-LABEL: memcpy_p0_p3_sz31_align_2_2:
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT:    ds_read_u8 v9, v2 offset:30
 ; CHECK-NEXT:    ds_read_b32 v8, v2 offset:24
+; CHECK-NEXT:    ds_read_u8 v9, v2 offset:30
 ; CHECK-NEXT:    ds_read_u16 v10, v2 offset:28
 ; CHECK-NEXT:    ds_read_b64 v[6:7], v2 offset:16
 ; CHECK-NEXT:    ds_read2_b64 v[2:5], v2 offset1:1
-; CHECK-NEXT:    s_waitcnt lgkmcnt(4)
+; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
 ; CHECK-NEXT:    flat_store_byte v[0:1], v9 offset:30
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
 ; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
@@ -896,19 +896,18 @@ define void @memcpy_p0_p5_sz31_align_1_1(ptr addrspace(0) align 1 %dst, ptr addr
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_byte v[0:1], v11 offset:30
+; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    flat_store_short v[0:1], v11 offset:28
+; CHECK-NEXT:    flat_store_byte v[0:1], v10 offset:30
 ; CHECK-NEXT:    flat_store_dwordx3 v[0:1], v[7:9] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
@@ -924,18 +923,18 @@ define void @memcpy_p0_p5_sz32_align_1_1(ptr addrspace(0) align 1 %dst, ptr addr
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x7
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:16
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    s_waitcnt vmcnt(1)
+; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10]
+; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
@@ -966,19 +965,18 @@ define void @memcpy_p0_p5_sz31_align_2_2(ptr addrspace(0) align 2 %dst, ptr addr
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_byte v[0:1], v11 offset:30
+; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    flat_store_short v[0:1], v11 offset:28
+; CHECK-NEXT:    flat_store_byte v[0:1], v10 offset:30
 ; CHECK-NEXT:    flat_store_dwordx3 v[0:1], v[7:9] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
@@ -994,18 +992,18 @@ define void @memcpy_p0_p5_sz32_align_2_2(ptr addrspace(0) align 2 %dst, ptr addr
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x7
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:16
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    s_waitcnt vmcnt(1)
+; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10]
+; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memintrinsic-unroll.ll b/llvm/test/CodeGen/AMDGPU/memintrinsic-unroll.ll
index ee52611fae5ca..9cc42ac448067 100644
--- a/llvm/test/CodeGen/AMDGPU/memintrinsic-unroll.ll
+++ b/llvm/test/CodeGen/AMDGPU/memintrinsic-unroll.ll
@@ -3583,104 +3583,102 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; CHECK-NEXT:  .LBB4_1: ; %load-store-loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    s_clause 0x3e
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:32
-; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:36
-; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:40
-; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:44
-; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:48
-; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:52
-; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:56
-; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:60
-; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:124
-; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:120
-; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:116
-; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:112
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:32
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:36
+; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:40
+; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:44
+; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:48
+; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:52
+; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:56
+; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:76
+; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:92
+; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:88
+; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:84
+; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:80
+; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:72
+; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:68
+; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:64
 ; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:108
+; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:124
+; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:120
+; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:116
+; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:112
 ; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:104
 ; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:100
 ; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:96
-; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:252
-; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:248
-; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:244
-; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:240
 ; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:236
+; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:252
+; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:248
+; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:244
+; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:240
 ; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:232
 ; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:228
 ; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:224
-; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:220
-; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:216
-; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:212
-; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:208
-; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:204
-; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:200
-; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:196
-; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:192
-; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:188
-; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:184
-; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:180
-; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:176
+; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:220
+; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:216
+; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:212
+; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:208
+; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:204
+; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:200
+; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:196
+; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:192
 ; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:172
+; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:188
+; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:184
+; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:180
+; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:176
 ; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:168
 ; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:164
 ; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:160
-; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:156
-; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:152
-; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:148
-; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:144
-; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:140
-; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:136
-; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:132
-; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:128
-; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:92
-; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:88
-; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:84
-; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:80
-; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:76
-; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:72
-; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:68
-; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen offset:64
-; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:156
+; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:152
+; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:148
+; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:144
+; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:140
+; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:136
+; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:132
+; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:128
+; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:12
 ; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
 ; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
 ; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
 ; CHECK-NEXT:    s_addc_u32 s5, s5, 0
 ; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0x100, v2
 ; CHECK-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]
-; CHECK-NEXT:    s_waitcnt vmcnt(41)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[23:26] offset:240
-; CHECK-NEXT:    s_waitcnt vmcnt(37)
+; CHECK-NEXT:    s_waitcnt vmcnt(35)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[31:34] offset:240
+; CHECK-NEXT:    s_waitcnt vmcnt(32)
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[27:30] offset:224
-; CHECK-NEXT:    s_waitcnt vmcnt(33)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[31:34] offset:208
-; CHECK-NEXT:    s_waitcnt vmcnt(29)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[35:38] offset:192
-; CHECK-NEXT:    s_waitcnt vmcnt(25)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[48:51] offset:176
-; CHECK-NEXT:    s_waitcnt vmcnt(21)
+; CHECK-NEXT:    s_waitcnt vmcnt(28)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[35:38] offset:208
+; CHECK-NEXT:    s_waitcnt vmcnt(24)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[48:51] offset:192
+; CHECK-NEXT:    s_waitcnt vmcnt(19)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:176
+; CHECK-NEXT:    s_waitcnt vmcnt(16)
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:160
-; CHECK-NEXT:    s_waitcnt vmcnt(17)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:144
-; CHECK-NEXT:    s_waitcnt vmcnt(13)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:128
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[15:18] offset:112
+; CHECK-NEXT:    s_waitcnt vmcnt(12)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:144
+; CHECK-NEXT:    s_waitcnt vmcnt(8)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[23:26] offset:112
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[19:22] offset:96
-; CHECK-NEXT:    s_waitcnt vmcnt(9)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:80
-; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:64
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:48
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[7:10] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[15:18] offset:80
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[7:10] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(1)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87]
 ; CHECK-NEXT:    s_and_b32 vcc_lo, exec_lo, s6
 ; CHECK-NEXT:    s_cbranch_vccnz .LBB4_1
 ; CHECK-NEXT:  ; %bb.2: ; %memcpy-split
@@ -3748,16 +3746,17 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:21
 ; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:22
 ; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:23
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:24
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:24
 ; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:25
 ; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:26
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:30
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:31
+; ALIGNED-NEXT:    buffer_load_ubyte v127, v2, s[0:3], 0 offen offset:19
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:28
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:29
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:30
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:31
 ; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:32
 ; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:33
 ; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:34
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:29
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:28
 ; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:27
 ; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:35
 ; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:36
@@ -3798,10 +3797,9 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:71
 ; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:76
 ; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:77
+; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:75
 ; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:78
 ; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:79
-; ALIGNED-NEXT:    buffer_load_ubyte v127, v2, s[0:3], 0 offen offset:19
-; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:75
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(57)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(56)
@@ -3811,46 +3809,46 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(54)
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(53)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(52)
 ; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(51)
 ; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(50)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(49)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(48)
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(47)
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(46)
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(45)
 ; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v3
-; ALIGNED-NEXT:    s_waitcnt vmcnt(45)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(44)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(43)
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v5
+; ALIGNED-NEXT:    s_waitcnt vmcnt(42)
 ; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v9, 8, v5
-; ALIGNED-NEXT:    s_waitcnt vmcnt(41)
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v9, 8, v8
+; ALIGNED-NEXT:    s_waitcnt vmcnt(40)
 ; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v8, 8, v6
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 8, v7
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 8, v6
 ; ALIGNED-NEXT:    v_lshl_or_b32 v6, v11, 8, v12
 ; ALIGNED-NEXT:    v_lshl_or_b32 v7, v15, 8, v14
 ; ALIGNED-NEXT:    v_lshl_or_b32 v8, v19, 8, v17
-; ALIGNED-NEXT:    s_waitcnt vmcnt(40)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(39)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v13
-; ALIGNED-NEXT:    s_waitcnt vmcnt(38)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(37)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v10, v20, 8, v18
-; ALIGNED-NEXT:    s_waitcnt vmcnt(36)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(35)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v11, v23, 8, v22
-; ALIGNED-NEXT:    s_waitcnt vmcnt(34)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(33)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v12, v28, 8, v25
-; ALIGNED-NEXT:    s_waitcnt vmcnt(32)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(31)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v13, v24, 8, v21
-; ALIGNED-NEXT:    s_waitcnt vmcnt(30)
-; ALIGNED-NEXT:    v_lshl_or_b32 v14, v27, 8, v26
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    s_waitcnt vmcnt(29)
+; ALIGNED-NEXT:    v_lshl_or_b32 v14, v27, 8, v26
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v6, 16, v5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v8, 16, v7
@@ -3858,27 +3856,27 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v6, v12, 16, v11
 ; ALIGNED-NEXT:    v_lshl_or_b32 v7, v14, 16, v13
 ; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(28)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(27)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v15, v31, 8, v30
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(26)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(25)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v34, 8, v33
 ; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(24)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(23)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v37, 8, v32
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(22)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(21)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v36, 8, v35
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(17)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(16)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v50, 8, v38
 ; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(15)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(14)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v5, v49, 8, v39
 ; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v6, v51, 8, v48
 ; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v7, v53, 8, v52
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 16, v15
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v1
@@ -3888,13 +3886,13 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v55, 8, v29
 ; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v67, 8, v66
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v64, 8, v54
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v68, 8, v65
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    s_clause 0x1
@@ -3903,13 +3901,13 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v70, 8, v69
 ; ALIGNED-NEXT:    s_clause 0x1
 ; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:83
 ; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:74
 ; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v80, 8, v71
 ; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
@@ -3955,9 +3953,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
 ; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:1152 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
 ; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:87
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
@@ -4240,7 +4236,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:148
 ; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:149
 ; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:150
-; ALIGNED-NEXT:    buffer_load_ubyte v123, v2, s[0:3], 0 offen offset:151
+; ALIGNED-NEXT:    buffer_load_ubyte v124, v2, s[0:3], 0 offen offset:151
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v126, 8, v125
@@ -4252,7 +4248,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1116 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v1, v123, 8, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v124, 8, v5
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 8, v3
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
@@ -4286,7 +4282,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_ubyte v74, v2, s[0:3], 0 offen offset:163
 ; ALIGNED-NEXT:    buffer_load_ubyte v79, v2, s[0:3], 0 offen offset:164
 ; ALIGNED-NEXT:    buffer_load_ubyte v75, v2, s[0:3], 0 offen offset:165
-; ALIGNED-NEXT:    buffer_load_ubyte v76, v2, s[0:3], 0 offen offset:166
+; ALIGNED-NEXT:    buffer_load_ubyte v77, v2, s[0:3], 0 offen offset:166
 ; ALIGNED-NEXT:    buffer_load_ubyte v72, v2, s[0:3], 0 offen offset:167
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v78, 8, v89
@@ -4294,7 +4290,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v74, 8, v73
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v1, v72, 8, v76
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v72, 8, v77
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1156 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v75, 8, v79
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
@@ -4303,20 +4299,20 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_ubyte v63, v2, s[0:3], 0 offen offset:172
 ; ALIGNED-NEXT:    buffer_load_ubyte v61, v2, s[0:3], 0 offen offset:173
 ; ALIGNED-NEXT:    buffer_load_ubyte v62, v2, s[0:3], 0 offen offset:174
-; ALIGNED-NEXT:    buffer_load_ubyte v60, v2, s[0:3], 0 offen offset:175
+; ALIGNED-NEXT:    buffer_load_ubyte v59, v2, s[0:3], 0 offen offset:175
 ; ALIGNED-NEXT:    buffer_load_ubyte v57, v2, s[0:3], 0 offen offset:171
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v61, 8, v63
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v1, v60, 8, v62
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v59, 8, v62
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1164 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x2
-; ALIGNED-NEXT:    buffer_load_ubyte v59, v2, s[0:3], 0 offen offset:168
+; ALIGNED-NEXT:    buffer_load_ubyte v58, v2, s[0:3], 0 offen offset:168
 ; ALIGNED-NEXT:    buffer_load_ubyte v56, v2, s[0:3], 0 offen offset:169
 ; ALIGNED-NEXT:    buffer_load_ubyte v47, v2, s[0:3], 0 offen offset:170
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v56, 8, v59
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v56, 8, v58
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v57, 8, v47
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
@@ -4326,7 +4322,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_ubyte v43, v2, s[0:3], 0 offen offset:177
 ; ALIGNED-NEXT:    buffer_load_ubyte v119, v2, s[0:3], 0 offen offset:178
 ; ALIGNED-NEXT:    buffer_load_ubyte v40, v2, s[0:3], 0 offen offset:179
-; ALIGNED-NEXT:    buffer_load_ubyte v45, v2, s[0:3], 0 offen offset:180
+; ALIGNED-NEXT:    buffer_load_ubyte v44, v2, s[0:3], 0 offen offset:180
 ; ALIGNED-NEXT:    buffer_load_ubyte v41, v2, s[0:3], 0 offen offset:181
 ; ALIGNED-NEXT:    buffer_load_ubyte v42, v2, s[0:3], 0 offen offset:182
 ; ALIGNED-NEXT:    buffer_load_ubyte v118, v2, s[0:3], 0 offen offset:183
@@ -4338,7 +4334,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v118, 8, v42
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v41, 8, v45
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v41, 8, v44
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1176 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x4
@@ -4373,15 +4369,16 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_ubyte v96, v2, s[0:3], 0 offen offset:198
 ; ALIGNED-NEXT:    buffer_load_ubyte v85, v2, s[0:3], 0 offen offset:199
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v98, 8, v100
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v98, 8, v100
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v86, 8, v87
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v3, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v86, 8, v87
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
+; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v97, 8, v99
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v85, 8, v96
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v85, 8, v96
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v97, 8, v99
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v3, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1192 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x4
 ; ALIGNED-NEXT:    buffer_load_ubyte v83, v2, s[0:3], 0 offen offset:204
@@ -4492,23 +4489,23 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:248
 ; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:249
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:250
-; ALIGNED-NEXT:    v_lshl_or_b32 v124, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v123, v4, 16, v3
 ; ALIGNED-NEXT:    s_clause 0x5
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen
 ; ALIGNED-NEXT:    buffer_load_ubyte v94, v2, s[0:3], 0 offen offset:2
 ; ALIGNED-NEXT:    buffer_load_ubyte v88, v2, s[0:3], 0 offen offset:4
 ; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:5
 ; ALIGNED-NEXT:    buffer_load_ubyte v92, v2, s[0:3], 0 offen offset:6
-; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:7
+; ALIGNED-NEXT:    buffer_load_ubyte v104, v2, s[0:3], 0 offen offset:7
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(28)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v25, 8, v27
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(26)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v24, 8, v26
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(14)
-; ALIGNED-NEXT:    v_lshl_or_b32 v44, v12, 8, v16
+; ALIGNED-NEXT:    v_lshl_or_b32 v45, v12, 8, v16
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    v_lshl_or_b32 v58, v8, 8, v10
-; ALIGNED-NEXT:    v_lshl_or_b32 v104, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v60, v8, 8, v10
+; ALIGNED-NEXT:    v_lshl_or_b32 v95, v4, 16, v3
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v21, 8, v22
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v23, 8, v20
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
@@ -4517,35 +4514,35 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_store_dword v90, off, s[0:3], s32 offset:1096 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
 ; ALIGNED-NEXT:    buffer_store_dword v92, off, s[0:3], s32 offset:1100 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v77, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v76, v4, 16, v3
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v17, 8, v19
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v13
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v95, off, s[0:3], s32 offset:1108 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v104, off, s[0:3], s32 offset:1108 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v101, v4, 16, v3
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v15, 8, v18
-; ALIGNED-NEXT:    v_lshl_or_b32 v84, v44, 16, v4
-; ALIGNED-NEXT:    v_lshl_or_b32 v44, v9, 8, v11
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v58, 16, v44
-; ALIGNED-NEXT:    v_lshl_or_b32 v44, v5, 8, v6
-; ALIGNED-NEXT:    v_lshl_or_b32 v58, v7, 8, v1
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v58, 16, v44
+; ALIGNED-NEXT:    v_lshl_or_b32 v84, v45, 16, v4
+; ALIGNED-NEXT:    v_lshl_or_b32 v45, v9, 8, v11
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v60, 16, v45
+; ALIGNED-NEXT:    v_lshl_or_b32 v45, v5, 8, v6
+; ALIGNED-NEXT:    v_lshl_or_b32 v60, v7, 8, v1
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v60, 16, v45
 ; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v44, v2, s[0:3], 0 offen offset:1
-; ALIGNED-NEXT:    buffer_load_ubyte v58, v2, s[0:3], 0 offen offset:3
+; ALIGNED-NEXT:    buffer_load_ubyte v45, v2, s[0:3], 0 offen offset:1
+; ALIGNED-NEXT:    buffer_load_ubyte v60, v2, s[0:3], 0 offen offset:3
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1068 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v94, off, s[0:3], s32 offset:1092 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:1076 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:1076 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:1080 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v44, v44, 8, v0
-; ALIGNED-NEXT:    v_lshl_or_b32 v58, v58, 8, v94
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v58, 16, v44
-; ALIGNED-NEXT:    v_lshl_or_b32 v44, v90, 8, v88
-; ALIGNED-NEXT:    v_lshl_or_b32 v58, v95, 8, v92
+; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:1080 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v45, v45, 8, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v60, v60, 8, v94
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v60, 16, v45
+; ALIGNED-NEXT:    v_lshl_or_b32 v45, v90, 8, v88
+; ALIGNED-NEXT:    v_lshl_or_b32 v60, v104, 8, v92
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v58, 16, v44
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v60, 16, v45
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x4
 ; ALIGNED-NEXT:    buffer_load_ubyte v122, v2, s[0:3], 0 offen offset:12
@@ -4554,34 +4551,34 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_ubyte v110, v2, s[0:3], 0 offen offset:15
 ; ALIGNED-NEXT:    buffer_load_ubyte v94, v2, s[0:3], 0 offen offset:11
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    v_lshl_or_b32 v44, v111, 8, v122
+; ALIGNED-NEXT:    v_lshl_or_b32 v45, v111, 8, v122
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v58, v110, 8, v120
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v58, 16, v44
+; ALIGNED-NEXT:    v_lshl_or_b32 v60, v110, 8, v120
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v60, 16, v45
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1140 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x2
-; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:8
+; ALIGNED-NEXT:    buffer_load_ubyte v104, v2, s[0:3], 0 offen offset:8
 ; ALIGNED-NEXT:    buffer_load_ubyte v92, v2, s[0:3], 0 offen offset:9
 ; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:10
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v44, v92, 8, v95
+; ALIGNED-NEXT:    v_lshl_or_b32 v45, v92, 8, v104
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v58, v94, 8, v90
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v58, 16, v44
+; ALIGNED-NEXT:    v_lshl_or_b32 v60, v94, 8, v90
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v60, 16, v45
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x2
+; ALIGNED-NEXT:    buffer_load_ubyte v60, v2, s[0:3], 0 offen offset:18
 ; ALIGNED-NEXT:    buffer_load_ubyte v88, v2, s[0:3], 0 offen offset:16
-; ALIGNED-NEXT:    buffer_load_ubyte v44, v2, s[0:3], 0 offen offset:18
-; ALIGNED-NEXT:    buffer_load_ubyte v58, v2, s[0:3], 0 offen offset:17
+; ALIGNED-NEXT:    buffer_load_ubyte v45, v2, s[0:3], 0 offen offset:17
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:232
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:236
 ; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:228
 ; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:224
 ; ALIGNED-NEXT:    v_add_nc_u32_e32 v2, 0x100, v2
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v127, 8, v44
+; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v127, 8, v60
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v127, v58, 8, v88
+; ALIGNED-NEXT:    v_lshl_or_b32 v127, v45, 8, v88
 ; ALIGNED-NEXT:    v_lshl_or_b32 v127, v0, 16, v127
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
@@ -4606,9 +4603,9 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v16 offset:246
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v18 offset:244
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v19 offset:240
-; ALIGNED-NEXT:    buffer_store_dword v77, off, s[0:3], s32 offset:248
-; ALIGNED-NEXT:    buffer_store_dword v104, off, s[0:3], s32 offset:252
-; ALIGNED-NEXT:    buffer_store_dword v124, off, s[0:3], s32 offset:244
+; ALIGNED-NEXT:    buffer_store_dword v76, off, s[0:3], s32 offset:248
+; ALIGNED-NEXT:    buffer_store_dword v95, off, s[0:3], s32 offset:252
+; ALIGNED-NEXT:    buffer_store_dword v123, off, s[0:3], s32 offset:244
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1220 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
 ; ALIGNED-NEXT:    v_cmp_gt_u64_e64 s6, 0x800, s[4:5]
@@ -4713,7 +4710,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v118 offset:183
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v41 offset:181
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v42 offset:182
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v45 offset:180
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v44 offset:180
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v46 offset:176
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
@@ -4730,17 +4727,17 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v47 offset:170
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v57 offset:171
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v56 offset:169
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v60 offset:175
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v59 offset:175
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v61 offset:173
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v62 offset:174
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v63 offset:172
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v59 offset:168
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v58 offset:168
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v73 offset:162
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v74 offset:163
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v78 offset:161
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v72 offset:167
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v75 offset:165
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v76 offset:166
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v77 offset:166
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v79 offset:164
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v89 offset:160
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Reload
@@ -4768,7 +4765,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:145
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v123 offset:151
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v124 offset:151
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1112 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:149
@@ -5235,11 +5232,11 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:24
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v44 offset:18
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v60 offset:18
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:19
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v58 offset:17
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v45 offset:17
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:23
@@ -5272,7 +5269,7 @@ define void @memcpy_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v110 offset:15
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v120 offset:14
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v122 offset:12
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v95 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v104 offset:8
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:2
@@ -12461,97 +12458,97 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; CHECK-NEXT:  .LBB9_1: ; %memmove_fwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    s_clause 0x3e
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:32
-; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:36
-; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:40
-; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:44
-; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:48
-; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:52
-; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:56
-; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:32
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:36
+; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:40
+; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:44
+; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:48
+; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:52
+; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:56
+; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:76
 ; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:92
 ; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:88
 ; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:84
 ; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:80
-; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:124
-; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:120
-; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:116
-; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:112
-; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:108
-; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:104
-; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:100
-; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:96
-; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:156
-; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:152
-; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:148
-; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:144
+; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:72
+; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:68
+; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:64
+; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:108
+; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:124
+; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:120
+; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:116
+; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:112
+; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:104
+; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:100
+; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:96
+; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:172
 ; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:188
 ; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:184
 ; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:180
 ; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:176
-; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:172
-; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:168
-; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:164
-; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:160
+; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:168
+; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:164
+; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:160
+; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:204
 ; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:220
 ; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:216
 ; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:212
 ; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:208
-; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:252
-; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:248
-; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:244
-; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:240
-; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:236
-; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:232
-; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:228
-; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:224
-; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:204
-; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:200
-; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:196
-; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:192
-; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:140
-; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:136
-; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:132
-; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:128
-; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:76
-; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:72
-; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:68
-; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen offset:64
-; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:200
+; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:196
+; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:192
+; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:236
+; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:252
+; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:248
+; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:244
+; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:240
+; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:232
+; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:228
+; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:224
+; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:140
+; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:156
+; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:152
+; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:148
+; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:144
+; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:136
+; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:132
+; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:128
+; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:12
 ; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
 ; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
 ; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
 ; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0x100, v2
 ; CHECK-NEXT:    s_addc_u32 s5, s5, 0
-; CHECK-NEXT:    s_waitcnt vmcnt(20)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:240
+; CHECK-NEXT:    s_waitcnt vmcnt(19)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:240
 ; CHECK-NEXT:    s_waitcnt vmcnt(16)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:224
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:224
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[48:51] offset:208
-; CHECK-NEXT:    s_waitcnt vmcnt(12)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:192
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[35:38] offset:192
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[31:34] offset:176
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[35:38] offset:160
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[27:30] offset:144
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[27:30] offset:160
+; CHECK-NEXT:    s_waitcnt vmcnt(11)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:144
 ; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:128
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[19:22] offset:112
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[23:26] offset:96
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[23:26] offset:112
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[19:22] offset:96
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[15:18] offset:80
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:64
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:48
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[7:10] offset:32
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[7:10] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:32
+; CHECK-NEXT:    s_waitcnt vmcnt(1)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87]
 ; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0x800
 ; CHECK-NEXT:    s_cbranch_scc1 .LBB9_1
 ; CHECK-NEXT:  .LBB9_2: ; %Flow10
@@ -12565,103 +12562,101 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; CHECK-NEXT:  .LBB9_4: ; %memmove_bwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    s_clause 0x3e
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:32
-; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:36
-; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:40
-; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:44
-; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:48
-; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:52
-; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:56
-; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:60
-; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:124
-; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:120
-; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:116
-; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:112
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:32
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:36
+; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:40
+; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:44
+; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:48
+; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:52
+; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:56
+; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:76
+; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:92
+; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:88
+; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:84
+; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:80
+; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:72
+; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:68
+; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:64
 ; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:108
+; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:124
+; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:120
+; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:116
+; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:112
 ; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:104
 ; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:100
 ; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:96
-; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:252
-; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:248
-; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:244
-; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:240
 ; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:236
+; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:252
+; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:248
+; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:244
+; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:240
 ; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:232
 ; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:228
 ; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:224
-; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:220
-; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:216
-; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:212
-; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:208
 ; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:204
+; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:220
+; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:216
+; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:212
+; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:208
 ; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:200
 ; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:196
 ; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:192
-; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:188
-; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:184
-; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:180
-; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:176
 ; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:172
+; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:188
+; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:184
+; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:180
+; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:176
 ; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:168
 ; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:164
 ; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:160
-; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:156
-; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:152
-; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:148
-; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:144
-; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:140
-; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:136
-; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:132
-; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:128
-; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:92
-; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:88
-; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:84
-; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:80
-; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:76
-; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:72
-; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:68
-; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen offset:64
-; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:156
+; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:152
+; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:148
+; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:144
+; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:140
+; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:136
+; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:132
+; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:128
+; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:12
 ; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
 ; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
 ; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0xffffff00, v2
 ; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff00
 ; CHECK-NEXT:    s_addc_u32 s5, s5, -1
-; CHECK-NEXT:    s_waitcnt vmcnt(41)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[23:26] offset:240
-; CHECK-NEXT:    s_waitcnt vmcnt(37)
+; CHECK-NEXT:    s_waitcnt vmcnt(35)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[31:34] offset:240
+; CHECK-NEXT:    s_waitcnt vmcnt(32)
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[27:30] offset:224
-; CHECK-NEXT:    s_waitcnt vmcnt(33)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[31:34] offset:208
-; CHECK-NEXT:    s_waitcnt vmcnt(29)
+; CHECK-NEXT:    s_waitcnt vmcnt(27)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[48:51] offset:208
+; CHECK-NEXT:    s_waitcnt vmcnt(24)
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[35:38] offset:192
-; CHECK-NEXT:    s_waitcnt vmcnt(25)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[48:51] offset:176
-; CHECK-NEXT:    s_waitcnt vmcnt(21)
+; CHECK-NEXT:    s_waitcnt vmcnt(19)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:176
+; CHECK-NEXT:    s_waitcnt vmcnt(16)
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:160
-; CHECK-NEXT:    s_waitcnt vmcnt(17)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:144
-; CHECK-NEXT:    s_waitcnt vmcnt(13)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:128
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[15:18] offset:112
+; CHECK-NEXT:    s_waitcnt vmcnt(12)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:144
+; CHECK-NEXT:    s_waitcnt vmcnt(8)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[23:26] offset:112
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[19:22] offset:96
-; CHECK-NEXT:    s_waitcnt vmcnt(9)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:80
-; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:64
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:48
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[7:10] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[15:18] offset:80
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[7:10] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(1)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87]
 ; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; CHECK-NEXT:    s_cbranch_scc0 .LBB9_4
 ; CHECK-NEXT:  .LBB9_5: ; %Flow11
@@ -12736,16 +12731,17 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:21
 ; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:22
 ; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:23
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:24
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:24
 ; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:25
 ; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:26
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:30
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:31
+; ALIGNED-NEXT:    buffer_load_ubyte v127, v2, s[0:3], 0 offen offset:19
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:28
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:29
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:30
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:31
 ; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:32
 ; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:33
 ; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:34
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:29
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:28
 ; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:27
 ; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:35
 ; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:36
@@ -12768,16 +12764,16 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:53
 ; ALIGNED-NEXT:    buffer_load_ubyte v33, v2, s[0:3], 0 offen offset:54
 ; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:55
-; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:56
+; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:56
 ; ALIGNED-NEXT:    buffer_load_ubyte v50, v2, s[0:3], 0 offen offset:57
 ; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:58
+; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:60
+; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:61
 ; ALIGNED-NEXT:    buffer_load_ubyte v38, v2, s[0:3], 0 offen offset:62
-; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:63
+; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:63
 ; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:64
 ; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:65
 ; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:66
-; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:61
-; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:60
 ; ALIGNED-NEXT:    buffer_load_ubyte v51, v2, s[0:3], 0 offen offset:59
 ; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:67
 ; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:68
@@ -12786,10 +12782,9 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:71
 ; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:76
 ; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:77
+; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:75
 ; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:78
 ; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:79
-; ALIGNED-NEXT:    buffer_load_ubyte v127, v2, s[0:3], 0 offen offset:19
-; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:75
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(57)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(56)
@@ -12799,46 +12794,46 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(54)
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(53)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(52)
 ; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(51)
 ; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(50)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(49)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(48)
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(47)
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(46)
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(45)
 ; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v3
-; ALIGNED-NEXT:    s_waitcnt vmcnt(45)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(44)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(43)
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v5
+; ALIGNED-NEXT:    s_waitcnt vmcnt(42)
 ; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v9, 8, v5
-; ALIGNED-NEXT:    s_waitcnt vmcnt(41)
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v9, 8, v8
+; ALIGNED-NEXT:    s_waitcnt vmcnt(40)
 ; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v8, 8, v6
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 8, v7
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 8, v6
 ; ALIGNED-NEXT:    v_lshl_or_b32 v6, v11, 8, v12
 ; ALIGNED-NEXT:    v_lshl_or_b32 v7, v15, 8, v14
 ; ALIGNED-NEXT:    v_lshl_or_b32 v8, v19, 8, v17
-; ALIGNED-NEXT:    s_waitcnt vmcnt(40)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(39)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v13
-; ALIGNED-NEXT:    s_waitcnt vmcnt(38)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(37)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v10, v20, 8, v18
-; ALIGNED-NEXT:    s_waitcnt vmcnt(36)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(35)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v11, v23, 8, v22
-; ALIGNED-NEXT:    s_waitcnt vmcnt(34)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(33)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v12, v27, 8, v25
-; ALIGNED-NEXT:    s_waitcnt vmcnt(32)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(31)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v13, v24, 8, v21
-; ALIGNED-NEXT:    s_waitcnt vmcnt(30)
-; ALIGNED-NEXT:    v_lshl_or_b32 v14, v28, 8, v26
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    s_waitcnt vmcnt(29)
+; ALIGNED-NEXT:    v_lshl_or_b32 v14, v28, 8, v26
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v6, 16, v5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v8, 16, v7
@@ -12846,26 +12841,27 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v6, v12, 16, v11
 ; ALIGNED-NEXT:    v_lshl_or_b32 v7, v14, 16, v13
 ; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(28)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(27)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v15, v30, 8, v29
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(26)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(25)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v32, 8, v34
 ; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(24)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(23)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v36, 8, v31
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(22)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(21)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v35, 8, v33
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(16)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v48, 8, v37
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v39, 8, v38
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(14)
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v49, 8, v38
 ; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v6, v50, 8, v49
+; ALIGNED-NEXT:    v_lshl_or_b32 v6, v50, 8, v39
 ; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v7, v51, 8, v52
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 16, v15
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v1
@@ -12875,13 +12871,13 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v54, 8, v53
 ; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v55, 8, v65
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v66, 8, v64
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v68, 8, v67
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    s_clause 0x1
@@ -12890,13 +12886,13 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v70, 8, v69
 ; ALIGNED-NEXT:    s_clause 0x1
 ; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:83
 ; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:74
 ; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v80, 8, v71
 ; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
@@ -12922,11 +12918,11 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
@@ -12942,9 +12938,8 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
 ; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:1000 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:1404 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:1400 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:87
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
 ; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1032 ; 4-byte Folded Spill
@@ -13227,39 +13222,39 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v120, 8, v111
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1400 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1404 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x2
 ; ALIGNED-NEXT:    buffer_load_ubyte v121, v2, s[0:3], 0 offen offset:149
 ; ALIGNED-NEXT:    buffer_load_ubyte v122, v2, s[0:3], 0 offen offset:150
-; ALIGNED-NEXT:    buffer_load_ubyte v109, v2, s[0:3], 0 offen offset:151
+; ALIGNED-NEXT:    buffer_load_ubyte v110, v2, s[0:3], 0 offen offset:151
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v121, 8, v3
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v1, v109, 8, v122
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v110, 8, v122
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1412 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x4
 ; ALIGNED-NEXT:    buffer_load_ubyte v108, v2, s[0:3], 0 offen offset:156
 ; ALIGNED-NEXT:    buffer_load_ubyte v105, v2, s[0:3], 0 offen offset:157
-; ALIGNED-NEXT:    buffer_load_ubyte v107, v2, s[0:3], 0 offen offset:158
+; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:158
 ; ALIGNED-NEXT:    buffer_load_ubyte v104, v2, s[0:3], 0 offen offset:159
-; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:155
+; ALIGNED-NEXT:    buffer_load_ubyte v94, v2, s[0:3], 0 offen offset:155
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v105, 8, v108
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v1, v104, 8, v107
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v104, 8, v106
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1416 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x2
-; ALIGNED-NEXT:    buffer_load_ubyte v93, v2, s[0:3], 0 offen offset:152
+; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:152
 ; ALIGNED-NEXT:    buffer_load_ubyte v92, v2, s[0:3], 0 offen offset:153
 ; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:154
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v92, 8, v93
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v92, 8, v95
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v1, v95, 8, v90
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v94, 8, v90
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x7
@@ -13268,8 +13263,8 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_ubyte v73, v2, s[0:3], 0 offen offset:162
 ; ALIGNED-NEXT:    buffer_load_ubyte v74, v2, s[0:3], 0 offen offset:163
 ; ALIGNED-NEXT:    buffer_load_ubyte v88, v2, s[0:3], 0 offen offset:164
-; ALIGNED-NEXT:    buffer_load_ubyte v75, v2, s[0:3], 0 offen offset:165
-; ALIGNED-NEXT:    buffer_load_ubyte v76, v2, s[0:3], 0 offen offset:166
+; ALIGNED-NEXT:    buffer_load_ubyte v76, v2, s[0:3], 0 offen offset:165
+; ALIGNED-NEXT:    buffer_load_ubyte v75, v2, s[0:3], 0 offen offset:166
 ; ALIGNED-NEXT:    buffer_load_ubyte v72, v2, s[0:3], 0 offen offset:167
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v79, 8, v89
@@ -13277,9 +13272,9 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v74, 8, v73
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v1, v72, 8, v76
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v72, 8, v75
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1424 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v75, 8, v88
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v76, 8, v88
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1428 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x4
@@ -13356,15 +13351,16 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_ubyte v98, v2, s[0:3], 0 offen offset:198
 ; ALIGNED-NEXT:    buffer_load_ubyte v87, v2, s[0:3], 0 offen offset:199
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v100, 8, v102
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v100, 8, v102
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v96, 8, v97
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v3, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v96, 8, v97
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
+; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v99, 8, v101
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v87, 8, v98
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v87, 8, v98
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v99, 8, v101
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v3, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1460 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x4
 ; ALIGNED-NEXT:    buffer_load_ubyte v85, v2, s[0:3], 0 offen offset:204
@@ -13380,10 +13376,10 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1464 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x2
 ; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:200
-; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:201
+; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:201
 ; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:202
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v70, 8, v80
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v71, 8, v80
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v81, 8, v69
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
@@ -13475,11 +13471,11 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:248
 ; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:249
 ; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:250
-; ALIGNED-NEXT:    v_lshl_or_b32 v110, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v109, v4, 16, v3
 ; ALIGNED-NEXT:    s_clause 0x4
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:3
-; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:4
+; ALIGNED-NEXT:    buffer_load_ubyte v107, v2, s[0:3], 0 offen offset:4
 ; ALIGNED-NEXT:    buffer_load_ubyte v123, v2, s[0:3], 0 offen offset:5
 ; ALIGNED-NEXT:    buffer_load_ubyte v125, v2, s[0:3], 0 offen offset:6
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(27)
@@ -13490,7 +13486,7 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v77, v13, 8, v16
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v91, v9, 8, v10
-; ALIGNED-NEXT:    v_lshl_or_b32 v94, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v93, v4, 16, v3
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v22, 8, v24
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v23, 8, v21
 ; ALIGNED-NEXT:    v_lshl_or_b32 v78, v4, 16, v3
@@ -13502,14 +13498,14 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1292 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v106, off, s[0:3], s32 offset:1296 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v107, off, s[0:3], s32 offset:1296 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
 ; ALIGNED-NEXT:    buffer_store_dword v123, off, s[0:3], s32 offset:1304 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v86, v77, 16, v4
 ; ALIGNED-NEXT:    v_lshl_or_b32 v77, v11, 8, v12
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
 ; ALIGNED-NEXT:    buffer_store_dword v125, off, s[0:3], s32 offset:1308 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v71, v91, 16, v77
+; ALIGNED-NEXT:    v_lshl_or_b32 v70, v91, 16, v77
 ; ALIGNED-NEXT:    v_lshl_or_b32 v77, v6, 8, v8
 ; ALIGNED-NEXT:    v_lshl_or_b32 v91, v7, 8, v5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v4, v91, 16, v77
@@ -13527,7 +13523,7 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v91, v0, 8, v91
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:12
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v91, 16, v77
-; ALIGNED-NEXT:    v_lshl_or_b32 v77, v123, 8, v106
+; ALIGNED-NEXT:    v_lshl_or_b32 v77, v123, 8, v107
 ; ALIGNED-NEXT:    v_lshl_or_b32 v91, v3, 8, v125
 ; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:13
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1316 ; 4-byte Folded Spill
@@ -13560,21 +13556,21 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v91, 16, v77
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x2
-; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:16
-; ALIGNED-NEXT:    buffer_load_ubyte v77, v2, s[0:3], 0 offen offset:18
-; ALIGNED-NEXT:    buffer_load_ubyte v91, v2, s[0:3], 0 offen offset:17
+; ALIGNED-NEXT:    buffer_load_ubyte v91, v2, s[0:3], 0 offen offset:18
+; ALIGNED-NEXT:    buffer_load_ubyte v107, v2, s[0:3], 0 offen offset:16
+; ALIGNED-NEXT:    buffer_load_ubyte v77, v2, s[0:3], 0 offen offset:17
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:232
-; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:236
+; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:236
 ; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:228
 ; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:224
 ; ALIGNED-NEXT:    s_clause 0x1
 ; ALIGNED-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704
 ; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:708
 ; ALIGNED-NEXT:    v_add_nc_u32_e32 v2, 0x100, v2
-; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v127, 8, v77
+; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v127, 8, v91
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    v_lshl_or_b32 v127, v91, 8, v106
+; ALIGNED-NEXT:    v_lshl_or_b32 v127, v77, 8, v107
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
 ; ALIGNED-NEXT:    v_add_co_u32 v3, vcc_lo, v3, s4
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
@@ -13596,8 +13592,8 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v19 offset:244
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v20 offset:240
 ; ALIGNED-NEXT:    buffer_store_dword v78, off, s[0:3], s32 offset:248
-; ALIGNED-NEXT:    buffer_store_dword v94, off, s[0:3], s32 offset:252
-; ALIGNED-NEXT:    buffer_store_dword v110, off, s[0:3], s32 offset:244
+; ALIGNED-NEXT:    buffer_store_dword v93, off, s[0:3], s32 offset:252
+; ALIGNED-NEXT:    buffer_store_dword v109, off, s[0:3], s32 offset:244
 ; ALIGNED-NEXT:    v_lshl_or_b32 v127, v0, 16, v127
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1488 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
@@ -13663,7 +13659,7 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v69 offset:202
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v81 offset:203
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v70 offset:201
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v71 offset:201
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v82 offset:207
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v83 offset:205
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v84 offset:206
@@ -13729,8 +13725,8 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v74 offset:163
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v79 offset:161
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v72 offset:167
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v75 offset:165
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v76 offset:166
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v76 offset:165
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v75 offset:166
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v88 offset:164
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v89 offset:160
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Reload
@@ -13746,20 +13742,20 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v90 offset:154
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v95 offset:155
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v94 offset:155
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v92 offset:153
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v104 offset:159
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v105 offset:157
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v107 offset:158
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v106 offset:158
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v108 offset:156
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v93 offset:152
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v95 offset:152
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v111 offset:146
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v120 offset:147
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v124 offset:145
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v109 offset:151
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v110 offset:151
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v121 offset:149
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v122 offset:150
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:148
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1396 ; 4-byte Folded Reload
@@ -14219,11 +14215,11 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:24
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v77 offset:18
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v91 offset:18
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:19
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v91 offset:17
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v77 offset:17
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:23
@@ -14236,7 +14232,7 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:20
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v106 offset:16
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v107 offset:16
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408
@@ -14305,16 +14301,17 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v4, s[0:3], 0 offen offset:21
 ; ALIGNED-NEXT:    buffer_load_ubyte v2, v4, s[0:3], 0 offen offset:22
 ; ALIGNED-NEXT:    buffer_load_ubyte v3, v4, s[0:3], 0 offen offset:23
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v4, s[0:3], 0 offen offset:24
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v4, s[0:3], 0 offen offset:24
 ; ALIGNED-NEXT:    buffer_load_ubyte v10, v4, s[0:3], 0 offen offset:25
 ; ALIGNED-NEXT:    buffer_load_ubyte v12, v4, s[0:3], 0 offen offset:26
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v4, s[0:3], 0 offen offset:30
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v4, s[0:3], 0 offen offset:31
+; ALIGNED-NEXT:    buffer_load_ubyte v126, v4, s[0:3], 0 offen offset:19
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v4, s[0:3], 0 offen offset:28
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v4, s[0:3], 0 offen offset:29
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v4, s[0:3], 0 offen offset:30
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v4, s[0:3], 0 offen offset:31
 ; ALIGNED-NEXT:    buffer_load_ubyte v14, v4, s[0:3], 0 offen offset:32
 ; ALIGNED-NEXT:    buffer_load_ubyte v15, v4, s[0:3], 0 offen offset:33
 ; ALIGNED-NEXT:    buffer_load_ubyte v17, v4, s[0:3], 0 offen offset:34
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v4, s[0:3], 0 offen offset:29
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v4, s[0:3], 0 offen offset:28
 ; ALIGNED-NEXT:    buffer_load_ubyte v11, v4, s[0:3], 0 offen offset:27
 ; ALIGNED-NEXT:    buffer_load_ubyte v19, v4, s[0:3], 0 offen offset:35
 ; ALIGNED-NEXT:    buffer_load_ubyte v13, v4, s[0:3], 0 offen offset:36
@@ -14355,10 +14352,9 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_ubyte v68, v4, s[0:3], 0 offen offset:71
 ; ALIGNED-NEXT:    buffer_load_ubyte v69, v4, s[0:3], 0 offen offset:76
 ; ALIGNED-NEXT:    buffer_load_ubyte v70, v4, s[0:3], 0 offen offset:77
+; ALIGNED-NEXT:    buffer_load_ubyte v81, v4, s[0:3], 0 offen offset:75
 ; ALIGNED-NEXT:    buffer_load_ubyte v71, v4, s[0:3], 0 offen offset:78
 ; ALIGNED-NEXT:    buffer_load_ubyte v80, v4, s[0:3], 0 offen offset:79
-; ALIGNED-NEXT:    buffer_load_ubyte v126, v4, s[0:3], 0 offen offset:19
-; ALIGNED-NEXT:    buffer_load_ubyte v81, v4, s[0:3], 0 offen offset:75
 ; ALIGNED-NEXT:    buffer_load_ubyte v125, v4, s[0:3], 0 offen offset:151
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(58)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
@@ -14369,46 +14365,46 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(55)
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(54)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(53)
 ; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(52)
 ; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(51)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(50)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(49)
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(48)
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(47)
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(46)
 ; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v2
-; ALIGNED-NEXT:    s_waitcnt vmcnt(46)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(45)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(44)
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v7, 8, v5
+; ALIGNED-NEXT:    s_waitcnt vmcnt(43)
 ; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v2, v9, 8, v5
-; ALIGNED-NEXT:    s_waitcnt vmcnt(42)
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v9, 8, v8
+; ALIGNED-NEXT:    s_waitcnt vmcnt(41)
 ; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v8, 8, v6
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 8, v7
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 8, v6
 ; ALIGNED-NEXT:    v_lshl_or_b32 v6, v11, 8, v12
 ; ALIGNED-NEXT:    v_lshl_or_b32 v7, v15, 8, v14
 ; ALIGNED-NEXT:    v_lshl_or_b32 v8, v19, 8, v17
-; ALIGNED-NEXT:    s_waitcnt vmcnt(41)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(40)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v13
-; ALIGNED-NEXT:    s_waitcnt vmcnt(39)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(38)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v10, v20, 8, v18
-; ALIGNED-NEXT:    s_waitcnt vmcnt(37)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(36)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v11, v23, 8, v22
-; ALIGNED-NEXT:    s_waitcnt vmcnt(35)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(34)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v12, v28, 8, v25
-; ALIGNED-NEXT:    s_waitcnt vmcnt(33)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(32)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v13, v24, 8, v21
-; ALIGNED-NEXT:    s_waitcnt vmcnt(31)
-; ALIGNED-NEXT:    v_lshl_or_b32 v14, v27, 8, v26
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    s_waitcnt vmcnt(30)
+; ALIGNED-NEXT:    v_lshl_or_b32 v14, v27, 8, v26
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
 ; ALIGNED-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
@@ -14416,27 +14412,27 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v6, v12, 16, v11
 ; ALIGNED-NEXT:    v_lshl_or_b32 v7, v14, 16, v13
 ; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(29)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(28)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v15, v31, 8, v30
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(27)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(26)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v34, 8, v33
 ; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(25)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(24)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v37, 8, v32
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(23)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(22)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v2, v36, 8, v35
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(18)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(17)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v50, 8, v38
 ; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(16)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(15)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v5, v49, 8, v39
 ; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v6, v51, 8, v48
 ; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v7, v53, 8, v52
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 16, v15
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
@@ -14446,13 +14442,13 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v55, 8, v29
 ; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v67, 8, v66
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v2, v64, 8, v54
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v68, 8, v65
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    s_clause 0x1
@@ -14461,13 +14457,13 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v70, 8, v69
 ; ALIGNED-NEXT:    s_clause 0x1
 ; ALIGNED-NEXT:    buffer_load_ubyte v3, v4, s[0:3], 0 offen offset:83
 ; ALIGNED-NEXT:    buffer_load_ubyte v2, v4, s[0:3], 0 offen offset:74
 ; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:988 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
+; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v1, v80, 8, v71
 ; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
@@ -14513,9 +14509,7 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
 ; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:1416 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
 ; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:1000 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v8, v4, s[0:3], 0 offen offset:87
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
@@ -14840,21 +14834,21 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    s_clause 0x7
 ; ALIGNED-NEXT:    buffer_load_ubyte v89, v4, s[0:3], 0 offen offset:160
 ; ALIGNED-NEXT:    buffer_load_ubyte v79, v4, s[0:3], 0 offen offset:161
-; ALIGNED-NEXT:    buffer_load_ubyte v73, v4, s[0:3], 0 offen offset:162
+; ALIGNED-NEXT:    buffer_load_ubyte v75, v4, s[0:3], 0 offen offset:162
 ; ALIGNED-NEXT:    buffer_load_ubyte v74, v4, s[0:3], 0 offen offset:163
 ; ALIGNED-NEXT:    buffer_load_ubyte v88, v4, s[0:3], 0 offen offset:164
-; ALIGNED-NEXT:    buffer_load_ubyte v75, v4, s[0:3], 0 offen offset:165
-; ALIGNED-NEXT:    buffer_load_ubyte v77, v4, s[0:3], 0 offen offset:166
+; ALIGNED-NEXT:    buffer_load_ubyte v77, v4, s[0:3], 0 offen offset:165
+; ALIGNED-NEXT:    buffer_load_ubyte v76, v4, s[0:3], 0 offen offset:166
 ; ALIGNED-NEXT:    buffer_load_ubyte v72, v4, s[0:3], 0 offen offset:167
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v79, 8, v89
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    v_lshl_or_b32 v1, v74, 8, v73
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v74, 8, v75
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v1, v72, 8, v77
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v72, 8, v76
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v75, 8, v88
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v77, 8, v88
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1424 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x4
@@ -15064,7 +15058,7 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v104, v3, 16, v2
 ; ALIGNED-NEXT:    v_lshl_or_b32 v2, v21, 8, v22
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v23, 8, v20
-; ALIGNED-NEXT:    v_lshl_or_b32 v76, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v73, v3, 16, v2
 ; ALIGNED-NEXT:    v_lshl_or_b32 v2, v17, 8, v19
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v14, 8, v13
 ; ALIGNED-NEXT:    v_lshl_or_b32 v101, v3, 16, v2
@@ -15129,9 +15123,9 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    v_lshl_or_b32 v0, v57, 16, v43
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1412 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x2
+; ALIGNED-NEXT:    buffer_load_ubyte v57, v4, s[0:3], 0 offen offset:18
 ; ALIGNED-NEXT:    buffer_load_ubyte v78, v4, s[0:3], 0 offen offset:16
-; ALIGNED-NEXT:    buffer_load_ubyte v43, v4, s[0:3], 0 offen offset:18
-; ALIGNED-NEXT:    buffer_load_ubyte v57, v4, s[0:3], 0 offen offset:17
+; ALIGNED-NEXT:    buffer_load_ubyte v43, v4, s[0:3], 0 offen offset:17
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:488
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:492
 ; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:484
@@ -15140,10 +15134,10 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:704
 ; ALIGNED-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:708
 ; ALIGNED-NEXT:    v_add_nc_u32_e32 v4, 0xffffff00, v4
-; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v126, 8, v43
+; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v126, 8, v57
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    v_lshl_or_b32 v126, v57, 8, v78
+; ALIGNED-NEXT:    v_lshl_or_b32 v126, v43, 8, v78
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
 ; ALIGNED-NEXT:    v_add_co_u32 v2, vcc_lo, v2, s4
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
@@ -15164,7 +15158,7 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v16 offset:246
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v18 offset:244
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v19 offset:240
-; ALIGNED-NEXT:    buffer_store_dword v76, off, s[0:3], s32 offset:504
+; ALIGNED-NEXT:    buffer_store_dword v73, off, s[0:3], s32 offset:504
 ; ALIGNED-NEXT:    buffer_store_dword v104, off, s[0:3], s32 offset:508
 ; ALIGNED-NEXT:    buffer_store_dword v123, off, s[0:3], s32 offset:500
 ; ALIGNED-NEXT:    v_lshl_or_b32 v126, v0, 16, v126
@@ -15294,12 +15288,12 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v62 offset:174
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v63 offset:172
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v59 offset:168
-; ALIGNED-NEXT:    flat_store_byte v[2:3], v73 offset:162
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v75 offset:162
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v74 offset:163
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v79 offset:161
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v72 offset:167
-; ALIGNED-NEXT:    flat_store_byte v[2:3], v75 offset:165
-; ALIGNED-NEXT:    flat_store_byte v[2:3], v77 offset:166
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v77 offset:165
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v76 offset:166
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v88 offset:164
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v89 offset:160
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Reload
@@ -15798,11 +15792,11 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:24
-; ALIGNED-NEXT:    flat_store_byte v[2:3], v43 offset:18
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v57 offset:18
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1416 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:19
-; ALIGNED-NEXT:    flat_store_byte v[2:3], v57 offset:17
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v43 offset:17
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:23
diff --git a/llvm/test/CodeGen/AMDGPU/memmove-param-combinations.ll b/llvm/test/CodeGen/AMDGPU/memmove-param-combinations.ll
index 4e5688adcd6bb..f08ea27040fb5 100644
--- a/llvm/test/CodeGen/AMDGPU/memmove-param-combinations.ll
+++ b/llvm/test/CodeGen/AMDGPU/memmove-param-combinations.ll
@@ -485,12 +485,12 @@ define void @memmove_p0_p3_sz31_align_1_1(ptr addrspace(0) align 1 %dst, ptr add
 ; CHECK-LABEL: memmove_p0_p3_sz31_align_1_1:
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT:    ds_read_u8 v9, v2 offset:30
 ; CHECK-NEXT:    ds_read_b32 v8, v2 offset:24
+; CHECK-NEXT:    ds_read_u8 v9, v2 offset:30
 ; CHECK-NEXT:    ds_read_u16 v10, v2 offset:28
 ; CHECK-NEXT:    ds_read_b64 v[6:7], v2 offset:16
 ; CHECK-NEXT:    ds_read2_b64 v[2:5], v2 offset1:1
-; CHECK-NEXT:    s_waitcnt lgkmcnt(4)
+; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
 ; CHECK-NEXT:    flat_store_byte v[0:1], v9 offset:30
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
 ; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
@@ -540,12 +540,12 @@ define void @memmove_p0_p3_sz31_align_2_2(ptr addrspace(0) align 2 %dst, ptr add
 ; CHECK-LABEL: memmove_p0_p3_sz31_align_2_2:
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT:    ds_read_u8 v9, v2 offset:30
 ; CHECK-NEXT:    ds_read_b32 v8, v2 offset:24
+; CHECK-NEXT:    ds_read_u8 v9, v2 offset:30
 ; CHECK-NEXT:    ds_read_u16 v10, v2 offset:28
 ; CHECK-NEXT:    ds_read_b64 v[6:7], v2 offset:16
 ; CHECK-NEXT:    ds_read2_b64 v[2:5], v2 offset1:1
-; CHECK-NEXT:    s_waitcnt lgkmcnt(4)
+; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
 ; CHECK-NEXT:    flat_store_byte v[0:1], v9 offset:30
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
 ; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
@@ -939,19 +939,18 @@ define void @memmove_p0_p5_sz31_align_1_1(ptr addrspace(0) align 1 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_byte v[0:1], v11 offset:30
+; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    flat_store_short v[0:1], v11 offset:28
+; CHECK-NEXT:    flat_store_byte v[0:1], v10 offset:30
 ; CHECK-NEXT:    flat_store_dwordx3 v[0:1], v[7:9] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
@@ -967,18 +966,18 @@ define void @memmove_p0_p5_sz32_align_1_1(ptr addrspace(0) align 1 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x7
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:16
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    s_waitcnt vmcnt(1)
+; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10]
+; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
@@ -1009,19 +1008,18 @@ define void @memmove_p0_p5_sz31_align_2_2(ptr addrspace(0) align 2 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_byte v[0:1], v11 offset:30
+; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    flat_store_short v[0:1], v11 offset:28
+; CHECK-NEXT:    flat_store_byte v[0:1], v10 offset:30
 ; CHECK-NEXT:    flat_store_dwordx3 v[0:1], v[7:9] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
@@ -1037,18 +1035,18 @@ define void @memmove_p0_p5_sz32_align_2_2(ptr addrspace(0) align 2 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x7
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:16
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    s_waitcnt vmcnt(1)
+; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10]
+; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 entry:
@@ -1079,20 +1077,20 @@ define void @memmove_p0_p5_sz31_align_8_8(ptr addrspace(0) align 8 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(6)
+; CHECK-NEXT:    s_waitcnt vmcnt(3)
 ; CHECK-NEXT:    flat_store_dwordx3 v[0:1], v[7:9] offset:16
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_byte v[0:1], v11 offset:30
-; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
+; CHECK-NEXT:    flat_store_byte v[0:1], v10 offset:30
+; CHECK-NEXT:    s_waitcnt vmcnt(1)
+; CHECK-NEXT:    flat_store_short v[0:1], v11 offset:28
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1149,20 +1147,20 @@ define void @memmove_p0_p5_sz31_align_16_16(ptr addrspace(0) align 16 %dst, ptr
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(6)
+; CHECK-NEXT:    s_waitcnt vmcnt(3)
 ; CHECK-NEXT:    flat_store_dwordx3 v[0:1], v[7:9] offset:16
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_byte v[0:1], v11 offset:30
-; CHECK-NEXT:    flat_store_short v[0:1], v10 offset:28
+; CHECK-NEXT:    flat_store_byte v[0:1], v10 offset:30
+; CHECK-NEXT:    s_waitcnt vmcnt(1)
+; CHECK-NEXT:    flat_store_short v[0:1], v11 offset:28
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2079,20 +2077,18 @@ define void @memmove_p1_p5_sz31_align_1_1(ptr addrspace(1) align 1 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
 ; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    global_store_short v[0:1], v10, off offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    global_store_byte v[0:1], v11, off offset:30
 ; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    global_store_short v[0:1], v11, off offset:28
+; CHECK-NEXT:    global_store_byte v[0:1], v10, off offset:30
 ; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[3:6], off
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    global_store_dwordx3 v[0:1], v[7:9], off offset:16
@@ -2147,20 +2143,18 @@ define void @memmove_p1_p5_sz31_align_2_2(ptr addrspace(1) align 2 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
 ; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    global_store_short v[0:1], v10, off offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    global_store_byte v[0:1], v11, off offset:30
 ; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    global_store_short v[0:1], v11, off offset:28
+; CHECK-NEXT:    global_store_byte v[0:1], v10, off offset:30
 ; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[3:6], off
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    global_store_dwordx3 v[0:1], v[7:9], off offset:16
@@ -2215,20 +2209,18 @@ define void @memmove_p1_p5_sz31_align_8_8(ptr addrspace(1) align 8 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
 ; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    global_store_short v[0:1], v10, off offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    global_store_byte v[0:1], v11, off offset:30
 ; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    global_store_short v[0:1], v11, off offset:28
+; CHECK-NEXT:    global_store_byte v[0:1], v10, off offset:30
 ; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[3:6], off
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    global_store_dwordx3 v[0:1], v[7:9], off offset:16
@@ -2283,20 +2275,18 @@ define void @memmove_p1_p5_sz31_align_16_16(ptr addrspace(1) align 16 %dst, ptr
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
-; CHECK-NEXT:    buffer_load_ushort v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
 ; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_ushort v11, v2, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    global_store_short v[0:1], v10, off offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    global_store_byte v[0:1], v11, off offset:30
 ; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    global_store_short v[0:1], v11, off offset:28
+; CHECK-NEXT:    global_store_byte v[0:1], v10, off offset:30
 ; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[3:6], off
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    global_store_dwordx3 v[0:1], v[7:9], off offset:16
@@ -3266,21 +3256,20 @@ define void @memmove_p3_p5_sz31_align_1_1(ptr addrspace(3) align 1 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
-; CHECK-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_ushort v9, v1, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v10, v1, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_ubyte v8, v1, s[0:3], 0 offen offset:30
 ; CHECK-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_ushort v10, v1, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:12
 ; CHECK-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    ds_write_b32 v0, v8 offset:24
-; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    ds_write_b16 v0, v9 offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(6)
-; CHECK-NEXT:    ds_write_b8 v0, v10 offset:30
+; CHECK-NEXT:    s_waitcnt vmcnt(4)
+; CHECK-NEXT:    ds_write_b32 v0, v9 offset:24
+; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    ds_write_b16 v0, v10 offset:28
+; CHECK-NEXT:    ds_write_b8 v0, v8 offset:30
 ; CHECK-NEXT:    s_waitcnt vmcnt(2)
 ; CHECK-NEXT:    ds_write2_b64 v0, v[2:3], v[4:5] offset1:1
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
@@ -3339,21 +3328,20 @@ define void @memmove_p3_p5_sz31_align_2_2(ptr addrspace(3) align 2 %dst, ptr add
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
-; CHECK-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_ushort v9, v1, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v10, v1, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_ubyte v8, v1, s[0:3], 0 offen offset:30
 ; CHECK-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen
 ; CHECK-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_ushort v10, v1, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:12
 ; CHECK-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    ds_write_b32 v0, v8 offset:24
-; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    ds_write_b16 v0, v9 offset:28
-; CHECK-NEXT:    s_waitcnt vmcnt(6)
-; CHECK-NEXT:    ds_write_b8 v0, v10 offset:30
+; CHECK-NEXT:    s_waitcnt vmcnt(4)
+; CHECK-NEXT:    ds_write_b32 v0, v9 offset:24
+; CHECK-NEXT:    s_waitcnt vmcnt(3)
+; CHECK-NEXT:    ds_write_b16 v0, v10 offset:28
+; CHECK-NEXT:    ds_write_b8 v0, v8 offset:30
 ; CHECK-NEXT:    s_waitcnt vmcnt(2)
 ; CHECK-NEXT:    ds_write2_b64 v0, v[2:3], v[4:5] offset1:1
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
@@ -3485,22 +3473,21 @@ define void @memmove_p3_p5_sz31_align_16_16(ptr addrspace(3) align 16 %dst, ptr
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    s_clause 0x8
-; CHECK-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_ushort v9, v1, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_ubyte v10, v1, s[0:3], 0 offen offset:30
+; CHECK-NEXT:    buffer_load_ubyte v6, v1, s[0:3], 0 offen offset:30
 ; CHECK-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:24
 ; CHECK-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_ushort v10, v1, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    s_waitcnt vmcnt(6)
-; CHECK-NEXT:    ds_write2_b32 v0, v7, v8 offset0:5 offset1:6
-; CHECK-NEXT:    ds_write_b32 v0, v6 offset:16
-; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    ds_write_b16 v0, v9 offset:28
 ; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    ds_write_b8 v0, v10 offset:30
+; CHECK-NEXT:    ds_write2_b32 v0, v8, v9 offset0:5 offset1:6
+; CHECK-NEXT:    ds_write_b32 v0, v7 offset:16
+; CHECK-NEXT:    s_waitcnt vmcnt(2)
+; CHECK-NEXT:    ds_write_b16 v0, v10 offset:28
+; CHECK-NEXT:    ds_write_b8 v0, v6 offset:30
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    ds_write_b128 v0, v[2:5]
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/mul.ll b/llvm/test/CodeGen/AMDGPU/mul.ll
index 0f47a31f52dcb..b5e7589cbd134 100644
--- a/llvm/test/CodeGen/AMDGPU/mul.ll
+++ b/llvm/test/CodeGen/AMDGPU/mul.ll
@@ -2689,45 +2689,45 @@ define amdgpu_kernel void @s_mul_i128(ptr addrspace(1) %out, [8 x i32], i128 %a,
 ;
 ; GFX9-LABEL: s_mul_i128:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x4c
-; GFX9-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x7c
+; GFX9-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x7c
+; GFX9-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x4c
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; GFX9-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX9-NEXT:    s_mov_b32 s2, -1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mul_i32 s4, s12, s11
-; GFX9-NEXT:    s_mul_hi_u32 s5, s12, s10
-; GFX9-NEXT:    s_mul_i32 s6, s14, s9
-; GFX9-NEXT:    s_mul_hi_u32 s7, s14, s8
+; GFX9-NEXT:    s_mul_i32 s4, s8, s15
+; GFX9-NEXT:    s_mul_hi_u32 s5, s8, s14
+; GFX9-NEXT:    s_mul_i32 s6, s10, s13
+; GFX9-NEXT:    s_mul_hi_u32 s7, s10, s12
 ; GFX9-NEXT:    s_add_i32 s4, s5, s4
-; GFX9-NEXT:    s_mul_i32 s5, s13, s10
+; GFX9-NEXT:    s_mul_i32 s5, s9, s14
 ; GFX9-NEXT:    s_add_i32 s6, s7, s6
-; GFX9-NEXT:    s_mul_i32 s7, s15, s8
+; GFX9-NEXT:    s_mul_i32 s7, s11, s12
 ; GFX9-NEXT:    s_add_i32 s4, s4, s5
-; GFX9-NEXT:    s_mul_i32 s5, s12, s10
+; GFX9-NEXT:    s_mul_i32 s5, s8, s14
 ; GFX9-NEXT:    s_add_i32 s6, s6, s7
-; GFX9-NEXT:    s_mul_i32 s7, s14, s8
+; GFX9-NEXT:    s_mul_i32 s7, s10, s12
 ; GFX9-NEXT:    s_add_u32 s7, s7, s5
 ; GFX9-NEXT:    s_addc_u32 s6, s6, s4
-; GFX9-NEXT:    s_mul_i32 s14, s9, s12
-; GFX9-NEXT:    s_mul_hi_u32 s15, s8, s12
-; GFX9-NEXT:    s_mul_hi_u32 s11, s9, s12
+; GFX9-NEXT:    s_mul_i32 s14, s13, s8
+; GFX9-NEXT:    s_mul_hi_u32 s15, s12, s8
+; GFX9-NEXT:    s_mul_hi_u32 s11, s13, s8
 ; GFX9-NEXT:    s_add_u32 s14, s14, s15
-; GFX9-NEXT:    s_mul_i32 s5, s8, s13
+; GFX9-NEXT:    s_mul_i32 s5, s12, s9
 ; GFX9-NEXT:    s_addc_u32 s11, s11, 0
-; GFX9-NEXT:    s_mul_hi_u32 s10, s8, s13
+; GFX9-NEXT:    s_mul_hi_u32 s10, s12, s9
 ; GFX9-NEXT:    s_add_u32 s5, s5, s14
 ; GFX9-NEXT:    s_addc_u32 s10, s10, 0
 ; GFX9-NEXT:    s_add_u32 s10, s11, s10
 ; GFX9-NEXT:    s_addc_u32 s11, 0, 0
-; GFX9-NEXT:    s_mul_hi_u32 s14, s9, s13
-; GFX9-NEXT:    s_mul_i32 s9, s9, s13
+; GFX9-NEXT:    s_mul_hi_u32 s14, s13, s9
+; GFX9-NEXT:    s_mul_i32 s9, s13, s9
 ; GFX9-NEXT:    s_add_u32 s9, s9, s10
 ; GFX9-NEXT:    s_addc_u32 s10, s14, s11
 ; GFX9-NEXT:    s_mov_b32 s4, 0
 ; GFX9-NEXT:    s_add_u32 s9, s9, s7
 ; GFX9-NEXT:    s_addc_u32 s10, s10, s6
-; GFX9-NEXT:    s_mul_i32 s6, s8, s12
+; GFX9-NEXT:    s_mul_i32 s6, s12, s8
 ; GFX9-NEXT:    s_mov_b32 s7, s4
 ; GFX9-NEXT:    s_or_b64 s[4:5], s[6:7], s[4:5]
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s4
diff --git a/llvm/test/CodeGen/AMDGPU/or.ll b/llvm/test/CodeGen/AMDGPU/or.ll
index cc9650b9a7309..1abd2e6b60f2f 100644
--- a/llvm/test/CodeGen/AMDGPU/or.ll
+++ b/llvm/test/CodeGen/AMDGPU/or.ll
@@ -355,15 +355,15 @@ define amdgpu_kernel void @scalar_or_literal_multi_use_i64(ptr addrspace(1) %out
 ;
 ; GFX8-LABEL: scalar_or_literal_multi_use_i64:
 ; GFX8:       ; %bb.0:
-; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; GFX8-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x4c
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; GFX8-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x74
 ; GFX8-NEXT:    s_movk_i32 s8, 0x3039
 ; GFX8-NEXT:    s_mov_b32 s9, 0xf237b
-; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    s_or_b64 s[6:7], s[6:7], s[8:9]
 ; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
 ; GFX8-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX8-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
diff --git a/llvm/test/CodeGen/AMDGPU/permute_i8.ll b/llvm/test/CodeGen/AMDGPU/permute_i8.ll
index b4e5fa088b533..a4ddfee115fa6 100644
--- a/llvm/test/CodeGen/AMDGPU/permute_i8.ll
+++ b/llvm/test/CodeGen/AMDGPU/permute_i8.ll
@@ -3656,14 +3656,14 @@ define hidden void @extract_v6i16(ptr addrspace(1) %in0, ptr addrspace(1) %in1,
 ; GFX9-LABEL: extract_v6i16:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_load_ushort v2, v[0:1], off offset:6
-; GFX9-NEXT:    global_load_ushort v3, v[0:1], off
-; GFX9-NEXT:    global_load_ushort v8, v[0:1], off offset:4
+; GFX9-NEXT:    global_load_ushort v2, v[0:1], off offset:4
+; GFX9-NEXT:    global_load_ushort v3, v[0:1], off offset:6
+; GFX9-NEXT:    global_load_ushort v8, v[0:1], off
 ; GFX9-NEXT:    global_load_ushort v9, v[0:1], off offset:2
-; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 16, v8
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_lshl_or_b32 v0, v3, 16, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_lshl_or_b32 v1, v9, 16, v3
+; GFX9-NEXT:    v_lshl_or_b32 v1, v9, 16, v8
 ; GFX9-NEXT:    global_store_dword v[4:5], v1, off
 ; GFX9-NEXT:    global_store_dword v[6:7], v0, off
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/pr51516.mir b/llvm/test/CodeGen/AMDGPU/pr51516.mir
index cc299789af8dd..69983faf2b154 100644
--- a/llvm/test/CodeGen/AMDGPU/pr51516.mir
+++ b/llvm/test/CodeGen/AMDGPU/pr51516.mir
@@ -1,3 +1,4 @@
+# NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
 # RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -amdgpu-disable-unclustered-high-rp-reschedule -verify-misched -start-before=machine-scheduler -stop-after=virtregrewriter,2 -o - %s | FileCheck -check-prefix=GCN %s
 # RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -amdgpu-disable-unclustered-high-rp-reschedule -amdgpu-use-amdgpu-trackers=1 -verify-misched -start-before=machine-scheduler -stop-after=virtregrewriter,2 -o - %s | FileCheck -check-prefix=GCN-GCNTRACKER %s
 
@@ -116,3 +117,6 @@ body:             |
     S_ENDPGM 0
 
 ...
+## NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+# GCN: {{.*}}
+# GCN-GCNTRACKER: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/promote-constOffset-to-imm.ll b/llvm/test/CodeGen/AMDGPU/promote-constOffset-to-imm.ll
index b1a618ae4d5c5..c4842c1f4f523 100644
--- a/llvm/test/CodeGen/AMDGPU/promote-constOffset-to-imm.ll
+++ b/llvm/test/CodeGen/AMDGPU/promote-constOffset-to-imm.ll
@@ -1193,9 +1193,9 @@ define amdgpu_kernel void @Address32(ptr addrspace(1) %buffer) {
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[6:7]
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 7, v0
 ; GFX10-NEXT:    v_mov_b32_e32 v2, 2
-; GFX10-NEXT:    v_and_b32_e32 v8, 0xffff8000, v1
+; GFX10-NEXT:    v_and_b32_e32 v10, 0xffff8000, v1
 ; GFX10-NEXT:    v_lshlrev_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX10-NEXT:    v_add_co_u32 v1, s0, s34, v8
+; GFX10-NEXT:    v_add_co_u32 v1, s0, s34, v10
 ; GFX10-NEXT:    v_add_co_ci_u32_e64 v2, s0, s35, 0, s0
 ; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v1, v0
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v2, vcc_lo
@@ -1203,38 +1203,38 @@ define amdgpu_kernel void @Address32(ptr addrspace(1) %buffer) {
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v1, vcc_lo
 ; GFX10-NEXT:    v_add_co_u32 v4, vcc_lo, v0, 0x1000
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT:    s_clause 0x3
+; GFX10-NEXT:    global_load_dword v11, v[0:1], off
+; GFX10-NEXT:    global_load_dword v12, v[0:1], off offset:1024
+; GFX10-NEXT:    global_load_dword v13, v[4:5], off offset:-2048
+; GFX10-NEXT:    global_load_dword v14, v[2:3], off offset:1024
 ; GFX10-NEXT:    v_add_co_u32 v6, vcc_lo, 0x1000, v0
-; GFX10-NEXT:    s_clause 0x4
-; GFX10-NEXT:    global_load_dword v9, v[0:1], off
-; GFX10-NEXT:    global_load_dword v10, v[0:1], off offset:1024
-; GFX10-NEXT:    global_load_dword v11, v[2:3], off offset:1024
-; GFX10-NEXT:    global_load_dword v12, v[4:5], off offset:-2048
-; GFX10-NEXT:    global_load_dword v13, v[4:5], off
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v7, vcc_lo, 0, v1, vcc_lo
 ; GFX10-NEXT:    v_add_co_u32 v2, vcc_lo, 0x1800, v0
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v1, vcc_lo
-; GFX10-NEXT:    v_add_co_u32 v4, vcc_lo, v0, 0x2000
-; GFX10-NEXT:    v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
-; GFX10-NEXT:    s_clause 0x1
-; GFX10-NEXT:    global_load_dword v14, v[6:7], off offset:1024
-; GFX10-NEXT:    global_load_dword v15, v[2:3], off offset:1024
+; GFX10-NEXT:    v_add_co_u32 v8, vcc_lo, v0, 0x2000
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v9, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT:    s_clause 0x2
+; GFX10-NEXT:    global_load_dword v15, v[4:5], off
+; GFX10-NEXT:    global_load_dword v16, v[6:7], off offset:1024
+; GFX10-NEXT:    global_load_dword v17, v[2:3], off offset:1024
 ; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, 0x2000, v0
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
 ; GFX10-NEXT:    s_clause 0x2
-; GFX10-NEXT:    global_load_dword v2, v[4:5], off offset:-2048
-; GFX10-NEXT:    global_load_dword v3, v[4:5], off
-; GFX10-NEXT:    global_load_dword v6, v[0:1], off offset:1024
+; GFX10-NEXT:    global_load_dword v2, v[8:9], off offset:-2048
+; GFX10-NEXT:    global_load_dword v3, v[8:9], off
+; GFX10-NEXT:    global_load_dword v4, v[0:1], off offset:1024
 ; GFX10-NEXT:    s_waitcnt vmcnt(8)
-; GFX10-NEXT:    v_add_nc_u32_e32 v0, v10, v9
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v12, v11
 ; GFX10-NEXT:    s_waitcnt vmcnt(6)
-; GFX10-NEXT:    v_add3_u32 v0, v12, v0, v11
-; GFX10-NEXT:    s_waitcnt vmcnt(4)
 ; GFX10-NEXT:    v_add3_u32 v0, v13, v0, v14
+; GFX10-NEXT:    s_waitcnt vmcnt(4)
+; GFX10-NEXT:    v_add3_u32 v0, v15, v0, v16
 ; GFX10-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-NEXT:    v_add3_u32 v0, v2, v0, v15
+; GFX10-NEXT:    v_add3_u32 v0, v2, v0, v17
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    v_add3_u32 v0, v3, v0, v6
-; GFX10-NEXT:    global_store_dword v8, v0, s[34:35]
+; GFX10-NEXT:    v_add3_u32 v0, v3, v0, v4
+; GFX10-NEXT:    global_store_dword v10, v0, s[34:35]
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: Address32:
@@ -1375,19 +1375,19 @@ define amdgpu_kernel void @Offset64(ptr addrspace(1)  %buffer) {
 ; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v1, v0
 ; GFX8-NEXT:    v_addc_u32_e32 v4, vcc, 0, v2, vcc
 ; GFX8-NEXT:    s_movk_i32 s0, 0xf000
-; GFX8-NEXT:    v_add_u32_e32 v7, vcc, s0, v3
-; GFX8-NEXT:    v_addc_u32_e32 v8, vcc, 0, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v5, vcc, s0, v3
+; GFX8-NEXT:    v_addc_u32_e32 v6, vcc, 0, v4, vcc
 ; GFX8-NEXT:    s_movk_i32 s0, 0xf800
-; GFX8-NEXT:    flat_load_dwordx2 v[5:6], v[3:4]
-; GFX8-NEXT:    flat_load_dwordx2 v[7:8], v[7:8]
+; GFX8-NEXT:    flat_load_dwordx2 v[7:8], v[3:4]
+; GFX8-NEXT:    flat_load_dwordx2 v[5:6], v[5:6]
 ; GFX8-NEXT:    v_add_u32_e32 v9, vcc, s0, v3
 ; GFX8-NEXT:    v_addc_u32_e32 v10, vcc, 0, v4, vcc
 ; GFX8-NEXT:    flat_load_dwordx2 v[9:10], v[9:10]
 ; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v4
 ; GFX8-NEXT:    flat_load_dwordx2 v[3:4], v[3:4]
 ; GFX8-NEXT:    s_waitcnt vmcnt(2)
-; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v7, v5
-; GFX8-NEXT:    v_addc_u32_e32 v5, vcc, v8, v6, vcc
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v5, v7
+; GFX8-NEXT:    v_addc_u32_e32 v5, vcc, v6, v8, vcc
 ; GFX8-NEXT:    s_waitcnt vmcnt(1)
 ; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v9, v0
 ; GFX8-NEXT:    v_addc_u32_e32 v5, vcc, v10, v5, vcc
@@ -1429,14 +1429,14 @@ define amdgpu_kernel void @Offset64(ptr addrspace(1)  %buffer) {
 ; GFX9-NEXT:    s_movk_i32 s0, 0xf000
 ; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s0, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v1, vcc
-; GFX9-NEXT:    global_load_dwordx2 v[4:5], v[2:3], off
-; GFX9-NEXT:    global_load_dwordx2 v[6:7], v[0:1], off
+; GFX9-NEXT:    global_load_dwordx2 v[4:5], v[0:1], off
+; GFX9-NEXT:    global_load_dwordx2 v[6:7], v[2:3], off
 ; GFX9-NEXT:    global_load_dwordx2 v[8:9], v[2:3], off offset:2048
 ; GFX9-NEXT:    v_add_u32_e32 v1, 1, v1
 ; GFX9-NEXT:    global_load_dwordx2 v[0:1], v[0:1], off
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v4, v6
-; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, v5, v7, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v6, v4
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, v7, v5, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v8, v2
 ; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, v9, v3, vcc
@@ -1521,15 +1521,15 @@ define amdgpu_kernel void @Offset64(ptr addrspace(1)  %buffer) {
 ; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, 0xfffff000, v0
 ; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v1, vcc_lo
 ; GFX11-NEXT:    s_clause 0x2
-; GFX11-NEXT:    global_load_b64 v[4:5], v[2:3], off
-; GFX11-NEXT:    global_load_b64 v[6:7], v[0:1], off
+; GFX11-NEXT:    global_load_b64 v[4:5], v[0:1], off
+; GFX11-NEXT:    global_load_b64 v[6:7], v[2:3], off
 ; GFX11-NEXT:    global_load_b64 v[2:3], v[2:3], off offset:2048
 ; GFX11-NEXT:    v_add_nc_u32_e32 v1, 1, v1
 ; GFX11-NEXT:    global_load_b64 v[0:1], v[0:1], off
 ; GFX11-NEXT:    s_waitcnt vmcnt(2)
-; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v4, v6
+; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v6, v4
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, v5, v7, vcc_lo
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, v7, v5, vcc_lo
 ; GFX11-NEXT:    s_waitcnt vmcnt(1)
 ; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, v4
 ; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, v3, v5, vcc_lo
@@ -1686,26 +1686,26 @@ define amdgpu_kernel void @p32Offset64(ptr addrspace(1)  %buffer) {
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[6:7]
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 7, v0
 ; GFX10-NEXT:    v_mov_b32_e32 v2, 2
-; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff8000, v1
+; GFX10-NEXT:    v_and_b32_e32 v6, 0xffff8000, v1
 ; GFX10-NEXT:    v_lshlrev_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
-; GFX10-NEXT:    v_add_co_u32 v1, s0, s34, v4
+; GFX10-NEXT:    v_add_co_u32 v1, s0, s34, v6
 ; GFX10-NEXT:    v_add_co_ci_u32_e64 v2, s0, s35, 0, s0
 ; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v1, v0
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v2, vcc_lo
 ; GFX10-NEXT:    v_add_co_u32 v2, vcc_lo, v0, 0x80000000
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v1, vcc_lo
-; GFX10-NEXT:    global_load_dword v5, v[0:1], off
-; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, 0x7ffff800, v0
-; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
-; GFX10-NEXT:    s_clause 0x2
-; GFX10-NEXT:    global_load_dword v6, v[2:3], off offset:-2048
-; GFX10-NEXT:    global_load_dword v7, v[2:3], off
-; GFX10-NEXT:    global_load_dword v8, v[0:1], off offset:1024
+; GFX10-NEXT:    v_add_co_u32 v4, vcc_lo, 0x7ffff800, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT:    s_clause 0x3
+; GFX10-NEXT:    global_load_dword v7, v[0:1], off
+; GFX10-NEXT:    global_load_dword v8, v[2:3], off offset:-2048
+; GFX10-NEXT:    global_load_dword v9, v[2:3], off
+; GFX10-NEXT:    global_load_dword v10, v[4:5], off offset:1024
 ; GFX10-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-NEXT:    v_add_nc_u32_e32 v0, v6, v5
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v8, v7
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    v_add3_u32 v0, v8, v0, v7
-; GFX10-NEXT:    global_store_dword v4, v0, s[34:35]
+; GFX10-NEXT:    v_add3_u32 v0, v10, v0, v9
+; GFX10-NEXT:    global_store_dword v6, v0, s[34:35]
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: p32Offset64:
@@ -2160,25 +2160,25 @@ define amdgpu_kernel void @ReverseOrder(ptr addrspace(1) %buffer) {
 ; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v2, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
 ; GFX9-NEXT:    s_movk_i32 s0, 0x3000
-; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, s0, v0
-; GFX9-NEXT:    global_load_dwordx2 v[2:3], v[0:1], off
-; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
-; GFX9-NEXT:    global_load_dwordx2 v[6:7], v[4:5], off offset:2048
-; GFX9-NEXT:    global_load_dwordx2 v[8:9], v[4:5], off
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s0, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v1, vcc
+; GFX9-NEXT:    global_load_dwordx2 v[4:5], v[0:1], off
+; GFX9-NEXT:    global_load_dwordx2 v[6:7], v[2:3], off offset:2048
+; GFX9-NEXT:    global_load_dwordx2 v[8:9], v[2:3], off
 ; GFX9-NEXT:    s_movk_i32 s0, 0x2000
-; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, s0, v0
-; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
-; GFX9-NEXT:    global_load_dwordx2 v[10:11], v[4:5], off offset:2048
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s0, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v1, vcc
+; GFX9-NEXT:    global_load_dwordx2 v[10:11], v[2:3], off offset:2048
 ; GFX9-NEXT:    s_movk_i32 s0, 0x1000
 ; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, s0, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v1, vcc
 ; GFX9-NEXT:    global_load_dwordx2 v[14:15], v[12:13], off
-; GFX9-NEXT:    global_load_dwordx2 v[16:17], v[4:5], off
+; GFX9-NEXT:    global_load_dwordx2 v[16:17], v[2:3], off
 ; GFX9-NEXT:    global_load_dwordx2 v[18:19], v[12:13], off offset:2048
 ; GFX9-NEXT:    global_load_dwordx2 v[20:21], v[0:1], off offset:2048
 ; GFX9-NEXT:    s_waitcnt vmcnt(6)
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v6, v2
-; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v7, v3, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v6, v4
+; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v7, v5, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(5)
 ; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v8, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v9, v1, vcc
diff --git a/llvm/test/CodeGen/AMDGPU/select.f16.ll b/llvm/test/CodeGen/AMDGPU/select.f16.ll
index 2863fccac9fbc..51de691e0eccc 100644
--- a/llvm/test/CodeGen/AMDGPU/select.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/select.f16.ll
@@ -3214,11 +3214,11 @@ define <32 x half> @v_vselect_v32f16(<32 x half> %a, <32 x half> %b, <32 x i32>
 ; VI-NEXT:    v_lshrrev_b32_e32 v55, 16, v2
 ; VI-NEXT:    v_lshrrev_b32_e32 v43, 16, v18
 ; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:108
-; VI-NEXT:    s_waitcnt vmcnt(5)
+; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:92
+; VI-NEXT:    s_waitcnt vmcnt(6)
 ; VI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v49
 ; VI-NEXT:    v_cndmask_b32_e32 v49, v43, v55, vcc
 ; VI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:100
-; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:92
 ; VI-NEXT:    v_lshrrev_b32_e32 v43, 16, v1
 ; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
 ; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:84
@@ -3249,10 +3249,9 @@ define <32 x half> @v_vselect_v32f16(<32 x half> %a, <32 x half> %b, <32 x i32>
 ; VI-NEXT:    s_waitcnt vmcnt(13)
 ; VI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v44
 ; VI-NEXT:    v_cndmask_b32_e32 v13, v29, v13, vcc
-; VI-NEXT:    s_waitcnt vmcnt(12)
+; VI-NEXT:    s_waitcnt vmcnt(11)
 ; VI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v55
 ; VI-NEXT:    v_cndmask_b32_e32 v12, v28, v12, vcc
-; VI-NEXT:    s_waitcnt vmcnt(11)
 ; VI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v45
 ; VI-NEXT:    v_cndmask_b32_e32 v11, v27, v11, vcc
 ; VI-NEXT:    s_waitcnt vmcnt(10)
diff --git a/llvm/test/CodeGen/AMDGPU/shl.ll b/llvm/test/CodeGen/AMDGPU/shl.ll
index 593cff712004a..a82a6a8a4c367 100644
--- a/llvm/test/CodeGen/AMDGPU/shl.ll
+++ b/llvm/test/CodeGen/AMDGPU/shl.ll
@@ -878,19 +878,19 @@ define amdgpu_kernel void @shl_v4i64(ptr addrspace(1) %out, ptr addrspace(1) %in
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    s_mov_b32 s8, s2
 ; SI-NEXT:    s_mov_b32 s9, s3
-; SI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:16
-; SI-NEXT:    buffer_load_dwordx4 v[4:7], off, s[8:11], 0 offset:48
+; SI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:48
+; SI-NEXT:    buffer_load_dwordx4 v[3:6], off, s[8:11], 0 offset:16
 ; SI-NEXT:    buffer_load_dwordx4 v[7:10], off, s[8:11], 0
 ; SI-NEXT:    buffer_load_dwordx4 v[11:14], off, s[8:11], 0 offset:32
 ; SI-NEXT:    s_mov_b32 s4, s0
 ; SI-NEXT:    s_mov_b32 s5, s1
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshl_b64 v[2:3], v[2:3], v6
-; SI-NEXT:    v_lshl_b64 v[0:1], v[0:1], v4
+; SI-NEXT:    v_lshl_b64 v[5:6], v[5:6], v2
+; SI-NEXT:    v_lshl_b64 v[3:4], v[3:4], v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshl_b64 v[9:10], v[9:10], v13
 ; SI-NEXT:    v_lshl_b64 v[7:8], v[7:8], v11
-; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0 offset:16
+; SI-NEXT:    buffer_store_dwordx4 v[3:6], off, s[4:7], 0 offset:16
 ; SI-NEXT:    buffer_store_dwordx4 v[7:10], off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/sra.ll b/llvm/test/CodeGen/AMDGPU/sra.ll
index ce89b2a962eea..0b49b9c815da5 100644
--- a/llvm/test/CodeGen/AMDGPU/sra.ll
+++ b/llvm/test/CodeGen/AMDGPU/sra.ll
@@ -563,19 +563,19 @@ define amdgpu_kernel void @ashr_v4i64(ptr addrspace(1) %out, ptr addrspace(1) %i
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    s_mov_b32 s8, s2
 ; SI-NEXT:    s_mov_b32 s9, s3
-; SI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:16
-; SI-NEXT:    buffer_load_dwordx4 v[4:7], off, s[8:11], 0 offset:48
+; SI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:48
+; SI-NEXT:    buffer_load_dwordx4 v[3:6], off, s[8:11], 0 offset:16
 ; SI-NEXT:    buffer_load_dwordx4 v[7:10], off, s[8:11], 0
 ; SI-NEXT:    buffer_load_dwordx4 v[11:14], off, s[8:11], 0 offset:32
 ; SI-NEXT:    s_mov_b32 s4, s0
 ; SI-NEXT:    s_mov_b32 s5, s1
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_ashr_i64 v[2:3], v[2:3], v6
-; SI-NEXT:    v_ashr_i64 v[0:1], v[0:1], v4
+; SI-NEXT:    v_ashr_i64 v[5:6], v[5:6], v2
+; SI-NEXT:    v_ashr_i64 v[3:4], v[3:4], v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_ashr_i64 v[9:10], v[9:10], v13
 ; SI-NEXT:    v_ashr_i64 v[7:8], v[7:8], v11
-; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0 offset:16
+; SI-NEXT:    buffer_store_dwordx4 v[3:6], off, s[4:7], 0 offset:16
 ; SI-NEXT:    buffer_store_dwordx4 v[7:10], off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -589,19 +589,19 @@ define amdgpu_kernel void @ashr_v4i64(ptr addrspace(1) %out, ptr addrspace(1) %i
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_mov_b32 s8, s6
 ; VI-NEXT:    s_mov_b32 s9, s7
-; VI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:16
-; VI-NEXT:    buffer_load_dwordx4 v[4:7], off, s[8:11], 0 offset:48
+; VI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:48
+; VI-NEXT:    buffer_load_dwordx4 v[3:6], off, s[8:11], 0 offset:16
 ; VI-NEXT:    buffer_load_dwordx4 v[7:10], off, s[8:11], 0
 ; VI-NEXT:    buffer_load_dwordx4 v[11:14], off, s[8:11], 0 offset:32
 ; VI-NEXT:    s_mov_b32 s0, s4
 ; VI-NEXT:    s_mov_b32 s1, s5
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_ashrrev_i64 v[2:3], v6, v[2:3]
-; VI-NEXT:    v_ashrrev_i64 v[0:1], v4, v[0:1]
+; VI-NEXT:    v_ashrrev_i64 v[5:6], v2, v[5:6]
+; VI-NEXT:    v_ashrrev_i64 v[3:4], v0, v[3:4]
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_ashrrev_i64 v[9:10], v13, v[9:10]
 ; VI-NEXT:    v_ashrrev_i64 v[7:8], v11, v[7:8]
-; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
+; VI-NEXT:    buffer_store_dwordx4 v[3:6], off, s[0:3], 0 offset:16
 ; VI-NEXT:    buffer_store_dwordx4 v[7:10], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/srl.ll b/llvm/test/CodeGen/AMDGPU/srl.ll
index badb1f6fe9847..239de43baa457 100644
--- a/llvm/test/CodeGen/AMDGPU/srl.ll
+++ b/llvm/test/CodeGen/AMDGPU/srl.ll
@@ -266,19 +266,19 @@ define amdgpu_kernel void @lshr_v4i64(ptr addrspace(1) %out, ptr addrspace(1) %i
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    s_mov_b32 s8, s2
 ; SI-NEXT:    s_mov_b32 s9, s3
-; SI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:16
-; SI-NEXT:    buffer_load_dwordx4 v[4:7], off, s[8:11], 0 offset:48
+; SI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:48
+; SI-NEXT:    buffer_load_dwordx4 v[3:6], off, s[8:11], 0 offset:16
 ; SI-NEXT:    buffer_load_dwordx4 v[7:10], off, s[8:11], 0
 ; SI-NEXT:    buffer_load_dwordx4 v[11:14], off, s[8:11], 0 offset:32
 ; SI-NEXT:    s_mov_b32 s4, s0
 ; SI-NEXT:    s_mov_b32 s5, s1
 ; SI-NEXT:    s_waitcnt vmcnt(2)
-; SI-NEXT:    v_lshr_b64 v[2:3], v[2:3], v6
-; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], v4
+; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], v2
+; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], v13
 ; SI-NEXT:    v_lshr_b64 v[7:8], v[7:8], v11
-; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0 offset:16
+; SI-NEXT:    buffer_store_dwordx4 v[3:6], off, s[4:7], 0 offset:16
 ; SI-NEXT:    buffer_store_dwordx4 v[7:10], off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/store-local.128.ll b/llvm/test/CodeGen/AMDGPU/store-local.128.ll
index dcf3643756cb2..2efa022efd70f 100644
--- a/llvm/test/CodeGen/AMDGPU/store-local.128.ll
+++ b/llvm/test/CodeGen/AMDGPU/store-local.128.ll
@@ -279,36 +279,37 @@ define amdgpu_kernel void @store_lds_v4i32_align1(ptr addrspace(3) %out, <4 x i3
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x10
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s1
 ; GFX11-NEXT:    s_lshr_b32 s4, s3, 8
-; GFX11-NEXT:    s_lshr_b32 s3, s3, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s4
 ; GFX11-NEXT:    s_lshr_b32 s5, s2, 8
-; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s1
 ; GFX11-NEXT:    s_lshr_b32 s2, s2, 24
 ; GFX11-NEXT:    s_lshr_b32 s6, s1, 8
-; GFX11-NEXT:    v_dual_mov_b32 v6, s3 :: v_dual_mov_b32 v7, s5
-; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s6
-; GFX11-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s4
+; GFX11-NEXT:    s_lshr_b32 s3, s3, 24
 ; GFX11-NEXT:    s_lshr_b32 s1, s1, 24
 ; GFX11-NEXT:    s_lshr_b32 s7, s0, 8
 ; GFX11-NEXT:    s_lshr_b32 s0, s0, 24
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s6
+; GFX11-NEXT:    v_dual_mov_b32 v6, s3 :: v_dual_mov_b32 v7, s5
+; GFX11-NEXT:    v_mov_b32_e32 v10, s1
 ; GFX11-NEXT:    ds_store_b8 v0, v2 offset:8
 ; GFX11-NEXT:    ds_store_b8_d16_hi v0, v2 offset:10
 ; GFX11-NEXT:    ds_store_b8 v0, v1 offset:12
-; GFX11-NEXT:    ds_store_b8 v0, v4
-; GFX11-NEXT:    ds_store_b8_d16_hi v0, v4 offset:2
-; GFX11-NEXT:    ds_store_b8 v0, v3 offset:4
-; GFX11-NEXT:    ds_store_b8 v0, v5 offset:13
 ; GFX11-NEXT:    ds_store_b8_d16_hi v0, v1 offset:14
+; GFX11-NEXT:    ds_store_b8 v0, v5 offset:13
 ; GFX11-NEXT:    ds_store_b8 v0, v6 offset:15
-; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v10, s7
-; GFX11-NEXT:    v_mov_b32_e32 v11, s0
 ; GFX11-NEXT:    ds_store_b8 v0, v7 offset:9
 ; GFX11-NEXT:    ds_store_b8 v0, v8 offset:11
-; GFX11-NEXT:    ds_store_b8 v0, v9 offset:5
+; GFX11-NEXT:    v_dual_mov_b32 v1, s7 :: v_dual_mov_b32 v2, s0
+; GFX11-NEXT:    ds_store_b8 v0, v4
+; GFX11-NEXT:    ds_store_b8_d16_hi v0, v4 offset:2
+; GFX11-NEXT:    ds_store_b8 v0, v3 offset:4
 ; GFX11-NEXT:    ds_store_b8_d16_hi v0, v3 offset:6
-; GFX11-NEXT:    ds_store_b8 v0, v1 offset:7
-; GFX11-NEXT:    ds_store_b8 v0, v10 offset:1
-; GFX11-NEXT:    ds_store_b8 v0, v11 offset:3
+; GFX11-NEXT:    ds_store_b8 v0, v9 offset:5
+; GFX11-NEXT:    ds_store_b8 v0, v10 offset:7
+; GFX11-NEXT:    ds_store_b8 v0, v1 offset:1
+; GFX11-NEXT:    ds_store_b8 v0, v2 offset:3
 ; GFX11-NEXT:    s_endpgm
   store <4 x i32> %x, ptr addrspace(3) %out, align 1
   ret void
@@ -420,17 +421,17 @@ define amdgpu_kernel void @store_lds_v4i32_align2(ptr addrspace(3) %out, <4 x i3
 ; GFX11-NEXT:    s_load_b32 s6, s[4:5], 0x0
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x10
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s3
-; GFX11-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v3, s1
-; GFX11-NEXT:    v_mov_b32_e32 v4, s2
-; GFX11-NEXT:    ds_store_b16_d16_hi v0, v1 offset:14
-; GFX11-NEXT:    ds_store_b16 v0, v2
-; GFX11-NEXT:    ds_store_b16 v0, v3 offset:4
-; GFX11-NEXT:    ds_store_b16 v0, v4 offset:8
-; GFX11-NEXT:    ds_store_b16 v0, v1 offset:12
-; GFX11-NEXT:    ds_store_b16_d16_hi v0, v4 offset:10
-; GFX11-NEXT:    ds_store_b16_d16_hi v0, v3 offset:6
-; GFX11-NEXT:    ds_store_b16_d16_hi v0, v2 offset:2
+; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s0
+; GFX11-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v3, s2
+; GFX11-NEXT:    v_mov_b32_e32 v4, s3
+; GFX11-NEXT:    ds_store_b16 v0, v1
+; GFX11-NEXT:    ds_store_b16 v0, v2 offset:4
+; GFX11-NEXT:    ds_store_b16 v0, v3 offset:8
+; GFX11-NEXT:    ds_store_b16 v0, v4 offset:12
+; GFX11-NEXT:    ds_store_b16_d16_hi v0, v4 offset:14
+; GFX11-NEXT:    ds_store_b16_d16_hi v0, v3 offset:10
+; GFX11-NEXT:    ds_store_b16_d16_hi v0, v2 offset:6
+; GFX11-NEXT:    ds_store_b16_d16_hi v0, v1 offset:2
 ; GFX11-NEXT:    s_endpgm
   store <4 x i32> %x, ptr addrspace(3) %out, align 2
   ret void
diff --git a/llvm/test/CodeGen/AMDGPU/store-local.96.ll b/llvm/test/CodeGen/AMDGPU/store-local.96.ll
index 70906d8474aa5..03a7ec4883ff8 100644
--- a/llvm/test/CodeGen/AMDGPU/store-local.96.ll
+++ b/llvm/test/CodeGen/AMDGPU/store-local.96.ll
@@ -239,25 +239,26 @@ define amdgpu_kernel void @store_lds_v3i32_align1(ptr addrspace(3) %out, <3 x i3
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x10
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s2
-; GFX11-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v3, s0
 ; GFX11-NEXT:    s_lshr_b32 s3, s2, 8
 ; GFX11-NEXT:    s_lshr_b32 s2, s2, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v4, s3 :: v_dual_mov_b32 v5, s2
+; GFX11-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v3, s0
 ; GFX11-NEXT:    s_lshr_b32 s4, s1, 8
 ; GFX11-NEXT:    s_lshr_b32 s1, s1, 24
 ; GFX11-NEXT:    s_lshr_b32 s5, s0, 8
 ; GFX11-NEXT:    s_lshr_b32 s0, s0, 24
-; GFX11-NEXT:    v_dual_mov_b32 v4, s3 :: v_dual_mov_b32 v5, s2
 ; GFX11-NEXT:    v_dual_mov_b32 v6, s4 :: v_dual_mov_b32 v7, s1
 ; GFX11-NEXT:    v_dual_mov_b32 v8, s5 :: v_dual_mov_b32 v9, s0
 ; GFX11-NEXT:    ds_store_b8 v0, v1 offset:8
+; GFX11-NEXT:    ds_store_b8_d16_hi v0, v1 offset:10
+; GFX11-NEXT:    ds_store_b8 v0, v4 offset:9
+; GFX11-NEXT:    ds_store_b8 v0, v5 offset:11
 ; GFX11-NEXT:    ds_store_b8 v0, v3
 ; GFX11-NEXT:    ds_store_b8_d16_hi v0, v3 offset:2
 ; GFX11-NEXT:    ds_store_b8 v0, v2 offset:4
-; GFX11-NEXT:    ds_store_b8 v0, v4 offset:9
-; GFX11-NEXT:    ds_store_b8_d16_hi v0, v1 offset:10
-; GFX11-NEXT:    ds_store_b8 v0, v5 offset:11
-; GFX11-NEXT:    ds_store_b8 v0, v6 offset:5
 ; GFX11-NEXT:    ds_store_b8_d16_hi v0, v2 offset:6
+; GFX11-NEXT:    ds_store_b8 v0, v6 offset:5
 ; GFX11-NEXT:    ds_store_b8 v0, v7 offset:7
 ; GFX11-NEXT:    ds_store_b8 v0, v8 offset:1
 ; GFX11-NEXT:    ds_store_b8 v0, v9 offset:3
@@ -356,14 +357,14 @@ define amdgpu_kernel void @store_lds_v3i32_align2(ptr addrspace(3) %out, <3 x i3
 ; GFX11-NEXT:    s_load_b32 s6, s[4:5], 0x0
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x10
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s2
-; GFX11-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v3, s1
-; GFX11-NEXT:    ds_store_b16_d16_hi v0, v1 offset:10
-; GFX11-NEXT:    ds_store_b16 v0, v2
-; GFX11-NEXT:    ds_store_b16 v0, v3 offset:4
-; GFX11-NEXT:    ds_store_b16 v0, v1 offset:8
-; GFX11-NEXT:    ds_store_b16_d16_hi v0, v3 offset:6
-; GFX11-NEXT:    ds_store_b16_d16_hi v0, v2 offset:2
+; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s0
+; GFX11-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v3, s2
+; GFX11-NEXT:    ds_store_b16 v0, v1
+; GFX11-NEXT:    ds_store_b16 v0, v2 offset:4
+; GFX11-NEXT:    ds_store_b16 v0, v3 offset:8
+; GFX11-NEXT:    ds_store_b16_d16_hi v0, v3 offset:10
+; GFX11-NEXT:    ds_store_b16_d16_hi v0, v2 offset:6
+; GFX11-NEXT:    ds_store_b16_d16_hi v0, v1 offset:2
 ; GFX11-NEXT:    s_endpgm
   store <3 x i32> %x, ptr addrspace(3) %out, align 2
   ret void
diff --git a/llvm/test/CodeGen/AMDGPU/sub.ll b/llvm/test/CodeGen/AMDGPU/sub.ll
index a3aeea8a145cd..ec065b4daa376 100644
--- a/llvm/test/CodeGen/AMDGPU/sub.ll
+++ b/llvm/test/CodeGen/AMDGPU/sub.ll
@@ -967,20 +967,20 @@ define amdgpu_kernel void @v_test_sub_v4i64(ptr addrspace(1) %out, ptr addrspace
 ;
 ; GFX9-LABEL: v_test_sub_v4i64:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 5, v0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_load_dwordx4 v[0:3], v16, s[2:3]
-; GFX9-NEXT:    global_load_dwordx4 v[4:7], v16, s[6:7]
+; GFX9-NEXT:    global_load_dwordx4 v[0:3], v16, s[6:7]
+; GFX9-NEXT:    global_load_dwordx4 v[4:7], v16, s[2:3]
 ; GFX9-NEXT:    global_load_dwordx4 v[8:11], v16, s[2:3] offset:16
 ; GFX9-NEXT:    global_load_dwordx4 v[12:15], v16, s[6:7] offset:16
 ; GFX9-NEXT:    v_mov_b32_e32 v16, 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_sub_co_u32_e32 v2, vcc, v2, v6
-; GFX9-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v7, vcc
-; GFX9-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v4
-; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v5, vcc
+; GFX9-NEXT:    v_sub_co_u32_e32 v2, vcc, v6, v2
+; GFX9-NEXT:    v_subb_co_u32_e32 v3, vcc, v7, v3, vcc
+; GFX9-NEXT:    v_sub_co_u32_e32 v0, vcc, v4, v0
+; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v5, v1, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_sub_co_u32_e32 v6, vcc, v10, v14
 ; GFX9-NEXT:    v_subb_co_u32_e32 v7, vcc, v11, v15, vcc
@@ -993,22 +993,22 @@ define amdgpu_kernel void @v_test_sub_v4i64(ptr addrspace(1) %out, ptr addrspace
 ; GFX12-LABEL: v_test_sub_v4i64:
 ; GFX12:       ; %bb.0:
 ; GFX12-NEXT:    s_clause 0x1
+; GFX12-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
 ; GFX12-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX12-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
 ; GFX12-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
 ; GFX12-NEXT:    v_mov_b32_e32 v16, 0
 ; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v12, 5, v0
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
 ; GFX12-NEXT:    s_clause 0x3
-; GFX12-NEXT:    global_load_b128 v[0:3], v12, s[2:3]
-; GFX12-NEXT:    global_load_b128 v[4:7], v12, s[4:5]
+; GFX12-NEXT:    global_load_b128 v[0:3], v12, s[6:7]
+; GFX12-NEXT:    global_load_b128 v[4:7], v12, s[2:3]
 ; GFX12-NEXT:    global_load_b128 v[8:11], v12, s[2:3] offset:16
-; GFX12-NEXT:    global_load_b128 v[12:15], v12, s[4:5] offset:16
+; GFX12-NEXT:    global_load_b128 v[12:15], v12, s[6:7] offset:16
 ; GFX12-NEXT:    s_wait_loadcnt 0x2
-; GFX12-NEXT:    v_sub_co_u32 v2, vcc_lo, v2, v6
+; GFX12-NEXT:    v_sub_co_u32 v2, vcc_lo, v6, v2
 ; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT:    v_sub_co_ci_u32_e64 v3, null, v3, v7, vcc_lo
+; GFX12-NEXT:    v_sub_co_ci_u32_e64 v3, null, v7, v3, vcc_lo
 ; GFX12-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-NEXT:    v_sub_co_u32 v10, vcc_lo, v10, v14
 ; GFX12-NEXT:    s_wait_alu 0xfffd
@@ -1016,9 +1016,9 @@ define amdgpu_kernel void @v_test_sub_v4i64(ptr addrspace(1) %out, ptr addrspace
 ; GFX12-NEXT:    v_sub_co_u32 v8, vcc_lo, v8, v12
 ; GFX12-NEXT:    s_wait_alu 0xfffd
 ; GFX12-NEXT:    v_sub_co_ci_u32_e64 v9, null, v9, v13, vcc_lo
-; GFX12-NEXT:    v_sub_co_u32 v0, vcc_lo, v0, v4
+; GFX12-NEXT:    v_sub_co_u32 v0, vcc_lo, v4, v0
 ; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    v_sub_co_ci_u32_e64 v1, null, v1, v5, vcc_lo
+; GFX12-NEXT:    v_sub_co_ci_u32_e64 v1, null, v5, v1, vcc_lo
 ; GFX12-NEXT:    s_clause 0x1
 ; GFX12-NEXT:    global_store_b128 v16, v[8:11], s[0:1] offset:16
 ; GFX12-NEXT:    global_store_b128 v16, v[0:3], s[0:1]
diff --git a/llvm/test/CodeGen/AMDGPU/udivrem.ll b/llvm/test/CodeGen/AMDGPU/udivrem.ll
index a56346f3bb45b..74e536f813716 100644
--- a/llvm/test/CodeGen/AMDGPU/udivrem.ll
+++ b/llvm/test/CodeGen/AMDGPU/udivrem.ll
@@ -37,22 +37,23 @@ define amdgpu_kernel void @test_udivrem(ptr addrspace(1) %out0, [8 x i32], ptr a
 ; GFX6-LABEL: test_udivrem:
 ; GFX6:       ; %bb.0:
 ; GFX6-NEXT:    s_load_dword s8, s[4:5], 0x26
-; GFX6-NEXT:    s_load_dword s9, s[4:5], 0x1d
 ; GFX6-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX6-NEXT:    s_mov_b32 s2, -1
 ; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
 ; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX6-NEXT:    v_cvt_f32_u32_e32 v0, s8
 ; GFX6-NEXT:    s_sub_i32 s0, 0, s8
-; GFX6-NEXT:    s_mov_b32 s7, s3
 ; GFX6-NEXT:    v_rcp_iflag_f32_e32 v0, v0
 ; GFX6-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
 ; GFX6-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GFX6-NEXT:    v_mul_lo_u32 v1, s0, v0
+; GFX6-NEXT:    s_load_dword s9, s[4:5], 0x1d
 ; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
 ; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x13
 ; GFX6-NEXT:    v_mul_hi_u32 v1, v0, v1
 ; GFX6-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX6-NEXT:    v_mul_hi_u32 v0, s9, v0
 ; GFX6-NEXT:    v_readfirstlane_b32 s10, v0
 ; GFX6-NEXT:    s_mul_i32 s10, s10, s8
@@ -69,7 +70,6 @@ define amdgpu_kernel void @test_udivrem(ptr addrspace(1) %out0, [8 x i32], ptr a
 ; GFX6-NEXT:    s_cselect_b64 vcc, -1, 0
 ; GFX6-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
 ; GFX6-NEXT:    s_cselect_b32 s8, s10, s9
-; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX6-NEXT:    s_waitcnt expcnt(0)
 ; GFX6-NEXT:    v_mov_b32_e32 v0, s8
@@ -79,7 +79,6 @@ define amdgpu_kernel void @test_udivrem(ptr addrspace(1) %out0, [8 x i32], ptr a
 ; GFX8-LABEL: test_udivrem:
 ; GFX8:       ; %bb.0:
 ; GFX8-NEXT:    s_load_dword s6, s[4:5], 0x98
-; GFX8-NEXT:    s_load_dword s7, s[4:5], 0x74
 ; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s6
 ; GFX8-NEXT:    s_sub_i32 s0, 0, s6
@@ -87,6 +86,7 @@ define amdgpu_kernel void @test_udivrem(ptr addrspace(1) %out0, [8 x i32], ptr a
 ; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
 ; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GFX8-NEXT:    v_mul_lo_u32 v1, s0, v0
+; GFX8-NEXT:    s_load_dword s7, s[4:5], 0x74
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; GFX8-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x4c
 ; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/wide-scalar-shift-by-byte-multiple-legalization.ll b/llvm/test/CodeGen/RISCV/GlobalISel/wide-scalar-shift-by-byte-multiple-legalization.ll
index 5278c951ecddf..69519c00f88ea 100644
--- a/llvm/test/CodeGen/RISCV/GlobalISel/wide-scalar-shift-by-byte-multiple-legalization.ll
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/wide-scalar-shift-by-byte-multiple-legalization.ll
@@ -1481,38 +1481,38 @@ define void @shl_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ;
 ; RV32I-LABEL: shl_16bytes:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
+; RV32I-NEXT:    lbu a3, 0(a0)
+; RV32I-NEXT:    lbu a4, 1(a0)
 ; RV32I-NEXT:    lbu a5, 2(a0)
 ; RV32I-NEXT:    lbu a6, 3(a0)
-; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 4(a0)
-; RV32I-NEXT:    lbu a7, 5(a0)
-; RV32I-NEXT:    lbu t0, 6(a0)
-; RV32I-NEXT:    lbu t1, 7(a0)
+; RV32I-NEXT:    lbu a7, 4(a0)
+; RV32I-NEXT:    lbu t0, 5(a0)
+; RV32I-NEXT:    lbu t1, 6(a0)
+; RV32I-NEXT:    lbu t2, 7(a0)
+; RV32I-NEXT:    slli a4, a4, 8
+; RV32I-NEXT:    or a3, a4, a3
 ; RV32I-NEXT:    slli a6, a6, 8
-; RV32I-NEXT:    slli a7, a7, 8
-; RV32I-NEXT:    or a5, a6, a5
-; RV32I-NEXT:    or a4, a7, a4
-; RV32I-NEXT:    lbu a6, 0(a1)
-; RV32I-NEXT:    lbu a7, 1(a1)
-; RV32I-NEXT:    slli t1, t1, 8
-; RV32I-NEXT:    or t0, t1, t0
+; RV32I-NEXT:    slli t0, t0, 8
+; RV32I-NEXT:    slli t2, t2, 8
+; RV32I-NEXT:    or a4, a6, a5
+; RV32I-NEXT:    or a5, t0, a7
+; RV32I-NEXT:    or a7, t2, t1
+; RV32I-NEXT:    lbu a6, 1(a1)
+; RV32I-NEXT:    lbu t0, 0(a1)
 ; RV32I-NEXT:    lbu t1, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a7, a7, 8
-; RV32I-NEXT:    or a7, a7, a6
+; RV32I-NEXT:    slli a6, a6, 8
+; RV32I-NEXT:    or t2, a6, t0
 ; RV32I-NEXT:    li a6, 64
 ; RV32I-NEXT:    slli a1, a1, 8
 ; RV32I-NEXT:    or a1, a1, t1
 ; RV32I-NEXT:    li t1, 32
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli t2, t0, 16
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a7, a7, 16
 ; RV32I-NEXT:    slli a1, a1, 16
-; RV32I-NEXT:    or t0, a5, a3
-; RV32I-NEXT:    or a4, t2, a4
-; RV32I-NEXT:    or a5, a1, a7
+; RV32I-NEXT:    or t0, a4, a3
+; RV32I-NEXT:    or a4, a7, a5
+; RV32I-NEXT:    or a5, a1, t2
 ; RV32I-NEXT:    slli a5, a5, 3
 ; RV32I-NEXT:    neg t3, a5
 ; RV32I-NEXT:    srl t4, t0, t3
@@ -1810,38 +1810,38 @@ define void @shl_16bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) nounw
 ;
 ; RV32I-LABEL: shl_16bytes_wordOff:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
+; RV32I-NEXT:    lbu a3, 0(a0)
+; RV32I-NEXT:    lbu a4, 1(a0)
 ; RV32I-NEXT:    lbu a5, 2(a0)
 ; RV32I-NEXT:    lbu a6, 3(a0)
-; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 4(a0)
-; RV32I-NEXT:    lbu a7, 5(a0)
-; RV32I-NEXT:    lbu t0, 6(a0)
-; RV32I-NEXT:    lbu t1, 7(a0)
+; RV32I-NEXT:    lbu a7, 4(a0)
+; RV32I-NEXT:    lbu t0, 5(a0)
+; RV32I-NEXT:    lbu t1, 6(a0)
+; RV32I-NEXT:    lbu t2, 7(a0)
+; RV32I-NEXT:    slli a4, a4, 8
+; RV32I-NEXT:    or a3, a4, a3
 ; RV32I-NEXT:    slli a6, a6, 8
-; RV32I-NEXT:    slli a7, a7, 8
-; RV32I-NEXT:    or a5, a6, a5
-; RV32I-NEXT:    or a4, a7, a4
-; RV32I-NEXT:    lbu a6, 0(a1)
-; RV32I-NEXT:    lbu a7, 1(a1)
-; RV32I-NEXT:    slli t1, t1, 8
-; RV32I-NEXT:    or t0, t1, t0
+; RV32I-NEXT:    slli t0, t0, 8
+; RV32I-NEXT:    slli t2, t2, 8
+; RV32I-NEXT:    or a4, a6, a5
+; RV32I-NEXT:    or a5, t0, a7
+; RV32I-NEXT:    or a7, t2, t1
+; RV32I-NEXT:    lbu a6, 1(a1)
+; RV32I-NEXT:    lbu t0, 0(a1)
 ; RV32I-NEXT:    lbu t1, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a7, a7, 8
-; RV32I-NEXT:    or a7, a7, a6
+; RV32I-NEXT:    slli a6, a6, 8
+; RV32I-NEXT:    or t2, a6, t0
 ; RV32I-NEXT:    li a6, 64
 ; RV32I-NEXT:    slli a1, a1, 8
 ; RV32I-NEXT:    or a1, a1, t1
 ; RV32I-NEXT:    li t1, 32
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli t2, t0, 16
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a7, a7, 16
 ; RV32I-NEXT:    slli a1, a1, 16
-; RV32I-NEXT:    or t0, a5, a3
-; RV32I-NEXT:    or a4, t2, a4
-; RV32I-NEXT:    or a5, a1, a7
+; RV32I-NEXT:    or t0, a4, a3
+; RV32I-NEXT:    or a4, a7, a5
+; RV32I-NEXT:    or a5, a1, t2
 ; RV32I-NEXT:    slli a5, a5, 5
 ; RV32I-NEXT:    neg t3, a5
 ; RV32I-NEXT:    srl t4, t0, t3
@@ -5781,26 +5781,26 @@ define void @shl_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    lbu a3, 15(a0)
 ; RV32I-NEXT:    slli a7, a7, 8
 ; RV32I-NEXT:    slli t1, t1, 8
+; RV32I-NEXT:    slli t3, t3, 8
 ; RV32I-NEXT:    or a5, a7, a5
 ; RV32I-NEXT:    or a7, t1, t0
-; RV32I-NEXT:    lbu t0, 0(a1)
+; RV32I-NEXT:    or t0, t3, t2
 ; RV32I-NEXT:    lbu t1, 1(a1)
-; RV32I-NEXT:    slli t3, t3, 8
-; RV32I-NEXT:    or t2, t3, t2
+; RV32I-NEXT:    lbu t2, 0(a1)
 ; RV32I-NEXT:    lbu t3, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
 ; RV32I-NEXT:    slli t1, t1, 8
-; RV32I-NEXT:    or t0, t1, t0
+; RV32I-NEXT:    or t1, t1, t2
 ; RV32I-NEXT:    li s9, 64
 ; RV32I-NEXT:    slli a1, a1, 8
 ; RV32I-NEXT:    or a1, a1, t3
 ; RV32I-NEXT:    li t4, 32
 ; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli t2, t2, 16
+; RV32I-NEXT:    slli t0, t0, 16
 ; RV32I-NEXT:    slli a1, a1, 16
 ; RV32I-NEXT:    or t3, a5, a4
-; RV32I-NEXT:    or a5, t2, a7
-; RV32I-NEXT:    or a4, a1, t0
+; RV32I-NEXT:    or a5, t0, a7
+; RV32I-NEXT:    or a4, a1, t1
 ; RV32I-NEXT:    slli a4, a4, 3
 ; RV32I-NEXT:    neg s10, a4
 ; RV32I-NEXT:    srl t5, t3, s10
@@ -6695,26 +6695,26 @@ define void @shl_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) nounw
 ; RV32I-NEXT:    lbu a3, 15(a0)
 ; RV32I-NEXT:    slli a7, a7, 8
 ; RV32I-NEXT:    slli t1, t1, 8
+; RV32I-NEXT:    slli t3, t3, 8
 ; RV32I-NEXT:    or a5, a7, a5
 ; RV32I-NEXT:    or a7, t1, t0
-; RV32I-NEXT:    lbu t0, 0(a1)
+; RV32I-NEXT:    or t0, t3, t2
 ; RV32I-NEXT:    lbu t1, 1(a1)
-; RV32I-NEXT:    slli t3, t3, 8
-; RV32I-NEXT:    or t2, t3, t2
+; RV32I-NEXT:    lbu t2, 0(a1)
 ; RV32I-NEXT:    lbu t3, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
 ; RV32I-NEXT:    slli t1, t1, 8
-; RV32I-NEXT:    or t0, t1, t0
+; RV32I-NEXT:    or t1, t1, t2
 ; RV32I-NEXT:    li s9, 64
 ; RV32I-NEXT:    slli a1, a1, 8
 ; RV32I-NEXT:    or a1, a1, t3
 ; RV32I-NEXT:    li t4, 32
 ; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli t2, t2, 16
+; RV32I-NEXT:    slli t0, t0, 16
 ; RV32I-NEXT:    slli a1, a1, 16
 ; RV32I-NEXT:    or t3, a5, a4
-; RV32I-NEXT:    or a5, t2, a7
-; RV32I-NEXT:    or a4, a1, t0
+; RV32I-NEXT:    or a5, t0, a7
+; RV32I-NEXT:    or a4, a1, t1
 ; RV32I-NEXT:    slli a4, a4, 5
 ; RV32I-NEXT:    neg s10, a4
 ; RV32I-NEXT:    srl t5, t3, s10
@@ -7609,26 +7609,26 @@ define void @shl_32bytes_dwordOff(ptr %src.ptr, ptr %dwordOff.ptr, ptr %dst) nou
 ; RV32I-NEXT:    lbu a3, 15(a0)
 ; RV32I-NEXT:    slli a7, a7, 8
 ; RV32I-NEXT:    slli t1, t1, 8
+; RV32I-NEXT:    slli t3, t3, 8
 ; RV32I-NEXT:    or a5, a7, a5
 ; RV32I-NEXT:    or a7, t1, t0
-; RV32I-NEXT:    lbu t0, 0(a1)
+; RV32I-NEXT:    or t0, t3, t2
 ; RV32I-NEXT:    lbu t1, 1(a1)
-; RV32I-NEXT:    slli t3, t3, 8
-; RV32I-NEXT:    or t2, t3, t2
+; RV32I-NEXT:    lbu t2, 0(a1)
 ; RV32I-NEXT:    lbu t3, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
 ; RV32I-NEXT:    slli t1, t1, 8
-; RV32I-NEXT:    or t0, t1, t0
+; RV32I-NEXT:    or t1, t1, t2
 ; RV32I-NEXT:    li s9, 64
 ; RV32I-NEXT:    slli a1, a1, 8
 ; RV32I-NEXT:    or a1, a1, t3
 ; RV32I-NEXT:    li t4, 32
 ; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli t2, t2, 16
+; RV32I-NEXT:    slli t0, t0, 16
 ; RV32I-NEXT:    slli a1, a1, 16
 ; RV32I-NEXT:    or t3, a5, a4
-; RV32I-NEXT:    or a5, t2, a7
-; RV32I-NEXT:    or a4, a1, t0
+; RV32I-NEXT:    or a5, t0, a7
+; RV32I-NEXT:    or a4, a1, t1
 ; RV32I-NEXT:    slli a4, a4, 6
 ; RV32I-NEXT:    neg s10, a4
 ; RV32I-NEXT:    srl t5, t3, s10
diff --git a/llvm/test/CodeGen/RISCV/abds-neg.ll b/llvm/test/CodeGen/RISCV/abds-neg.ll
index c9a48acb8d14a..3fb0f2c53bdf0 100644
--- a/llvm/test/CodeGen/RISCV/abds-neg.ll
+++ b/llvm/test/CodeGen/RISCV/abds-neg.ll
@@ -625,9 +625,9 @@ define i128 @abd_ext_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:    lw a4, 4(a1)
 ; RV32I-NEXT:    lw a6, 8(a1)
 ; RV32I-NEXT:    lw t1, 12(a1)
+; RV32I-NEXT:    lw a1, 0(a2)
 ; RV32I-NEXT:    lw t0, 8(a2)
 ; RV32I-NEXT:    lw t2, 12(a2)
-; RV32I-NEXT:    lw a1, 0(a2)
 ; RV32I-NEXT:    lw a2, 4(a2)
 ; RV32I-NEXT:    sltu t3, t0, a6
 ; RV32I-NEXT:    mv t4, t3
@@ -744,9 +744,9 @@ define i128 @abd_ext_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:    lw a4, 4(a1)
 ; RV32ZBB-NEXT:    lw a6, 8(a1)
 ; RV32ZBB-NEXT:    lw t1, 12(a1)
+; RV32ZBB-NEXT:    lw a1, 0(a2)
 ; RV32ZBB-NEXT:    lw t0, 8(a2)
 ; RV32ZBB-NEXT:    lw t2, 12(a2)
-; RV32ZBB-NEXT:    lw a1, 0(a2)
 ; RV32ZBB-NEXT:    lw a2, 4(a2)
 ; RV32ZBB-NEXT:    sltu t3, t0, a6
 ; RV32ZBB-NEXT:    mv t4, t3
@@ -872,9 +872,9 @@ define i128 @abd_ext_i128_undef(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:    lw a4, 4(a1)
 ; RV32I-NEXT:    lw a6, 8(a1)
 ; RV32I-NEXT:    lw t1, 12(a1)
+; RV32I-NEXT:    lw a1, 0(a2)
 ; RV32I-NEXT:    lw t0, 8(a2)
 ; RV32I-NEXT:    lw t2, 12(a2)
-; RV32I-NEXT:    lw a1, 0(a2)
 ; RV32I-NEXT:    lw a2, 4(a2)
 ; RV32I-NEXT:    sltu t3, t0, a6
 ; RV32I-NEXT:    mv t4, t3
@@ -991,9 +991,9 @@ define i128 @abd_ext_i128_undef(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:    lw a4, 4(a1)
 ; RV32ZBB-NEXT:    lw a6, 8(a1)
 ; RV32ZBB-NEXT:    lw t1, 12(a1)
+; RV32ZBB-NEXT:    lw a1, 0(a2)
 ; RV32ZBB-NEXT:    lw t0, 8(a2)
 ; RV32ZBB-NEXT:    lw t2, 12(a2)
-; RV32ZBB-NEXT:    lw a1, 0(a2)
 ; RV32ZBB-NEXT:    lw a2, 4(a2)
 ; RV32ZBB-NEXT:    sltu t3, t0, a6
 ; RV32ZBB-NEXT:    mv t4, t3
@@ -1385,8 +1385,8 @@ define i128 @abd_minmax_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:    lw a6, 4(a2)
 ; RV32I-NEXT:    lw a7, 8(a2)
 ; RV32I-NEXT:    lw t0, 12(a2)
-; RV32I-NEXT:    lw a5, 12(a1)
 ; RV32I-NEXT:    lw a3, 4(a1)
+; RV32I-NEXT:    lw a5, 12(a1)
 ; RV32I-NEXT:    lw a4, 8(a1)
 ; RV32I-NEXT:    beq a5, t0, .LBB17_2
 ; RV32I-NEXT:  # %bb.1:
@@ -1512,8 +1512,8 @@ define i128 @abd_minmax_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:    lw a6, 4(a2)
 ; RV32ZBB-NEXT:    lw a7, 8(a2)
 ; RV32ZBB-NEXT:    lw t0, 12(a2)
-; RV32ZBB-NEXT:    lw a5, 12(a1)
 ; RV32ZBB-NEXT:    lw a3, 4(a1)
+; RV32ZBB-NEXT:    lw a5, 12(a1)
 ; RV32ZBB-NEXT:    lw a4, 8(a1)
 ; RV32ZBB-NEXT:    beq a5, t0, .LBB17_2
 ; RV32ZBB-NEXT:  # %bb.1:
@@ -1864,15 +1864,15 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:    lw a4, 4(a2)
 ; RV32I-NEXT:    lw a5, 8(a2)
 ; RV32I-NEXT:    lw a7, 12(a2)
-; RV32I-NEXT:    lw a6, 8(a1)
-; RV32I-NEXT:    lw t0, 12(a1)
 ; RV32I-NEXT:    lw a2, 0(a1)
+; RV32I-NEXT:    lw a6, 8(a1)
+; RV32I-NEXT:    lw t1, 12(a1)
 ; RV32I-NEXT:    lw a1, 4(a1)
-; RV32I-NEXT:    sltu t1, a6, a5
-; RV32I-NEXT:    mv t4, t1
-; RV32I-NEXT:    beq t0, a7, .LBB22_2
+; RV32I-NEXT:    sltu t0, a6, a5
+; RV32I-NEXT:    mv t4, t0
+; RV32I-NEXT:    beq t1, a7, .LBB22_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    slt t4, t0, a7
+; RV32I-NEXT:    slt t4, t1, a7
 ; RV32I-NEXT:  .LBB22_2:
 ; RV32I-NEXT:    sltu t2, a2, a3
 ; RV32I-NEXT:    mv t3, t2
@@ -1880,7 +1880,7 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:  # %bb.3:
 ; RV32I-NEXT:    sltu t3, a1, a4
 ; RV32I-NEXT:  .LBB22_4:
-; RV32I-NEXT:    xor t5, t0, a7
+; RV32I-NEXT:    xor t5, t1, a7
 ; RV32I-NEXT:    xor t6, a6, a5
 ; RV32I-NEXT:    or t5, t6, t5
 ; RV32I-NEXT:    mv t6, t3
@@ -1896,11 +1896,11 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:  .LBB22_8:
 ; RV32I-NEXT:    bnez t6, .LBB22_10
 ; RV32I-NEXT:  # %bb.9:
-; RV32I-NEXT:    sltu t1, a5, a6
-; RV32I-NEXT:    sub a7, a7, t0
+; RV32I-NEXT:    sltu t0, a5, a6
+; RV32I-NEXT:    sub a7, a7, t1
 ; RV32I-NEXT:    sub a5, a5, a6
 ; RV32I-NEXT:    sub a4, a4, a1
-; RV32I-NEXT:    sub a6, a7, t1
+; RV32I-NEXT:    sub a6, a7, t0
 ; RV32I-NEXT:    sltu a7, a5, t5
 ; RV32I-NEXT:    sub a1, a5, t5
 ; RV32I-NEXT:    sub a5, a4, t4
@@ -1908,10 +1908,10 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:    sub a2, a3, a2
 ; RV32I-NEXT:    j .LBB22_11
 ; RV32I-NEXT:  .LBB22_10:
-; RV32I-NEXT:    sub a7, t0, a7
+; RV32I-NEXT:    sub a7, t1, a7
 ; RV32I-NEXT:    sub a5, a6, a5
 ; RV32I-NEXT:    sub a4, a1, a4
-; RV32I-NEXT:    sub a6, a7, t1
+; RV32I-NEXT:    sub a6, a7, t0
 ; RV32I-NEXT:    sltu a7, a5, t3
 ; RV32I-NEXT:    sub a1, a5, t3
 ; RV32I-NEXT:    sub a5, a4, t2
@@ -1951,15 +1951,15 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:    lw a4, 4(a2)
 ; RV32ZBB-NEXT:    lw a5, 8(a2)
 ; RV32ZBB-NEXT:    lw a7, 12(a2)
-; RV32ZBB-NEXT:    lw a6, 8(a1)
-; RV32ZBB-NEXT:    lw t0, 12(a1)
 ; RV32ZBB-NEXT:    lw a2, 0(a1)
+; RV32ZBB-NEXT:    lw a6, 8(a1)
+; RV32ZBB-NEXT:    lw t1, 12(a1)
 ; RV32ZBB-NEXT:    lw a1, 4(a1)
-; RV32ZBB-NEXT:    sltu t1, a6, a5
-; RV32ZBB-NEXT:    mv t4, t1
-; RV32ZBB-NEXT:    beq t0, a7, .LBB22_2
+; RV32ZBB-NEXT:    sltu t0, a6, a5
+; RV32ZBB-NEXT:    mv t4, t0
+; RV32ZBB-NEXT:    beq t1, a7, .LBB22_2
 ; RV32ZBB-NEXT:  # %bb.1:
-; RV32ZBB-NEXT:    slt t4, t0, a7
+; RV32ZBB-NEXT:    slt t4, t1, a7
 ; RV32ZBB-NEXT:  .LBB22_2:
 ; RV32ZBB-NEXT:    sltu t2, a2, a3
 ; RV32ZBB-NEXT:    mv t3, t2
@@ -1967,7 +1967,7 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:  # %bb.3:
 ; RV32ZBB-NEXT:    sltu t3, a1, a4
 ; RV32ZBB-NEXT:  .LBB22_4:
-; RV32ZBB-NEXT:    xor t5, t0, a7
+; RV32ZBB-NEXT:    xor t5, t1, a7
 ; RV32ZBB-NEXT:    xor t6, a6, a5
 ; RV32ZBB-NEXT:    or t5, t6, t5
 ; RV32ZBB-NEXT:    mv t6, t3
@@ -1983,11 +1983,11 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:  .LBB22_8:
 ; RV32ZBB-NEXT:    bnez t6, .LBB22_10
 ; RV32ZBB-NEXT:  # %bb.9:
-; RV32ZBB-NEXT:    sltu t1, a5, a6
-; RV32ZBB-NEXT:    sub a7, a7, t0
+; RV32ZBB-NEXT:    sltu t0, a5, a6
+; RV32ZBB-NEXT:    sub a7, a7, t1
 ; RV32ZBB-NEXT:    sub a5, a5, a6
 ; RV32ZBB-NEXT:    sub a4, a4, a1
-; RV32ZBB-NEXT:    sub a6, a7, t1
+; RV32ZBB-NEXT:    sub a6, a7, t0
 ; RV32ZBB-NEXT:    sltu a7, a5, t5
 ; RV32ZBB-NEXT:    sub a1, a5, t5
 ; RV32ZBB-NEXT:    sub a5, a4, t4
@@ -1995,10 +1995,10 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:    sub a2, a3, a2
 ; RV32ZBB-NEXT:    j .LBB22_11
 ; RV32ZBB-NEXT:  .LBB22_10:
-; RV32ZBB-NEXT:    sub a7, t0, a7
+; RV32ZBB-NEXT:    sub a7, t1, a7
 ; RV32ZBB-NEXT:    sub a5, a6, a5
 ; RV32ZBB-NEXT:    sub a4, a1, a4
-; RV32ZBB-NEXT:    sub a6, a7, t1
+; RV32ZBB-NEXT:    sub a6, a7, t0
 ; RV32ZBB-NEXT:    sltu a7, a5, t3
 ; RV32ZBB-NEXT:    sub a1, a5, t3
 ; RV32ZBB-NEXT:    sub a5, a4, t2
diff --git a/llvm/test/CodeGen/RISCV/abds.ll b/llvm/test/CodeGen/RISCV/abds.ll
index 56e6dacff9748..efb4e1a6f15d6 100644
--- a/llvm/test/CodeGen/RISCV/abds.ll
+++ b/llvm/test/CodeGen/RISCV/abds.ll
@@ -536,73 +536,73 @@ define i128 @abd_ext_i128(i128 %a, i128 %b) nounwind {
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lw a3, 0(a1)
 ; RV32I-NEXT:    lw a4, 4(a1)
-; RV32I-NEXT:    lw a6, 8(a1)
-; RV32I-NEXT:    lw t0, 12(a1)
-; RV32I-NEXT:    lw a7, 8(a2)
+; RV32I-NEXT:    lw a5, 8(a1)
+; RV32I-NEXT:    lw a7, 12(a1)
+; RV32I-NEXT:    lw a1, 0(a2)
+; RV32I-NEXT:    lw a6, 8(a2)
 ; RV32I-NEXT:    lw t1, 12(a2)
-; RV32I-NEXT:    lw a5, 0(a2)
-; RV32I-NEXT:    lw a1, 4(a2)
-; RV32I-NEXT:    sltu a2, a7, a6
-; RV32I-NEXT:    mv t4, a2
-; RV32I-NEXT:    beq t0, t1, .LBB11_2
+; RV32I-NEXT:    lw a2, 4(a2)
+; RV32I-NEXT:    sltu t0, a6, a5
+; RV32I-NEXT:    mv t4, t0
+; RV32I-NEXT:    beq a7, t1, .LBB11_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    slt t4, t1, t0
+; RV32I-NEXT:    slt t4, t1, a7
 ; RV32I-NEXT:  .LBB11_2:
-; RV32I-NEXT:    sltu t2, a5, a3
-; RV32I-NEXT:    sltu t5, a1, a4
+; RV32I-NEXT:    sltu t2, a1, a3
+; RV32I-NEXT:    sltu t5, a2, a4
 ; RV32I-NEXT:    mv t3, t2
-; RV32I-NEXT:    beq a4, a1, .LBB11_4
+; RV32I-NEXT:    beq a4, a2, .LBB11_4
 ; RV32I-NEXT:  # %bb.3:
 ; RV32I-NEXT:    mv t3, t5
 ; RV32I-NEXT:  .LBB11_4:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    xor t6, t0, t1
-; RV32I-NEXT:    xor s0, a6, a7
+; RV32I-NEXT:    xor t6, a7, t1
+; RV32I-NEXT:    xor s0, a5, a6
 ; RV32I-NEXT:    or t6, s0, t6
 ; RV32I-NEXT:    beqz t6, .LBB11_6
 ; RV32I-NEXT:  # %bb.5:
 ; RV32I-NEXT:    mv t3, t4
 ; RV32I-NEXT:  .LBB11_6:
 ; RV32I-NEXT:    mv t4, t2
-; RV32I-NEXT:    beq a1, a4, .LBB11_8
+; RV32I-NEXT:    beq a2, a4, .LBB11_8
 ; RV32I-NEXT:  # %bb.7:
 ; RV32I-NEXT:    mv t4, t5
 ; RV32I-NEXT:  .LBB11_8:
-; RV32I-NEXT:    sltu t5, a3, a5
+; RV32I-NEXT:    sltu t5, a3, a1
 ; RV32I-NEXT:    mv t6, t5
-; RV32I-NEXT:    beq a4, a1, .LBB11_10
+; RV32I-NEXT:    beq a4, a2, .LBB11_10
 ; RV32I-NEXT:  # %bb.9:
-; RV32I-NEXT:    sltu t6, a4, a1
+; RV32I-NEXT:    sltu t6, a4, a2
 ; RV32I-NEXT:  .LBB11_10:
 ; RV32I-NEXT:    bnez t3, .LBB11_12
 ; RV32I-NEXT:  # %bb.11:
-; RV32I-NEXT:    sub t0, t1, t0
-; RV32I-NEXT:    sub a6, a7, a6
-; RV32I-NEXT:    sub a3, a5, a3
-; RV32I-NEXT:    sub a1, a1, a4
-; RV32I-NEXT:    sub a4, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t4
-; RV32I-NEXT:    sub a2, a1, t2
-; RV32I-NEXT:    sub a1, a4, a5
-; RV32I-NEXT:    sub a4, a6, t4
+; RV32I-NEXT:    sub a7, t1, a7
+; RV32I-NEXT:    sub a5, a6, a5
+; RV32I-NEXT:    sub a1, a1, a3
+; RV32I-NEXT:    sub a2, a2, a4
+; RV32I-NEXT:    sub a4, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t4
+; RV32I-NEXT:    sub a3, a2, t2
+; RV32I-NEXT:    sub a2, a4, a6
+; RV32I-NEXT:    sub a4, a5, t4
 ; RV32I-NEXT:    j .LBB11_13
 ; RV32I-NEXT:  .LBB11_12:
-; RV32I-NEXT:    sltu a2, a6, a7
-; RV32I-NEXT:    sub t0, t0, t1
-; RV32I-NEXT:    sub a6, a6, a7
-; RV32I-NEXT:    sub a3, a3, a5
-; RV32I-NEXT:    sub a4, a4, a1
-; RV32I-NEXT:    sub a1, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t6
-; RV32I-NEXT:    sub a2, a4, t5
-; RV32I-NEXT:    sub a1, a1, a5
-; RV32I-NEXT:    sub a4, a6, t6
+; RV32I-NEXT:    sltu t0, a5, a6
+; RV32I-NEXT:    sub a7, a7, t1
+; RV32I-NEXT:    sub a5, a5, a6
+; RV32I-NEXT:    sub a1, a3, a1
+; RV32I-NEXT:    sub a4, a4, a2
+; RV32I-NEXT:    sub a2, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t6
+; RV32I-NEXT:    sub a3, a4, t5
+; RV32I-NEXT:    sub a2, a2, a6
+; RV32I-NEXT:    sub a4, a5, t6
 ; RV32I-NEXT:  .LBB11_13:
-; RV32I-NEXT:    sw a3, 0(a0)
-; RV32I-NEXT:    sw a2, 4(a0)
+; RV32I-NEXT:    sw a1, 0(a0)
+; RV32I-NEXT:    sw a3, 4(a0)
 ; RV32I-NEXT:    sw a4, 8(a0)
-; RV32I-NEXT:    sw a1, 12(a0)
+; RV32I-NEXT:    sw a2, 12(a0)
 ; RV32I-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32I-NEXT:    addi sp, sp, 16
 ; RV32I-NEXT:    ret
@@ -632,73 +632,73 @@ define i128 @abd_ext_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB:       # %bb.0:
 ; RV32ZBB-NEXT:    lw a3, 0(a1)
 ; RV32ZBB-NEXT:    lw a4, 4(a1)
-; RV32ZBB-NEXT:    lw a6, 8(a1)
-; RV32ZBB-NEXT:    lw t0, 12(a1)
-; RV32ZBB-NEXT:    lw a7, 8(a2)
+; RV32ZBB-NEXT:    lw a5, 8(a1)
+; RV32ZBB-NEXT:    lw a7, 12(a1)
+; RV32ZBB-NEXT:    lw a1, 0(a2)
+; RV32ZBB-NEXT:    lw a6, 8(a2)
 ; RV32ZBB-NEXT:    lw t1, 12(a2)
-; RV32ZBB-NEXT:    lw a5, 0(a2)
-; RV32ZBB-NEXT:    lw a1, 4(a2)
-; RV32ZBB-NEXT:    sltu a2, a7, a6
-; RV32ZBB-NEXT:    mv t4, a2
-; RV32ZBB-NEXT:    beq t0, t1, .LBB11_2
+; RV32ZBB-NEXT:    lw a2, 4(a2)
+; RV32ZBB-NEXT:    sltu t0, a6, a5
+; RV32ZBB-NEXT:    mv t4, t0
+; RV32ZBB-NEXT:    beq a7, t1, .LBB11_2
 ; RV32ZBB-NEXT:  # %bb.1:
-; RV32ZBB-NEXT:    slt t4, t1, t0
+; RV32ZBB-NEXT:    slt t4, t1, a7
 ; RV32ZBB-NEXT:  .LBB11_2:
-; RV32ZBB-NEXT:    sltu t2, a5, a3
-; RV32ZBB-NEXT:    sltu t5, a1, a4
+; RV32ZBB-NEXT:    sltu t2, a1, a3
+; RV32ZBB-NEXT:    sltu t5, a2, a4
 ; RV32ZBB-NEXT:    mv t3, t2
-; RV32ZBB-NEXT:    beq a4, a1, .LBB11_4
+; RV32ZBB-NEXT:    beq a4, a2, .LBB11_4
 ; RV32ZBB-NEXT:  # %bb.3:
 ; RV32ZBB-NEXT:    mv t3, t5
 ; RV32ZBB-NEXT:  .LBB11_4:
 ; RV32ZBB-NEXT:    addi sp, sp, -16
 ; RV32ZBB-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32ZBB-NEXT:    xor t6, t0, t1
-; RV32ZBB-NEXT:    xor s0, a6, a7
+; RV32ZBB-NEXT:    xor t6, a7, t1
+; RV32ZBB-NEXT:    xor s0, a5, a6
 ; RV32ZBB-NEXT:    or t6, s0, t6
 ; RV32ZBB-NEXT:    beqz t6, .LBB11_6
 ; RV32ZBB-NEXT:  # %bb.5:
 ; RV32ZBB-NEXT:    mv t3, t4
 ; RV32ZBB-NEXT:  .LBB11_6:
 ; RV32ZBB-NEXT:    mv t4, t2
-; RV32ZBB-NEXT:    beq a1, a4, .LBB11_8
+; RV32ZBB-NEXT:    beq a2, a4, .LBB11_8
 ; RV32ZBB-NEXT:  # %bb.7:
 ; RV32ZBB-NEXT:    mv t4, t5
 ; RV32ZBB-NEXT:  .LBB11_8:
-; RV32ZBB-NEXT:    sltu t5, a3, a5
+; RV32ZBB-NEXT:    sltu t5, a3, a1
 ; RV32ZBB-NEXT:    mv t6, t5
-; RV32ZBB-NEXT:    beq a4, a1, .LBB11_10
+; RV32ZBB-NEXT:    beq a4, a2, .LBB11_10
 ; RV32ZBB-NEXT:  # %bb.9:
-; RV32ZBB-NEXT:    sltu t6, a4, a1
+; RV32ZBB-NEXT:    sltu t6, a4, a2
 ; RV32ZBB-NEXT:  .LBB11_10:
 ; RV32ZBB-NEXT:    bnez t3, .LBB11_12
 ; RV32ZBB-NEXT:  # %bb.11:
-; RV32ZBB-NEXT:    sub t0, t1, t0
-; RV32ZBB-NEXT:    sub a6, a7, a6
-; RV32ZBB-NEXT:    sub a3, a5, a3
-; RV32ZBB-NEXT:    sub a1, a1, a4
-; RV32ZBB-NEXT:    sub a4, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t4
-; RV32ZBB-NEXT:    sub a2, a1, t2
-; RV32ZBB-NEXT:    sub a1, a4, a5
-; RV32ZBB-NEXT:    sub a4, a6, t4
+; RV32ZBB-NEXT:    sub a7, t1, a7
+; RV32ZBB-NEXT:    sub a5, a6, a5
+; RV32ZBB-NEXT:    sub a1, a1, a3
+; RV32ZBB-NEXT:    sub a2, a2, a4
+; RV32ZBB-NEXT:    sub a4, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t4
+; RV32ZBB-NEXT:    sub a3, a2, t2
+; RV32ZBB-NEXT:    sub a2, a4, a6
+; RV32ZBB-NEXT:    sub a4, a5, t4
 ; RV32ZBB-NEXT:    j .LBB11_13
 ; RV32ZBB-NEXT:  .LBB11_12:
-; RV32ZBB-NEXT:    sltu a2, a6, a7
-; RV32ZBB-NEXT:    sub t0, t0, t1
-; RV32ZBB-NEXT:    sub a6, a6, a7
-; RV32ZBB-NEXT:    sub a3, a3, a5
-; RV32ZBB-NEXT:    sub a4, a4, a1
-; RV32ZBB-NEXT:    sub a1, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t6
-; RV32ZBB-NEXT:    sub a2, a4, t5
-; RV32ZBB-NEXT:    sub a1, a1, a5
-; RV32ZBB-NEXT:    sub a4, a6, t6
+; RV32ZBB-NEXT:    sltu t0, a5, a6
+; RV32ZBB-NEXT:    sub a7, a7, t1
+; RV32ZBB-NEXT:    sub a5, a5, a6
+; RV32ZBB-NEXT:    sub a1, a3, a1
+; RV32ZBB-NEXT:    sub a4, a4, a2
+; RV32ZBB-NEXT:    sub a2, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t6
+; RV32ZBB-NEXT:    sub a3, a4, t5
+; RV32ZBB-NEXT:    sub a2, a2, a6
+; RV32ZBB-NEXT:    sub a4, a5, t6
 ; RV32ZBB-NEXT:  .LBB11_13:
-; RV32ZBB-NEXT:    sw a3, 0(a0)
-; RV32ZBB-NEXT:    sw a2, 4(a0)
+; RV32ZBB-NEXT:    sw a1, 0(a0)
+; RV32ZBB-NEXT:    sw a3, 4(a0)
 ; RV32ZBB-NEXT:    sw a4, 8(a0)
-; RV32ZBB-NEXT:    sw a1, 12(a0)
+; RV32ZBB-NEXT:    sw a2, 12(a0)
 ; RV32ZBB-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32ZBB-NEXT:    addi sp, sp, 16
 ; RV32ZBB-NEXT:    ret
@@ -736,73 +736,73 @@ define i128 @abd_ext_i128_undef(i128 %a, i128 %b) nounwind {
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lw a3, 0(a1)
 ; RV32I-NEXT:    lw a4, 4(a1)
-; RV32I-NEXT:    lw a6, 8(a1)
-; RV32I-NEXT:    lw t0, 12(a1)
-; RV32I-NEXT:    lw a7, 8(a2)
+; RV32I-NEXT:    lw a5, 8(a1)
+; RV32I-NEXT:    lw a7, 12(a1)
+; RV32I-NEXT:    lw a1, 0(a2)
+; RV32I-NEXT:    lw a6, 8(a2)
 ; RV32I-NEXT:    lw t1, 12(a2)
-; RV32I-NEXT:    lw a5, 0(a2)
-; RV32I-NEXT:    lw a1, 4(a2)
-; RV32I-NEXT:    sltu a2, a7, a6
-; RV32I-NEXT:    mv t4, a2
-; RV32I-NEXT:    beq t0, t1, .LBB12_2
+; RV32I-NEXT:    lw a2, 4(a2)
+; RV32I-NEXT:    sltu t0, a6, a5
+; RV32I-NEXT:    mv t4, t0
+; RV32I-NEXT:    beq a7, t1, .LBB12_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    slt t4, t1, t0
+; RV32I-NEXT:    slt t4, t1, a7
 ; RV32I-NEXT:  .LBB12_2:
-; RV32I-NEXT:    sltu t2, a5, a3
-; RV32I-NEXT:    sltu t5, a1, a4
+; RV32I-NEXT:    sltu t2, a1, a3
+; RV32I-NEXT:    sltu t5, a2, a4
 ; RV32I-NEXT:    mv t3, t2
-; RV32I-NEXT:    beq a4, a1, .LBB12_4
+; RV32I-NEXT:    beq a4, a2, .LBB12_4
 ; RV32I-NEXT:  # %bb.3:
 ; RV32I-NEXT:    mv t3, t5
 ; RV32I-NEXT:  .LBB12_4:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    xor t6, t0, t1
-; RV32I-NEXT:    xor s0, a6, a7
+; RV32I-NEXT:    xor t6, a7, t1
+; RV32I-NEXT:    xor s0, a5, a6
 ; RV32I-NEXT:    or t6, s0, t6
 ; RV32I-NEXT:    beqz t6, .LBB12_6
 ; RV32I-NEXT:  # %bb.5:
 ; RV32I-NEXT:    mv t3, t4
 ; RV32I-NEXT:  .LBB12_6:
 ; RV32I-NEXT:    mv t4, t2
-; RV32I-NEXT:    beq a1, a4, .LBB12_8
+; RV32I-NEXT:    beq a2, a4, .LBB12_8
 ; RV32I-NEXT:  # %bb.7:
 ; RV32I-NEXT:    mv t4, t5
 ; RV32I-NEXT:  .LBB12_8:
-; RV32I-NEXT:    sltu t5, a3, a5
+; RV32I-NEXT:    sltu t5, a3, a1
 ; RV32I-NEXT:    mv t6, t5
-; RV32I-NEXT:    beq a4, a1, .LBB12_10
+; RV32I-NEXT:    beq a4, a2, .LBB12_10
 ; RV32I-NEXT:  # %bb.9:
-; RV32I-NEXT:    sltu t6, a4, a1
+; RV32I-NEXT:    sltu t6, a4, a2
 ; RV32I-NEXT:  .LBB12_10:
 ; RV32I-NEXT:    bnez t3, .LBB12_12
 ; RV32I-NEXT:  # %bb.11:
-; RV32I-NEXT:    sub t0, t1, t0
-; RV32I-NEXT:    sub a6, a7, a6
-; RV32I-NEXT:    sub a3, a5, a3
-; RV32I-NEXT:    sub a1, a1, a4
-; RV32I-NEXT:    sub a4, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t4
-; RV32I-NEXT:    sub a2, a1, t2
-; RV32I-NEXT:    sub a1, a4, a5
-; RV32I-NEXT:    sub a4, a6, t4
+; RV32I-NEXT:    sub a7, t1, a7
+; RV32I-NEXT:    sub a5, a6, a5
+; RV32I-NEXT:    sub a1, a1, a3
+; RV32I-NEXT:    sub a2, a2, a4
+; RV32I-NEXT:    sub a4, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t4
+; RV32I-NEXT:    sub a3, a2, t2
+; RV32I-NEXT:    sub a2, a4, a6
+; RV32I-NEXT:    sub a4, a5, t4
 ; RV32I-NEXT:    j .LBB12_13
 ; RV32I-NEXT:  .LBB12_12:
-; RV32I-NEXT:    sltu a2, a6, a7
-; RV32I-NEXT:    sub t0, t0, t1
-; RV32I-NEXT:    sub a6, a6, a7
-; RV32I-NEXT:    sub a3, a3, a5
-; RV32I-NEXT:    sub a4, a4, a1
-; RV32I-NEXT:    sub a1, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t6
-; RV32I-NEXT:    sub a2, a4, t5
-; RV32I-NEXT:    sub a1, a1, a5
-; RV32I-NEXT:    sub a4, a6, t6
+; RV32I-NEXT:    sltu t0, a5, a6
+; RV32I-NEXT:    sub a7, a7, t1
+; RV32I-NEXT:    sub a5, a5, a6
+; RV32I-NEXT:    sub a1, a3, a1
+; RV32I-NEXT:    sub a4, a4, a2
+; RV32I-NEXT:    sub a2, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t6
+; RV32I-NEXT:    sub a3, a4, t5
+; RV32I-NEXT:    sub a2, a2, a6
+; RV32I-NEXT:    sub a4, a5, t6
 ; RV32I-NEXT:  .LBB12_13:
-; RV32I-NEXT:    sw a3, 0(a0)
-; RV32I-NEXT:    sw a2, 4(a0)
+; RV32I-NEXT:    sw a1, 0(a0)
+; RV32I-NEXT:    sw a3, 4(a0)
 ; RV32I-NEXT:    sw a4, 8(a0)
-; RV32I-NEXT:    sw a1, 12(a0)
+; RV32I-NEXT:    sw a2, 12(a0)
 ; RV32I-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32I-NEXT:    addi sp, sp, 16
 ; RV32I-NEXT:    ret
@@ -832,73 +832,73 @@ define i128 @abd_ext_i128_undef(i128 %a, i128 %b) nounwind {
 ; RV32ZBB:       # %bb.0:
 ; RV32ZBB-NEXT:    lw a3, 0(a1)
 ; RV32ZBB-NEXT:    lw a4, 4(a1)
-; RV32ZBB-NEXT:    lw a6, 8(a1)
-; RV32ZBB-NEXT:    lw t0, 12(a1)
-; RV32ZBB-NEXT:    lw a7, 8(a2)
+; RV32ZBB-NEXT:    lw a5, 8(a1)
+; RV32ZBB-NEXT:    lw a7, 12(a1)
+; RV32ZBB-NEXT:    lw a1, 0(a2)
+; RV32ZBB-NEXT:    lw a6, 8(a2)
 ; RV32ZBB-NEXT:    lw t1, 12(a2)
-; RV32ZBB-NEXT:    lw a5, 0(a2)
-; RV32ZBB-NEXT:    lw a1, 4(a2)
-; RV32ZBB-NEXT:    sltu a2, a7, a6
-; RV32ZBB-NEXT:    mv t4, a2
-; RV32ZBB-NEXT:    beq t0, t1, .LBB12_2
+; RV32ZBB-NEXT:    lw a2, 4(a2)
+; RV32ZBB-NEXT:    sltu t0, a6, a5
+; RV32ZBB-NEXT:    mv t4, t0
+; RV32ZBB-NEXT:    beq a7, t1, .LBB12_2
 ; RV32ZBB-NEXT:  # %bb.1:
-; RV32ZBB-NEXT:    slt t4, t1, t0
+; RV32ZBB-NEXT:    slt t4, t1, a7
 ; RV32ZBB-NEXT:  .LBB12_2:
-; RV32ZBB-NEXT:    sltu t2, a5, a3
-; RV32ZBB-NEXT:    sltu t5, a1, a4
+; RV32ZBB-NEXT:    sltu t2, a1, a3
+; RV32ZBB-NEXT:    sltu t5, a2, a4
 ; RV32ZBB-NEXT:    mv t3, t2
-; RV32ZBB-NEXT:    beq a4, a1, .LBB12_4
+; RV32ZBB-NEXT:    beq a4, a2, .LBB12_4
 ; RV32ZBB-NEXT:  # %bb.3:
 ; RV32ZBB-NEXT:    mv t3, t5
 ; RV32ZBB-NEXT:  .LBB12_4:
 ; RV32ZBB-NEXT:    addi sp, sp, -16
 ; RV32ZBB-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32ZBB-NEXT:    xor t6, t0, t1
-; RV32ZBB-NEXT:    xor s0, a6, a7
+; RV32ZBB-NEXT:    xor t6, a7, t1
+; RV32ZBB-NEXT:    xor s0, a5, a6
 ; RV32ZBB-NEXT:    or t6, s0, t6
 ; RV32ZBB-NEXT:    beqz t6, .LBB12_6
 ; RV32ZBB-NEXT:  # %bb.5:
 ; RV32ZBB-NEXT:    mv t3, t4
 ; RV32ZBB-NEXT:  .LBB12_6:
 ; RV32ZBB-NEXT:    mv t4, t2
-; RV32ZBB-NEXT:    beq a1, a4, .LBB12_8
+; RV32ZBB-NEXT:    beq a2, a4, .LBB12_8
 ; RV32ZBB-NEXT:  # %bb.7:
 ; RV32ZBB-NEXT:    mv t4, t5
 ; RV32ZBB-NEXT:  .LBB12_8:
-; RV32ZBB-NEXT:    sltu t5, a3, a5
+; RV32ZBB-NEXT:    sltu t5, a3, a1
 ; RV32ZBB-NEXT:    mv t6, t5
-; RV32ZBB-NEXT:    beq a4, a1, .LBB12_10
+; RV32ZBB-NEXT:    beq a4, a2, .LBB12_10
 ; RV32ZBB-NEXT:  # %bb.9:
-; RV32ZBB-NEXT:    sltu t6, a4, a1
+; RV32ZBB-NEXT:    sltu t6, a4, a2
 ; RV32ZBB-NEXT:  .LBB12_10:
 ; RV32ZBB-NEXT:    bnez t3, .LBB12_12
 ; RV32ZBB-NEXT:  # %bb.11:
-; RV32ZBB-NEXT:    sub t0, t1, t0
-; RV32ZBB-NEXT:    sub a6, a7, a6
-; RV32ZBB-NEXT:    sub a3, a5, a3
-; RV32ZBB-NEXT:    sub a1, a1, a4
-; RV32ZBB-NEXT:    sub a4, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t4
-; RV32ZBB-NEXT:    sub a2, a1, t2
-; RV32ZBB-NEXT:    sub a1, a4, a5
-; RV32ZBB-NEXT:    sub a4, a6, t4
+; RV32ZBB-NEXT:    sub a7, t1, a7
+; RV32ZBB-NEXT:    sub a5, a6, a5
+; RV32ZBB-NEXT:    sub a1, a1, a3
+; RV32ZBB-NEXT:    sub a2, a2, a4
+; RV32ZBB-NEXT:    sub a4, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t4
+; RV32ZBB-NEXT:    sub a3, a2, t2
+; RV32ZBB-NEXT:    sub a2, a4, a6
+; RV32ZBB-NEXT:    sub a4, a5, t4
 ; RV32ZBB-NEXT:    j .LBB12_13
 ; RV32ZBB-NEXT:  .LBB12_12:
-; RV32ZBB-NEXT:    sltu a2, a6, a7
-; RV32ZBB-NEXT:    sub t0, t0, t1
-; RV32ZBB-NEXT:    sub a6, a6, a7
-; RV32ZBB-NEXT:    sub a3, a3, a5
-; RV32ZBB-NEXT:    sub a4, a4, a1
-; RV32ZBB-NEXT:    sub a1, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t6
-; RV32ZBB-NEXT:    sub a2, a4, t5
-; RV32ZBB-NEXT:    sub a1, a1, a5
-; RV32ZBB-NEXT:    sub a4, a6, t6
+; RV32ZBB-NEXT:    sltu t0, a5, a6
+; RV32ZBB-NEXT:    sub a7, a7, t1
+; RV32ZBB-NEXT:    sub a5, a5, a6
+; RV32ZBB-NEXT:    sub a1, a3, a1
+; RV32ZBB-NEXT:    sub a4, a4, a2
+; RV32ZBB-NEXT:    sub a2, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t6
+; RV32ZBB-NEXT:    sub a3, a4, t5
+; RV32ZBB-NEXT:    sub a2, a2, a6
+; RV32ZBB-NEXT:    sub a4, a5, t6
 ; RV32ZBB-NEXT:  .LBB12_13:
-; RV32ZBB-NEXT:    sw a3, 0(a0)
-; RV32ZBB-NEXT:    sw a2, 4(a0)
+; RV32ZBB-NEXT:    sw a1, 0(a0)
+; RV32ZBB-NEXT:    sw a3, 4(a0)
 ; RV32ZBB-NEXT:    sw a4, 8(a0)
-; RV32ZBB-NEXT:    sw a1, 12(a0)
+; RV32ZBB-NEXT:    sw a2, 12(a0)
 ; RV32ZBB-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32ZBB-NEXT:    addi sp, sp, 16
 ; RV32ZBB-NEXT:    ret
@@ -1125,73 +1125,73 @@ define i128 @abd_minmax_i128(i128 %a, i128 %b) nounwind {
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lw a3, 0(a1)
 ; RV32I-NEXT:    lw a4, 4(a1)
-; RV32I-NEXT:    lw a6, 8(a1)
-; RV32I-NEXT:    lw t0, 12(a1)
-; RV32I-NEXT:    lw a7, 8(a2)
+; RV32I-NEXT:    lw a5, 8(a1)
+; RV32I-NEXT:    lw a7, 12(a1)
+; RV32I-NEXT:    lw a1, 0(a2)
+; RV32I-NEXT:    lw a6, 8(a2)
 ; RV32I-NEXT:    lw t1, 12(a2)
-; RV32I-NEXT:    lw a5, 0(a2)
-; RV32I-NEXT:    lw a1, 4(a2)
-; RV32I-NEXT:    sltu a2, a7, a6
-; RV32I-NEXT:    mv t4, a2
-; RV32I-NEXT:    beq t0, t1, .LBB17_2
+; RV32I-NEXT:    lw a2, 4(a2)
+; RV32I-NEXT:    sltu t0, a6, a5
+; RV32I-NEXT:    mv t4, t0
+; RV32I-NEXT:    beq a7, t1, .LBB17_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    slt t4, t1, t0
+; RV32I-NEXT:    slt t4, t1, a7
 ; RV32I-NEXT:  .LBB17_2:
-; RV32I-NEXT:    sltu t2, a5, a3
-; RV32I-NEXT:    sltu t5, a1, a4
+; RV32I-NEXT:    sltu t2, a1, a3
+; RV32I-NEXT:    sltu t5, a2, a4
 ; RV32I-NEXT:    mv t3, t2
-; RV32I-NEXT:    beq a4, a1, .LBB17_4
+; RV32I-NEXT:    beq a4, a2, .LBB17_4
 ; RV32I-NEXT:  # %bb.3:
 ; RV32I-NEXT:    mv t3, t5
 ; RV32I-NEXT:  .LBB17_4:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    xor t6, t0, t1
-; RV32I-NEXT:    xor s0, a6, a7
+; RV32I-NEXT:    xor t6, a7, t1
+; RV32I-NEXT:    xor s0, a5, a6
 ; RV32I-NEXT:    or t6, s0, t6
 ; RV32I-NEXT:    beqz t6, .LBB17_6
 ; RV32I-NEXT:  # %bb.5:
 ; RV32I-NEXT:    mv t3, t4
 ; RV32I-NEXT:  .LBB17_6:
 ; RV32I-NEXT:    mv t4, t2
-; RV32I-NEXT:    beq a1, a4, .LBB17_8
+; RV32I-NEXT:    beq a2, a4, .LBB17_8
 ; RV32I-NEXT:  # %bb.7:
 ; RV32I-NEXT:    mv t4, t5
 ; RV32I-NEXT:  .LBB17_8:
-; RV32I-NEXT:    sltu t5, a3, a5
+; RV32I-NEXT:    sltu t5, a3, a1
 ; RV32I-NEXT:    mv t6, t5
-; RV32I-NEXT:    beq a4, a1, .LBB17_10
+; RV32I-NEXT:    beq a4, a2, .LBB17_10
 ; RV32I-NEXT:  # %bb.9:
-; RV32I-NEXT:    sltu t6, a4, a1
+; RV32I-NEXT:    sltu t6, a4, a2
 ; RV32I-NEXT:  .LBB17_10:
 ; RV32I-NEXT:    bnez t3, .LBB17_12
 ; RV32I-NEXT:  # %bb.11:
-; RV32I-NEXT:    sub t0, t1, t0
-; RV32I-NEXT:    sub a6, a7, a6
-; RV32I-NEXT:    sub a3, a5, a3
-; RV32I-NEXT:    sub a1, a1, a4
-; RV32I-NEXT:    sub a4, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t4
-; RV32I-NEXT:    sub a2, a1, t2
-; RV32I-NEXT:    sub a1, a4, a5
-; RV32I-NEXT:    sub a4, a6, t4
+; RV32I-NEXT:    sub a7, t1, a7
+; RV32I-NEXT:    sub a5, a6, a5
+; RV32I-NEXT:    sub a1, a1, a3
+; RV32I-NEXT:    sub a2, a2, a4
+; RV32I-NEXT:    sub a4, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t4
+; RV32I-NEXT:    sub a3, a2, t2
+; RV32I-NEXT:    sub a2, a4, a6
+; RV32I-NEXT:    sub a4, a5, t4
 ; RV32I-NEXT:    j .LBB17_13
 ; RV32I-NEXT:  .LBB17_12:
-; RV32I-NEXT:    sltu a2, a6, a7
-; RV32I-NEXT:    sub t0, t0, t1
-; RV32I-NEXT:    sub a6, a6, a7
-; RV32I-NEXT:    sub a3, a3, a5
-; RV32I-NEXT:    sub a4, a4, a1
-; RV32I-NEXT:    sub a1, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t6
-; RV32I-NEXT:    sub a2, a4, t5
-; RV32I-NEXT:    sub a1, a1, a5
-; RV32I-NEXT:    sub a4, a6, t6
+; RV32I-NEXT:    sltu t0, a5, a6
+; RV32I-NEXT:    sub a7, a7, t1
+; RV32I-NEXT:    sub a5, a5, a6
+; RV32I-NEXT:    sub a1, a3, a1
+; RV32I-NEXT:    sub a4, a4, a2
+; RV32I-NEXT:    sub a2, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t6
+; RV32I-NEXT:    sub a3, a4, t5
+; RV32I-NEXT:    sub a2, a2, a6
+; RV32I-NEXT:    sub a4, a5, t6
 ; RV32I-NEXT:  .LBB17_13:
-; RV32I-NEXT:    sw a3, 0(a0)
-; RV32I-NEXT:    sw a2, 4(a0)
+; RV32I-NEXT:    sw a1, 0(a0)
+; RV32I-NEXT:    sw a3, 4(a0)
 ; RV32I-NEXT:    sw a4, 8(a0)
-; RV32I-NEXT:    sw a1, 12(a0)
+; RV32I-NEXT:    sw a2, 12(a0)
 ; RV32I-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32I-NEXT:    addi sp, sp, 16
 ; RV32I-NEXT:    ret
@@ -1221,73 +1221,73 @@ define i128 @abd_minmax_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB:       # %bb.0:
 ; RV32ZBB-NEXT:    lw a3, 0(a1)
 ; RV32ZBB-NEXT:    lw a4, 4(a1)
-; RV32ZBB-NEXT:    lw a6, 8(a1)
-; RV32ZBB-NEXT:    lw t0, 12(a1)
-; RV32ZBB-NEXT:    lw a7, 8(a2)
+; RV32ZBB-NEXT:    lw a5, 8(a1)
+; RV32ZBB-NEXT:    lw a7, 12(a1)
+; RV32ZBB-NEXT:    lw a1, 0(a2)
+; RV32ZBB-NEXT:    lw a6, 8(a2)
 ; RV32ZBB-NEXT:    lw t1, 12(a2)
-; RV32ZBB-NEXT:    lw a5, 0(a2)
-; RV32ZBB-NEXT:    lw a1, 4(a2)
-; RV32ZBB-NEXT:    sltu a2, a7, a6
-; RV32ZBB-NEXT:    mv t4, a2
-; RV32ZBB-NEXT:    beq t0, t1, .LBB17_2
+; RV32ZBB-NEXT:    lw a2, 4(a2)
+; RV32ZBB-NEXT:    sltu t0, a6, a5
+; RV32ZBB-NEXT:    mv t4, t0
+; RV32ZBB-NEXT:    beq a7, t1, .LBB17_2
 ; RV32ZBB-NEXT:  # %bb.1:
-; RV32ZBB-NEXT:    slt t4, t1, t0
+; RV32ZBB-NEXT:    slt t4, t1, a7
 ; RV32ZBB-NEXT:  .LBB17_2:
-; RV32ZBB-NEXT:    sltu t2, a5, a3
-; RV32ZBB-NEXT:    sltu t5, a1, a4
+; RV32ZBB-NEXT:    sltu t2, a1, a3
+; RV32ZBB-NEXT:    sltu t5, a2, a4
 ; RV32ZBB-NEXT:    mv t3, t2
-; RV32ZBB-NEXT:    beq a4, a1, .LBB17_4
+; RV32ZBB-NEXT:    beq a4, a2, .LBB17_4
 ; RV32ZBB-NEXT:  # %bb.3:
 ; RV32ZBB-NEXT:    mv t3, t5
 ; RV32ZBB-NEXT:  .LBB17_4:
 ; RV32ZBB-NEXT:    addi sp, sp, -16
 ; RV32ZBB-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32ZBB-NEXT:    xor t6, t0, t1
-; RV32ZBB-NEXT:    xor s0, a6, a7
+; RV32ZBB-NEXT:    xor t6, a7, t1
+; RV32ZBB-NEXT:    xor s0, a5, a6
 ; RV32ZBB-NEXT:    or t6, s0, t6
 ; RV32ZBB-NEXT:    beqz t6, .LBB17_6
 ; RV32ZBB-NEXT:  # %bb.5:
 ; RV32ZBB-NEXT:    mv t3, t4
 ; RV32ZBB-NEXT:  .LBB17_6:
 ; RV32ZBB-NEXT:    mv t4, t2
-; RV32ZBB-NEXT:    beq a1, a4, .LBB17_8
+; RV32ZBB-NEXT:    beq a2, a4, .LBB17_8
 ; RV32ZBB-NEXT:  # %bb.7:
 ; RV32ZBB-NEXT:    mv t4, t5
 ; RV32ZBB-NEXT:  .LBB17_8:
-; RV32ZBB-NEXT:    sltu t5, a3, a5
+; RV32ZBB-NEXT:    sltu t5, a3, a1
 ; RV32ZBB-NEXT:    mv t6, t5
-; RV32ZBB-NEXT:    beq a4, a1, .LBB17_10
+; RV32ZBB-NEXT:    beq a4, a2, .LBB17_10
 ; RV32ZBB-NEXT:  # %bb.9:
-; RV32ZBB-NEXT:    sltu t6, a4, a1
+; RV32ZBB-NEXT:    sltu t6, a4, a2
 ; RV32ZBB-NEXT:  .LBB17_10:
 ; RV32ZBB-NEXT:    bnez t3, .LBB17_12
 ; RV32ZBB-NEXT:  # %bb.11:
-; RV32ZBB-NEXT:    sub t0, t1, t0
-; RV32ZBB-NEXT:    sub a6, a7, a6
-; RV32ZBB-NEXT:    sub a3, a5, a3
-; RV32ZBB-NEXT:    sub a1, a1, a4
-; RV32ZBB-NEXT:    sub a4, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t4
-; RV32ZBB-NEXT:    sub a2, a1, t2
-; RV32ZBB-NEXT:    sub a1, a4, a5
-; RV32ZBB-NEXT:    sub a4, a6, t4
+; RV32ZBB-NEXT:    sub a7, t1, a7
+; RV32ZBB-NEXT:    sub a5, a6, a5
+; RV32ZBB-NEXT:    sub a1, a1, a3
+; RV32ZBB-NEXT:    sub a2, a2, a4
+; RV32ZBB-NEXT:    sub a4, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t4
+; RV32ZBB-NEXT:    sub a3, a2, t2
+; RV32ZBB-NEXT:    sub a2, a4, a6
+; RV32ZBB-NEXT:    sub a4, a5, t4
 ; RV32ZBB-NEXT:    j .LBB17_13
 ; RV32ZBB-NEXT:  .LBB17_12:
-; RV32ZBB-NEXT:    sltu a2, a6, a7
-; RV32ZBB-NEXT:    sub t0, t0, t1
-; RV32ZBB-NEXT:    sub a6, a6, a7
-; RV32ZBB-NEXT:    sub a3, a3, a5
-; RV32ZBB-NEXT:    sub a4, a4, a1
-; RV32ZBB-NEXT:    sub a1, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t6
-; RV32ZBB-NEXT:    sub a2, a4, t5
-; RV32ZBB-NEXT:    sub a1, a1, a5
-; RV32ZBB-NEXT:    sub a4, a6, t6
+; RV32ZBB-NEXT:    sltu t0, a5, a6
+; RV32ZBB-NEXT:    sub a7, a7, t1
+; RV32ZBB-NEXT:    sub a5, a5, a6
+; RV32ZBB-NEXT:    sub a1, a3, a1
+; RV32ZBB-NEXT:    sub a4, a4, a2
+; RV32ZBB-NEXT:    sub a2, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t6
+; RV32ZBB-NEXT:    sub a3, a4, t5
+; RV32ZBB-NEXT:    sub a2, a2, a6
+; RV32ZBB-NEXT:    sub a4, a5, t6
 ; RV32ZBB-NEXT:  .LBB17_13:
-; RV32ZBB-NEXT:    sw a3, 0(a0)
-; RV32ZBB-NEXT:    sw a2, 4(a0)
+; RV32ZBB-NEXT:    sw a1, 0(a0)
+; RV32ZBB-NEXT:    sw a3, 4(a0)
 ; RV32ZBB-NEXT:    sw a4, 8(a0)
-; RV32ZBB-NEXT:    sw a1, 12(a0)
+; RV32ZBB-NEXT:    sw a2, 12(a0)
 ; RV32ZBB-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32ZBB-NEXT:    addi sp, sp, 16
 ; RV32ZBB-NEXT:    ret
@@ -1516,73 +1516,73 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lw a3, 0(a1)
 ; RV32I-NEXT:    lw a4, 4(a1)
-; RV32I-NEXT:    lw a6, 8(a1)
-; RV32I-NEXT:    lw t0, 12(a1)
-; RV32I-NEXT:    lw a7, 8(a2)
+; RV32I-NEXT:    lw a5, 8(a1)
+; RV32I-NEXT:    lw a7, 12(a1)
+; RV32I-NEXT:    lw a1, 0(a2)
+; RV32I-NEXT:    lw a6, 8(a2)
 ; RV32I-NEXT:    lw t1, 12(a2)
-; RV32I-NEXT:    lw a5, 0(a2)
-; RV32I-NEXT:    lw a1, 4(a2)
-; RV32I-NEXT:    sltu a2, a7, a6
-; RV32I-NEXT:    mv t4, a2
-; RV32I-NEXT:    beq t0, t1, .LBB22_2
+; RV32I-NEXT:    lw a2, 4(a2)
+; RV32I-NEXT:    sltu t0, a6, a5
+; RV32I-NEXT:    mv t4, t0
+; RV32I-NEXT:    beq a7, t1, .LBB22_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    slt t4, t1, t0
+; RV32I-NEXT:    slt t4, t1, a7
 ; RV32I-NEXT:  .LBB22_2:
-; RV32I-NEXT:    sltu t2, a5, a3
-; RV32I-NEXT:    sltu t5, a1, a4
+; RV32I-NEXT:    sltu t2, a1, a3
+; RV32I-NEXT:    sltu t5, a2, a4
 ; RV32I-NEXT:    mv t3, t2
-; RV32I-NEXT:    beq a4, a1, .LBB22_4
+; RV32I-NEXT:    beq a4, a2, .LBB22_4
 ; RV32I-NEXT:  # %bb.3:
 ; RV32I-NEXT:    mv t3, t5
 ; RV32I-NEXT:  .LBB22_4:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    xor t6, t0, t1
-; RV32I-NEXT:    xor s0, a6, a7
+; RV32I-NEXT:    xor t6, a7, t1
+; RV32I-NEXT:    xor s0, a5, a6
 ; RV32I-NEXT:    or t6, s0, t6
 ; RV32I-NEXT:    beqz t6, .LBB22_6
 ; RV32I-NEXT:  # %bb.5:
 ; RV32I-NEXT:    mv t3, t4
 ; RV32I-NEXT:  .LBB22_6:
 ; RV32I-NEXT:    mv t4, t2
-; RV32I-NEXT:    beq a1, a4, .LBB22_8
+; RV32I-NEXT:    beq a2, a4, .LBB22_8
 ; RV32I-NEXT:  # %bb.7:
 ; RV32I-NEXT:    mv t4, t5
 ; RV32I-NEXT:  .LBB22_8:
-; RV32I-NEXT:    sltu t5, a3, a5
+; RV32I-NEXT:    sltu t5, a3, a1
 ; RV32I-NEXT:    mv t6, t5
-; RV32I-NEXT:    beq a4, a1, .LBB22_10
+; RV32I-NEXT:    beq a4, a2, .LBB22_10
 ; RV32I-NEXT:  # %bb.9:
-; RV32I-NEXT:    sltu t6, a4, a1
+; RV32I-NEXT:    sltu t6, a4, a2
 ; RV32I-NEXT:  .LBB22_10:
 ; RV32I-NEXT:    bnez t3, .LBB22_12
 ; RV32I-NEXT:  # %bb.11:
-; RV32I-NEXT:    sub t0, t1, t0
-; RV32I-NEXT:    sub a6, a7, a6
-; RV32I-NEXT:    sub a3, a5, a3
-; RV32I-NEXT:    sub a1, a1, a4
-; RV32I-NEXT:    sub a4, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t4
-; RV32I-NEXT:    sub a2, a1, t2
-; RV32I-NEXT:    sub a1, a4, a5
-; RV32I-NEXT:    sub a4, a6, t4
+; RV32I-NEXT:    sub a7, t1, a7
+; RV32I-NEXT:    sub a5, a6, a5
+; RV32I-NEXT:    sub a1, a1, a3
+; RV32I-NEXT:    sub a2, a2, a4
+; RV32I-NEXT:    sub a4, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t4
+; RV32I-NEXT:    sub a3, a2, t2
+; RV32I-NEXT:    sub a2, a4, a6
+; RV32I-NEXT:    sub a4, a5, t4
 ; RV32I-NEXT:    j .LBB22_13
 ; RV32I-NEXT:  .LBB22_12:
-; RV32I-NEXT:    sltu a2, a6, a7
-; RV32I-NEXT:    sub t0, t0, t1
-; RV32I-NEXT:    sub a6, a6, a7
-; RV32I-NEXT:    sub a3, a3, a5
-; RV32I-NEXT:    sub a4, a4, a1
-; RV32I-NEXT:    sub a1, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t6
-; RV32I-NEXT:    sub a2, a4, t5
-; RV32I-NEXT:    sub a1, a1, a5
-; RV32I-NEXT:    sub a4, a6, t6
+; RV32I-NEXT:    sltu t0, a5, a6
+; RV32I-NEXT:    sub a7, a7, t1
+; RV32I-NEXT:    sub a5, a5, a6
+; RV32I-NEXT:    sub a1, a3, a1
+; RV32I-NEXT:    sub a4, a4, a2
+; RV32I-NEXT:    sub a2, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t6
+; RV32I-NEXT:    sub a3, a4, t5
+; RV32I-NEXT:    sub a2, a2, a6
+; RV32I-NEXT:    sub a4, a5, t6
 ; RV32I-NEXT:  .LBB22_13:
-; RV32I-NEXT:    sw a3, 0(a0)
-; RV32I-NEXT:    sw a2, 4(a0)
+; RV32I-NEXT:    sw a1, 0(a0)
+; RV32I-NEXT:    sw a3, 4(a0)
 ; RV32I-NEXT:    sw a4, 8(a0)
-; RV32I-NEXT:    sw a1, 12(a0)
+; RV32I-NEXT:    sw a2, 12(a0)
 ; RV32I-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32I-NEXT:    addi sp, sp, 16
 ; RV32I-NEXT:    ret
@@ -1612,73 +1612,73 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB:       # %bb.0:
 ; RV32ZBB-NEXT:    lw a3, 0(a1)
 ; RV32ZBB-NEXT:    lw a4, 4(a1)
-; RV32ZBB-NEXT:    lw a6, 8(a1)
-; RV32ZBB-NEXT:    lw t0, 12(a1)
-; RV32ZBB-NEXT:    lw a7, 8(a2)
+; RV32ZBB-NEXT:    lw a5, 8(a1)
+; RV32ZBB-NEXT:    lw a7, 12(a1)
+; RV32ZBB-NEXT:    lw a1, 0(a2)
+; RV32ZBB-NEXT:    lw a6, 8(a2)
 ; RV32ZBB-NEXT:    lw t1, 12(a2)
-; RV32ZBB-NEXT:    lw a5, 0(a2)
-; RV32ZBB-NEXT:    lw a1, 4(a2)
-; RV32ZBB-NEXT:    sltu a2, a7, a6
-; RV32ZBB-NEXT:    mv t4, a2
-; RV32ZBB-NEXT:    beq t0, t1, .LBB22_2
+; RV32ZBB-NEXT:    lw a2, 4(a2)
+; RV32ZBB-NEXT:    sltu t0, a6, a5
+; RV32ZBB-NEXT:    mv t4, t0
+; RV32ZBB-NEXT:    beq a7, t1, .LBB22_2
 ; RV32ZBB-NEXT:  # %bb.1:
-; RV32ZBB-NEXT:    slt t4, t1, t0
+; RV32ZBB-NEXT:    slt t4, t1, a7
 ; RV32ZBB-NEXT:  .LBB22_2:
-; RV32ZBB-NEXT:    sltu t2, a5, a3
-; RV32ZBB-NEXT:    sltu t5, a1, a4
+; RV32ZBB-NEXT:    sltu t2, a1, a3
+; RV32ZBB-NEXT:    sltu t5, a2, a4
 ; RV32ZBB-NEXT:    mv t3, t2
-; RV32ZBB-NEXT:    beq a4, a1, .LBB22_4
+; RV32ZBB-NEXT:    beq a4, a2, .LBB22_4
 ; RV32ZBB-NEXT:  # %bb.3:
 ; RV32ZBB-NEXT:    mv t3, t5
 ; RV32ZBB-NEXT:  .LBB22_4:
 ; RV32ZBB-NEXT:    addi sp, sp, -16
 ; RV32ZBB-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32ZBB-NEXT:    xor t6, t0, t1
-; RV32ZBB-NEXT:    xor s0, a6, a7
+; RV32ZBB-NEXT:    xor t6, a7, t1
+; RV32ZBB-NEXT:    xor s0, a5, a6
 ; RV32ZBB-NEXT:    or t6, s0, t6
 ; RV32ZBB-NEXT:    beqz t6, .LBB22_6
 ; RV32ZBB-NEXT:  # %bb.5:
 ; RV32ZBB-NEXT:    mv t3, t4
 ; RV32ZBB-NEXT:  .LBB22_6:
 ; RV32ZBB-NEXT:    mv t4, t2
-; RV32ZBB-NEXT:    beq a1, a4, .LBB22_8
+; RV32ZBB-NEXT:    beq a2, a4, .LBB22_8
 ; RV32ZBB-NEXT:  # %bb.7:
 ; RV32ZBB-NEXT:    mv t4, t5
 ; RV32ZBB-NEXT:  .LBB22_8:
-; RV32ZBB-NEXT:    sltu t5, a3, a5
+; RV32ZBB-NEXT:    sltu t5, a3, a1
 ; RV32ZBB-NEXT:    mv t6, t5
-; RV32ZBB-NEXT:    beq a4, a1, .LBB22_10
+; RV32ZBB-NEXT:    beq a4, a2, .LBB22_10
 ; RV32ZBB-NEXT:  # %bb.9:
-; RV32ZBB-NEXT:    sltu t6, a4, a1
+; RV32ZBB-NEXT:    sltu t6, a4, a2
 ; RV32ZBB-NEXT:  .LBB22_10:
 ; RV32ZBB-NEXT:    bnez t3, .LBB22_12
 ; RV32ZBB-NEXT:  # %bb.11:
-; RV32ZBB-NEXT:    sub t0, t1, t0
-; RV32ZBB-NEXT:    sub a6, a7, a6
-; RV32ZBB-NEXT:    sub a3, a5, a3
-; RV32ZBB-NEXT:    sub a1, a1, a4
-; RV32ZBB-NEXT:    sub a4, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t4
-; RV32ZBB-NEXT:    sub a2, a1, t2
-; RV32ZBB-NEXT:    sub a1, a4, a5
-; RV32ZBB-NEXT:    sub a4, a6, t4
+; RV32ZBB-NEXT:    sub a7, t1, a7
+; RV32ZBB-NEXT:    sub a5, a6, a5
+; RV32ZBB-NEXT:    sub a1, a1, a3
+; RV32ZBB-NEXT:    sub a2, a2, a4
+; RV32ZBB-NEXT:    sub a4, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t4
+; RV32ZBB-NEXT:    sub a3, a2, t2
+; RV32ZBB-NEXT:    sub a2, a4, a6
+; RV32ZBB-NEXT:    sub a4, a5, t4
 ; RV32ZBB-NEXT:    j .LBB22_13
 ; RV32ZBB-NEXT:  .LBB22_12:
-; RV32ZBB-NEXT:    sltu a2, a6, a7
-; RV32ZBB-NEXT:    sub t0, t0, t1
-; RV32ZBB-NEXT:    sub a6, a6, a7
-; RV32ZBB-NEXT:    sub a3, a3, a5
-; RV32ZBB-NEXT:    sub a4, a4, a1
-; RV32ZBB-NEXT:    sub a1, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t6
-; RV32ZBB-NEXT:    sub a2, a4, t5
-; RV32ZBB-NEXT:    sub a1, a1, a5
-; RV32ZBB-NEXT:    sub a4, a6, t6
+; RV32ZBB-NEXT:    sltu t0, a5, a6
+; RV32ZBB-NEXT:    sub a7, a7, t1
+; RV32ZBB-NEXT:    sub a5, a5, a6
+; RV32ZBB-NEXT:    sub a1, a3, a1
+; RV32ZBB-NEXT:    sub a4, a4, a2
+; RV32ZBB-NEXT:    sub a2, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t6
+; RV32ZBB-NEXT:    sub a3, a4, t5
+; RV32ZBB-NEXT:    sub a2, a2, a6
+; RV32ZBB-NEXT:    sub a4, a5, t6
 ; RV32ZBB-NEXT:  .LBB22_13:
-; RV32ZBB-NEXT:    sw a3, 0(a0)
-; RV32ZBB-NEXT:    sw a2, 4(a0)
+; RV32ZBB-NEXT:    sw a1, 0(a0)
+; RV32ZBB-NEXT:    sw a3, 4(a0)
 ; RV32ZBB-NEXT:    sw a4, 8(a0)
-; RV32ZBB-NEXT:    sw a1, 12(a0)
+; RV32ZBB-NEXT:    sw a2, 12(a0)
 ; RV32ZBB-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32ZBB-NEXT:    addi sp, sp, 16
 ; RV32ZBB-NEXT:    ret
@@ -2539,73 +2539,73 @@ define i128 @abd_select_i128(i128 %a, i128 %b) nounwind {
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lw a3, 0(a1)
 ; RV32I-NEXT:    lw a4, 4(a1)
-; RV32I-NEXT:    lw a6, 8(a1)
-; RV32I-NEXT:    lw t0, 12(a1)
-; RV32I-NEXT:    lw a7, 8(a2)
+; RV32I-NEXT:    lw a5, 8(a1)
+; RV32I-NEXT:    lw a7, 12(a1)
+; RV32I-NEXT:    lw a1, 0(a2)
+; RV32I-NEXT:    lw a6, 8(a2)
 ; RV32I-NEXT:    lw t1, 12(a2)
-; RV32I-NEXT:    lw a5, 0(a2)
-; RV32I-NEXT:    lw a1, 4(a2)
-; RV32I-NEXT:    sltu a2, a7, a6
-; RV32I-NEXT:    mv t4, a2
-; RV32I-NEXT:    beq t0, t1, .LBB38_2
+; RV32I-NEXT:    lw a2, 4(a2)
+; RV32I-NEXT:    sltu t0, a6, a5
+; RV32I-NEXT:    mv t4, t0
+; RV32I-NEXT:    beq a7, t1, .LBB38_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    slt t4, t1, t0
+; RV32I-NEXT:    slt t4, t1, a7
 ; RV32I-NEXT:  .LBB38_2:
-; RV32I-NEXT:    sltu t2, a5, a3
-; RV32I-NEXT:    sltu t5, a1, a4
+; RV32I-NEXT:    sltu t2, a1, a3
+; RV32I-NEXT:    sltu t5, a2, a4
 ; RV32I-NEXT:    mv t3, t2
-; RV32I-NEXT:    beq a4, a1, .LBB38_4
+; RV32I-NEXT:    beq a4, a2, .LBB38_4
 ; RV32I-NEXT:  # %bb.3:
 ; RV32I-NEXT:    mv t3, t5
 ; RV32I-NEXT:  .LBB38_4:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    xor t6, t0, t1
-; RV32I-NEXT:    xor s0, a6, a7
+; RV32I-NEXT:    xor t6, a7, t1
+; RV32I-NEXT:    xor s0, a5, a6
 ; RV32I-NEXT:    or t6, s0, t6
 ; RV32I-NEXT:    beqz t6, .LBB38_6
 ; RV32I-NEXT:  # %bb.5:
 ; RV32I-NEXT:    mv t3, t4
 ; RV32I-NEXT:  .LBB38_6:
 ; RV32I-NEXT:    mv t4, t2
-; RV32I-NEXT:    beq a1, a4, .LBB38_8
+; RV32I-NEXT:    beq a2, a4, .LBB38_8
 ; RV32I-NEXT:  # %bb.7:
 ; RV32I-NEXT:    mv t4, t5
 ; RV32I-NEXT:  .LBB38_8:
-; RV32I-NEXT:    sltu t5, a3, a5
+; RV32I-NEXT:    sltu t5, a3, a1
 ; RV32I-NEXT:    mv t6, t5
-; RV32I-NEXT:    beq a4, a1, .LBB38_10
+; RV32I-NEXT:    beq a4, a2, .LBB38_10
 ; RV32I-NEXT:  # %bb.9:
-; RV32I-NEXT:    sltu t6, a4, a1
+; RV32I-NEXT:    sltu t6, a4, a2
 ; RV32I-NEXT:  .LBB38_10:
 ; RV32I-NEXT:    bnez t3, .LBB38_12
 ; RV32I-NEXT:  # %bb.11:
-; RV32I-NEXT:    sub t0, t1, t0
-; RV32I-NEXT:    sub a6, a7, a6
-; RV32I-NEXT:    sub a3, a5, a3
-; RV32I-NEXT:    sub a1, a1, a4
-; RV32I-NEXT:    sub a4, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t4
-; RV32I-NEXT:    sub a2, a1, t2
-; RV32I-NEXT:    sub a1, a4, a5
-; RV32I-NEXT:    sub a4, a6, t4
+; RV32I-NEXT:    sub a7, t1, a7
+; RV32I-NEXT:    sub a5, a6, a5
+; RV32I-NEXT:    sub a1, a1, a3
+; RV32I-NEXT:    sub a2, a2, a4
+; RV32I-NEXT:    sub a4, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t4
+; RV32I-NEXT:    sub a3, a2, t2
+; RV32I-NEXT:    sub a2, a4, a6
+; RV32I-NEXT:    sub a4, a5, t4
 ; RV32I-NEXT:    j .LBB38_13
 ; RV32I-NEXT:  .LBB38_12:
-; RV32I-NEXT:    sltu a2, a6, a7
-; RV32I-NEXT:    sub t0, t0, t1
-; RV32I-NEXT:    sub a6, a6, a7
-; RV32I-NEXT:    sub a3, a3, a5
-; RV32I-NEXT:    sub a4, a4, a1
-; RV32I-NEXT:    sub a1, t0, a2
-; RV32I-NEXT:    sltu a5, a6, t6
-; RV32I-NEXT:    sub a2, a4, t5
-; RV32I-NEXT:    sub a1, a1, a5
-; RV32I-NEXT:    sub a4, a6, t6
+; RV32I-NEXT:    sltu t0, a5, a6
+; RV32I-NEXT:    sub a7, a7, t1
+; RV32I-NEXT:    sub a5, a5, a6
+; RV32I-NEXT:    sub a1, a3, a1
+; RV32I-NEXT:    sub a4, a4, a2
+; RV32I-NEXT:    sub a2, a7, t0
+; RV32I-NEXT:    sltu a6, a5, t6
+; RV32I-NEXT:    sub a3, a4, t5
+; RV32I-NEXT:    sub a2, a2, a6
+; RV32I-NEXT:    sub a4, a5, t6
 ; RV32I-NEXT:  .LBB38_13:
-; RV32I-NEXT:    sw a3, 0(a0)
-; RV32I-NEXT:    sw a2, 4(a0)
+; RV32I-NEXT:    sw a1, 0(a0)
+; RV32I-NEXT:    sw a3, 4(a0)
 ; RV32I-NEXT:    sw a4, 8(a0)
-; RV32I-NEXT:    sw a1, 12(a0)
+; RV32I-NEXT:    sw a2, 12(a0)
 ; RV32I-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32I-NEXT:    addi sp, sp, 16
 ; RV32I-NEXT:    ret
@@ -2635,73 +2635,73 @@ define i128 @abd_select_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB:       # %bb.0:
 ; RV32ZBB-NEXT:    lw a3, 0(a1)
 ; RV32ZBB-NEXT:    lw a4, 4(a1)
-; RV32ZBB-NEXT:    lw a6, 8(a1)
-; RV32ZBB-NEXT:    lw t0, 12(a1)
-; RV32ZBB-NEXT:    lw a7, 8(a2)
+; RV32ZBB-NEXT:    lw a5, 8(a1)
+; RV32ZBB-NEXT:    lw a7, 12(a1)
+; RV32ZBB-NEXT:    lw a1, 0(a2)
+; RV32ZBB-NEXT:    lw a6, 8(a2)
 ; RV32ZBB-NEXT:    lw t1, 12(a2)
-; RV32ZBB-NEXT:    lw a5, 0(a2)
-; RV32ZBB-NEXT:    lw a1, 4(a2)
-; RV32ZBB-NEXT:    sltu a2, a7, a6
-; RV32ZBB-NEXT:    mv t4, a2
-; RV32ZBB-NEXT:    beq t0, t1, .LBB38_2
+; RV32ZBB-NEXT:    lw a2, 4(a2)
+; RV32ZBB-NEXT:    sltu t0, a6, a5
+; RV32ZBB-NEXT:    mv t4, t0
+; RV32ZBB-NEXT:    beq a7, t1, .LBB38_2
 ; RV32ZBB-NEXT:  # %bb.1:
-; RV32ZBB-NEXT:    slt t4, t1, t0
+; RV32ZBB-NEXT:    slt t4, t1, a7
 ; RV32ZBB-NEXT:  .LBB38_2:
-; RV32ZBB-NEXT:    sltu t2, a5, a3
-; RV32ZBB-NEXT:    sltu t5, a1, a4
+; RV32ZBB-NEXT:    sltu t2, a1, a3
+; RV32ZBB-NEXT:    sltu t5, a2, a4
 ; RV32ZBB-NEXT:    mv t3, t2
-; RV32ZBB-NEXT:    beq a4, a1, .LBB38_4
+; RV32ZBB-NEXT:    beq a4, a2, .LBB38_4
 ; RV32ZBB-NEXT:  # %bb.3:
 ; RV32ZBB-NEXT:    mv t3, t5
 ; RV32ZBB-NEXT:  .LBB38_4:
 ; RV32ZBB-NEXT:    addi sp, sp, -16
 ; RV32ZBB-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
-; RV32ZBB-NEXT:    xor t6, t0, t1
-; RV32ZBB-NEXT:    xor s0, a6, a7
+; RV32ZBB-NEXT:    xor t6, a7, t1
+; RV32ZBB-NEXT:    xor s0, a5, a6
 ; RV32ZBB-NEXT:    or t6, s0, t6
 ; RV32ZBB-NEXT:    beqz t6, .LBB38_6
 ; RV32ZBB-NEXT:  # %bb.5:
 ; RV32ZBB-NEXT:    mv t3, t4
 ; RV32ZBB-NEXT:  .LBB38_6:
 ; RV32ZBB-NEXT:    mv t4, t2
-; RV32ZBB-NEXT:    beq a1, a4, .LBB38_8
+; RV32ZBB-NEXT:    beq a2, a4, .LBB38_8
 ; RV32ZBB-NEXT:  # %bb.7:
 ; RV32ZBB-NEXT:    mv t4, t5
 ; RV32ZBB-NEXT:  .LBB38_8:
-; RV32ZBB-NEXT:    sltu t5, a3, a5
+; RV32ZBB-NEXT:    sltu t5, a3, a1
 ; RV32ZBB-NEXT:    mv t6, t5
-; RV32ZBB-NEXT:    beq a4, a1, .LBB38_10
+; RV32ZBB-NEXT:    beq a4, a2, .LBB38_10
 ; RV32ZBB-NEXT:  # %bb.9:
-; RV32ZBB-NEXT:    sltu t6, a4, a1
+; RV32ZBB-NEXT:    sltu t6, a4, a2
 ; RV32ZBB-NEXT:  .LBB38_10:
 ; RV32ZBB-NEXT:    bnez t3, .LBB38_12
 ; RV32ZBB-NEXT:  # %bb.11:
-; RV32ZBB-NEXT:    sub t0, t1, t0
-; RV32ZBB-NEXT:    sub a6, a7, a6
-; RV32ZBB-NEXT:    sub a3, a5, a3
-; RV32ZBB-NEXT:    sub a1, a1, a4
-; RV32ZBB-NEXT:    sub a4, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t4
-; RV32ZBB-NEXT:    sub a2, a1, t2
-; RV32ZBB-NEXT:    sub a1, a4, a5
-; RV32ZBB-NEXT:    sub a4, a6, t4
+; RV32ZBB-NEXT:    sub a7, t1, a7
+; RV32ZBB-NEXT:    sub a5, a6, a5
+; RV32ZBB-NEXT:    sub a1, a1, a3
+; RV32ZBB-NEXT:    sub a2, a2, a4
+; RV32ZBB-NEXT:    sub a4, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t4
+; RV32ZBB-NEXT:    sub a3, a2, t2
+; RV32ZBB-NEXT:    sub a2, a4, a6
+; RV32ZBB-NEXT:    sub a4, a5, t4
 ; RV32ZBB-NEXT:    j .LBB38_13
 ; RV32ZBB-NEXT:  .LBB38_12:
-; RV32ZBB-NEXT:    sltu a2, a6, a7
-; RV32ZBB-NEXT:    sub t0, t0, t1
-; RV32ZBB-NEXT:    sub a6, a6, a7
-; RV32ZBB-NEXT:    sub a3, a3, a5
-; RV32ZBB-NEXT:    sub a4, a4, a1
-; RV32ZBB-NEXT:    sub a1, t0, a2
-; RV32ZBB-NEXT:    sltu a5, a6, t6
-; RV32ZBB-NEXT:    sub a2, a4, t5
-; RV32ZBB-NEXT:    sub a1, a1, a5
-; RV32ZBB-NEXT:    sub a4, a6, t6
+; RV32ZBB-NEXT:    sltu t0, a5, a6
+; RV32ZBB-NEXT:    sub a7, a7, t1
+; RV32ZBB-NEXT:    sub a5, a5, a6
+; RV32ZBB-NEXT:    sub a1, a3, a1
+; RV32ZBB-NEXT:    sub a4, a4, a2
+; RV32ZBB-NEXT:    sub a2, a7, t0
+; RV32ZBB-NEXT:    sltu a6, a5, t6
+; RV32ZBB-NEXT:    sub a3, a4, t5
+; RV32ZBB-NEXT:    sub a2, a2, a6
+; RV32ZBB-NEXT:    sub a4, a5, t6
 ; RV32ZBB-NEXT:  .LBB38_13:
-; RV32ZBB-NEXT:    sw a3, 0(a0)
-; RV32ZBB-NEXT:    sw a2, 4(a0)
+; RV32ZBB-NEXT:    sw a1, 0(a0)
+; RV32ZBB-NEXT:    sw a3, 4(a0)
 ; RV32ZBB-NEXT:    sw a4, 8(a0)
-; RV32ZBB-NEXT:    sw a1, 12(a0)
+; RV32ZBB-NEXT:    sw a2, 12(a0)
 ; RV32ZBB-NEXT:    lw s0, 12(sp) # 4-byte Folded Reload
 ; RV32ZBB-NEXT:    addi sp, sp, 16
 ; RV32ZBB-NEXT:    ret
diff --git a/llvm/test/CodeGen/RISCV/abdu-neg.ll b/llvm/test/CodeGen/RISCV/abdu-neg.ll
index 911db598eb831..713b52f53e3d9 100644
--- a/llvm/test/CodeGen/RISCV/abdu-neg.ll
+++ b/llvm/test/CodeGen/RISCV/abdu-neg.ll
@@ -1318,8 +1318,8 @@ define i128 @abd_minmax_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:    lw a6, 4(a2)
 ; RV32I-NEXT:    lw a7, 8(a2)
 ; RV32I-NEXT:    lw t0, 12(a2)
-; RV32I-NEXT:    lw a5, 12(a1)
 ; RV32I-NEXT:    lw a3, 4(a1)
+; RV32I-NEXT:    lw a5, 12(a1)
 ; RV32I-NEXT:    lw a4, 8(a1)
 ; RV32I-NEXT:    beq a5, t0, .LBB17_2
 ; RV32I-NEXT:  # %bb.1:
@@ -1445,8 +1445,8 @@ define i128 @abd_minmax_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:    lw a6, 4(a2)
 ; RV32ZBB-NEXT:    lw a7, 8(a2)
 ; RV32ZBB-NEXT:    lw t0, 12(a2)
-; RV32ZBB-NEXT:    lw a5, 12(a1)
 ; RV32ZBB-NEXT:    lw a3, 4(a1)
+; RV32ZBB-NEXT:    lw a5, 12(a1)
 ; RV32ZBB-NEXT:    lw a4, 8(a1)
 ; RV32ZBB-NEXT:    beq a5, t0, .LBB17_2
 ; RV32ZBB-NEXT:  # %bb.1:
@@ -1767,15 +1767,15 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:    lw a4, 4(a2)
 ; RV32I-NEXT:    lw a5, 8(a2)
 ; RV32I-NEXT:    lw a7, 12(a2)
-; RV32I-NEXT:    lw a6, 8(a1)
-; RV32I-NEXT:    lw t0, 12(a1)
 ; RV32I-NEXT:    lw a2, 0(a1)
+; RV32I-NEXT:    lw a6, 8(a1)
+; RV32I-NEXT:    lw t1, 12(a1)
 ; RV32I-NEXT:    lw a1, 4(a1)
-; RV32I-NEXT:    sltu t1, a6, a5
-; RV32I-NEXT:    mv t4, t1
-; RV32I-NEXT:    beq t0, a7, .LBB22_2
+; RV32I-NEXT:    sltu t0, a6, a5
+; RV32I-NEXT:    mv t4, t0
+; RV32I-NEXT:    beq t1, a7, .LBB22_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    sltu t4, t0, a7
+; RV32I-NEXT:    sltu t4, t1, a7
 ; RV32I-NEXT:  .LBB22_2:
 ; RV32I-NEXT:    sltu t2, a2, a3
 ; RV32I-NEXT:    mv t3, t2
@@ -1783,7 +1783,7 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:  # %bb.3:
 ; RV32I-NEXT:    sltu t3, a1, a4
 ; RV32I-NEXT:  .LBB22_4:
-; RV32I-NEXT:    xor t5, t0, a7
+; RV32I-NEXT:    xor t5, t1, a7
 ; RV32I-NEXT:    xor t6, a6, a5
 ; RV32I-NEXT:    or t5, t6, t5
 ; RV32I-NEXT:    mv t6, t3
@@ -1799,11 +1799,11 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:  .LBB22_8:
 ; RV32I-NEXT:    bnez t6, .LBB22_10
 ; RV32I-NEXT:  # %bb.9:
-; RV32I-NEXT:    sltu t1, a5, a6
-; RV32I-NEXT:    sub a7, a7, t0
+; RV32I-NEXT:    sltu t0, a5, a6
+; RV32I-NEXT:    sub a7, a7, t1
 ; RV32I-NEXT:    sub a5, a5, a6
 ; RV32I-NEXT:    sub a4, a4, a1
-; RV32I-NEXT:    sub a6, a7, t1
+; RV32I-NEXT:    sub a6, a7, t0
 ; RV32I-NEXT:    sltu a7, a5, t5
 ; RV32I-NEXT:    sub a1, a5, t5
 ; RV32I-NEXT:    sub a5, a4, t4
@@ -1811,10 +1811,10 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32I-NEXT:    sub a2, a3, a2
 ; RV32I-NEXT:    j .LBB22_11
 ; RV32I-NEXT:  .LBB22_10:
-; RV32I-NEXT:    sub a7, t0, a7
+; RV32I-NEXT:    sub a7, t1, a7
 ; RV32I-NEXT:    sub a5, a6, a5
 ; RV32I-NEXT:    sub a4, a1, a4
-; RV32I-NEXT:    sub a6, a7, t1
+; RV32I-NEXT:    sub a6, a7, t0
 ; RV32I-NEXT:    sltu a7, a5, t3
 ; RV32I-NEXT:    sub a1, a5, t3
 ; RV32I-NEXT:    sub a5, a4, t2
@@ -1854,15 +1854,15 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:    lw a4, 4(a2)
 ; RV32ZBB-NEXT:    lw a5, 8(a2)
 ; RV32ZBB-NEXT:    lw a7, 12(a2)
-; RV32ZBB-NEXT:    lw a6, 8(a1)
-; RV32ZBB-NEXT:    lw t0, 12(a1)
 ; RV32ZBB-NEXT:    lw a2, 0(a1)
+; RV32ZBB-NEXT:    lw a6, 8(a1)
+; RV32ZBB-NEXT:    lw t1, 12(a1)
 ; RV32ZBB-NEXT:    lw a1, 4(a1)
-; RV32ZBB-NEXT:    sltu t1, a6, a5
-; RV32ZBB-NEXT:    mv t4, t1
-; RV32ZBB-NEXT:    beq t0, a7, .LBB22_2
+; RV32ZBB-NEXT:    sltu t0, a6, a5
+; RV32ZBB-NEXT:    mv t4, t0
+; RV32ZBB-NEXT:    beq t1, a7, .LBB22_2
 ; RV32ZBB-NEXT:  # %bb.1:
-; RV32ZBB-NEXT:    sltu t4, t0, a7
+; RV32ZBB-NEXT:    sltu t4, t1, a7
 ; RV32ZBB-NEXT:  .LBB22_2:
 ; RV32ZBB-NEXT:    sltu t2, a2, a3
 ; RV32ZBB-NEXT:    mv t3, t2
@@ -1870,7 +1870,7 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:  # %bb.3:
 ; RV32ZBB-NEXT:    sltu t3, a1, a4
 ; RV32ZBB-NEXT:  .LBB22_4:
-; RV32ZBB-NEXT:    xor t5, t0, a7
+; RV32ZBB-NEXT:    xor t5, t1, a7
 ; RV32ZBB-NEXT:    xor t6, a6, a5
 ; RV32ZBB-NEXT:    or t5, t6, t5
 ; RV32ZBB-NEXT:    mv t6, t3
@@ -1886,11 +1886,11 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:  .LBB22_8:
 ; RV32ZBB-NEXT:    bnez t6, .LBB22_10
 ; RV32ZBB-NEXT:  # %bb.9:
-; RV32ZBB-NEXT:    sltu t1, a5, a6
-; RV32ZBB-NEXT:    sub a7, a7, t0
+; RV32ZBB-NEXT:    sltu t0, a5, a6
+; RV32ZBB-NEXT:    sub a7, a7, t1
 ; RV32ZBB-NEXT:    sub a5, a5, a6
 ; RV32ZBB-NEXT:    sub a4, a4, a1
-; RV32ZBB-NEXT:    sub a6, a7, t1
+; RV32ZBB-NEXT:    sub a6, a7, t0
 ; RV32ZBB-NEXT:    sltu a7, a5, t5
 ; RV32ZBB-NEXT:    sub a1, a5, t5
 ; RV32ZBB-NEXT:    sub a5, a4, t4
@@ -1898,10 +1898,10 @@ define i128 @abd_cmp_i128(i128 %a, i128 %b) nounwind {
 ; RV32ZBB-NEXT:    sub a2, a3, a2
 ; RV32ZBB-NEXT:    j .LBB22_11
 ; RV32ZBB-NEXT:  .LBB22_10:
-; RV32ZBB-NEXT:    sub a7, t0, a7
+; RV32ZBB-NEXT:    sub a7, t1, a7
 ; RV32ZBB-NEXT:    sub a5, a6, a5
 ; RV32ZBB-NEXT:    sub a4, a1, a4
-; RV32ZBB-NEXT:    sub a6, a7, t1
+; RV32ZBB-NEXT:    sub a6, a7, t0
 ; RV32ZBB-NEXT:    sltu a7, a5, t3
 ; RV32ZBB-NEXT:    sub a1, a5, t3
 ; RV32ZBB-NEXT:    sub a5, a4, t2
diff --git a/llvm/test/CodeGen/RISCV/add-before-shl.ll b/llvm/test/CodeGen/RISCV/add-before-shl.ll
index b6ff3c9060af5..35a39b89a2cb7 100644
--- a/llvm/test/CodeGen/RISCV/add-before-shl.ll
+++ b/llvm/test/CodeGen/RISCV/add-before-shl.ll
@@ -200,26 +200,26 @@ define i128 @add_wide_operand(i128 %a) nounwind {
 ;
 ; RV32C-LABEL: add_wide_operand:
 ; RV32C:       # %bb.0:
+; RV32C-NEXT:    c.lw a2, 0(a1)
 ; RV32C-NEXT:    c.lw a4, 12(a1)
-; RV32C-NEXT:    c.lw a3, 0(a1)
-; RV32C-NEXT:    c.lw a2, 4(a1)
+; RV32C-NEXT:    c.lw a3, 4(a1)
 ; RV32C-NEXT:    c.lw a1, 8(a1)
 ; RV32C-NEXT:    c.lui a5, 16
 ; RV32C-NEXT:    add a6, a4, a5
-; RV32C-NEXT:    srli a5, a3, 29
-; RV32C-NEXT:    slli a4, a2, 3
+; RV32C-NEXT:    srli a5, a2, 29
+; RV32C-NEXT:    slli a4, a3, 3
 ; RV32C-NEXT:    c.or a4, a5
 ; RV32C-NEXT:    srli a5, a1, 29
-; RV32C-NEXT:    c.srli a2, 29
+; RV32C-NEXT:    c.srli a3, 29
 ; RV32C-NEXT:    c.slli a1, 3
-; RV32C-NEXT:    c.slli a3, 3
+; RV32C-NEXT:    c.slli a2, 3
 ; RV32C-NEXT:    c.slli a6, 3
-; RV32C-NEXT:    c.or a1, a2
-; RV32C-NEXT:    or a2, a6, a5
-; RV32C-NEXT:    c.sw a3, 0(a0)
+; RV32C-NEXT:    c.or a1, a3
+; RV32C-NEXT:    or a3, a6, a5
+; RV32C-NEXT:    c.sw a2, 0(a0)
 ; RV32C-NEXT:    c.sw a4, 4(a0)
 ; RV32C-NEXT:    c.sw a1, 8(a0)
-; RV32C-NEXT:    c.sw a2, 12(a0)
+; RV32C-NEXT:    c.sw a3, 12(a0)
 ; RV32C-NEXT:    c.jr ra
 ;
 ; RV64C-LABEL: add_wide_operand:
diff --git a/llvm/test/CodeGen/RISCV/fold-mem-offset.ll b/llvm/test/CodeGen/RISCV/fold-mem-offset.ll
index f4072ffa1e3df..e6f1a08cc1879 100644
--- a/llvm/test/CodeGen/RISCV/fold-mem-offset.ll
+++ b/llvm/test/CodeGen/RISCV/fold-mem-offset.ll
@@ -213,12 +213,12 @@ define i64 @test_sh3add_uw(ptr %p, i32 signext %x, i32 signext %y) {
 ; RV32I-NEXT:    slli a2, a2, 3
 ; RV32I-NEXT:    add a1, a0, a1
 ; RV32I-NEXT:    add a0, a0, a2
-; RV32I-NEXT:    lw a2, 404(a0)
-; RV32I-NEXT:    lw a3, 400(a1)
+; RV32I-NEXT:    lw a2, 400(a1)
 ; RV32I-NEXT:    lw a1, 404(a1)
+; RV32I-NEXT:    lw a3, 404(a0)
 ; RV32I-NEXT:    lw a4, 400(a0)
-; RV32I-NEXT:    add a1, a2, a1
-; RV32I-NEXT:    add a0, a4, a3
+; RV32I-NEXT:    add a1, a3, a1
+; RV32I-NEXT:    add a0, a4, a2
 ; RV32I-NEXT:    sltu a2, a0, a4
 ; RV32I-NEXT:    add a1, a1, a2
 ; RV32I-NEXT:    ret
@@ -240,12 +240,12 @@ define i64 @test_sh3add_uw(ptr %p, i32 signext %x, i32 signext %y) {
 ; RV32ZBA:       # %bb.0: # %entry
 ; RV32ZBA-NEXT:    sh3add a1, a1, a0
 ; RV32ZBA-NEXT:    sh3add a0, a2, a0
-; RV32ZBA-NEXT:    lw a2, 404(a0)
-; RV32ZBA-NEXT:    lw a3, 400(a1)
+; RV32ZBA-NEXT:    lw a2, 400(a1)
 ; RV32ZBA-NEXT:    lw a1, 404(a1)
+; RV32ZBA-NEXT:    lw a3, 404(a0)
 ; RV32ZBA-NEXT:    lw a4, 400(a0)
-; RV32ZBA-NEXT:    add a1, a2, a1
-; RV32ZBA-NEXT:    add a0, a4, a3
+; RV32ZBA-NEXT:    add a1, a3, a1
+; RV32ZBA-NEXT:    add a0, a4, a2
 ; RV32ZBA-NEXT:    sltu a2, a0, a4
 ; RV32ZBA-NEXT:    add a1, a1, a2
 ; RV32ZBA-NEXT:    ret
diff --git a/llvm/test/CodeGen/RISCV/legalize-fneg.ll b/llvm/test/CodeGen/RISCV/legalize-fneg.ll
index f60b77b92c09e..9e66eb7a2ae6c 100644
--- a/llvm/test/CodeGen/RISCV/legalize-fneg.ll
+++ b/llvm/test/CodeGen/RISCV/legalize-fneg.ll
@@ -56,16 +56,16 @@ entry:
 define void @test3(ptr %a, ptr %b) nounwind {
 ; RV32-LABEL: test3:
 ; RV32:       # %bb.0: # %entry
-; RV32-NEXT:    lw a2, 12(a1)
-; RV32-NEXT:    lw a3, 0(a1)
+; RV32-NEXT:    lw a2, 0(a1)
+; RV32-NEXT:    lw a3, 12(a1)
 ; RV32-NEXT:    lw a4, 4(a1)
 ; RV32-NEXT:    lw a1, 8(a1)
 ; RV32-NEXT:    lui a5, 524288
-; RV32-NEXT:    xor a2, a2, a5
-; RV32-NEXT:    sw a3, 0(a0)
+; RV32-NEXT:    xor a3, a3, a5
+; RV32-NEXT:    sw a2, 0(a0)
 ; RV32-NEXT:    sw a4, 4(a0)
 ; RV32-NEXT:    sw a1, 8(a0)
-; RV32-NEXT:    sw a2, 12(a0)
+; RV32-NEXT:    sw a3, 12(a0)
 ; RV32-NEXT:    ret
 ;
 ; RV64-LABEL: test3:
diff --git a/llvm/test/CodeGen/RISCV/memcmp-optsize.ll b/llvm/test/CodeGen/RISCV/memcmp-optsize.ll
index fc714e3faef43..38cd51c074594 100644
--- a/llvm/test/CodeGen/RISCV/memcmp-optsize.ll
+++ b/llvm/test/CodeGen/RISCV/memcmp-optsize.ll
@@ -4411,25 +4411,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind optsize {
 ; CHECK-ALIGNED-RV32-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV32:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV32-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV32-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV32-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV32-NEXT:    lbu a1, 3(a1)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a4, 3(a1)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV32-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV32-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV32-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV32-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV32-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV32-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV32-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV32-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV32-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV32-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV32-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV32-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV32-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV32-NEXT:    lbu a0, 3(a0)
-; CHECK-ALIGNED-RV32-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV32-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV32-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV32-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV32-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV32-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV32-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV32-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV32-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV32-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV32-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV32-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV32-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV32-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV32-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV32-NEXT:    ret
@@ -4437,25 +4437,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind optsize {
 ; CHECK-ALIGNED-RV64-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV64:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV64-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV64-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV64-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV64-NEXT:    lb a1, 3(a1)
+; CHECK-ALIGNED-RV64-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV64-NEXT:    lb a4, 3(a1)
+; CHECK-ALIGNED-RV64-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV64-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV64-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV64-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV64-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV64-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV64-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV64-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV64-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV64-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV64-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV64-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV64-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV64-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV64-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV64-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV64-NEXT:    lb a0, 3(a0)
-; CHECK-ALIGNED-RV64-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV64-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV64-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV64-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV64-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV64-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV64-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV64-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV64-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV64-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV64-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV64-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV64-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV64-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV64-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV64-NEXT:    ret
@@ -4463,25 +4463,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind optsize {
 ; CHECK-ALIGNED-RV32-ZBB-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV32-ZBB:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a1, 3(a1)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a4, 3(a1)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a0, 3(a0)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    ret
@@ -4489,25 +4489,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind optsize {
 ; CHECK-ALIGNED-RV64-ZBB-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV64-ZBB:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lb a1, 3(a1)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lb a4, 3(a1)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    lb a0, 3(a0)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    ret
@@ -4559,25 +4559,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind optsize {
 ; CHECK-ALIGNED-RV32-V-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV32-V:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV32-V-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a1, 3(a1)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a4, 3(a1)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV32-V-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV32-V-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV32-V-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV32-V-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV32-V-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV32-V-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV32-V-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV32-V-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV32-V-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV32-V-NEXT:    lbu a0, 3(a0)
-; CHECK-ALIGNED-RV32-V-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV32-V-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV32-V-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV32-V-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV32-V-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV32-V-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV32-V-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV32-V-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV32-V-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV32-V-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV32-V-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV32-V-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV32-V-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV32-V-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV32-V-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV32-V-NEXT:    ret
@@ -4585,25 +4585,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind optsize {
 ; CHECK-ALIGNED-RV64-V-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV64-V:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV64-V-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV64-V-NEXT:    lb a1, 3(a1)
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV64-V-NEXT:    lb a4, 3(a1)
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV64-V-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV64-V-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV64-V-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV64-V-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV64-V-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV64-V-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV64-V-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV64-V-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV64-V-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV64-V-NEXT:    lb a0, 3(a0)
-; CHECK-ALIGNED-RV64-V-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV64-V-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV64-V-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV64-V-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV64-V-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV64-V-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV64-V-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV64-V-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV64-V-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV64-V-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV64-V-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV64-V-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV64-V-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV64-V-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV64-V-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV64-V-NEXT:    ret
diff --git a/llvm/test/CodeGen/RISCV/memcmp.ll b/llvm/test/CodeGen/RISCV/memcmp.ll
index ddfbd649a43b8..df9d781a4536d 100644
--- a/llvm/test/CodeGen/RISCV/memcmp.ll
+++ b/llvm/test/CodeGen/RISCV/memcmp.ll
@@ -5981,25 +5981,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind {
 ; CHECK-ALIGNED-RV32-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV32:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV32-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV32-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV32-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV32-NEXT:    lbu a1, 3(a1)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a4, 3(a1)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV32-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV32-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV32-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV32-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV32-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV32-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV32-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV32-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV32-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV32-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV32-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV32-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV32-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV32-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV32-NEXT:    lbu a0, 3(a0)
-; CHECK-ALIGNED-RV32-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV32-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV32-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV32-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV32-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV32-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV32-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV32-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV32-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV32-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV32-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV32-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV32-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV32-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV32-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV32-NEXT:    ret
@@ -6007,25 +6007,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind {
 ; CHECK-ALIGNED-RV64-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV64:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV64-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV64-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV64-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV64-NEXT:    lb a1, 3(a1)
+; CHECK-ALIGNED-RV64-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV64-NEXT:    lb a4, 3(a1)
+; CHECK-ALIGNED-RV64-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV64-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV64-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV64-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV64-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV64-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV64-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV64-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV64-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV64-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV64-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV64-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV64-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV64-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV64-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV64-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV64-NEXT:    lb a0, 3(a0)
-; CHECK-ALIGNED-RV64-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV64-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV64-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV64-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV64-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV64-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV64-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV64-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV64-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV64-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV64-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV64-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV64-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV64-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV64-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV64-NEXT:    ret
@@ -6033,25 +6033,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind {
 ; CHECK-ALIGNED-RV32-ZBB-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV32-ZBB:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a1, 3(a1)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a4, 3(a1)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    lbu a0, 3(a0)
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV32-ZBB-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV32-ZBB-NEXT:    ret
@@ -6059,25 +6059,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind {
 ; CHECK-ALIGNED-RV64-ZBB-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV64-ZBB:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lb a1, 3(a1)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lb a4, 3(a1)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    lb a0, 3(a0)
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV64-ZBB-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV64-ZBB-NEXT:    ret
@@ -6129,25 +6129,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind {
 ; CHECK-ALIGNED-RV32-V-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV32-V:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV32-V-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a1, 3(a1)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a4, 3(a1)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV32-V-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV32-V-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV32-V-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV32-V-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV32-V-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV32-V-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV32-V-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV32-V-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV32-V-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV32-V-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV32-V-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV32-V-NEXT:    lbu a0, 3(a0)
-; CHECK-ALIGNED-RV32-V-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV32-V-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV32-V-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV32-V-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV32-V-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV32-V-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV32-V-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV32-V-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV32-V-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV32-V-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV32-V-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV32-V-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV32-V-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV32-V-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV32-V-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV32-V-NEXT:    ret
@@ -6155,25 +6155,25 @@ define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind {
 ; CHECK-ALIGNED-RV64-V-LABEL: memcmp_eq_zero:
 ; CHECK-ALIGNED-RV64-V:       # %bb.0: # %entry
 ; CHECK-ALIGNED-RV64-V-NEXT:    lbu a2, 1(a1)
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a3, 0(a1)
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a4, 2(a1)
-; CHECK-ALIGNED-RV64-V-NEXT:    lb a1, 3(a1)
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a3, 2(a1)
+; CHECK-ALIGNED-RV64-V-NEXT:    lb a4, 3(a1)
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a1, 0(a1)
 ; CHECK-ALIGNED-RV64-V-NEXT:    slli a2, a2, 8
-; CHECK-ALIGNED-RV64-V-NEXT:    or a2, a2, a3
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a3, 0(a0)
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a5, 1(a0)
-; CHECK-ALIGNED-RV64-V-NEXT:    slli a4, a4, 16
-; CHECK-ALIGNED-RV64-V-NEXT:    slli a1, a1, 24
-; CHECK-ALIGNED-RV64-V-NEXT:    or a1, a1, a4
-; CHECK-ALIGNED-RV64-V-NEXT:    lbu a4, 2(a0)
+; CHECK-ALIGNED-RV64-V-NEXT:    slli a3, a3, 16
+; CHECK-ALIGNED-RV64-V-NEXT:    slli a4, a4, 24
+; CHECK-ALIGNED-RV64-V-NEXT:    or a1, a2, a1
+; CHECK-ALIGNED-RV64-V-NEXT:    or a3, a4, a3
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a2, 1(a0)
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a4, 0(a0)
+; CHECK-ALIGNED-RV64-V-NEXT:    lbu a5, 2(a0)
 ; CHECK-ALIGNED-RV64-V-NEXT:    lb a0, 3(a0)
-; CHECK-ALIGNED-RV64-V-NEXT:    slli a5, a5, 8
-; CHECK-ALIGNED-RV64-V-NEXT:    or a3, a5, a3
-; CHECK-ALIGNED-RV64-V-NEXT:    slli a4, a4, 16
+; CHECK-ALIGNED-RV64-V-NEXT:    slli a2, a2, 8
+; CHECK-ALIGNED-RV64-V-NEXT:    or a2, a2, a4
+; CHECK-ALIGNED-RV64-V-NEXT:    slli a5, a5, 16
 ; CHECK-ALIGNED-RV64-V-NEXT:    slli a0, a0, 24
-; CHECK-ALIGNED-RV64-V-NEXT:    or a0, a0, a4
-; CHECK-ALIGNED-RV64-V-NEXT:    or a1, a1, a2
-; CHECK-ALIGNED-RV64-V-NEXT:    or a0, a0, a3
+; CHECK-ALIGNED-RV64-V-NEXT:    or a0, a0, a5
+; CHECK-ALIGNED-RV64-V-NEXT:    or a1, a3, a1
+; CHECK-ALIGNED-RV64-V-NEXT:    or a0, a0, a2
 ; CHECK-ALIGNED-RV64-V-NEXT:    xor a0, a0, a1
 ; CHECK-ALIGNED-RV64-V-NEXT:    seqz a0, a0
 ; CHECK-ALIGNED-RV64-V-NEXT:    ret
diff --git a/llvm/test/CodeGen/RISCV/rv32zbb.ll b/llvm/test/CodeGen/RISCV/rv32zbb.ll
index 62bc7b3336a5c..8dd63015971d0 100644
--- a/llvm/test/CodeGen/RISCV/rv32zbb.ll
+++ b/llvm/test/CodeGen/RISCV/rv32zbb.ll
@@ -1590,8 +1590,8 @@ define i128 @sub_if_uge_i128(i128 %x, i128 %y) {
 ; CHECK-NEXT:    lw a7, 4(a2)
 ; CHECK-NEXT:    lw a6, 8(a2)
 ; CHECK-NEXT:    lw t0, 12(a2)
-; CHECK-NEXT:    lw a4, 12(a1)
 ; CHECK-NEXT:    lw a3, 4(a1)
+; CHECK-NEXT:    lw a4, 12(a1)
 ; CHECK-NEXT:    lw a5, 8(a1)
 ; CHECK-NEXT:    beq a4, t0, .LBB53_2
 ; CHECK-NEXT:  # %bb.1:
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-elen.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-elen.ll
index e13f4f4b50b0f..651894a1bb661 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-elen.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-elen.ll
@@ -26,26 +26,26 @@ define void @add_v4i32(ptr %x, ptr %y) {
 define void @add_v2i64(ptr %x, ptr %y) {
 ; RV32-LABEL: add_v2i64:
 ; RV32:       # %bb.0:
-; RV32-NEXT:    lw a2, 0(a1)
-; RV32-NEXT:    lw a3, 4(a1)
-; RV32-NEXT:    lw a4, 0(a0)
-; RV32-NEXT:    lw a5, 4(a0)
-; RV32-NEXT:    lw a6, 8(a0)
-; RV32-NEXT:    lw a7, 12(a0)
+; RV32-NEXT:    lw a2, 0(a0)
+; RV32-NEXT:    lw a3, 4(a0)
+; RV32-NEXT:    lw a4, 8(a0)
+; RV32-NEXT:    lw a5, 12(a0)
+; RV32-NEXT:    lw a6, 4(a1)
+; RV32-NEXT:    lw a7, 0(a1)
 ; RV32-NEXT:    lw t0, 12(a1)
 ; RV32-NEXT:    lw a1, 8(a1)
-; RV32-NEXT:    add a3, a5, a3
-; RV32-NEXT:    add a2, a4, a2
-; RV32-NEXT:    add a7, a7, t0
-; RV32-NEXT:    add a1, a6, a1
-; RV32-NEXT:    sltu a4, a2, a4
-; RV32-NEXT:    sltu a5, a1, a6
-; RV32-NEXT:    add a3, a3, a4
-; RV32-NEXT:    add a5, a7, a5
-; RV32-NEXT:    sw a2, 0(a0)
-; RV32-NEXT:    sw a3, 4(a0)
+; RV32-NEXT:    add a3, a3, a6
+; RV32-NEXT:    add a7, a2, a7
+; RV32-NEXT:    add a5, a5, t0
+; RV32-NEXT:    add a1, a4, a1
+; RV32-NEXT:    sltu a2, a7, a2
+; RV32-NEXT:    sltu a4, a1, a4
+; RV32-NEXT:    add a2, a3, a2
+; RV32-NEXT:    add a4, a5, a4
+; RV32-NEXT:    sw a7, 0(a0)
+; RV32-NEXT:    sw a2, 4(a0)
 ; RV32-NEXT:    sw a1, 8(a0)
-; RV32-NEXT:    sw a5, 12(a0)
+; RV32-NEXT:    sw a4, 12(a0)
 ; RV32-NEXT:    ret
 ;
 ; RV64-LABEL: add_v2i64:
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
index a5eba788feeff..1fa96d3c07ca9 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
@@ -1406,28 +1406,28 @@ define <16 x i8> @buildvec_v16i8_loads_contigous(ptr %p) {
 ; RV32VB-NEXT:    slli a7, a7, 16
 ; RV32VB-NEXT:    slli t0, t0, 24
 ; RV32VB-NEXT:    slli a5, a5, 8
-; RV32VB-NEXT:    or a7, t0, a7
-; RV32VB-NEXT:    or a4, a4, a5
-; RV32VB-NEXT:    lbu a5, 12(a0)
-; RV32VB-NEXT:    lbu t0, 13(a0)
 ; RV32VB-NEXT:    slli a6, a6, 16
 ; RV32VB-NEXT:    slli t1, t1, 24
-; RV32VB-NEXT:    or a6, t1, a6
+; RV32VB-NEXT:    or a7, t0, a7
+; RV32VB-NEXT:    or a4, a4, a5
+; RV32VB-NEXT:    or a5, t1, a6
+; RV32VB-NEXT:    lbu a6, 13(a0)
+; RV32VB-NEXT:    lbu t0, 12(a0)
 ; RV32VB-NEXT:    lbu t1, 14(a0)
 ; RV32VB-NEXT:    lbu a0, 15(a0)
-; RV32VB-NEXT:    slli t0, t0, 8
-; RV32VB-NEXT:    or a5, a5, t0
+; RV32VB-NEXT:    slli a6, a6, 8
+; RV32VB-NEXT:    or a6, t0, a6
 ; RV32VB-NEXT:    slli t1, t1, 16
 ; RV32VB-NEXT:    slli a0, a0, 24
 ; RV32VB-NEXT:    or a0, a0, t1
 ; RV32VB-NEXT:    or a1, a1, a3
 ; RV32VB-NEXT:    or a2, a2, a7
-; RV32VB-NEXT:    or a3, a4, a6
-; RV32VB-NEXT:    or a0, a5, a0
+; RV32VB-NEXT:    or a4, a4, a5
+; RV32VB-NEXT:    or a0, a6, a0
 ; RV32VB-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
 ; RV32VB-NEXT:    vmv.v.x v8, a1
 ; RV32VB-NEXT:    vslide1down.vx v8, v8, a2
-; RV32VB-NEXT:    vslide1down.vx v8, v8, a3
+; RV32VB-NEXT:    vslide1down.vx v8, v8, a4
 ; RV32VB-NEXT:    vslide1down.vx v8, v8, a0
 ; RV32VB-NEXT:    ret
 ;
@@ -1767,36 +1767,36 @@ define <16 x i8> @buildvec_v16i8_loads_gather(ptr %p) {
 ; RV32VB-NEXT:    lbu a7, 55(a0)
 ; RV32VB-NEXT:    lbu t0, 75(a0)
 ; RV32VB-NEXT:    lbu t1, 82(a0)
-; RV32VB-NEXT:    lbu t2, 154(a0)
-; RV32VB-NEXT:    lbu t3, 161(a0)
 ; RV32VB-NEXT:    slli a2, a2, 8
 ; RV32VB-NEXT:    slli a3, a3, 16
 ; RV32VB-NEXT:    slli a4, a4, 24
 ; RV32VB-NEXT:    slli a7, a7, 8
-; RV32VB-NEXT:    slli a5, a5, 16
-; RV32VB-NEXT:    slli t0, t0, 24
 ; RV32VB-NEXT:    or a1, a1, a2
 ; RV32VB-NEXT:    or a3, a4, a3
 ; RV32VB-NEXT:    or a2, a6, a7
-; RV32VB-NEXT:    or a4, t0, a5
-; RV32VB-NEXT:    lbu a5, 93(a0)
+; RV32VB-NEXT:    lbu a4, 93(a0)
 ; RV32VB-NEXT:    lbu a6, 105(a0)
 ; RV32VB-NEXT:    lbu a7, 124(a0)
-; RV32VB-NEXT:    lbu t0, 144(a0)
-; RV32VB-NEXT:    slli a5, a5, 8
-; RV32VB-NEXT:    or a5, t1, a5
+; RV32VB-NEXT:    lbu t2, 144(a0)
+; RV32VB-NEXT:    slli a5, a5, 16
+; RV32VB-NEXT:    slli t0, t0, 24
+; RV32VB-NEXT:    slli a4, a4, 8
+; RV32VB-NEXT:    or a5, t0, a5
+; RV32VB-NEXT:    or a4, t1, a4
+; RV32VB-NEXT:    lbu t0, 161(a0)
+; RV32VB-NEXT:    lbu t1, 154(a0)
 ; RV32VB-NEXT:    lbu a0, 163(a0)
 ; RV32VB-NEXT:    slli a6, a6, 16
-; RV32VB-NEXT:    slli t3, t3, 24
-; RV32VB-NEXT:    or a6, t3, a6
+; RV32VB-NEXT:    slli t0, t0, 24
+; RV32VB-NEXT:    or a6, t0, a6
 ; RV32VB-NEXT:    slli a0, a0, 8
 ; RV32VB-NEXT:    or a0, a7, a0
-; RV32VB-NEXT:    slli t0, t0, 16
-; RV32VB-NEXT:    slli t2, t2, 24
-; RV32VB-NEXT:    or a7, t2, t0
+; RV32VB-NEXT:    slli t2, t2, 16
+; RV32VB-NEXT:    slli t1, t1, 24
+; RV32VB-NEXT:    or a7, t1, t2
 ; RV32VB-NEXT:    or a1, a1, a3
-; RV32VB-NEXT:    or a2, a2, a4
-; RV32VB-NEXT:    or a3, a5, a6
+; RV32VB-NEXT:    or a2, a2, a5
+; RV32VB-NEXT:    or a3, a4, a6
 ; RV32VB-NEXT:    or a0, a0, a7
 ; RV32VB-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
 ; RV32VB-NEXT:    vmv.v.x v8, a1
@@ -1893,50 +1893,50 @@ define <16 x i8> @buildvec_v16i8_loads_gather(ptr %p) {
 ;
 ; RVA22U64-LABEL: buildvec_v16i8_loads_gather:
 ; RVA22U64:       # %bb.0:
-; RVA22U64-NEXT:    lbu a6, 0(a0)
-; RVA22U64-NEXT:    lbu a7, 1(a0)
-; RVA22U64-NEXT:    lbu t0, 22(a0)
-; RVA22U64-NEXT:    lbu t1, 31(a0)
-; RVA22U64-NEXT:    lbu t2, 623(a0)
-; RVA22U64-NEXT:    lbu a1, 44(a0)
-; RVA22U64-NEXT:    lbu a2, 55(a0)
-; RVA22U64-NEXT:    lbu a3, 75(a0)
-; RVA22U64-NEXT:    lbu t3, 82(a0)
-; RVA22U64-NEXT:    lbu t4, 154(a0)
-; RVA22U64-NEXT:    lbu t5, 161(a0)
-; RVA22U64-NEXT:    slli a7, a7, 8
-; RVA22U64-NEXT:    slli t0, t0, 16
-; RVA22U64-NEXT:    slli t1, t1, 24
-; RVA22U64-NEXT:    slli a1, a1, 32
-; RVA22U64-NEXT:    slli a2, a2, 40
-; RVA22U64-NEXT:    slli t2, t2, 48
-; RVA22U64-NEXT:    slli a3, a3, 56
-; RVA22U64-NEXT:    or a6, a6, a7
-; RVA22U64-NEXT:    or t0, t1, t0
-; RVA22U64-NEXT:    or a7, a2, a1
-; RVA22U64-NEXT:    or a2, a3, t2
-; RVA22U64-NEXT:    lbu a3, 93(a0)
-; RVA22U64-NEXT:    lbu a5, 105(a0)
-; RVA22U64-NEXT:    lbu a1, 124(a0)
-; RVA22U64-NEXT:    lbu a4, 144(a0)
-; RVA22U64-NEXT:    slli a3, a3, 8
-; RVA22U64-NEXT:    or a3, t3, a3
+; RVA22U64-NEXT:    lbu a7, 0(a0)
+; RVA22U64-NEXT:    lbu a2, 1(a0)
+; RVA22U64-NEXT:    lbu a3, 22(a0)
+; RVA22U64-NEXT:    lbu a4, 31(a0)
+; RVA22U64-NEXT:    lbu a6, 623(a0)
+; RVA22U64-NEXT:    lbu a5, 44(a0)
+; RVA22U64-NEXT:    lbu a1, 55(a0)
+; RVA22U64-NEXT:    lbu t0, 75(a0)
+; RVA22U64-NEXT:    lbu t1, 82(a0)
+; RVA22U64-NEXT:    slli a2, a2, 8
+; RVA22U64-NEXT:    slli a3, a3, 16
+; RVA22U64-NEXT:    slli a4, a4, 24
+; RVA22U64-NEXT:    slli a5, a5, 32
+; RVA22U64-NEXT:    slli a1, a1, 40
+; RVA22U64-NEXT:    or a7, a7, a2
+; RVA22U64-NEXT:    or t3, a4, a3
+; RVA22U64-NEXT:    or t2, a1, a5
+; RVA22U64-NEXT:    lbu a4, 93(a0)
+; RVA22U64-NEXT:    lbu t4, 105(a0)
+; RVA22U64-NEXT:    lbu a2, 124(a0)
+; RVA22U64-NEXT:    lbu t5, 144(a0)
+; RVA22U64-NEXT:    slli a6, a6, 48
+; RVA22U64-NEXT:    slli t0, t0, 56
+; RVA22U64-NEXT:    slli a4, a4, 8
+; RVA22U64-NEXT:    or a3, t0, a6
+; RVA22U64-NEXT:    or a4, t1, a4
+; RVA22U64-NEXT:    lbu a5, 161(a0)
+; RVA22U64-NEXT:    lbu a1, 154(a0)
 ; RVA22U64-NEXT:    lbu a0, 163(a0)
-; RVA22U64-NEXT:    slli a5, a5, 16
-; RVA22U64-NEXT:    slli t5, t5, 24
-; RVA22U64-NEXT:    or a5, t5, a5
-; RVA22U64-NEXT:    slli a1, a1, 32
+; RVA22U64-NEXT:    slli t4, t4, 16
+; RVA22U64-NEXT:    slli a5, a5, 24
+; RVA22U64-NEXT:    or a5, a5, t4
+; RVA22U64-NEXT:    slli a2, a2, 32
 ; RVA22U64-NEXT:    slli a0, a0, 40
+; RVA22U64-NEXT:    or a0, a0, a2
+; RVA22U64-NEXT:    slli t5, t5, 48
+; RVA22U64-NEXT:    slli a1, a1, 56
+; RVA22U64-NEXT:    or a1, a1, t5
+; RVA22U64-NEXT:    or a2, a7, t3
+; RVA22U64-NEXT:    or a3, a3, t2
+; RVA22U64-NEXT:    or a4, a4, a5
 ; RVA22U64-NEXT:    or a0, a0, a1
-; RVA22U64-NEXT:    slli a4, a4, 48
-; RVA22U64-NEXT:    slli t4, t4, 56
-; RVA22U64-NEXT:    or a1, t4, a4
-; RVA22U64-NEXT:    or a4, a6, t0
-; RVA22U64-NEXT:    or a2, a2, a7
-; RVA22U64-NEXT:    or a3, a3, a5
-; RVA22U64-NEXT:    or a0, a0, a1
-; RVA22U64-NEXT:    or a2, a2, a4
-; RVA22U64-NEXT:    or a0, a0, a3
+; RVA22U64-NEXT:    or a2, a2, a3
+; RVA22U64-NEXT:    or a0, a0, a4
 ; RVA22U64-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
 ; RVA22U64-NEXT:    vmv.v.x v8, a2
 ; RVA22U64-NEXT:    vslide1down.vx v8, v8, a0
@@ -2111,24 +2111,24 @@ define <16 x i8> @buildvec_v16i8_undef_low_half(ptr %p) {
 ;
 ; RV32VB-LABEL: buildvec_v16i8_undef_low_half:
 ; RV32VB:       # %bb.0:
-; RV32VB-NEXT:    lbu a1, 82(a0)
-; RV32VB-NEXT:    lbu a2, 93(a0)
-; RV32VB-NEXT:    lbu a3, 144(a0)
-; RV32VB-NEXT:    lbu a4, 154(a0)
-; RV32VB-NEXT:    lbu a5, 161(a0)
-; RV32VB-NEXT:    lbu a6, 105(a0)
-; RV32VB-NEXT:    lbu a7, 124(a0)
-; RV32VB-NEXT:    slli a2, a2, 8
-; RV32VB-NEXT:    or a1, a1, a2
+; RV32VB-NEXT:    lbu a1, 93(a0)
+; RV32VB-NEXT:    lbu a2, 82(a0)
+; RV32VB-NEXT:    lbu a3, 105(a0)
+; RV32VB-NEXT:    lbu a4, 124(a0)
+; RV32VB-NEXT:    slli a1, a1, 8
+; RV32VB-NEXT:    or a1, a2, a1
+; RV32VB-NEXT:    lbu a2, 161(a0)
+; RV32VB-NEXT:    lbu a5, 144(a0)
+; RV32VB-NEXT:    lbu a6, 154(a0)
 ; RV32VB-NEXT:    lbu a0, 163(a0)
-; RV32VB-NEXT:    slli a6, a6, 16
-; RV32VB-NEXT:    slli a5, a5, 24
-; RV32VB-NEXT:    or a2, a5, a6
-; RV32VB-NEXT:    slli a0, a0, 8
-; RV32VB-NEXT:    or a0, a7, a0
 ; RV32VB-NEXT:    slli a3, a3, 16
-; RV32VB-NEXT:    slli a4, a4, 24
-; RV32VB-NEXT:    or a3, a4, a3
+; RV32VB-NEXT:    slli a2, a2, 24
+; RV32VB-NEXT:    or a2, a2, a3
+; RV32VB-NEXT:    slli a0, a0, 8
+; RV32VB-NEXT:    or a0, a4, a0
+; RV32VB-NEXT:    slli a5, a5, 16
+; RV32VB-NEXT:    slli a6, a6, 24
+; RV32VB-NEXT:    or a3, a6, a5
 ; RV32VB-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
 ; RV32VB-NEXT:    vmv.v.i v8, 0
 ; RV32VB-NEXT:    or a1, a1, a2
@@ -2186,28 +2186,28 @@ define <16 x i8> @buildvec_v16i8_undef_low_half(ptr %p) {
 ;
 ; RVA22U64-LABEL: buildvec_v16i8_undef_low_half:
 ; RVA22U64:       # %bb.0:
-; RVA22U64-NEXT:    lbu a6, 82(a0)
-; RVA22U64-NEXT:    lbu a7, 93(a0)
-; RVA22U64-NEXT:    lbu t0, 144(a0)
-; RVA22U64-NEXT:    lbu a4, 154(a0)
-; RVA22U64-NEXT:    lbu a5, 161(a0)
-; RVA22U64-NEXT:    lbu a1, 105(a0)
-; RVA22U64-NEXT:    lbu a2, 124(a0)
-; RVA22U64-NEXT:    slli a7, a7, 8
-; RVA22U64-NEXT:    or a3, a6, a7
+; RVA22U64-NEXT:    lbu a1, 93(a0)
+; RVA22U64-NEXT:    lbu a2, 82(a0)
+; RVA22U64-NEXT:    lbu a3, 105(a0)
+; RVA22U64-NEXT:    lbu a4, 124(a0)
+; RVA22U64-NEXT:    slli a1, a1, 8
+; RVA22U64-NEXT:    or a6, a2, a1
+; RVA22U64-NEXT:    lbu a2, 161(a0)
+; RVA22U64-NEXT:    lbu a5, 144(a0)
+; RVA22U64-NEXT:    lbu a1, 154(a0)
 ; RVA22U64-NEXT:    lbu a0, 163(a0)
-; RVA22U64-NEXT:    slli a1, a1, 16
-; RVA22U64-NEXT:    slli a5, a5, 24
-; RVA22U64-NEXT:    or a1, a1, a5
-; RVA22U64-NEXT:    slli a2, a2, 32
+; RVA22U64-NEXT:    slli a3, a3, 16
+; RVA22U64-NEXT:    slli a2, a2, 24
+; RVA22U64-NEXT:    or a2, a2, a3
+; RVA22U64-NEXT:    slli a4, a4, 32
 ; RVA22U64-NEXT:    slli a0, a0, 40
-; RVA22U64-NEXT:    or a0, a0, a2
-; RVA22U64-NEXT:    slli t0, t0, 48
-; RVA22U64-NEXT:    slli a4, a4, 56
-; RVA22U64-NEXT:    or a2, a4, t0
-; RVA22U64-NEXT:    or a1, a1, a3
-; RVA22U64-NEXT:    or a0, a0, a2
+; RVA22U64-NEXT:    or a0, a0, a4
+; RVA22U64-NEXT:    slli a5, a5, 48
+; RVA22U64-NEXT:    slli a1, a1, 56
+; RVA22U64-NEXT:    or a1, a1, a5
+; RVA22U64-NEXT:    or a2, a6, a2
 ; RVA22U64-NEXT:    or a0, a0, a1
+; RVA22U64-NEXT:    or a0, a0, a2
 ; RVA22U64-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
 ; RVA22U64-NEXT:    vmv.v.i v8, 0
 ; RVA22U64-NEXT:    vslide1down.vx v8, v8, a0
@@ -2520,30 +2520,30 @@ define <16 x i8> @buildvec_v16i8_undef_edges(ptr %p) {
 ;
 ; RV32VB-LABEL: buildvec_v16i8_undef_edges:
 ; RV32VB:       # %bb.0:
-; RV32VB-NEXT:    lbu a1, 55(a0)
-; RV32VB-NEXT:    lbu a2, 31(a0)
-; RV32VB-NEXT:    lbu a3, 44(a0)
-; RV32VB-NEXT:    lbu a4, 623(a0)
-; RV32VB-NEXT:    lbu a5, 75(a0)
-; RV32VB-NEXT:    slli a1, a1, 8
-; RV32VB-NEXT:    or a1, a3, a1
-; RV32VB-NEXT:    lbu a3, 82(a0)
-; RV32VB-NEXT:    lbu a6, 93(a0)
-; RV32VB-NEXT:    slli a4, a4, 16
-; RV32VB-NEXT:    slli a5, a5, 24
-; RV32VB-NEXT:    or a4, a5, a4
-; RV32VB-NEXT:    lbu a5, 105(a0)
+; RV32VB-NEXT:    lbu a1, 623(a0)
+; RV32VB-NEXT:    lbu a2, 55(a0)
+; RV32VB-NEXT:    lbu a3, 31(a0)
+; RV32VB-NEXT:    lbu a4, 75(a0)
+; RV32VB-NEXT:    lbu a5, 44(a0)
+; RV32VB-NEXT:    slli a2, a2, 8
+; RV32VB-NEXT:    slli a1, a1, 16
+; RV32VB-NEXT:    slli a4, a4, 24
+; RV32VB-NEXT:    or a2, a5, a2
+; RV32VB-NEXT:    or a1, a4, a1
+; RV32VB-NEXT:    lbu a4, 93(a0)
+; RV32VB-NEXT:    lbu a5, 82(a0)
+; RV32VB-NEXT:    lbu a6, 105(a0)
 ; RV32VB-NEXT:    lbu a0, 161(a0)
-; RV32VB-NEXT:    slli a6, a6, 8
-; RV32VB-NEXT:    or a3, a3, a6
-; RV32VB-NEXT:    slli a5, a5, 16
+; RV32VB-NEXT:    slli a4, a4, 8
+; RV32VB-NEXT:    or a4, a5, a4
+; RV32VB-NEXT:    slli a6, a6, 16
 ; RV32VB-NEXT:    slli a0, a0, 24
-; RV32VB-NEXT:    or a0, a0, a5
-; RV32VB-NEXT:    slli a2, a2, 24
-; RV32VB-NEXT:    or a1, a1, a4
-; RV32VB-NEXT:    or a0, a3, a0
+; RV32VB-NEXT:    or a0, a0, a6
+; RV32VB-NEXT:    slli a3, a3, 24
+; RV32VB-NEXT:    or a1, a2, a1
+; RV32VB-NEXT:    or a0, a4, a0
 ; RV32VB-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
-; RV32VB-NEXT:    vmv.v.x v8, a2
+; RV32VB-NEXT:    vmv.v.x v8, a3
 ; RV32VB-NEXT:    vslide1down.vx v8, v8, a1
 ; RV32VB-NEXT:    vslide1down.vx v8, v8, a0
 ; RV32VB-NEXT:    vslide1down.vx v8, v8, zero
@@ -2607,32 +2607,32 @@ define <16 x i8> @buildvec_v16i8_undef_edges(ptr %p) {
 ;
 ; RVA22U64-LABEL: buildvec_v16i8_undef_edges:
 ; RVA22U64:       # %bb.0:
-; RVA22U64-NEXT:    lbu a6, 31(a0)
+; RVA22U64-NEXT:    lbu a1, 623(a0)
 ; RVA22U64-NEXT:    lbu a2, 44(a0)
 ; RVA22U64-NEXT:    lbu a3, 55(a0)
-; RVA22U64-NEXT:    lbu a4, 623(a0)
+; RVA22U64-NEXT:    lbu a6, 31(a0)
 ; RVA22U64-NEXT:    lbu a5, 75(a0)
 ; RVA22U64-NEXT:    slli a2, a2, 32
 ; RVA22U64-NEXT:    slli a3, a3, 40
-; RVA22U64-NEXT:    or a2, a2, a3
-; RVA22U64-NEXT:    lbu a3, 82(a0)
-; RVA22U64-NEXT:    lbu a1, 93(a0)
-; RVA22U64-NEXT:    slli a4, a4, 48
+; RVA22U64-NEXT:    slli a1, a1, 48
 ; RVA22U64-NEXT:    slli a5, a5, 56
-; RVA22U64-NEXT:    or a4, a4, a5
-; RVA22U64-NEXT:    lbu a5, 105(a0)
+; RVA22U64-NEXT:    or a2, a2, a3
+; RVA22U64-NEXT:    or a1, a1, a5
+; RVA22U64-NEXT:    lbu a3, 93(a0)
+; RVA22U64-NEXT:    lbu a5, 82(a0)
+; RVA22U64-NEXT:    lbu a4, 105(a0)
 ; RVA22U64-NEXT:    lbu a0, 161(a0)
-; RVA22U64-NEXT:    slli a1, a1, 8
-; RVA22U64-NEXT:    or a1, a1, a3
-; RVA22U64-NEXT:    slli a5, a5, 16
+; RVA22U64-NEXT:    slli a3, a3, 8
+; RVA22U64-NEXT:    or a3, a3, a5
+; RVA22U64-NEXT:    slli a4, a4, 16
 ; RVA22U64-NEXT:    slli a0, a0, 24
-; RVA22U64-NEXT:    or a0, a0, a5
+; RVA22U64-NEXT:    or a0, a0, a4
 ; RVA22U64-NEXT:    slli a6, a6, 24
-; RVA22U64-NEXT:    or a2, a2, a4
-; RVA22U64-NEXT:    add.uw a2, a6, a2
-; RVA22U64-NEXT:    or a0, a0, a1
+; RVA22U64-NEXT:    or a1, a1, a2
+; RVA22U64-NEXT:    add.uw a1, a6, a1
+; RVA22U64-NEXT:    or a0, a0, a3
 ; RVA22U64-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
-; RVA22U64-NEXT:    vmv.v.x v8, a2
+; RVA22U64-NEXT:    vmv.v.x v8, a1
 ; RVA22U64-NEXT:    vslide1down.vx v8, v8, a0
 ; RVA22U64-NEXT:    ret
 ;
@@ -2794,26 +2794,26 @@ define <16 x i8> @buildvec_v16i8_loads_undef_scattered(ptr %p) {
 ; RV32VB-PACK-NEXT:    lbu a2, 1(a0)
 ; RV32VB-PACK-NEXT:    lbu a3, 44(a0)
 ; RV32VB-PACK-NEXT:    lbu a4, 55(a0)
-; RV32VB-PACK-NEXT:    lbu a5, 75(a0)
-; RV32VB-PACK-NEXT:    lbu a6, 82(a0)
-; RV32VB-PACK-NEXT:    lbu a7, 93(a0)
+; RV32VB-PACK-NEXT:    lbu a5, 82(a0)
+; RV32VB-PACK-NEXT:    lbu a6, 93(a0)
 ; RV32VB-PACK-NEXT:    packh a1, a1, a2
 ; RV32VB-PACK-NEXT:    lbu a2, 144(a0)
-; RV32VB-PACK-NEXT:    lbu t0, 154(a0)
+; RV32VB-PACK-NEXT:    lbu a7, 154(a0)
 ; RV32VB-PACK-NEXT:    packh a3, a3, a4
+; RV32VB-PACK-NEXT:    lbu a4, 75(a0)
 ; RV32VB-PACK-NEXT:    lbu a0, 124(a0)
-; RV32VB-PACK-NEXT:    packh a4, a6, a7
-; RV32VB-PACK-NEXT:    packh a2, a2, t0
-; RV32VB-PACK-NEXT:    packh a5, a0, a5
-; RV32VB-PACK-NEXT:    pack a3, a3, a5
-; RV32VB-PACK-NEXT:    packh a5, a0, a0
+; RV32VB-PACK-NEXT:    packh a5, a5, a6
+; RV32VB-PACK-NEXT:    packh a2, a2, a7
+; RV32VB-PACK-NEXT:    packh a4, a0, a4
+; RV32VB-PACK-NEXT:    pack a3, a3, a4
+; RV32VB-PACK-NEXT:    packh a4, a0, a0
 ; RV32VB-PACK-NEXT:    packh a0, a0, a0
 ; RV32VB-PACK-NEXT:    pack a0, a0, a2
-; RV32VB-PACK-NEXT:    pack a1, a1, a5
+; RV32VB-PACK-NEXT:    pack a1, a1, a4
 ; RV32VB-PACK-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
 ; RV32VB-PACK-NEXT:    vmv.v.x v8, a1
 ; RV32VB-PACK-NEXT:    vslide1down.vx v8, v8, a3
-; RV32VB-PACK-NEXT:    pack a1, a4, a5
+; RV32VB-PACK-NEXT:    pack a1, a5, a4
 ; RV32VB-PACK-NEXT:    vslide1down.vx v8, v8, a1
 ; RV32VB-PACK-NEXT:    vslide1down.vx v8, v8, a0
 ; RV32VB-PACK-NEXT:    ret
@@ -2888,23 +2888,23 @@ define <16 x i8> @buildvec_v16i8_loads_undef_scattered(ptr %p) {
 ; RVA22U64-PACK:       # %bb.0:
 ; RVA22U64-PACK-NEXT:    lbu a1, 0(a0)
 ; RVA22U64-PACK-NEXT:    lbu a2, 1(a0)
-; RVA22U64-PACK-NEXT:    lbu a7, 44(a0)
-; RVA22U64-PACK-NEXT:    lbu t0, 55(a0)
-; RVA22U64-PACK-NEXT:    lbu a6, 75(a0)
-; RVA22U64-PACK-NEXT:    lbu a5, 82(a0)
+; RVA22U64-PACK-NEXT:    lbu a6, 44(a0)
+; RVA22U64-PACK-NEXT:    lbu a7, 55(a0)
+; RVA22U64-PACK-NEXT:    lbu t1, 82(a0)
 ; RVA22U64-PACK-NEXT:    lbu a3, 93(a0)
-; RVA22U64-PACK-NEXT:    packh t1, a1, a2
+; RVA22U64-PACK-NEXT:    packh t0, a1, a2
 ; RVA22U64-PACK-NEXT:    lbu a2, 144(a0)
 ; RVA22U64-PACK-NEXT:    lbu a4, 154(a0)
-; RVA22U64-PACK-NEXT:    packh a1, a7, t0
+; RVA22U64-PACK-NEXT:    packh a1, a6, a7
+; RVA22U64-PACK-NEXT:    lbu a5, 75(a0)
 ; RVA22U64-PACK-NEXT:    lbu a0, 124(a0)
-; RVA22U64-PACK-NEXT:    packh a3, a5, a3
+; RVA22U64-PACK-NEXT:    packh a3, t1, a3
 ; RVA22U64-PACK-NEXT:    packh a2, a2, a4
-; RVA22U64-PACK-NEXT:    packh a4, a0, a6
+; RVA22U64-PACK-NEXT:    packh a4, a0, a5
 ; RVA22U64-PACK-NEXT:    packw a1, a1, a4
 ; RVA22U64-PACK-NEXT:    packh a4, a0, a0
 ; RVA22U64-PACK-NEXT:    packh a0, a0, a0
-; RVA22U64-PACK-NEXT:    packw a5, t1, a4
+; RVA22U64-PACK-NEXT:    packw a5, t0, a4
 ; RVA22U64-PACK-NEXT:    packw a0, a0, a2
 ; RVA22U64-PACK-NEXT:    packw a2, a3, a4
 ; RVA22U64-PACK-NEXT:    pack a1, a5, a1
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll
index a6c8fa5931cae..533b8b6864ebc 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll
@@ -14191,37 +14191,37 @@ define <8 x i16> @mgather_strided_unaligned(ptr %base) {
 ; RV64ZVE32F-NEXT:    lbu t0, 13(a0)
 ; RV64ZVE32F-NEXT:    slli a2, a2, 8
 ; RV64ZVE32F-NEXT:    slli a4, a4, 8
-; RV64ZVE32F-NEXT:    slli a6, a6, 8
 ; RV64ZVE32F-NEXT:    or a1, a2, a1
 ; RV64ZVE32F-NEXT:    or a3, a4, a3
-; RV64ZVE32F-NEXT:    or a2, a6, a5
-; RV64ZVE32F-NEXT:    lbu a4, 16(a0)
-; RV64ZVE32F-NEXT:    lbu a5, 17(a0)
-; RV64ZVE32F-NEXT:    lbu a6, 20(a0)
-; RV64ZVE32F-NEXT:    lbu t1, 21(a0)
+; RV64ZVE32F-NEXT:    lbu a2, 16(a0)
+; RV64ZVE32F-NEXT:    lbu a4, 17(a0)
+; RV64ZVE32F-NEXT:    lbu t1, 20(a0)
+; RV64ZVE32F-NEXT:    lbu t2, 21(a0)
+; RV64ZVE32F-NEXT:    slli a6, a6, 8
+; RV64ZVE32F-NEXT:    or a5, a6, a5
 ; RV64ZVE32F-NEXT:    slli t0, t0, 8
-; RV64ZVE32F-NEXT:    slli a5, a5, 8
-; RV64ZVE32F-NEXT:    or a7, t0, a7
-; RV64ZVE32F-NEXT:    or a4, a5, a4
-; RV64ZVE32F-NEXT:    lbu a5, 24(a0)
-; RV64ZVE32F-NEXT:    lbu t0, 25(a0)
-; RV64ZVE32F-NEXT:    slli t1, t1, 8
-; RV64ZVE32F-NEXT:    or a6, t1, a6
+; RV64ZVE32F-NEXT:    slli a4, a4, 8
+; RV64ZVE32F-NEXT:    slli t2, t2, 8
+; RV64ZVE32F-NEXT:    or a6, t0, a7
+; RV64ZVE32F-NEXT:    or a2, a4, a2
+; RV64ZVE32F-NEXT:    or a4, t2, t1
+; RV64ZVE32F-NEXT:    lbu a7, 25(a0)
+; RV64ZVE32F-NEXT:    lbu t0, 24(a0)
 ; RV64ZVE32F-NEXT:    lbu t1, 28(a0)
 ; RV64ZVE32F-NEXT:    lbu a0, 29(a0)
-; RV64ZVE32F-NEXT:    slli t0, t0, 8
-; RV64ZVE32F-NEXT:    or a5, t0, a5
+; RV64ZVE32F-NEXT:    slli a7, a7, 8
+; RV64ZVE32F-NEXT:    or a7, a7, t0
 ; RV64ZVE32F-NEXT:    vsetivli zero, 8, e16, m1, ta, mu
 ; RV64ZVE32F-NEXT:    vmv.v.i v0, 15
 ; RV64ZVE32F-NEXT:    slli a0, a0, 8
 ; RV64ZVE32F-NEXT:    or a0, a0, t1
 ; RV64ZVE32F-NEXT:    vmv.v.x v8, a1
-; RV64ZVE32F-NEXT:    vmv.v.x v9, a4
+; RV64ZVE32F-NEXT:    vmv.v.x v9, a2
 ; RV64ZVE32F-NEXT:    vslide1down.vx v8, v8, a3
-; RV64ZVE32F-NEXT:    vslide1down.vx v9, v9, a6
-; RV64ZVE32F-NEXT:    vslide1down.vx v8, v8, a2
-; RV64ZVE32F-NEXT:    vslide1down.vx v9, v9, a5
-; RV64ZVE32F-NEXT:    vslide1down.vx v10, v8, a7
+; RV64ZVE32F-NEXT:    vslide1down.vx v9, v9, a4
+; RV64ZVE32F-NEXT:    vslide1down.vx v8, v8, a5
+; RV64ZVE32F-NEXT:    vslide1down.vx v9, v9, a7
+; RV64ZVE32F-NEXT:    vslide1down.vx v10, v8, a6
 ; RV64ZVE32F-NEXT:    vslide1down.vx v8, v9, a0
 ; RV64ZVE32F-NEXT:    vslidedown.vi v8, v10, 4, v0.t
 ; RV64ZVE32F-NEXT:    ret
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-deinterleave2.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-deinterleave2.ll
index f3e240eead817..b6267bf481c85 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-deinterleave2.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-deinterleave2.ll
@@ -1651,67 +1651,67 @@ define <16 x i64> @unzip2a_dual_v16i64_exact(<16 x i64> %a, <16 x i64> %b) vscal
 ;
 ; ZVE32F-LABEL: unzip2a_dual_v16i64_exact:
 ; ZVE32F:       # %bb.0: # %entry
-; ZVE32F-NEXT:    ld a6, 64(a1)
-; ZVE32F-NEXT:    ld a4, 80(a1)
-; ZVE32F-NEXT:    ld a7, 96(a1)
-; ZVE32F-NEXT:    ld t0, 0(a2)
-; ZVE32F-NEXT:    ld a3, 16(a2)
-; ZVE32F-NEXT:    ld t1, 32(a2)
-; ZVE32F-NEXT:    ld a5, 112(a1)
-; ZVE32F-NEXT:    srli t2, a7, 32
-; ZVE32F-NEXT:    vsetivli zero, 8, e32, m1, ta, mu
-; ZVE32F-NEXT:    vmv.v.x v8, a6
-; ZVE32F-NEXT:    srli a6, a6, 32
-; ZVE32F-NEXT:    vmv.v.x v9, a7
-; ZVE32F-NEXT:    vslide1down.vx v8, v8, a6
-; ZVE32F-NEXT:    vslide1down.vx v9, v9, t2
 ; ZVE32F-NEXT:    ld a6, 0(a1)
-; ZVE32F-NEXT:    ld a7, 16(a1)
-; ZVE32F-NEXT:    ld t2, 32(a1)
-; ZVE32F-NEXT:    ld a1, 48(a1)
+; ZVE32F-NEXT:    ld a4, 16(a1)
+; ZVE32F-NEXT:    ld a7, 32(a1)
+; ZVE32F-NEXT:    ld a3, 48(a1)
+; ZVE32F-NEXT:    ld a5, 80(a1)
+; ZVE32F-NEXT:    ld t0, 96(a1)
+; ZVE32F-NEXT:    ld t1, 64(a1)
+; ZVE32F-NEXT:    ld a1, 112(a1)
+; ZVE32F-NEXT:    srli t2, a6, 32
+; ZVE32F-NEXT:    vsetivli zero, 8, e32, m1, ta, mu
+; ZVE32F-NEXT:    vmv.v.x v8, t0
+; ZVE32F-NEXT:    srli t0, t0, 32
+; ZVE32F-NEXT:    vmv.v.x v9, t1
+; ZVE32F-NEXT:    srli t1, t1, 32
 ; ZVE32F-NEXT:    vmv.v.x v10, a6
-; ZVE32F-NEXT:    srli a6, a6, 32
-; ZVE32F-NEXT:    vslide1down.vx v10, v10, a6
+; ZVE32F-NEXT:    vslide1down.vx v9, v9, t1
+; ZVE32F-NEXT:    vslide1down.vx v8, v8, t0
+; ZVE32F-NEXT:    vslide1down.vx v10, v10, t2
+; ZVE32F-NEXT:    ld t1, 32(a2)
+; ZVE32F-NEXT:    ld t0, 16(a2)
+; ZVE32F-NEXT:    ld t2, 0(a2)
 ; ZVE32F-NEXT:    ld a6, 48(a2)
 ; ZVE32F-NEXT:    vmv.v.x v11, t1
 ; ZVE32F-NEXT:    srli t1, t1, 32
-; ZVE32F-NEXT:    vmv.v.x v12, t0
-; ZVE32F-NEXT:    srli t0, t0, 32
-; ZVE32F-NEXT:    vmv.v.x v13, t2
+; ZVE32F-NEXT:    vmv.v.x v12, t2
 ; ZVE32F-NEXT:    srli t2, t2, 32
-; ZVE32F-NEXT:    vslide1down.vx v13, v13, t2
-; ZVE32F-NEXT:    vslide1down.vx v12, v12, t0
+; ZVE32F-NEXT:    vmv.v.x v13, a7
+; ZVE32F-NEXT:    srli a7, a7, 32
+; ZVE32F-NEXT:    vslide1down.vx v13, v13, a7
+; ZVE32F-NEXT:    vslide1down.vx v12, v12, t2
 ; ZVE32F-NEXT:    vslide1down.vx v11, v11, t1
-; ZVE32F-NEXT:    ld t0, 64(a2)
+; ZVE32F-NEXT:    ld a7, 64(a2)
 ; ZVE32F-NEXT:    ld t1, 80(a2)
 ; ZVE32F-NEXT:    ld t2, 96(a2)
 ; ZVE32F-NEXT:    ld a2, 112(a2)
-; ZVE32F-NEXT:    vmv.v.x v14, t0
-; ZVE32F-NEXT:    srli t0, t0, 32
-; ZVE32F-NEXT:    vslide1down.vx v14, v14, t0
+; ZVE32F-NEXT:    vmv.v.x v14, a7
+; ZVE32F-NEXT:    srli a7, a7, 32
+; ZVE32F-NEXT:    vslide1down.vx v14, v14, a7
 ; ZVE32F-NEXT:    vmv.v.x v15, t2
-; ZVE32F-NEXT:    srli t0, t2, 32
-; ZVE32F-NEXT:    vslide1down.vx v15, v15, t0
-; ZVE32F-NEXT:    vslide1down.vx v8, v8, a4
-; ZVE32F-NEXT:    srli a4, a4, 32
-; ZVE32F-NEXT:    vslide1down.vx v8, v8, a4
+; ZVE32F-NEXT:    srli a7, t2, 32
+; ZVE32F-NEXT:    vslide1down.vx v15, v15, a7
 ; ZVE32F-NEXT:    vslide1down.vx v9, v9, a5
 ; ZVE32F-NEXT:    srli a5, a5, 32
-; ZVE32F-NEXT:    vslide1down.vx v9, v9, a5
-; ZVE32F-NEXT:    vslide1down.vx v10, v10, a7
-; ZVE32F-NEXT:    srli a4, a7, 32
-; ZVE32F-NEXT:    vslide1down.vx v10, v10, a4
-; ZVE32F-NEXT:    vslide1down.vx v12, v12, a3
-; ZVE32F-NEXT:    srli a3, a3, 32
-; ZVE32F-NEXT:    vslide1down.vx v12, v12, a3
-; ZVE32F-NEXT:    vmv.v.i v0, 15
-; ZVE32F-NEXT:    vslide1down.vx v14, v14, t1
-; ZVE32F-NEXT:    srli a3, t1, 32
-; ZVE32F-NEXT:    vslide1down.vx v14, v14, a3
-; ZVE32F-NEXT:    vslidedown.vi v9, v8, 4, v0.t
-; ZVE32F-NEXT:    vslide1down.vx v8, v13, a1
-; ZVE32F-NEXT:    srli a1, a1, 32
+; ZVE32F-NEXT:    vslide1down.vx v16, v9, a5
 ; ZVE32F-NEXT:    vslide1down.vx v8, v8, a1
+; ZVE32F-NEXT:    srli a1, a1, 32
+; ZVE32F-NEXT:    vslide1down.vx v9, v8, a1
+; ZVE32F-NEXT:    vslide1down.vx v8, v10, a4
+; ZVE32F-NEXT:    srli a4, a4, 32
+; ZVE32F-NEXT:    vslide1down.vx v10, v8, a4
+; ZVE32F-NEXT:    vslide1down.vx v8, v12, t0
+; ZVE32F-NEXT:    srli a1, t0, 32
+; ZVE32F-NEXT:    vslide1down.vx v12, v8, a1
+; ZVE32F-NEXT:    vmv.v.i v0, 15
+; ZVE32F-NEXT:    vslide1down.vx v8, v14, t1
+; ZVE32F-NEXT:    srli a1, t1, 32
+; ZVE32F-NEXT:    vslide1down.vx v14, v8, a1
+; ZVE32F-NEXT:    vslidedown.vi v9, v16, 4, v0.t
+; ZVE32F-NEXT:    vslide1down.vx v8, v13, a3
+; ZVE32F-NEXT:    srli a3, a3, 32
+; ZVE32F-NEXT:    vslide1down.vx v8, v8, a3
 ; ZVE32F-NEXT:    vslidedown.vi v8, v10, 4, v0.t
 ; ZVE32F-NEXT:    vslide1down.vx v10, v11, a6
 ; ZVE32F-NEXT:    srli a1, a6, 32
diff --git a/llvm/test/CodeGen/RISCV/scmp.ll b/llvm/test/CodeGen/RISCV/scmp.ll
index a212714db53e0..56c876a2409d2 100644
--- a/llvm/test/CodeGen/RISCV/scmp.ll
+++ b/llvm/test/CodeGen/RISCV/scmp.ll
@@ -89,8 +89,8 @@ define i8 @scmp.8.128(i128 %x, i128 %y) nounwind {
 ; RV32I-NEXT:    lw a2, 4(a1)
 ; RV32I-NEXT:    lw a4, 8(a1)
 ; RV32I-NEXT:    lw a5, 12(a1)
-; RV32I-NEXT:    lw a6, 12(a0)
 ; RV32I-NEXT:    lw a3, 4(a0)
+; RV32I-NEXT:    lw a6, 12(a0)
 ; RV32I-NEXT:    lw a7, 8(a0)
 ; RV32I-NEXT:    beq a6, a5, .LBB4_2
 ; RV32I-NEXT:  # %bb.1:
diff --git a/llvm/test/CodeGen/RISCV/srem-vector-lkk.ll b/llvm/test/CodeGen/RISCV/srem-vector-lkk.ll
index 2be3324b86032..30ffaf6c7ceca 100644
--- a/llvm/test/CodeGen/RISCV/srem-vector-lkk.ll
+++ b/llvm/test/CodeGen/RISCV/srem-vector-lkk.ll
@@ -562,49 +562,49 @@ define <4 x i16> @combine_srem_sdiv(<4 x i16> %x) nounwind {
 ;
 ; RV64IM-LABEL: combine_srem_sdiv:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lh a2, 16(a1)
-; RV64IM-NEXT:    lh a3, 24(a1)
-; RV64IM-NEXT:    lui a4, %hi(.LCPI2_0)
-; RV64IM-NEXT:    ld a4, %lo(.LCPI2_0)(a4)
+; RV64IM-NEXT:    lui a2, %hi(.LCPI2_0)
+; RV64IM-NEXT:    ld a2, %lo(.LCPI2_0)(a2)
+; RV64IM-NEXT:    lh a3, 16(a1)
+; RV64IM-NEXT:    lh a4, 24(a1)
 ; RV64IM-NEXT:    lh a5, 0(a1)
 ; RV64IM-NEXT:    lh a1, 8(a1)
 ; RV64IM-NEXT:    li a6, 95
-; RV64IM-NEXT:    mulh a7, a3, a4
-; RV64IM-NEXT:    mulh t0, a2, a4
-; RV64IM-NEXT:    mulh t1, a1, a4
-; RV64IM-NEXT:    mulh a4, a5, a4
-; RV64IM-NEXT:    add a7, a7, a3
-; RV64IM-NEXT:    add t0, t0, a2
+; RV64IM-NEXT:    mulh a7, a4, a2
+; RV64IM-NEXT:    mulh t0, a3, a2
+; RV64IM-NEXT:    mulh t1, a1, a2
+; RV64IM-NEXT:    mulh a2, a5, a2
+; RV64IM-NEXT:    add a7, a7, a4
+; RV64IM-NEXT:    add t0, t0, a3
 ; RV64IM-NEXT:    add t1, t1, a1
-; RV64IM-NEXT:    add a4, a4, a5
+; RV64IM-NEXT:    add a2, a2, a5
 ; RV64IM-NEXT:    srli t2, a7, 63
 ; RV64IM-NEXT:    srai a7, a7, 6
 ; RV64IM-NEXT:    srli t3, t0, 63
 ; RV64IM-NEXT:    srai t0, t0, 6
 ; RV64IM-NEXT:    srli t4, t1, 63
 ; RV64IM-NEXT:    srai t1, t1, 6
-; RV64IM-NEXT:    srli t5, a4, 63
-; RV64IM-NEXT:    srai a4, a4, 6
+; RV64IM-NEXT:    srli t5, a2, 63
+; RV64IM-NEXT:    srai a2, a2, 6
 ; RV64IM-NEXT:    add a7, a7, t2
 ; RV64IM-NEXT:    add t0, t0, t3
 ; RV64IM-NEXT:    add t1, t1, t4
-; RV64IM-NEXT:    add a4, a4, t5
+; RV64IM-NEXT:    add a2, a2, t5
 ; RV64IM-NEXT:    mul t2, a7, a6
 ; RV64IM-NEXT:    mul t3, t0, a6
 ; RV64IM-NEXT:    mul t4, t1, a6
-; RV64IM-NEXT:    mul a6, a4, a6
-; RV64IM-NEXT:    add a4, a5, a4
+; RV64IM-NEXT:    mul a6, a2, a6
+; RV64IM-NEXT:    add a2, a5, a2
 ; RV64IM-NEXT:    add a1, a1, t1
-; RV64IM-NEXT:    add a2, a2, t0
-; RV64IM-NEXT:    add a3, a3, a7
-; RV64IM-NEXT:    subw a4, a4, a6
+; RV64IM-NEXT:    add a3, a3, t0
+; RV64IM-NEXT:    add a4, a4, a7
+; RV64IM-NEXT:    subw a2, a2, a6
 ; RV64IM-NEXT:    subw a1, a1, t4
-; RV64IM-NEXT:    subw a2, a2, t3
-; RV64IM-NEXT:    subw a3, a3, t2
-; RV64IM-NEXT:    sh a4, 0(a0)
+; RV64IM-NEXT:    subw a3, a3, t3
+; RV64IM-NEXT:    subw a4, a4, t2
+; RV64IM-NEXT:    sh a2, 0(a0)
 ; RV64IM-NEXT:    sh a1, 2(a0)
-; RV64IM-NEXT:    sh a2, 4(a0)
-; RV64IM-NEXT:    sh a3, 6(a0)
+; RV64IM-NEXT:    sh a3, 4(a0)
+; RV64IM-NEXT:    sh a4, 6(a0)
 ; RV64IM-NEXT:    ret
   %1 = srem <4 x i16> %x, <i16 95, i16 95, i16 95, i16 95>
   %2 = sdiv <4 x i16> %x, <i16 95, i16 95, i16 95, i16 95>
diff --git a/llvm/test/CodeGen/RISCV/ucmp.ll b/llvm/test/CodeGen/RISCV/ucmp.ll
index 50da56fbc5951..0a400b1c04a3f 100644
--- a/llvm/test/CodeGen/RISCV/ucmp.ll
+++ b/llvm/test/CodeGen/RISCV/ucmp.ll
@@ -89,8 +89,8 @@ define i8 @ucmp.8.128(i128 %x, i128 %y) nounwind {
 ; RV32I-NEXT:    lw a2, 4(a1)
 ; RV32I-NEXT:    lw a4, 8(a1)
 ; RV32I-NEXT:    lw a5, 12(a1)
-; RV32I-NEXT:    lw a6, 12(a0)
 ; RV32I-NEXT:    lw a3, 4(a0)
+; RV32I-NEXT:    lw a6, 12(a0)
 ; RV32I-NEXT:    lw a7, 8(a0)
 ; RV32I-NEXT:    beq a6, a5, .LBB4_2
 ; RV32I-NEXT:  # %bb.1:
diff --git a/llvm/test/CodeGen/RISCV/unaligned-load-store.ll b/llvm/test/CodeGen/RISCV/unaligned-load-store.ll
index 068b9f7620021..c9c49e8f7f532 100644
--- a/llvm/test/CodeGen/RISCV/unaligned-load-store.ll
+++ b/llvm/test/CodeGen/RISCV/unaligned-load-store.ll
@@ -133,49 +133,49 @@ define i64 @load_i64(ptr %p) {
 ; RV32I-LABEL: load_i64:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a1, 1(a0)
-; RV32I-NEXT:    lbu a2, 0(a0)
-; RV32I-NEXT:    lbu a3, 2(a0)
-; RV32I-NEXT:    lbu a4, 3(a0)
+; RV32I-NEXT:    lbu a2, 2(a0)
+; RV32I-NEXT:    lbu a3, 3(a0)
+; RV32I-NEXT:    lbu a4, 0(a0)
 ; RV32I-NEXT:    slli a1, a1, 8
-; RV32I-NEXT:    or a1, a1, a2
-; RV32I-NEXT:    lbu a2, 4(a0)
-; RV32I-NEXT:    lbu a5, 5(a0)
-; RV32I-NEXT:    slli a3, a3, 16
-; RV32I-NEXT:    slli a4, a4, 24
-; RV32I-NEXT:    or a3, a4, a3
-; RV32I-NEXT:    lbu a4, 6(a0)
+; RV32I-NEXT:    slli a2, a2, 16
+; RV32I-NEXT:    slli a3, a3, 24
+; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    or a2, a3, a2
+; RV32I-NEXT:    lbu a3, 5(a0)
+; RV32I-NEXT:    lbu a4, 4(a0)
+; RV32I-NEXT:    lbu a5, 6(a0)
 ; RV32I-NEXT:    lbu a0, 7(a0)
-; RV32I-NEXT:    slli a5, a5, 8
-; RV32I-NEXT:    or a2, a5, a2
-; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a3, a3, 8
+; RV32I-NEXT:    or a3, a3, a4
+; RV32I-NEXT:    slli a5, a5, 16
 ; RV32I-NEXT:    slli a0, a0, 24
-; RV32I-NEXT:    or a4, a0, a4
-; RV32I-NEXT:    or a0, a3, a1
-; RV32I-NEXT:    or a1, a4, a2
+; RV32I-NEXT:    or a5, a0, a5
+; RV32I-NEXT:    or a0, a2, a1
+; RV32I-NEXT:    or a1, a5, a3
 ; RV32I-NEXT:    ret
 ;
 ; RV64I-LABEL: load_i64:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    lbu a1, 1(a0)
-; RV64I-NEXT:    lbu a2, 0(a0)
-; RV64I-NEXT:    lbu a3, 2(a0)
-; RV64I-NEXT:    lbu a4, 3(a0)
+; RV64I-NEXT:    lbu a2, 2(a0)
+; RV64I-NEXT:    lbu a3, 3(a0)
+; RV64I-NEXT:    lbu a4, 0(a0)
 ; RV64I-NEXT:    slli a1, a1, 8
-; RV64I-NEXT:    or a1, a1, a2
-; RV64I-NEXT:    lbu a2, 4(a0)
-; RV64I-NEXT:    lbu a5, 5(a0)
-; RV64I-NEXT:    slli a3, a3, 16
-; RV64I-NEXT:    slli a4, a4, 24
-; RV64I-NEXT:    or a3, a4, a3
-; RV64I-NEXT:    lbu a4, 6(a0)
+; RV64I-NEXT:    slli a2, a2, 16
+; RV64I-NEXT:    slli a3, a3, 24
+; RV64I-NEXT:    or a1, a1, a4
+; RV64I-NEXT:    or a2, a3, a2
+; RV64I-NEXT:    lbu a3, 5(a0)
+; RV64I-NEXT:    lbu a4, 4(a0)
+; RV64I-NEXT:    lbu a5, 6(a0)
 ; RV64I-NEXT:    lbu a0, 7(a0)
-; RV64I-NEXT:    slli a5, a5, 8
-; RV64I-NEXT:    or a2, a5, a2
-; RV64I-NEXT:    slli a4, a4, 16
+; RV64I-NEXT:    slli a3, a3, 8
+; RV64I-NEXT:    or a3, a3, a4
+; RV64I-NEXT:    slli a5, a5, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, a4
-; RV64I-NEXT:    or a1, a3, a1
-; RV64I-NEXT:    or a0, a0, a2
+; RV64I-NEXT:    or a0, a0, a5
+; RV64I-NEXT:    or a1, a2, a1
+; RV64I-NEXT:    or a0, a0, a3
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a1
 ; RV64I-NEXT:    ret
diff --git a/llvm/test/CodeGen/RISCV/urem-vector-lkk.ll b/llvm/test/CodeGen/RISCV/urem-vector-lkk.ll
index e8e91f0baee14..3ef9f3f945108 100644
--- a/llvm/test/CodeGen/RISCV/urem-vector-lkk.ll
+++ b/llvm/test/CodeGen/RISCV/urem-vector-lkk.ll
@@ -489,33 +489,33 @@ define <4 x i16> @combine_urem_udiv(<4 x i16> %x) nounwind {
 ;
 ; RV64IM-LABEL: combine_urem_udiv:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lhu a2, 16(a1)
-; RV64IM-NEXT:    lhu a3, 24(a1)
-; RV64IM-NEXT:    lui a4, %hi(.LCPI2_0)
-; RV64IM-NEXT:    ld a4, %lo(.LCPI2_0)(a4)
+; RV64IM-NEXT:    lui a2, %hi(.LCPI2_0)
+; RV64IM-NEXT:    ld a2, %lo(.LCPI2_0)(a2)
+; RV64IM-NEXT:    lhu a3, 16(a1)
+; RV64IM-NEXT:    lhu a4, 24(a1)
 ; RV64IM-NEXT:    lhu a5, 0(a1)
 ; RV64IM-NEXT:    lhu a1, 8(a1)
 ; RV64IM-NEXT:    li a6, 95
-; RV64IM-NEXT:    mulhu a7, a3, a4
-; RV64IM-NEXT:    mulhu t0, a2, a4
-; RV64IM-NEXT:    mulhu t1, a1, a4
-; RV64IM-NEXT:    mulhu a4, a5, a4
+; RV64IM-NEXT:    mulhu a7, a4, a2
+; RV64IM-NEXT:    mulhu t0, a3, a2
+; RV64IM-NEXT:    mulhu t1, a1, a2
+; RV64IM-NEXT:    mulhu a2, a5, a2
 ; RV64IM-NEXT:    mul t2, a7, a6
 ; RV64IM-NEXT:    mul t3, t0, a6
 ; RV64IM-NEXT:    mul t4, t1, a6
-; RV64IM-NEXT:    mul a6, a4, a6
-; RV64IM-NEXT:    add a4, a5, a4
+; RV64IM-NEXT:    mul a6, a2, a6
+; RV64IM-NEXT:    add a2, a5, a2
 ; RV64IM-NEXT:    add a1, a1, t1
-; RV64IM-NEXT:    add a2, a2, t0
-; RV64IM-NEXT:    add a3, a3, a7
-; RV64IM-NEXT:    subw a4, a4, a6
+; RV64IM-NEXT:    add a3, a3, t0
+; RV64IM-NEXT:    add a4, a4, a7
+; RV64IM-NEXT:    subw a2, a2, a6
 ; RV64IM-NEXT:    subw a1, a1, t4
-; RV64IM-NEXT:    subw a2, a2, t3
-; RV64IM-NEXT:    subw a3, a3, t2
-; RV64IM-NEXT:    sh a4, 0(a0)
+; RV64IM-NEXT:    subw a3, a3, t3
+; RV64IM-NEXT:    subw a4, a4, t2
+; RV64IM-NEXT:    sh a2, 0(a0)
 ; RV64IM-NEXT:    sh a1, 2(a0)
-; RV64IM-NEXT:    sh a2, 4(a0)
-; RV64IM-NEXT:    sh a3, 6(a0)
+; RV64IM-NEXT:    sh a3, 4(a0)
+; RV64IM-NEXT:    sh a4, 6(a0)
 ; RV64IM-NEXT:    ret
   %1 = urem <4 x i16> %x, <i16 95, i16 95, i16 95, i16 95>
   %2 = udiv <4 x i16> %x, <i16 95, i16 95, i16 95, i16 95>
diff --git a/llvm/test/CodeGen/RISCV/wide-scalar-shift-by-byte-multiple-legalization.ll b/llvm/test/CodeGen/RISCV/wide-scalar-shift-by-byte-multiple-legalization.ll
index 2a0228f95f1cd..09b2eeb19a69c 100644
--- a/llvm/test/CodeGen/RISCV/wide-scalar-shift-by-byte-multiple-legalization.ll
+++ b/llvm/test/CodeGen/RISCV/wide-scalar-shift-by-byte-multiple-legalization.ll
@@ -30,25 +30,25 @@ define void @lshr_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: lshr_4bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
-; RV32I-NEXT:    lbu a5, 2(a0)
-; RV32I-NEXT:    lbu a0, 3(a0)
+; RV32I-NEXT:    lbu a4, 2(a0)
+; RV32I-NEXT:    lbu a5, 3(a0)
+; RV32I-NEXT:    lbu a0, 0(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a6, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a0, a0, 24
-; RV32I-NEXT:    or a0, a0, a5
-; RV32I-NEXT:    lbu a5, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a0, a3, a0
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a3, 1(a1)
+; RV32I-NEXT:    lbu a5, 0(a1)
+; RV32I-NEXT:    lbu a6, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a6, a6, 8
-; RV32I-NEXT:    or a4, a6, a4
-; RV32I-NEXT:    slli a5, a5, 16
+; RV32I-NEXT:    slli a3, a3, 8
+; RV32I-NEXT:    or a3, a3, a5
+; RV32I-NEXT:    slli a6, a6, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a5
-; RV32I-NEXT:    or a0, a0, a3
-; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    or a1, a1, a6
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:    or a1, a1, a3
 ; RV32I-NEXT:    slli a1, a1, 3
 ; RV32I-NEXT:    srl a0, a0, a1
 ; RV32I-NEXT:    srli a1, a0, 16
@@ -94,25 +94,25 @@ define void @shl_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: shl_4bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
-; RV32I-NEXT:    lbu a5, 2(a0)
-; RV32I-NEXT:    lbu a0, 3(a0)
+; RV32I-NEXT:    lbu a4, 2(a0)
+; RV32I-NEXT:    lbu a5, 3(a0)
+; RV32I-NEXT:    lbu a0, 0(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a6, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a0, a0, 24
-; RV32I-NEXT:    or a0, a0, a5
-; RV32I-NEXT:    lbu a5, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a0, a3, a0
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a3, 1(a1)
+; RV32I-NEXT:    lbu a5, 0(a1)
+; RV32I-NEXT:    lbu a6, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a6, a6, 8
-; RV32I-NEXT:    or a4, a6, a4
-; RV32I-NEXT:    slli a5, a5, 16
+; RV32I-NEXT:    slli a3, a3, 8
+; RV32I-NEXT:    or a3, a3, a5
+; RV32I-NEXT:    slli a6, a6, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a5
-; RV32I-NEXT:    or a0, a0, a3
-; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    or a1, a1, a6
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:    or a1, a1, a3
 ; RV32I-NEXT:    slli a1, a1, 3
 ; RV32I-NEXT:    sll a0, a0, a1
 ; RV32I-NEXT:    srli a1, a0, 16
@@ -158,25 +158,25 @@ define void @ashr_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: ashr_4bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
-; RV32I-NEXT:    lbu a5, 2(a0)
-; RV32I-NEXT:    lbu a0, 3(a0)
+; RV32I-NEXT:    lbu a4, 2(a0)
+; RV32I-NEXT:    lbu a5, 3(a0)
+; RV32I-NEXT:    lbu a0, 0(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a6, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a0, a0, 24
-; RV32I-NEXT:    or a0, a0, a5
-; RV32I-NEXT:    lbu a5, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a0, a3, a0
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a3, 1(a1)
+; RV32I-NEXT:    lbu a5, 0(a1)
+; RV32I-NEXT:    lbu a6, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a6, a6, 8
-; RV32I-NEXT:    or a4, a6, a4
-; RV32I-NEXT:    slli a5, a5, 16
+; RV32I-NEXT:    slli a3, a3, 8
+; RV32I-NEXT:    or a3, a3, a5
+; RV32I-NEXT:    slli a6, a6, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a5
-; RV32I-NEXT:    or a0, a0, a3
-; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    or a1, a1, a6
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:    or a1, a1, a3
 ; RV32I-NEXT:    slli a1, a1, 3
 ; RV32I-NEXT:    sra a0, a0, a1
 ; RV32I-NEXT:    srli a1, a0, 16
@@ -220,24 +220,24 @@ define void @lshr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli a0, a0, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 0(a1)
-; RV64I-NEXT:    lbu t1, 1(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t2, t2, 24
-; RV64I-NEXT:    or t0, t2, t0
+; RV64I-NEXT:    or a0, a0, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t2, t0
+; RV64I-NEXT:    lbu t0, 1(a1)
+; RV64I-NEXT:    lbu t1, 0(a1)
 ; RV64I-NEXT:    lbu t2, 2(a1)
 ; RV64I-NEXT:    lbu a1, 3(a1)
-; RV64I-NEXT:    slli t1, t1, 8
-; RV64I-NEXT:    or a7, t1, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
 ; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t2
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a0, a0, a5
-; RV64I-NEXT:    or a4, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a4, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    slli a1, a1, 3
 ; RV64I-NEXT:    slli a4, a4, 35
@@ -264,26 +264,26 @@ define void @lshr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: lshr_8bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 5(a0)
-; RV32I-NEXT:    lbu a4, 4(a0)
-; RV32I-NEXT:    lbu a5, 6(a0)
-; RV32I-NEXT:    lbu a6, 7(a0)
+; RV32I-NEXT:    lbu a4, 6(a0)
+; RV32I-NEXT:    lbu a5, 7(a0)
+; RV32I-NEXT:    lbu a6, 4(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a7, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a6, a6, 24
-; RV32I-NEXT:    or a5, a6, a5
-; RV32I-NEXT:    lbu a6, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a3, a3, a6
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a5, 1(a1)
+; RV32I-NEXT:    lbu a6, 0(a1)
+; RV32I-NEXT:    lbu a7, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a7, a7, 8
-; RV32I-NEXT:    or a4, a7, a4
-; RV32I-NEXT:    slli a6, a6, 16
+; RV32I-NEXT:    slli a5, a5, 8
+; RV32I-NEXT:    or a6, a5, a6
+; RV32I-NEXT:    slli a7, a7, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a6
-; RV32I-NEXT:    or a5, a5, a3
-; RV32I-NEXT:    or a1, a1, a4
-; RV32I-NEXT:    slli a4, a1, 3
+; RV32I-NEXT:    or a1, a1, a7
+; RV32I-NEXT:    or a5, a4, a3
+; RV32I-NEXT:    or a4, a1, a6
+; RV32I-NEXT:    slli a4, a4, 3
 ; RV32I-NEXT:    addi a3, a4, -32
 ; RV32I-NEXT:    srl a1, a5, a4
 ; RV32I-NEXT:    bltz a3, .LBB3_2
@@ -356,24 +356,24 @@ define void @shl_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli a0, a0, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 0(a1)
-; RV64I-NEXT:    lbu t1, 1(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t2, t2, 24
-; RV64I-NEXT:    or t0, t2, t0
+; RV64I-NEXT:    or a0, a0, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t2, t0
+; RV64I-NEXT:    lbu t0, 1(a1)
+; RV64I-NEXT:    lbu t1, 0(a1)
 ; RV64I-NEXT:    lbu t2, 2(a1)
 ; RV64I-NEXT:    lbu a1, 3(a1)
-; RV64I-NEXT:    slli t1, t1, 8
-; RV64I-NEXT:    or a7, t1, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
 ; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t2
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a0, a0, a5
-; RV64I-NEXT:    or a4, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a4, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    slli a1, a1, 3
 ; RV64I-NEXT:    slli a4, a4, 35
@@ -400,26 +400,26 @@ define void @shl_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: shl_8bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
-; RV32I-NEXT:    lbu a5, 2(a0)
-; RV32I-NEXT:    lbu a6, 3(a0)
+; RV32I-NEXT:    lbu a4, 2(a0)
+; RV32I-NEXT:    lbu a5, 3(a0)
+; RV32I-NEXT:    lbu a6, 0(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a7, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a6, a6, 24
-; RV32I-NEXT:    or a5, a6, a5
-; RV32I-NEXT:    lbu a6, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a3, a3, a6
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a5, 1(a1)
+; RV32I-NEXT:    lbu a6, 0(a1)
+; RV32I-NEXT:    lbu a7, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a7, a7, 8
-; RV32I-NEXT:    or a4, a7, a4
-; RV32I-NEXT:    slli a6, a6, 16
+; RV32I-NEXT:    slli a5, a5, 8
+; RV32I-NEXT:    or a6, a5, a6
+; RV32I-NEXT:    slli a7, a7, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a6
-; RV32I-NEXT:    or a5, a5, a3
-; RV32I-NEXT:    or a1, a1, a4
-; RV32I-NEXT:    slli a4, a1, 3
+; RV32I-NEXT:    or a1, a1, a7
+; RV32I-NEXT:    or a5, a4, a3
+; RV32I-NEXT:    or a4, a1, a6
+; RV32I-NEXT:    slli a4, a4, 3
 ; RV32I-NEXT:    addi a3, a4, -32
 ; RV32I-NEXT:    sll a1, a5, a4
 ; RV32I-NEXT:    bltz a3, .LBB4_2
@@ -492,24 +492,24 @@ define void @ashr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli a0, a0, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 0(a1)
-; RV64I-NEXT:    lbu t1, 1(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t2, t2, 24
-; RV64I-NEXT:    or t0, t2, t0
+; RV64I-NEXT:    or a0, a0, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t2, t0
+; RV64I-NEXT:    lbu t0, 1(a1)
+; RV64I-NEXT:    lbu t1, 0(a1)
 ; RV64I-NEXT:    lbu t2, 2(a1)
 ; RV64I-NEXT:    lbu a1, 3(a1)
-; RV64I-NEXT:    slli t1, t1, 8
-; RV64I-NEXT:    or a7, t1, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
 ; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t2
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a0, a0, a5
-; RV64I-NEXT:    or a4, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a4, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    slli a1, a1, 3
 ; RV64I-NEXT:    slli a4, a4, 35
@@ -535,21 +535,21 @@ define void @ashr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ;
 ; RV32I-LABEL: ashr_8bytes:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    lbu a3, 4(a0)
-; RV32I-NEXT:    lbu a4, 5(a0)
+; RV32I-NEXT:    lbu a3, 5(a0)
+; RV32I-NEXT:    lbu a4, 4(a0)
 ; RV32I-NEXT:    lbu a5, 6(a0)
 ; RV32I-NEXT:    lbu a6, 7(a0)
+; RV32I-NEXT:    slli a3, a3, 8
+; RV32I-NEXT:    or a3, a3, a4
+; RV32I-NEXT:    lbu a4, 1(a1)
 ; RV32I-NEXT:    lbu a7, 0(a1)
-; RV32I-NEXT:    lbu t0, 1(a1)
-; RV32I-NEXT:    slli a4, a4, 8
-; RV32I-NEXT:    or a3, a4, a3
-; RV32I-NEXT:    lbu a4, 2(a1)
+; RV32I-NEXT:    lbu t0, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli t0, t0, 8
-; RV32I-NEXT:    or a7, t0, a7
-; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a4, a4, 8
+; RV32I-NEXT:    or a7, a4, a7
+; RV32I-NEXT:    slli t0, t0, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    or a1, a1, t0
 ; RV32I-NEXT:    slli a4, a5, 16
 ; RV32I-NEXT:    slli a5, a6, 24
 ; RV32I-NEXT:    or a4, a5, a4
@@ -629,24 +629,24 @@ define void @lshr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli t2, t2, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or t1, t2, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 0(a1)
-; RV64I-NEXT:    lbu t2, 1(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t3, t3, 24
-; RV64I-NEXT:    or t0, t3, t0
+; RV64I-NEXT:    or t1, t2, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t3, t0
+; RV64I-NEXT:    lbu t0, 1(a1)
+; RV64I-NEXT:    lbu t2, 0(a1)
 ; RV64I-NEXT:    lbu t3, 2(a1)
 ; RV64I-NEXT:    lbu a1, 3(a1)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t2
 ; RV64I-NEXT:    slli t3, t3, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t3
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a4, t1, a5
-; RV64I-NEXT:    or a5, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a5, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a1, a1, 3
 ; RV64I-NEXT:    slli a6, a5, 35
@@ -660,31 +660,31 @@ define void @lshr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    j .LBB6_3
 ; RV64I-NEXT:  .LBB6_2:
 ; RV64I-NEXT:    lbu a6, 1(a0)
-; RV64I-NEXT:    lbu a7, 0(a0)
-; RV64I-NEXT:    lbu t0, 2(a0)
-; RV64I-NEXT:    lbu t1, 3(a0)
+; RV64I-NEXT:    lbu a7, 2(a0)
+; RV64I-NEXT:    lbu t0, 3(a0)
+; RV64I-NEXT:    lbu t1, 0(a0)
 ; RV64I-NEXT:    slli a6, a6, 8
-; RV64I-NEXT:    or a6, a6, a7
-; RV64I-NEXT:    lbu a7, 4(a0)
-; RV64I-NEXT:    lbu t2, 5(a0)
-; RV64I-NEXT:    slli t0, t0, 16
-; RV64I-NEXT:    slli t1, t1, 24
-; RV64I-NEXT:    or t0, t1, t0
-; RV64I-NEXT:    lbu t1, 6(a0)
+; RV64I-NEXT:    slli a7, a7, 16
+; RV64I-NEXT:    slli t0, t0, 24
+; RV64I-NEXT:    or a6, a6, t1
+; RV64I-NEXT:    or a7, t0, a7
+; RV64I-NEXT:    lbu t0, 5(a0)
+; RV64I-NEXT:    lbu t1, 4(a0)
+; RV64I-NEXT:    lbu t2, 6(a0)
 ; RV64I-NEXT:    lbu a0, 7(a0)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
-; RV64I-NEXT:    slli t1, t1, 16
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
+; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, t0, a6
-; RV64I-NEXT:    not t0, a4
+; RV64I-NEXT:    or a0, a0, t2
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    not a7, a4
 ; RV64I-NEXT:    slli a5, a5, 1
-; RV64I-NEXT:    or a0, a0, a7
+; RV64I-NEXT:    or a0, a0, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a6
 ; RV64I-NEXT:    srl a0, a0, a4
-; RV64I-NEXT:    sll a4, a5, t0
+; RV64I-NEXT:    sll a4, a5, a7
 ; RV64I-NEXT:    or a0, a0, a4
 ; RV64I-NEXT:  .LBB6_3:
 ; RV64I-NEXT:    srai a3, a3, 63
@@ -787,27 +787,27 @@ define void @lshr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    andi a1, a1, 12
 ; RV32I-NEXT:    add a1, t2, a1
 ; RV32I-NEXT:    andi a3, a0, 24
-; RV32I-NEXT:    lw a4, 0(a1)
-; RV32I-NEXT:    lw a5, 4(a1)
-; RV32I-NEXT:    lw a6, 8(a1)
 ; RV32I-NEXT:    xori a3, a3, 31
+; RV32I-NEXT:    lw a4, 4(a1)
+; RV32I-NEXT:    lw a5, 8(a1)
+; RV32I-NEXT:    lw a6, 0(a1)
 ; RV32I-NEXT:    lw a1, 12(a1)
-; RV32I-NEXT:    srl a7, a5, a0
-; RV32I-NEXT:    slli t0, a6, 1
-; RV32I-NEXT:    srl a4, a4, a0
-; RV32I-NEXT:    slli a5, a5, 1
+; RV32I-NEXT:    srl a7, a4, a0
+; RV32I-NEXT:    slli t0, a5, 1
 ; RV32I-NEXT:    srl a6, a6, a0
+; RV32I-NEXT:    slli a4, a4, 1
+; RV32I-NEXT:    srl a5, a5, a0
 ; RV32I-NEXT:    slli t1, a1, 1
 ; RV32I-NEXT:    srl a0, a1, a0
 ; RV32I-NEXT:    sll a1, t0, a3
-; RV32I-NEXT:    sll a5, a5, a3
+; RV32I-NEXT:    sll a4, a4, a3
 ; RV32I-NEXT:    sll a3, t1, a3
 ; RV32I-NEXT:    srli t0, a0, 16
 ; RV32I-NEXT:    srli t1, a0, 24
 ; RV32I-NEXT:    srli t2, a0, 8
 ; RV32I-NEXT:    or a1, a7, a1
-; RV32I-NEXT:    or a5, a4, a5
-; RV32I-NEXT:    or a3, a6, a3
+; RV32I-NEXT:    or a4, a6, a4
+; RV32I-NEXT:    or a3, a5, a3
 ; RV32I-NEXT:    sb a0, 12(a2)
 ; RV32I-NEXT:    sb t2, 13(a2)
 ; RV32I-NEXT:    sb t0, 14(a2)
@@ -815,18 +815,18 @@ define void @lshr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    srli a0, a3, 16
 ; RV32I-NEXT:    srli t0, a3, 24
 ; RV32I-NEXT:    srli a3, a3, 8
-; RV32I-NEXT:    srli t1, a5, 16
-; RV32I-NEXT:    srli t2, a5, 24
-; RV32I-NEXT:    srli a5, a5, 8
+; RV32I-NEXT:    srli t1, a4, 16
+; RV32I-NEXT:    srli t2, a4, 24
+; RV32I-NEXT:    srli a4, a4, 8
 ; RV32I-NEXT:    srli t3, a1, 16
 ; RV32I-NEXT:    srli t4, a1, 24
 ; RV32I-NEXT:    srli a1, a1, 8
-; RV32I-NEXT:    sb a6, 8(a2)
+; RV32I-NEXT:    sb a5, 8(a2)
 ; RV32I-NEXT:    sb a3, 9(a2)
 ; RV32I-NEXT:    sb a0, 10(a2)
 ; RV32I-NEXT:    sb t0, 11(a2)
-; RV32I-NEXT:    sb a4, 0(a2)
-; RV32I-NEXT:    sb a5, 1(a2)
+; RV32I-NEXT:    sb a6, 0(a2)
+; RV32I-NEXT:    sb a4, 1(a2)
 ; RV32I-NEXT:    sb t1, 2(a2)
 ; RV32I-NEXT:    sb t2, 3(a2)
 ; RV32I-NEXT:    sb a7, 4(a2)
@@ -868,24 +868,24 @@ define void @lshr_16bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli t2, t2, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or t1, t2, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 0(a1)
-; RV64I-NEXT:    lbu t2, 1(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t3, t3, 24
-; RV64I-NEXT:    or t0, t3, t0
+; RV64I-NEXT:    or t1, t2, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t3, t0
+; RV64I-NEXT:    lbu t0, 1(a1)
+; RV64I-NEXT:    lbu t2, 0(a1)
 ; RV64I-NEXT:    lbu t3, 2(a1)
 ; RV64I-NEXT:    lbu a1, 3(a1)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t2
 ; RV64I-NEXT:    slli t3, t3, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t3
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a4, t1, a5
-; RV64I-NEXT:    or a5, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a5, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a1, a1, 5
 ; RV64I-NEXT:    slli a6, a5, 37
@@ -899,31 +899,31 @@ define void @lshr_16bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    j .LBB7_3
 ; RV64I-NEXT:  .LBB7_2:
 ; RV64I-NEXT:    lbu a6, 1(a0)
-; RV64I-NEXT:    lbu a7, 0(a0)
-; RV64I-NEXT:    lbu t0, 2(a0)
-; RV64I-NEXT:    lbu t1, 3(a0)
+; RV64I-NEXT:    lbu a7, 2(a0)
+; RV64I-NEXT:    lbu t0, 3(a0)
+; RV64I-NEXT:    lbu t1, 0(a0)
 ; RV64I-NEXT:    slli a6, a6, 8
-; RV64I-NEXT:    or a6, a6, a7
-; RV64I-NEXT:    lbu a7, 4(a0)
-; RV64I-NEXT:    lbu t2, 5(a0)
-; RV64I-NEXT:    slli t0, t0, 16
-; RV64I-NEXT:    slli t1, t1, 24
-; RV64I-NEXT:    or t0, t1, t0
-; RV64I-NEXT:    lbu t1, 6(a0)
+; RV64I-NEXT:    slli a7, a7, 16
+; RV64I-NEXT:    slli t0, t0, 24
+; RV64I-NEXT:    or a6, a6, t1
+; RV64I-NEXT:    or a7, t0, a7
+; RV64I-NEXT:    lbu t0, 5(a0)
+; RV64I-NEXT:    lbu t1, 4(a0)
+; RV64I-NEXT:    lbu t2, 6(a0)
 ; RV64I-NEXT:    lbu a0, 7(a0)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
-; RV64I-NEXT:    slli t1, t1, 16
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
+; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, t0, a6
-; RV64I-NEXT:    not t0, a4
+; RV64I-NEXT:    or a0, a0, t2
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    not a7, a4
 ; RV64I-NEXT:    slli a5, a5, 1
-; RV64I-NEXT:    or a0, a0, a7
+; RV64I-NEXT:    or a0, a0, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a6
 ; RV64I-NEXT:    srl a0, a0, a4
-; RV64I-NEXT:    sll a4, a5, t0
+; RV64I-NEXT:    sll a4, a5, a7
 ; RV64I-NEXT:    or a0, a0, a4
 ; RV64I-NEXT:  .LBB7_3:
 ; RV64I-NEXT:    srai a3, a3, 63
@@ -1083,24 +1083,24 @@ define void @shl_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli t2, t2, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or t1, t2, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 0(a1)
-; RV64I-NEXT:    lbu t2, 1(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t3, t3, 24
-; RV64I-NEXT:    or t0, t3, t0
+; RV64I-NEXT:    or t1, t2, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t3, t0
+; RV64I-NEXT:    lbu t0, 1(a1)
+; RV64I-NEXT:    lbu t2, 0(a1)
 ; RV64I-NEXT:    lbu t3, 2(a1)
 ; RV64I-NEXT:    lbu a1, 3(a1)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t2
 ; RV64I-NEXT:    slli t3, t3, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t3
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a4, t1, a5
-; RV64I-NEXT:    or a5, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a5, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a1, a1, 3
 ; RV64I-NEXT:    slli a6, a5, 35
@@ -1114,31 +1114,31 @@ define void @shl_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    j .LBB8_3
 ; RV64I-NEXT:  .LBB8_2:
 ; RV64I-NEXT:    lbu a6, 9(a0)
-; RV64I-NEXT:    lbu a7, 8(a0)
-; RV64I-NEXT:    lbu t0, 10(a0)
-; RV64I-NEXT:    lbu t1, 11(a0)
+; RV64I-NEXT:    lbu a7, 10(a0)
+; RV64I-NEXT:    lbu t0, 11(a0)
+; RV64I-NEXT:    lbu t1, 8(a0)
 ; RV64I-NEXT:    slli a6, a6, 8
-; RV64I-NEXT:    or a6, a6, a7
-; RV64I-NEXT:    lbu a7, 12(a0)
-; RV64I-NEXT:    lbu t2, 13(a0)
-; RV64I-NEXT:    slli t0, t0, 16
-; RV64I-NEXT:    slli t1, t1, 24
-; RV64I-NEXT:    or t0, t1, t0
-; RV64I-NEXT:    lbu t1, 14(a0)
+; RV64I-NEXT:    slli a7, a7, 16
+; RV64I-NEXT:    slli t0, t0, 24
+; RV64I-NEXT:    or a6, a6, t1
+; RV64I-NEXT:    or a7, t0, a7
+; RV64I-NEXT:    lbu t0, 13(a0)
+; RV64I-NEXT:    lbu t1, 12(a0)
+; RV64I-NEXT:    lbu t2, 14(a0)
 ; RV64I-NEXT:    lbu a0, 15(a0)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
-; RV64I-NEXT:    slli t1, t1, 16
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
+; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, t0, a6
-; RV64I-NEXT:    not t0, a4
+; RV64I-NEXT:    or a0, a0, t2
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    not a7, a4
 ; RV64I-NEXT:    srli a5, a5, 1
-; RV64I-NEXT:    or a0, a0, a7
+; RV64I-NEXT:    or a0, a0, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a6
 ; RV64I-NEXT:    sll a0, a0, a4
-; RV64I-NEXT:    srl a4, a5, t0
+; RV64I-NEXT:    srl a4, a5, a7
 ; RV64I-NEXT:    or a0, a0, a4
 ; RV64I-NEXT:  .LBB8_3:
 ; RV64I-NEXT:    srai a3, a3, 63
@@ -1322,24 +1322,24 @@ define void @shl_16bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) nounw
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli t2, t2, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or t1, t2, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 0(a1)
-; RV64I-NEXT:    lbu t2, 1(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t3, t3, 24
-; RV64I-NEXT:    or t0, t3, t0
+; RV64I-NEXT:    or t1, t2, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t3, t0
+; RV64I-NEXT:    lbu t0, 1(a1)
+; RV64I-NEXT:    lbu t2, 0(a1)
 ; RV64I-NEXT:    lbu t3, 2(a1)
 ; RV64I-NEXT:    lbu a1, 3(a1)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t2
 ; RV64I-NEXT:    slli t3, t3, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t3
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a4, t1, a5
-; RV64I-NEXT:    or a5, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a5, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a1, a1, 5
 ; RV64I-NEXT:    slli a6, a5, 37
@@ -1353,31 +1353,31 @@ define void @shl_16bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) nounw
 ; RV64I-NEXT:    j .LBB9_3
 ; RV64I-NEXT:  .LBB9_2:
 ; RV64I-NEXT:    lbu a6, 9(a0)
-; RV64I-NEXT:    lbu a7, 8(a0)
-; RV64I-NEXT:    lbu t0, 10(a0)
-; RV64I-NEXT:    lbu t1, 11(a0)
+; RV64I-NEXT:    lbu a7, 10(a0)
+; RV64I-NEXT:    lbu t0, 11(a0)
+; RV64I-NEXT:    lbu t1, 8(a0)
 ; RV64I-NEXT:    slli a6, a6, 8
-; RV64I-NEXT:    or a6, a6, a7
-; RV64I-NEXT:    lbu a7, 12(a0)
-; RV64I-NEXT:    lbu t2, 13(a0)
-; RV64I-NEXT:    slli t0, t0, 16
-; RV64I-NEXT:    slli t1, t1, 24
-; RV64I-NEXT:    or t0, t1, t0
-; RV64I-NEXT:    lbu t1, 14(a0)
+; RV64I-NEXT:    slli a7, a7, 16
+; RV64I-NEXT:    slli t0, t0, 24
+; RV64I-NEXT:    or a6, a6, t1
+; RV64I-NEXT:    or a7, t0, a7
+; RV64I-NEXT:    lbu t0, 13(a0)
+; RV64I-NEXT:    lbu t1, 12(a0)
+; RV64I-NEXT:    lbu t2, 14(a0)
 ; RV64I-NEXT:    lbu a0, 15(a0)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
-; RV64I-NEXT:    slli t1, t1, 16
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
+; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, t0, a6
-; RV64I-NEXT:    not t0, a4
+; RV64I-NEXT:    or a0, a0, t2
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    not a7, a4
 ; RV64I-NEXT:    srli a5, a5, 1
-; RV64I-NEXT:    or a0, a0, a7
+; RV64I-NEXT:    or a0, a0, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a6
 ; RV64I-NEXT:    sll a0, a0, a4
-; RV64I-NEXT:    srl a4, a5, t0
+; RV64I-NEXT:    srl a4, a5, a7
 ; RV64I-NEXT:    or a0, a0, a4
 ; RV64I-NEXT:  .LBB9_3:
 ; RV64I-NEXT:    srai a3, a3, 63
@@ -1538,24 +1538,24 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli t2, t2, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or t1, t2, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 0(a1)
-; RV64I-NEXT:    lbu t2, 1(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t3, t3, 24
-; RV64I-NEXT:    or t0, t3, t0
+; RV64I-NEXT:    or t1, t2, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t3, t0
+; RV64I-NEXT:    lbu t0, 1(a1)
+; RV64I-NEXT:    lbu t2, 0(a1)
 ; RV64I-NEXT:    lbu t3, 2(a1)
 ; RV64I-NEXT:    lbu a1, 3(a1)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t2
 ; RV64I-NEXT:    slli t3, t3, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t3
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a5, t1, a5
-; RV64I-NEXT:    or a4, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a4, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a6, a5, 32
 ; RV64I-NEXT:    slli a1, a1, 3
 ; RV64I-NEXT:    slli a7, a4, 35
@@ -1571,31 +1571,31 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    j .LBB10_3
 ; RV64I-NEXT:  .LBB10_2:
 ; RV64I-NEXT:    lbu a5, 1(a0)
-; RV64I-NEXT:    lbu a6, 0(a0)
-; RV64I-NEXT:    lbu a7, 2(a0)
-; RV64I-NEXT:    lbu t0, 3(a0)
+; RV64I-NEXT:    lbu a6, 2(a0)
+; RV64I-NEXT:    lbu a7, 3(a0)
+; RV64I-NEXT:    lbu t0, 0(a0)
 ; RV64I-NEXT:    slli a5, a5, 8
-; RV64I-NEXT:    or a5, a5, a6
-; RV64I-NEXT:    lbu a6, 4(a0)
-; RV64I-NEXT:    lbu t1, 5(a0)
-; RV64I-NEXT:    slli a7, a7, 16
-; RV64I-NEXT:    slli t0, t0, 24
-; RV64I-NEXT:    or a7, t0, a7
-; RV64I-NEXT:    lbu t0, 6(a0)
+; RV64I-NEXT:    slli a6, a6, 16
+; RV64I-NEXT:    slli a7, a7, 24
+; RV64I-NEXT:    or a5, a5, t0
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    lbu a7, 5(a0)
+; RV64I-NEXT:    lbu t0, 4(a0)
+; RV64I-NEXT:    lbu t1, 6(a0)
 ; RV64I-NEXT:    lbu a0, 7(a0)
-; RV64I-NEXT:    slli t1, t1, 8
-; RV64I-NEXT:    or a6, t1, a6
-; RV64I-NEXT:    slli t0, t0, 16
+; RV64I-NEXT:    slli a7, a7, 8
+; RV64I-NEXT:    or a7, a7, t0
+; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, t0
-; RV64I-NEXT:    or a5, a7, a5
-; RV64I-NEXT:    not a7, a3
+; RV64I-NEXT:    or a0, a0, t1
+; RV64I-NEXT:    or a5, a6, a5
+; RV64I-NEXT:    not a6, a3
 ; RV64I-NEXT:    slli a4, a4, 1
-; RV64I-NEXT:    or a0, a0, a6
+; RV64I-NEXT:    or a0, a0, a7
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a5
 ; RV64I-NEXT:    srl a0, a0, a3
-; RV64I-NEXT:    sll a3, a4, a7
+; RV64I-NEXT:    sll a3, a4, a6
 ; RV64I-NEXT:    or a0, a0, a3
 ; RV64I-NEXT:  .LBB10_3:
 ; RV64I-NEXT:    srli a3, a1, 56
@@ -1662,20 +1662,20 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    slli t3, t3, 8
 ; RV32I-NEXT:    slli t4, t4, 16
 ; RV32I-NEXT:    slli t5, t5, 24
+; RV32I-NEXT:    slli t1, t1, 8
 ; RV32I-NEXT:    or a4, t3, a4
 ; RV32I-NEXT:    or t3, t5, t4
-; RV32I-NEXT:    lbu t4, 0(a1)
-; RV32I-NEXT:    lbu t5, 1(a1)
-; RV32I-NEXT:    slli t1, t1, 8
 ; RV32I-NEXT:    or t0, t1, t0
-; RV32I-NEXT:    lbu t1, 2(a1)
+; RV32I-NEXT:    lbu t1, 1(a1)
+; RV32I-NEXT:    lbu t4, 0(a1)
+; RV32I-NEXT:    lbu t5, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli t5, t5, 8
-; RV32I-NEXT:    or t4, t5, t4
-; RV32I-NEXT:    slli t1, t1, 16
+; RV32I-NEXT:    slli t1, t1, 8
+; RV32I-NEXT:    or t1, t1, t4
+; RV32I-NEXT:    slli t5, t5, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, t1
-; RV32I-NEXT:    mv t1, sp
+; RV32I-NEXT:    or a1, a1, t5
+; RV32I-NEXT:    mv t4, sp
 ; RV32I-NEXT:    slli t2, t2, 16
 ; RV32I-NEXT:    slli a0, a0, 24
 ; RV32I-NEXT:    or t2, a0, t2
@@ -1684,7 +1684,7 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    or a5, a7, a6
 ; RV32I-NEXT:    or a4, t3, a4
 ; RV32I-NEXT:    or a6, t2, t0
-; RV32I-NEXT:    or a1, a1, t4
+; RV32I-NEXT:    or a1, a1, t1
 ; RV32I-NEXT:    sw a0, 16(sp)
 ; RV32I-NEXT:    sw a0, 20(sp)
 ; RV32I-NEXT:    sw a0, 24(sp)
@@ -1695,29 +1695,29 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    sw a6, 12(sp)
 ; RV32I-NEXT:    slli a0, a1, 3
 ; RV32I-NEXT:    andi a1, a1, 12
-; RV32I-NEXT:    add a1, t1, a1
+; RV32I-NEXT:    add a1, t4, a1
 ; RV32I-NEXT:    andi a3, a0, 24
-; RV32I-NEXT:    lw a4, 0(a1)
-; RV32I-NEXT:    lw a5, 4(a1)
-; RV32I-NEXT:    lw a6, 8(a1)
 ; RV32I-NEXT:    xori a3, a3, 31
+; RV32I-NEXT:    lw a4, 4(a1)
+; RV32I-NEXT:    lw a5, 8(a1)
+; RV32I-NEXT:    lw a6, 0(a1)
 ; RV32I-NEXT:    lw a1, 12(a1)
-; RV32I-NEXT:    srl a7, a5, a0
-; RV32I-NEXT:    slli t0, a6, 1
-; RV32I-NEXT:    srl a4, a4, a0
-; RV32I-NEXT:    slli a5, a5, 1
+; RV32I-NEXT:    srl a7, a4, a0
+; RV32I-NEXT:    slli t0, a5, 1
 ; RV32I-NEXT:    srl a6, a6, a0
+; RV32I-NEXT:    slli a4, a4, 1
+; RV32I-NEXT:    srl a5, a5, a0
 ; RV32I-NEXT:    slli t1, a1, 1
 ; RV32I-NEXT:    sra a0, a1, a0
 ; RV32I-NEXT:    sll a1, t0, a3
-; RV32I-NEXT:    sll a5, a5, a3
+; RV32I-NEXT:    sll a4, a4, a3
 ; RV32I-NEXT:    sll a3, t1, a3
 ; RV32I-NEXT:    srli t0, a0, 16
 ; RV32I-NEXT:    srli t1, a0, 24
 ; RV32I-NEXT:    srli t2, a0, 8
 ; RV32I-NEXT:    or a1, a7, a1
-; RV32I-NEXT:    or a5, a4, a5
-; RV32I-NEXT:    or a3, a6, a3
+; RV32I-NEXT:    or a4, a6, a4
+; RV32I-NEXT:    or a3, a5, a3
 ; RV32I-NEXT:    sb a0, 12(a2)
 ; RV32I-NEXT:    sb t2, 13(a2)
 ; RV32I-NEXT:    sb t0, 14(a2)
@@ -1725,18 +1725,18 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    srli a0, a3, 16
 ; RV32I-NEXT:    srli t0, a3, 24
 ; RV32I-NEXT:    srli a3, a3, 8
-; RV32I-NEXT:    srli t1, a5, 16
-; RV32I-NEXT:    srli t2, a5, 24
-; RV32I-NEXT:    srli a5, a5, 8
+; RV32I-NEXT:    srli t1, a4, 16
+; RV32I-NEXT:    srli t2, a4, 24
+; RV32I-NEXT:    srli a4, a4, 8
 ; RV32I-NEXT:    srli t3, a1, 16
 ; RV32I-NEXT:    srli t4, a1, 24
 ; RV32I-NEXT:    srli a1, a1, 8
-; RV32I-NEXT:    sb a6, 8(a2)
+; RV32I-NEXT:    sb a5, 8(a2)
 ; RV32I-NEXT:    sb a3, 9(a2)
 ; RV32I-NEXT:    sb a0, 10(a2)
 ; RV32I-NEXT:    sb t0, 11(a2)
-; RV32I-NEXT:    sb a4, 0(a2)
-; RV32I-NEXT:    sb a5, 1(a2)
+; RV32I-NEXT:    sb a6, 0(a2)
+; RV32I-NEXT:    sb a4, 1(a2)
 ; RV32I-NEXT:    sb t1, 2(a2)
 ; RV32I-NEXT:    sb t2, 3(a2)
 ; RV32I-NEXT:    sb a7, 4(a2)
@@ -1778,24 +1778,24 @@ define void @ashr_16bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli t2, t2, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or t1, t2, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 0(a1)
-; RV64I-NEXT:    lbu t2, 1(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t3, t3, 24
-; RV64I-NEXT:    or t0, t3, t0
+; RV64I-NEXT:    or t1, t2, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t3, t0
+; RV64I-NEXT:    lbu t0, 1(a1)
+; RV64I-NEXT:    lbu t2, 0(a1)
 ; RV64I-NEXT:    lbu t3, 2(a1)
 ; RV64I-NEXT:    lbu a1, 3(a1)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t2
 ; RV64I-NEXT:    slli t3, t3, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t3
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a5, t1, a5
-; RV64I-NEXT:    or a4, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a4, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a6, a5, 32
 ; RV64I-NEXT:    slli a1, a1, 5
 ; RV64I-NEXT:    slli a7, a4, 37
@@ -1811,31 +1811,31 @@ define void @ashr_16bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    j .LBB11_3
 ; RV64I-NEXT:  .LBB11_2:
 ; RV64I-NEXT:    lbu a5, 1(a0)
-; RV64I-NEXT:    lbu a6, 0(a0)
-; RV64I-NEXT:    lbu a7, 2(a0)
-; RV64I-NEXT:    lbu t0, 3(a0)
+; RV64I-NEXT:    lbu a6, 2(a0)
+; RV64I-NEXT:    lbu a7, 3(a0)
+; RV64I-NEXT:    lbu t0, 0(a0)
 ; RV64I-NEXT:    slli a5, a5, 8
-; RV64I-NEXT:    or a5, a5, a6
-; RV64I-NEXT:    lbu a6, 4(a0)
-; RV64I-NEXT:    lbu t1, 5(a0)
-; RV64I-NEXT:    slli a7, a7, 16
-; RV64I-NEXT:    slli t0, t0, 24
-; RV64I-NEXT:    or a7, t0, a7
-; RV64I-NEXT:    lbu t0, 6(a0)
+; RV64I-NEXT:    slli a6, a6, 16
+; RV64I-NEXT:    slli a7, a7, 24
+; RV64I-NEXT:    or a5, a5, t0
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    lbu a7, 5(a0)
+; RV64I-NEXT:    lbu t0, 4(a0)
+; RV64I-NEXT:    lbu t1, 6(a0)
 ; RV64I-NEXT:    lbu a0, 7(a0)
-; RV64I-NEXT:    slli t1, t1, 8
-; RV64I-NEXT:    or a6, t1, a6
-; RV64I-NEXT:    slli t0, t0, 16
+; RV64I-NEXT:    slli a7, a7, 8
+; RV64I-NEXT:    or a7, a7, t0
+; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, t0
-; RV64I-NEXT:    or a5, a7, a5
-; RV64I-NEXT:    not a7, a3
+; RV64I-NEXT:    or a0, a0, t1
+; RV64I-NEXT:    or a5, a6, a5
+; RV64I-NEXT:    not a6, a3
 ; RV64I-NEXT:    slli a4, a4, 1
-; RV64I-NEXT:    or a0, a0, a6
+; RV64I-NEXT:    or a0, a0, a7
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a5
 ; RV64I-NEXT:    srl a0, a0, a3
-; RV64I-NEXT:    sll a3, a4, a7
+; RV64I-NEXT:    sll a3, a4, a6
 ; RV64I-NEXT:    or a0, a0, a3
 ; RV64I-NEXT:  .LBB11_3:
 ; RV64I-NEXT:    srli a3, a1, 56
@@ -2061,17 +2061,17 @@ define void @lshr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli s5, s5, 16
 ; RV64I-NEXT:    slli s6, s6, 24
 ; RV64I-NEXT:    slli s3, s3, 8
-; RV64I-NEXT:    or s5, s6, s5
-; RV64I-NEXT:    or s2, s3, s2
-; RV64I-NEXT:    lbu s3, 4(a1)
-; RV64I-NEXT:    lbu s6, 5(a1)
 ; RV64I-NEXT:    slli s4, s4, 16
 ; RV64I-NEXT:    slli s7, s7, 24
-; RV64I-NEXT:    or s4, s7, s4
+; RV64I-NEXT:    or s5, s6, s5
+; RV64I-NEXT:    or s2, s3, s2
+; RV64I-NEXT:    or s3, s7, s4
+; RV64I-NEXT:    lbu s4, 5(a1)
+; RV64I-NEXT:    lbu s6, 4(a1)
 ; RV64I-NEXT:    lbu s7, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli s6, s6, 8
-; RV64I-NEXT:    or s3, s6, s3
+; RV64I-NEXT:    slli s4, s4, 8
+; RV64I-NEXT:    or s4, s4, s6
 ; RV64I-NEXT:    sd zero, 32(sp)
 ; RV64I-NEXT:    sd zero, 40(sp)
 ; RV64I-NEXT:    sd zero, 48(sp)
@@ -2088,8 +2088,8 @@ define void @lshr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    or a0, a0, t5
 ; RV64I-NEXT:    or t0, s0, t6
 ; RV64I-NEXT:    or t1, s5, s1
-; RV64I-NEXT:    or t2, s4, s2
-; RV64I-NEXT:    or a1, a1, s3
+; RV64I-NEXT:    or t2, s3, s2
+; RV64I-NEXT:    or a1, a1, s4
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a6, a6, 32
 ; RV64I-NEXT:    slli a0, a0, 32
@@ -2108,22 +2108,22 @@ define void @lshr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    andi a1, a1, 24
 ; RV64I-NEXT:    add a1, s6, a1
 ; RV64I-NEXT:    andi a0, a4, 56
-; RV64I-NEXT:    ld a3, 0(a1)
-; RV64I-NEXT:    ld a5, 8(a1)
+; RV64I-NEXT:    xori a5, a0, 63
+; RV64I-NEXT:    ld a3, 8(a1)
 ; RV64I-NEXT:    ld a6, 16(a1)
-; RV64I-NEXT:    xori a7, a0, 63
+; RV64I-NEXT:    ld a7, 0(a1)
 ; RV64I-NEXT:    ld t0, 24(a1)
-; RV64I-NEXT:    srl a0, a5, a4
+; RV64I-NEXT:    srl a0, a3, a4
 ; RV64I-NEXT:    slli t1, a6, 1
-; RV64I-NEXT:    srl a1, a3, a4
-; RV64I-NEXT:    slli a5, a5, 1
+; RV64I-NEXT:    srl a1, a7, a4
+; RV64I-NEXT:    slli a7, a3, 1
 ; RV64I-NEXT:    srl a3, a6, a4
 ; RV64I-NEXT:    slli a6, t0, 1
 ; RV64I-NEXT:    srl t0, t0, a4
-; RV64I-NEXT:    sll a4, t1, a7
-; RV64I-NEXT:    sll a5, a5, a7
-; RV64I-NEXT:    sll a6, a6, a7
-; RV64I-NEXT:    srli a7, t0, 56
+; RV64I-NEXT:    sll a4, t1, a5
+; RV64I-NEXT:    sll a7, a7, a5
+; RV64I-NEXT:    sll a5, a6, a5
+; RV64I-NEXT:    srli a6, t0, 56
 ; RV64I-NEXT:    srli t1, t0, 48
 ; RV64I-NEXT:    srli t2, t0, 40
 ; RV64I-NEXT:    srli t3, t0, 32
@@ -2131,40 +2131,40 @@ define void @lshr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    srli t5, t0, 16
 ; RV64I-NEXT:    srli t6, t0, 8
 ; RV64I-NEXT:    or a4, a0, a4
-; RV64I-NEXT:    or a5, a1, a5
-; RV64I-NEXT:    or a6, a3, a6
+; RV64I-NEXT:    or a7, a1, a7
+; RV64I-NEXT:    or a5, a3, a5
 ; RV64I-NEXT:    sb t3, 28(a2)
 ; RV64I-NEXT:    sb t2, 29(a2)
 ; RV64I-NEXT:    sb t1, 30(a2)
-; RV64I-NEXT:    sb a7, 31(a2)
+; RV64I-NEXT:    sb a6, 31(a2)
 ; RV64I-NEXT:    sb t0, 24(a2)
 ; RV64I-NEXT:    sb t6, 25(a2)
 ; RV64I-NEXT:    sb t5, 26(a2)
 ; RV64I-NEXT:    sb t4, 27(a2)
-; RV64I-NEXT:    srli a7, a6, 56
-; RV64I-NEXT:    srli t0, a6, 48
-; RV64I-NEXT:    srli t1, a6, 40
-; RV64I-NEXT:    srli t2, a6, 32
-; RV64I-NEXT:    srli t3, a6, 24
-; RV64I-NEXT:    srli t4, a6, 16
-; RV64I-NEXT:    srli a6, a6, 8
-; RV64I-NEXT:    srli t5, a5, 56
-; RV64I-NEXT:    srli t6, a5, 48
-; RV64I-NEXT:    srli s0, a5, 40
-; RV64I-NEXT:    srli s1, a5, 32
-; RV64I-NEXT:    srli s2, a5, 24
-; RV64I-NEXT:    srli s3, a5, 16
+; RV64I-NEXT:    srli a6, a5, 56
+; RV64I-NEXT:    srli t0, a5, 48
+; RV64I-NEXT:    srli t1, a5, 40
+; RV64I-NEXT:    srli t2, a5, 32
+; RV64I-NEXT:    srli t3, a5, 24
+; RV64I-NEXT:    srli t4, a5, 16
 ; RV64I-NEXT:    srli a5, a5, 8
+; RV64I-NEXT:    srli t5, a7, 56
+; RV64I-NEXT:    srli t6, a7, 48
+; RV64I-NEXT:    srli s0, a7, 40
+; RV64I-NEXT:    srli s1, a7, 32
+; RV64I-NEXT:    srli s2, a7, 24
+; RV64I-NEXT:    srli s3, a7, 16
+; RV64I-NEXT:    srli a7, a7, 8
 ; RV64I-NEXT:    srli s4, a4, 56
 ; RV64I-NEXT:    srli s5, a4, 48
 ; RV64I-NEXT:    srli s6, a4, 40
 ; RV64I-NEXT:    sb t2, 20(a2)
 ; RV64I-NEXT:    sb t1, 21(a2)
 ; RV64I-NEXT:    sb t0, 22(a2)
-; RV64I-NEXT:    sb a7, 23(a2)
-; RV64I-NEXT:    srli a7, a4, 32
+; RV64I-NEXT:    sb a6, 23(a2)
+; RV64I-NEXT:    srli a6, a4, 32
 ; RV64I-NEXT:    sb a3, 16(a2)
-; RV64I-NEXT:    sb a6, 17(a2)
+; RV64I-NEXT:    sb a5, 17(a2)
 ; RV64I-NEXT:    sb t4, 18(a2)
 ; RV64I-NEXT:    sb t3, 19(a2)
 ; RV64I-NEXT:    srli a3, a4, 24
@@ -2172,19 +2172,19 @@ define void @lshr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    sb s0, 5(a2)
 ; RV64I-NEXT:    sb t6, 6(a2)
 ; RV64I-NEXT:    sb t5, 7(a2)
-; RV64I-NEXT:    srli a6, a4, 16
+; RV64I-NEXT:    srli a5, a4, 16
 ; RV64I-NEXT:    srli a4, a4, 8
 ; RV64I-NEXT:    sb a1, 0(a2)
-; RV64I-NEXT:    sb a5, 1(a2)
+; RV64I-NEXT:    sb a7, 1(a2)
 ; RV64I-NEXT:    sb s3, 2(a2)
 ; RV64I-NEXT:    sb s2, 3(a2)
-; RV64I-NEXT:    sb a7, 12(a2)
+; RV64I-NEXT:    sb a6, 12(a2)
 ; RV64I-NEXT:    sb s6, 13(a2)
 ; RV64I-NEXT:    sb s5, 14(a2)
 ; RV64I-NEXT:    sb s4, 15(a2)
 ; RV64I-NEXT:    sb a0, 8(a2)
 ; RV64I-NEXT:    sb a4, 9(a2)
-; RV64I-NEXT:    sb a6, 10(a2)
+; RV64I-NEXT:    sb a5, 10(a2)
 ; RV64I-NEXT:    sb a3, 11(a2)
 ; RV64I-NEXT:    ld s0, 152(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s1, 144(sp) # 8-byte Folded Reload
@@ -2539,17 +2539,17 @@ define void @lshr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    slli s5, s5, 16
 ; RV64I-NEXT:    slli s6, s6, 24
 ; RV64I-NEXT:    slli s3, s3, 8
-; RV64I-NEXT:    or s5, s6, s5
-; RV64I-NEXT:    or s2, s3, s2
-; RV64I-NEXT:    lbu s3, 4(a1)
-; RV64I-NEXT:    lbu s6, 5(a1)
 ; RV64I-NEXT:    slli s4, s4, 16
 ; RV64I-NEXT:    slli s7, s7, 24
-; RV64I-NEXT:    or s4, s7, s4
+; RV64I-NEXT:    or s5, s6, s5
+; RV64I-NEXT:    or s2, s3, s2
+; RV64I-NEXT:    or s3, s7, s4
+; RV64I-NEXT:    lbu s4, 5(a1)
+; RV64I-NEXT:    lbu s6, 4(a1)
 ; RV64I-NEXT:    lbu s7, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli s6, s6, 8
-; RV64I-NEXT:    or s3, s6, s3
+; RV64I-NEXT:    slli s4, s4, 8
+; RV64I-NEXT:    or s4, s4, s6
 ; RV64I-NEXT:    sd zero, 32(sp)
 ; RV64I-NEXT:    sd zero, 40(sp)
 ; RV64I-NEXT:    sd zero, 48(sp)
@@ -2566,8 +2566,8 @@ define void @lshr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    or a0, a0, t5
 ; RV64I-NEXT:    or t0, s0, t6
 ; RV64I-NEXT:    or t1, s5, s1
-; RV64I-NEXT:    or t2, s4, s2
-; RV64I-NEXT:    or a1, a1, s3
+; RV64I-NEXT:    or t2, s3, s2
+; RV64I-NEXT:    or a1, a1, s4
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a6, a6, 32
 ; RV64I-NEXT:    slli a0, a0, 32
@@ -2587,24 +2587,24 @@ define void @lshr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    andi a1, a1, 24
 ; RV64I-NEXT:    andi a0, a3, 32
 ; RV64I-NEXT:    add a1, s6, a1
-; RV64I-NEXT:    ld a4, 0(a1)
+; RV64I-NEXT:    xori a4, a0, 63
 ; RV64I-NEXT:    ld a5, 8(a1)
 ; RV64I-NEXT:    ld a6, 16(a1)
-; RV64I-NEXT:    xori a7, a0, 63
+; RV64I-NEXT:    ld a7, 0(a1)
 ; RV64I-NEXT:    ld t0, 24(a1)
 ; RV64I-NEXT:    srl a0, a5, a3
 ; RV64I-NEXT:    slli t1, a6, 1
-; RV64I-NEXT:    srl a1, a4, a3
+; RV64I-NEXT:    srl a1, a7, a3
 ; RV64I-NEXT:    slli a5, a5, 1
-; RV64I-NEXT:    srl a4, a6, a3
-; RV64I-NEXT:    slli a6, t0, 1
+; RV64I-NEXT:    srl a6, a6, a3
+; RV64I-NEXT:    slli a7, t0, 1
 ; RV64I-NEXT:    srl a3, t0, a3
-; RV64I-NEXT:    sll t0, t1, a7
-; RV64I-NEXT:    sll a5, a5, a7
-; RV64I-NEXT:    sll a6, a6, a7
-; RV64I-NEXT:    srli a7, a4, 24
-; RV64I-NEXT:    srli t1, a4, 16
-; RV64I-NEXT:    srli t2, a4, 8
+; RV64I-NEXT:    sll t0, t1, a4
+; RV64I-NEXT:    sll a5, a5, a4
+; RV64I-NEXT:    sll a4, a7, a4
+; RV64I-NEXT:    srli a7, a6, 24
+; RV64I-NEXT:    srli t1, a6, 16
+; RV64I-NEXT:    srli t2, a6, 8
 ; RV64I-NEXT:    srli t3, a3, 56
 ; RV64I-NEXT:    srli t4, a3, 48
 ; RV64I-NEXT:    srli t5, a3, 40
@@ -2616,12 +2616,12 @@ define void @lshr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    srli s4, a1, 16
 ; RV64I-NEXT:    srli s5, a1, 8
 ; RV64I-NEXT:    srli s6, a0, 24
-; RV64I-NEXT:    or a6, a4, a6
-; RV64I-NEXT:    sb a4, 16(a2)
+; RV64I-NEXT:    or a4, a6, a4
+; RV64I-NEXT:    sb a6, 16(a2)
 ; RV64I-NEXT:    sb t2, 17(a2)
 ; RV64I-NEXT:    sb t1, 18(a2)
 ; RV64I-NEXT:    sb a7, 19(a2)
-; RV64I-NEXT:    srli a4, a0, 16
+; RV64I-NEXT:    srli a6, a0, 16
 ; RV64I-NEXT:    sb t6, 28(a2)
 ; RV64I-NEXT:    sb t5, 29(a2)
 ; RV64I-NEXT:    sb t4, 30(a2)
@@ -2639,12 +2639,12 @@ define void @lshr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    sb s3, 3(a2)
 ; RV64I-NEXT:    sb a0, 8(a2)
 ; RV64I-NEXT:    sb a7, 9(a2)
-; RV64I-NEXT:    sb a4, 10(a2)
+; RV64I-NEXT:    sb a6, 10(a2)
 ; RV64I-NEXT:    sb s6, 11(a2)
-; RV64I-NEXT:    srli a0, a6, 56
-; RV64I-NEXT:    srli a1, a6, 48
-; RV64I-NEXT:    srli a3, a6, 40
-; RV64I-NEXT:    srli a4, a6, 32
+; RV64I-NEXT:    srli a0, a4, 56
+; RV64I-NEXT:    srli a1, a4, 48
+; RV64I-NEXT:    srli a3, a4, 40
+; RV64I-NEXT:    srli a4, a4, 32
 ; RV64I-NEXT:    srli a6, a5, 56
 ; RV64I-NEXT:    srli a7, a5, 48
 ; RV64I-NEXT:    srli t1, a5, 40
@@ -2797,13 +2797,13 @@ define void @lshr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV32I-NEXT:    sw t0, 12(sp)
 ; RV32I-NEXT:    sw t1, 16(sp)
 ; RV32I-NEXT:    sw a5, 20(sp)
-; RV32I-NEXT:    lw a6, 16(t6)
-; RV32I-NEXT:    lw a5, 20(t6)
-; RV32I-NEXT:    lw a7, 24(t6)
 ; RV32I-NEXT:    lw a1, 0(t6)
 ; RV32I-NEXT:    lw a0, 4(t6)
 ; RV32I-NEXT:    lw a4, 8(t6)
 ; RV32I-NEXT:    lw a3, 12(t6)
+; RV32I-NEXT:    lw a7, 24(t6)
+; RV32I-NEXT:    lw a5, 20(t6)
+; RV32I-NEXT:    lw a6, 16(t6)
 ; RV32I-NEXT:    lw t0, 28(t6)
 ; RV32I-NEXT:    srli t1, a7, 24
 ; RV32I-NEXT:    srli t2, a7, 16
@@ -3197,13 +3197,13 @@ define void @lshr_32bytes_dwordOff(ptr %src.ptr, ptr %dwordOff.ptr, ptr %dst) no
 ; RV32I-NEXT:    sw t0, 12(sp)
 ; RV32I-NEXT:    sw t1, 16(sp)
 ; RV32I-NEXT:    sw a5, 20(sp)
-; RV32I-NEXT:    lw a6, 16(t6)
-; RV32I-NEXT:    lw a5, 20(t6)
-; RV32I-NEXT:    lw a7, 24(t6)
 ; RV32I-NEXT:    lw a1, 0(t6)
 ; RV32I-NEXT:    lw a0, 4(t6)
 ; RV32I-NEXT:    lw a4, 8(t6)
 ; RV32I-NEXT:    lw a3, 12(t6)
+; RV32I-NEXT:    lw a7, 24(t6)
+; RV32I-NEXT:    lw a5, 20(t6)
+; RV32I-NEXT:    lw a6, 16(t6)
 ; RV32I-NEXT:    lw t0, 28(t6)
 ; RV32I-NEXT:    srli t1, a7, 24
 ; RV32I-NEXT:    srli t2, a7, 16
@@ -3376,17 +3376,17 @@ define void @shl_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli s5, s5, 16
 ; RV64I-NEXT:    slli s6, s6, 24
 ; RV64I-NEXT:    slli s3, s3, 8
-; RV64I-NEXT:    or s5, s6, s5
-; RV64I-NEXT:    or s2, s3, s2
-; RV64I-NEXT:    lbu s3, 4(a1)
-; RV64I-NEXT:    lbu s6, 5(a1)
 ; RV64I-NEXT:    slli s4, s4, 16
 ; RV64I-NEXT:    slli s7, s7, 24
-; RV64I-NEXT:    or s4, s7, s4
+; RV64I-NEXT:    or s5, s6, s5
+; RV64I-NEXT:    or s2, s3, s2
+; RV64I-NEXT:    or s3, s7, s4
+; RV64I-NEXT:    lbu s4, 5(a1)
+; RV64I-NEXT:    lbu s6, 4(a1)
 ; RV64I-NEXT:    lbu s7, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli s6, s6, 8
-; RV64I-NEXT:    or s3, s6, s3
+; RV64I-NEXT:    slli s4, s4, 8
+; RV64I-NEXT:    or s4, s4, s6
 ; RV64I-NEXT:    sd zero, 0(sp)
 ; RV64I-NEXT:    sd zero, 8(sp)
 ; RV64I-NEXT:    sd zero, 16(sp)
@@ -3403,8 +3403,8 @@ define void @shl_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    or a0, a0, t5
 ; RV64I-NEXT:    or t0, s0, t6
 ; RV64I-NEXT:    or t1, s5, s1
-; RV64I-NEXT:    or t2, s4, s2
-; RV64I-NEXT:    or a1, a1, s3
+; RV64I-NEXT:    or t2, s3, s2
+; RV64I-NEXT:    or a1, a1, s4
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a6, a6, 32
 ; RV64I-NEXT:    slli a0, a0, 32
@@ -3854,17 +3854,17 @@ define void @shl_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) nounw
 ; RV64I-NEXT:    slli s5, s5, 16
 ; RV64I-NEXT:    slli s6, s6, 24
 ; RV64I-NEXT:    slli s3, s3, 8
-; RV64I-NEXT:    or s5, s6, s5
-; RV64I-NEXT:    or s2, s3, s2
-; RV64I-NEXT:    lbu s3, 4(a1)
-; RV64I-NEXT:    lbu s6, 5(a1)
 ; RV64I-NEXT:    slli s4, s4, 16
 ; RV64I-NEXT:    slli s7, s7, 24
-; RV64I-NEXT:    or s4, s7, s4
+; RV64I-NEXT:    or s5, s6, s5
+; RV64I-NEXT:    or s2, s3, s2
+; RV64I-NEXT:    or s3, s7, s4
+; RV64I-NEXT:    lbu s4, 5(a1)
+; RV64I-NEXT:    lbu s6, 4(a1)
 ; RV64I-NEXT:    lbu s7, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli s6, s6, 8
-; RV64I-NEXT:    or s3, s6, s3
+; RV64I-NEXT:    slli s4, s4, 8
+; RV64I-NEXT:    or s4, s4, s6
 ; RV64I-NEXT:    sd zero, 0(sp)
 ; RV64I-NEXT:    sd zero, 8(sp)
 ; RV64I-NEXT:    sd zero, 16(sp)
@@ -3881,8 +3881,8 @@ define void @shl_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) nounw
 ; RV64I-NEXT:    or a0, a0, t5
 ; RV64I-NEXT:    or t0, s0, t6
 ; RV64I-NEXT:    or t1, s5, s1
-; RV64I-NEXT:    or t2, s4, s2
-; RV64I-NEXT:    or a1, a1, s3
+; RV64I-NEXT:    or t2, s3, s2
+; RV64I-NEXT:    or a1, a1, s4
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a6, a6, 32
 ; RV64I-NEXT:    slli a0, a0, 32
@@ -4112,13 +4112,13 @@ define void @shl_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) nounw
 ; RV32I-NEXT:    sw t0, 44(sp)
 ; RV32I-NEXT:    sw t1, 48(sp)
 ; RV32I-NEXT:    sw a5, 52(sp)
-; RV32I-NEXT:    lw a6, 16(t2)
-; RV32I-NEXT:    lw a5, 20(t2)
-; RV32I-NEXT:    lw a7, 24(t2)
 ; RV32I-NEXT:    lw a1, 0(t2)
 ; RV32I-NEXT:    lw a0, 4(t2)
 ; RV32I-NEXT:    lw a4, 8(t2)
 ; RV32I-NEXT:    lw a3, 12(t2)
+; RV32I-NEXT:    lw a7, 24(t2)
+; RV32I-NEXT:    lw a5, 20(t2)
+; RV32I-NEXT:    lw a6, 16(t2)
 ; RV32I-NEXT:    lw t0, 28(t2)
 ; RV32I-NEXT:    srli t1, a7, 24
 ; RV32I-NEXT:    srli t2, a7, 16
@@ -4512,13 +4512,13 @@ define void @shl_32bytes_dwordOff(ptr %src.ptr, ptr %dwordOff.ptr, ptr %dst) nou
 ; RV32I-NEXT:    sw t0, 44(sp)
 ; RV32I-NEXT:    sw t1, 48(sp)
 ; RV32I-NEXT:    sw a5, 52(sp)
-; RV32I-NEXT:    lw a6, 16(t2)
-; RV32I-NEXT:    lw a5, 20(t2)
-; RV32I-NEXT:    lw a7, 24(t2)
 ; RV32I-NEXT:    lw a1, 0(t2)
 ; RV32I-NEXT:    lw a0, 4(t2)
 ; RV32I-NEXT:    lw a4, 8(t2)
 ; RV32I-NEXT:    lw a3, 12(t2)
+; RV32I-NEXT:    lw a7, 24(t2)
+; RV32I-NEXT:    lw a5, 20(t2)
+; RV32I-NEXT:    lw a6, 16(t2)
 ; RV32I-NEXT:    lw t0, 28(t2)
 ; RV32I-NEXT:    srli t1, a7, 24
 ; RV32I-NEXT:    srli t2, a7, 16
@@ -4691,17 +4691,17 @@ define void @ashr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli s5, s5, 16
 ; RV64I-NEXT:    slli s6, s6, 24
 ; RV64I-NEXT:    slli s3, s3, 8
-; RV64I-NEXT:    or s5, s6, s5
-; RV64I-NEXT:    or s2, s3, s2
-; RV64I-NEXT:    lbu s3, 4(a1)
-; RV64I-NEXT:    lbu s6, 5(a1)
 ; RV64I-NEXT:    slli s4, s4, 16
 ; RV64I-NEXT:    slli s7, s7, 24
-; RV64I-NEXT:    or s4, s7, s4
+; RV64I-NEXT:    or s5, s6, s5
+; RV64I-NEXT:    or s2, s3, s2
+; RV64I-NEXT:    or s3, s7, s4
+; RV64I-NEXT:    lbu s4, 5(a1)
+; RV64I-NEXT:    lbu s6, 4(a1)
 ; RV64I-NEXT:    lbu s7, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli s6, s6, 8
-; RV64I-NEXT:    or s3, s6, s3
+; RV64I-NEXT:    slli s4, s4, 8
+; RV64I-NEXT:    or s4, s4, s6
 ; RV64I-NEXT:    slli s7, s7, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, s7
@@ -4714,8 +4714,8 @@ define void @ashr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    or a0, a0, t5
 ; RV64I-NEXT:    or t0, s0, t6
 ; RV64I-NEXT:    or t1, s5, s1
-; RV64I-NEXT:    or t2, s4, s2
-; RV64I-NEXT:    or a1, a1, s3
+; RV64I-NEXT:    or t2, s3, s2
+; RV64I-NEXT:    or a1, a1, s4
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a6, a6, 32
 ; RV64I-NEXT:    slli a0, a0, 32
@@ -4739,22 +4739,22 @@ define void @ashr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    andi a1, a1, 24
 ; RV64I-NEXT:    add a1, s6, a1
 ; RV64I-NEXT:    andi a0, a4, 56
-; RV64I-NEXT:    ld a3, 0(a1)
-; RV64I-NEXT:    ld a5, 8(a1)
+; RV64I-NEXT:    xori a5, a0, 63
+; RV64I-NEXT:    ld a3, 8(a1)
 ; RV64I-NEXT:    ld a6, 16(a1)
-; RV64I-NEXT:    xori a7, a0, 63
+; RV64I-NEXT:    ld a7, 0(a1)
 ; RV64I-NEXT:    ld t0, 24(a1)
-; RV64I-NEXT:    srl a0, a5, a4
+; RV64I-NEXT:    srl a0, a3, a4
 ; RV64I-NEXT:    slli t1, a6, 1
-; RV64I-NEXT:    srl a1, a3, a4
-; RV64I-NEXT:    slli a5, a5, 1
+; RV64I-NEXT:    srl a1, a7, a4
+; RV64I-NEXT:    slli a7, a3, 1
 ; RV64I-NEXT:    srl a3, a6, a4
 ; RV64I-NEXT:    slli a6, t0, 1
 ; RV64I-NEXT:    sra t0, t0, a4
-; RV64I-NEXT:    sll a4, t1, a7
-; RV64I-NEXT:    sll a5, a5, a7
-; RV64I-NEXT:    sll a6, a6, a7
-; RV64I-NEXT:    srli a7, t0, 56
+; RV64I-NEXT:    sll a4, t1, a5
+; RV64I-NEXT:    sll a7, a7, a5
+; RV64I-NEXT:    sll a5, a6, a5
+; RV64I-NEXT:    srli a6, t0, 56
 ; RV64I-NEXT:    srli t1, t0, 48
 ; RV64I-NEXT:    srli t2, t0, 40
 ; RV64I-NEXT:    srli t3, t0, 32
@@ -4762,40 +4762,40 @@ define void @ashr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    srli t5, t0, 16
 ; RV64I-NEXT:    srli t6, t0, 8
 ; RV64I-NEXT:    or a4, a0, a4
-; RV64I-NEXT:    or a5, a1, a5
-; RV64I-NEXT:    or a6, a3, a6
+; RV64I-NEXT:    or a7, a1, a7
+; RV64I-NEXT:    or a5, a3, a5
 ; RV64I-NEXT:    sb t3, 28(a2)
 ; RV64I-NEXT:    sb t2, 29(a2)
 ; RV64I-NEXT:    sb t1, 30(a2)
-; RV64I-NEXT:    sb a7, 31(a2)
+; RV64I-NEXT:    sb a6, 31(a2)
 ; RV64I-NEXT:    sb t0, 24(a2)
 ; RV64I-NEXT:    sb t6, 25(a2)
 ; RV64I-NEXT:    sb t5, 26(a2)
 ; RV64I-NEXT:    sb t4, 27(a2)
-; RV64I-NEXT:    srli a7, a6, 56
-; RV64I-NEXT:    srli t0, a6, 48
-; RV64I-NEXT:    srli t1, a6, 40
-; RV64I-NEXT:    srli t2, a6, 32
-; RV64I-NEXT:    srli t3, a6, 24
-; RV64I-NEXT:    srli t4, a6, 16
-; RV64I-NEXT:    srli a6, a6, 8
-; RV64I-NEXT:    srli t5, a5, 56
-; RV64I-NEXT:    srli t6, a5, 48
-; RV64I-NEXT:    srli s0, a5, 40
-; RV64I-NEXT:    srli s1, a5, 32
-; RV64I-NEXT:    srli s2, a5, 24
-; RV64I-NEXT:    srli s3, a5, 16
+; RV64I-NEXT:    srli a6, a5, 56
+; RV64I-NEXT:    srli t0, a5, 48
+; RV64I-NEXT:    srli t1, a5, 40
+; RV64I-NEXT:    srli t2, a5, 32
+; RV64I-NEXT:    srli t3, a5, 24
+; RV64I-NEXT:    srli t4, a5, 16
 ; RV64I-NEXT:    srli a5, a5, 8
+; RV64I-NEXT:    srli t5, a7, 56
+; RV64I-NEXT:    srli t6, a7, 48
+; RV64I-NEXT:    srli s0, a7, 40
+; RV64I-NEXT:    srli s1, a7, 32
+; RV64I-NEXT:    srli s2, a7, 24
+; RV64I-NEXT:    srli s3, a7, 16
+; RV64I-NEXT:    srli a7, a7, 8
 ; RV64I-NEXT:    srli s4, a4, 56
 ; RV64I-NEXT:    srli s5, a4, 48
 ; RV64I-NEXT:    srli s6, a4, 40
 ; RV64I-NEXT:    sb t2, 20(a2)
 ; RV64I-NEXT:    sb t1, 21(a2)
 ; RV64I-NEXT:    sb t0, 22(a2)
-; RV64I-NEXT:    sb a7, 23(a2)
-; RV64I-NEXT:    srli a7, a4, 32
+; RV64I-NEXT:    sb a6, 23(a2)
+; RV64I-NEXT:    srli a6, a4, 32
 ; RV64I-NEXT:    sb a3, 16(a2)
-; RV64I-NEXT:    sb a6, 17(a2)
+; RV64I-NEXT:    sb a5, 17(a2)
 ; RV64I-NEXT:    sb t4, 18(a2)
 ; RV64I-NEXT:    sb t3, 19(a2)
 ; RV64I-NEXT:    srli a3, a4, 24
@@ -4803,19 +4803,19 @@ define void @ashr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    sb s0, 5(a2)
 ; RV64I-NEXT:    sb t6, 6(a2)
 ; RV64I-NEXT:    sb t5, 7(a2)
-; RV64I-NEXT:    srli a6, a4, 16
+; RV64I-NEXT:    srli a5, a4, 16
 ; RV64I-NEXT:    srli a4, a4, 8
 ; RV64I-NEXT:    sb a1, 0(a2)
-; RV64I-NEXT:    sb a5, 1(a2)
+; RV64I-NEXT:    sb a7, 1(a2)
 ; RV64I-NEXT:    sb s3, 2(a2)
 ; RV64I-NEXT:    sb s2, 3(a2)
-; RV64I-NEXT:    sb a7, 12(a2)
+; RV64I-NEXT:    sb a6, 12(a2)
 ; RV64I-NEXT:    sb s6, 13(a2)
 ; RV64I-NEXT:    sb s5, 14(a2)
 ; RV64I-NEXT:    sb s4, 15(a2)
 ; RV64I-NEXT:    sb a0, 8(a2)
 ; RV64I-NEXT:    sb a4, 9(a2)
-; RV64I-NEXT:    sb a6, 10(a2)
+; RV64I-NEXT:    sb a5, 10(a2)
 ; RV64I-NEXT:    sb a3, 11(a2)
 ; RV64I-NEXT:    ld s0, 152(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s1, 144(sp) # 8-byte Folded Reload
@@ -5171,17 +5171,17 @@ define void @ashr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    slli s5, s5, 16
 ; RV64I-NEXT:    slli s6, s6, 24
 ; RV64I-NEXT:    slli s3, s3, 8
-; RV64I-NEXT:    or s5, s6, s5
-; RV64I-NEXT:    or s2, s3, s2
-; RV64I-NEXT:    lbu s3, 4(a1)
-; RV64I-NEXT:    lbu s6, 5(a1)
 ; RV64I-NEXT:    slli s4, s4, 16
 ; RV64I-NEXT:    slli s7, s7, 24
-; RV64I-NEXT:    or s4, s7, s4
+; RV64I-NEXT:    or s5, s6, s5
+; RV64I-NEXT:    or s2, s3, s2
+; RV64I-NEXT:    or s3, s7, s4
+; RV64I-NEXT:    lbu s4, 5(a1)
+; RV64I-NEXT:    lbu s6, 4(a1)
 ; RV64I-NEXT:    lbu s7, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli s6, s6, 8
-; RV64I-NEXT:    or s3, s6, s3
+; RV64I-NEXT:    slli s4, s4, 8
+; RV64I-NEXT:    or s4, s4, s6
 ; RV64I-NEXT:    slli s7, s7, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, s7
@@ -5194,8 +5194,8 @@ define void @ashr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    or a0, a0, t5
 ; RV64I-NEXT:    or t0, s0, t6
 ; RV64I-NEXT:    or t1, s5, s1
-; RV64I-NEXT:    or t2, s4, s2
-; RV64I-NEXT:    or a1, a1, s3
+; RV64I-NEXT:    or t2, s3, s2
+; RV64I-NEXT:    or a1, a1, s4
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a6, a6, 32
 ; RV64I-NEXT:    slli a0, a0, 32
@@ -5220,24 +5220,24 @@ define void @ashr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    andi a1, a1, 24
 ; RV64I-NEXT:    andi a0, a3, 32
 ; RV64I-NEXT:    add a1, s6, a1
-; RV64I-NEXT:    ld a4, 0(a1)
+; RV64I-NEXT:    xori a4, a0, 63
 ; RV64I-NEXT:    ld a5, 8(a1)
 ; RV64I-NEXT:    ld a6, 16(a1)
-; RV64I-NEXT:    xori a7, a0, 63
+; RV64I-NEXT:    ld a7, 0(a1)
 ; RV64I-NEXT:    ld t0, 24(a1)
 ; RV64I-NEXT:    srl a0, a5, a3
 ; RV64I-NEXT:    slli t1, a6, 1
-; RV64I-NEXT:    srl a1, a4, a3
+; RV64I-NEXT:    srl a1, a7, a3
 ; RV64I-NEXT:    slli a5, a5, 1
-; RV64I-NEXT:    srl a4, a6, a3
-; RV64I-NEXT:    slli a6, t0, 1
+; RV64I-NEXT:    srl a6, a6, a3
+; RV64I-NEXT:    slli a7, t0, 1
 ; RV64I-NEXT:    sra a3, t0, a3
-; RV64I-NEXT:    sll t0, t1, a7
-; RV64I-NEXT:    sll a5, a5, a7
-; RV64I-NEXT:    sll a6, a6, a7
-; RV64I-NEXT:    srli a7, a4, 24
-; RV64I-NEXT:    srli t1, a4, 16
-; RV64I-NEXT:    srli t2, a4, 8
+; RV64I-NEXT:    sll t0, t1, a4
+; RV64I-NEXT:    sll a5, a5, a4
+; RV64I-NEXT:    sll a4, a7, a4
+; RV64I-NEXT:    srli a7, a6, 24
+; RV64I-NEXT:    srli t1, a6, 16
+; RV64I-NEXT:    srli t2, a6, 8
 ; RV64I-NEXT:    srli t3, a3, 56
 ; RV64I-NEXT:    srli t4, a3, 48
 ; RV64I-NEXT:    srli t5, a3, 40
@@ -5249,12 +5249,12 @@ define void @ashr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    srli s4, a1, 16
 ; RV64I-NEXT:    srli s5, a1, 8
 ; RV64I-NEXT:    srli s6, a0, 24
-; RV64I-NEXT:    or a6, a4, a6
-; RV64I-NEXT:    sb a4, 16(a2)
+; RV64I-NEXT:    or a4, a6, a4
+; RV64I-NEXT:    sb a6, 16(a2)
 ; RV64I-NEXT:    sb t2, 17(a2)
 ; RV64I-NEXT:    sb t1, 18(a2)
 ; RV64I-NEXT:    sb a7, 19(a2)
-; RV64I-NEXT:    srli a4, a0, 16
+; RV64I-NEXT:    srli a6, a0, 16
 ; RV64I-NEXT:    sb t6, 28(a2)
 ; RV64I-NEXT:    sb t5, 29(a2)
 ; RV64I-NEXT:    sb t4, 30(a2)
@@ -5272,12 +5272,12 @@ define void @ashr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV64I-NEXT:    sb s3, 3(a2)
 ; RV64I-NEXT:    sb a0, 8(a2)
 ; RV64I-NEXT:    sb a7, 9(a2)
-; RV64I-NEXT:    sb a4, 10(a2)
+; RV64I-NEXT:    sb a6, 10(a2)
 ; RV64I-NEXT:    sb s6, 11(a2)
-; RV64I-NEXT:    srli a0, a6, 56
-; RV64I-NEXT:    srli a1, a6, 48
-; RV64I-NEXT:    srli a3, a6, 40
-; RV64I-NEXT:    srli a4, a6, 32
+; RV64I-NEXT:    srli a0, a4, 56
+; RV64I-NEXT:    srli a1, a4, 48
+; RV64I-NEXT:    srli a3, a4, 40
+; RV64I-NEXT:    srli a4, a4, 32
 ; RV64I-NEXT:    srli a6, a5, 56
 ; RV64I-NEXT:    srli a7, a5, 48
 ; RV64I-NEXT:    srli t1, a5, 40
@@ -5431,13 +5431,13 @@ define void @ashr_32bytes_wordOff(ptr %src.ptr, ptr %wordOff.ptr, ptr %dst) noun
 ; RV32I-NEXT:    sw t0, 12(sp)
 ; RV32I-NEXT:    sw t1, 16(sp)
 ; RV32I-NEXT:    sw a5, 20(sp)
-; RV32I-NEXT:    lw a6, 16(s6)
-; RV32I-NEXT:    lw a5, 20(s6)
-; RV32I-NEXT:    lw a7, 24(s6)
 ; RV32I-NEXT:    lw a1, 0(s6)
 ; RV32I-NEXT:    lw a0, 4(s6)
 ; RV32I-NEXT:    lw a4, 8(s6)
 ; RV32I-NEXT:    lw a3, 12(s6)
+; RV32I-NEXT:    lw a7, 24(s6)
+; RV32I-NEXT:    lw a5, 20(s6)
+; RV32I-NEXT:    lw a6, 16(s6)
 ; RV32I-NEXT:    lw t0, 28(s6)
 ; RV32I-NEXT:    srli t1, a7, 24
 ; RV32I-NEXT:    srli t2, a7, 16
@@ -5833,13 +5833,13 @@ define void @ashr_32bytes_dwordOff(ptr %src.ptr, ptr %dwordOff.ptr, ptr %dst) no
 ; RV32I-NEXT:    sw t0, 12(sp)
 ; RV32I-NEXT:    sw t1, 16(sp)
 ; RV32I-NEXT:    sw a5, 20(sp)
-; RV32I-NEXT:    lw a6, 16(s6)
-; RV32I-NEXT:    lw a5, 20(s6)
-; RV32I-NEXT:    lw a7, 24(s6)
 ; RV32I-NEXT:    lw a1, 0(s6)
 ; RV32I-NEXT:    lw a0, 4(s6)
 ; RV32I-NEXT:    lw a4, 8(s6)
 ; RV32I-NEXT:    lw a3, 12(s6)
+; RV32I-NEXT:    lw a7, 24(s6)
+; RV32I-NEXT:    lw a5, 20(s6)
+; RV32I-NEXT:    lw a6, 16(s6)
 ; RV32I-NEXT:    lw t0, 28(s6)
 ; RV32I-NEXT:    srli t1, a7, 24
 ; RV32I-NEXT:    srli t2, a7, 16
diff --git a/llvm/test/CodeGen/RISCV/wide-scalar-shift-legalization.ll b/llvm/test/CodeGen/RISCV/wide-scalar-shift-legalization.ll
index 78f63c72d0469..cd7f30d8f5898 100644
--- a/llvm/test/CodeGen/RISCV/wide-scalar-shift-legalization.ll
+++ b/llvm/test/CodeGen/RISCV/wide-scalar-shift-legalization.ll
@@ -29,25 +29,25 @@ define void @lshr_4bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: lshr_4bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
-; RV32I-NEXT:    lbu a5, 2(a0)
-; RV32I-NEXT:    lbu a0, 3(a0)
+; RV32I-NEXT:    lbu a4, 2(a0)
+; RV32I-NEXT:    lbu a5, 3(a0)
+; RV32I-NEXT:    lbu a0, 0(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a6, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a0, a0, 24
-; RV32I-NEXT:    or a0, a0, a5
-; RV32I-NEXT:    lbu a5, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a0, a3, a0
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a3, 1(a1)
+; RV32I-NEXT:    lbu a5, 0(a1)
+; RV32I-NEXT:    lbu a6, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a6, a6, 8
-; RV32I-NEXT:    or a4, a6, a4
-; RV32I-NEXT:    slli a5, a5, 16
+; RV32I-NEXT:    slli a3, a3, 8
+; RV32I-NEXT:    or a3, a3, a5
+; RV32I-NEXT:    slli a6, a6, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a5
-; RV32I-NEXT:    or a0, a0, a3
-; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    or a1, a1, a6
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:    or a1, a1, a3
 ; RV32I-NEXT:    srl a0, a0, a1
 ; RV32I-NEXT:    srli a1, a0, 16
 ; RV32I-NEXT:    srli a3, a0, 24
@@ -90,25 +90,25 @@ define void @shl_4bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: shl_4bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
-; RV32I-NEXT:    lbu a5, 2(a0)
-; RV32I-NEXT:    lbu a0, 3(a0)
+; RV32I-NEXT:    lbu a4, 2(a0)
+; RV32I-NEXT:    lbu a5, 3(a0)
+; RV32I-NEXT:    lbu a0, 0(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a6, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a0, a0, 24
-; RV32I-NEXT:    or a0, a0, a5
-; RV32I-NEXT:    lbu a5, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a0, a3, a0
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a3, 1(a1)
+; RV32I-NEXT:    lbu a5, 0(a1)
+; RV32I-NEXT:    lbu a6, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a6, a6, 8
-; RV32I-NEXT:    or a4, a6, a4
-; RV32I-NEXT:    slli a5, a5, 16
+; RV32I-NEXT:    slli a3, a3, 8
+; RV32I-NEXT:    or a3, a3, a5
+; RV32I-NEXT:    slli a6, a6, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a5
-; RV32I-NEXT:    or a0, a0, a3
-; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    or a1, a1, a6
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:    or a1, a1, a3
 ; RV32I-NEXT:    sll a0, a0, a1
 ; RV32I-NEXT:    srli a1, a0, 16
 ; RV32I-NEXT:    srli a3, a0, 24
@@ -151,25 +151,25 @@ define void @ashr_4bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: ashr_4bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
-; RV32I-NEXT:    lbu a5, 2(a0)
-; RV32I-NEXT:    lbu a0, 3(a0)
+; RV32I-NEXT:    lbu a4, 2(a0)
+; RV32I-NEXT:    lbu a5, 3(a0)
+; RV32I-NEXT:    lbu a0, 0(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a6, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a0, a0, 24
-; RV32I-NEXT:    or a0, a0, a5
-; RV32I-NEXT:    lbu a5, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a0, a3, a0
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a3, 1(a1)
+; RV32I-NEXT:    lbu a5, 0(a1)
+; RV32I-NEXT:    lbu a6, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a6, a6, 8
-; RV32I-NEXT:    or a4, a6, a4
-; RV32I-NEXT:    slli a5, a5, 16
+; RV32I-NEXT:    slli a3, a3, 8
+; RV32I-NEXT:    or a3, a3, a5
+; RV32I-NEXT:    slli a6, a6, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a5
-; RV32I-NEXT:    or a0, a0, a3
-; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    or a1, a1, a6
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:    or a1, a1, a3
 ; RV32I-NEXT:    sra a0, a0, a1
 ; RV32I-NEXT:    srli a1, a0, 16
 ; RV32I-NEXT:    srli a3, a0, 24
@@ -211,24 +211,24 @@ define void @lshr_8bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli a0, a0, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 4(a1)
-; RV64I-NEXT:    lbu t1, 5(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t2, t2, 24
-; RV64I-NEXT:    or t0, t2, t0
+; RV64I-NEXT:    or a0, a0, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t2, t0
+; RV64I-NEXT:    lbu t0, 5(a1)
+; RV64I-NEXT:    lbu t1, 4(a1)
 ; RV64I-NEXT:    lbu t2, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli t1, t1, 8
-; RV64I-NEXT:    or a7, t1, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
 ; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t2
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a0, a0, a5
-; RV64I-NEXT:    or a4, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a4, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    slli a1, a1, 32
 ; RV64I-NEXT:    or a0, a0, a3
@@ -254,25 +254,25 @@ define void @lshr_8bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: lshr_8bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 5(a0)
-; RV32I-NEXT:    lbu a4, 4(a0)
-; RV32I-NEXT:    lbu a5, 6(a0)
-; RV32I-NEXT:    lbu a6, 7(a0)
+; RV32I-NEXT:    lbu a4, 6(a0)
+; RV32I-NEXT:    lbu a5, 7(a0)
+; RV32I-NEXT:    lbu a6, 4(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a7, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a6, a6, 24
-; RV32I-NEXT:    or a5, a6, a5
-; RV32I-NEXT:    lbu a6, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a3, a3, a6
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a5, 1(a1)
+; RV32I-NEXT:    lbu a6, 0(a1)
+; RV32I-NEXT:    lbu a7, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a7, a7, 8
-; RV32I-NEXT:    or a4, a7, a4
-; RV32I-NEXT:    slli a6, a6, 16
+; RV32I-NEXT:    slli a5, a5, 8
+; RV32I-NEXT:    or a6, a5, a6
+; RV32I-NEXT:    slli a7, a7, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a6
-; RV32I-NEXT:    or a5, a5, a3
-; RV32I-NEXT:    or a4, a1, a4
+; RV32I-NEXT:    or a1, a1, a7
+; RV32I-NEXT:    or a5, a4, a3
+; RV32I-NEXT:    or a4, a1, a6
 ; RV32I-NEXT:    addi a3, a4, -32
 ; RV32I-NEXT:    srl a1, a5, a4
 ; RV32I-NEXT:    bltz a3, .LBB3_2
@@ -344,24 +344,24 @@ define void @shl_8bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli a0, a0, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 4(a1)
-; RV64I-NEXT:    lbu t1, 5(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t2, t2, 24
-; RV64I-NEXT:    or t0, t2, t0
+; RV64I-NEXT:    or a0, a0, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t2, t0
+; RV64I-NEXT:    lbu t0, 5(a1)
+; RV64I-NEXT:    lbu t1, 4(a1)
 ; RV64I-NEXT:    lbu t2, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli t1, t1, 8
-; RV64I-NEXT:    or a7, t1, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
 ; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t2
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a0, a0, a5
-; RV64I-NEXT:    or a4, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a4, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    slli a1, a1, 32
 ; RV64I-NEXT:    or a0, a0, a3
@@ -387,25 +387,25 @@ define void @shl_8bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-LABEL: shl_8bytes:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lbu a3, 1(a0)
-; RV32I-NEXT:    lbu a4, 0(a0)
-; RV32I-NEXT:    lbu a5, 2(a0)
-; RV32I-NEXT:    lbu a6, 3(a0)
+; RV32I-NEXT:    lbu a4, 2(a0)
+; RV32I-NEXT:    lbu a5, 3(a0)
+; RV32I-NEXT:    lbu a6, 0(a0)
 ; RV32I-NEXT:    slli a3, a3, 8
-; RV32I-NEXT:    or a3, a3, a4
-; RV32I-NEXT:    lbu a4, 0(a1)
-; RV32I-NEXT:    lbu a7, 1(a1)
-; RV32I-NEXT:    slli a5, a5, 16
-; RV32I-NEXT:    slli a6, a6, 24
-; RV32I-NEXT:    or a5, a6, a5
-; RV32I-NEXT:    lbu a6, 2(a1)
+; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a5, a5, 24
+; RV32I-NEXT:    or a3, a3, a6
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    lbu a5, 1(a1)
+; RV32I-NEXT:    lbu a6, 0(a1)
+; RV32I-NEXT:    lbu a7, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli a7, a7, 8
-; RV32I-NEXT:    or a4, a7, a4
-; RV32I-NEXT:    slli a6, a6, 16
+; RV32I-NEXT:    slli a5, a5, 8
+; RV32I-NEXT:    or a6, a5, a6
+; RV32I-NEXT:    slli a7, a7, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a6
-; RV32I-NEXT:    or a5, a5, a3
-; RV32I-NEXT:    or a4, a1, a4
+; RV32I-NEXT:    or a1, a1, a7
+; RV32I-NEXT:    or a5, a4, a3
+; RV32I-NEXT:    or a4, a1, a6
 ; RV32I-NEXT:    addi a3, a4, -32
 ; RV32I-NEXT:    sll a1, a5, a4
 ; RV32I-NEXT:    bltz a3, .LBB4_2
@@ -477,24 +477,24 @@ define void @ashr_8bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli a0, a0, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 4(a1)
-; RV64I-NEXT:    lbu t1, 5(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t2, t2, 24
-; RV64I-NEXT:    or t0, t2, t0
+; RV64I-NEXT:    or a0, a0, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t2, t0
+; RV64I-NEXT:    lbu t0, 5(a1)
+; RV64I-NEXT:    lbu t1, 4(a1)
 ; RV64I-NEXT:    lbu t2, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli t1, t1, 8
-; RV64I-NEXT:    or a7, t1, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
 ; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t2
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a0, a0, a5
-; RV64I-NEXT:    or a4, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a4, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    slli a1, a1, 32
 ; RV64I-NEXT:    or a0, a0, a3
@@ -519,21 +519,21 @@ define void @ashr_8bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ;
 ; RV32I-LABEL: ashr_8bytes:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    lbu a3, 4(a0)
-; RV32I-NEXT:    lbu a4, 5(a0)
+; RV32I-NEXT:    lbu a3, 5(a0)
+; RV32I-NEXT:    lbu a4, 4(a0)
 ; RV32I-NEXT:    lbu a5, 6(a0)
 ; RV32I-NEXT:    lbu a6, 7(a0)
+; RV32I-NEXT:    slli a3, a3, 8
+; RV32I-NEXT:    or a3, a3, a4
+; RV32I-NEXT:    lbu a4, 1(a1)
 ; RV32I-NEXT:    lbu a7, 0(a1)
-; RV32I-NEXT:    lbu t0, 1(a1)
-; RV32I-NEXT:    slli a4, a4, 8
-; RV32I-NEXT:    or a3, a4, a3
-; RV32I-NEXT:    lbu a4, 2(a1)
+; RV32I-NEXT:    lbu t0, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli t0, t0, 8
-; RV32I-NEXT:    or a7, t0, a7
-; RV32I-NEXT:    slli a4, a4, 16
+; RV32I-NEXT:    slli a4, a4, 8
+; RV32I-NEXT:    or a7, a4, a7
+; RV32I-NEXT:    slli t0, t0, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    or a1, a1, t0
 ; RV32I-NEXT:    slli a4, a5, 16
 ; RV32I-NEXT:    slli a5, a6, 24
 ; RV32I-NEXT:    or a4, a5, a4
@@ -611,24 +611,24 @@ define void @lshr_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli t2, t2, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or t1, t2, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 4(a1)
-; RV64I-NEXT:    lbu t2, 5(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t3, t3, 24
-; RV64I-NEXT:    or t0, t3, t0
+; RV64I-NEXT:    or t1, t2, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t3, t0
+; RV64I-NEXT:    lbu t0, 5(a1)
+; RV64I-NEXT:    lbu t2, 4(a1)
 ; RV64I-NEXT:    lbu t3, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t2
 ; RV64I-NEXT:    slli t3, t3, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t3
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a4, t1, a5
-; RV64I-NEXT:    or a6, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a1, a1, 32
 ; RV64I-NEXT:    or a5, a4, a3
@@ -641,31 +641,31 @@ define void @lshr_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    j .LBB6_3
 ; RV64I-NEXT:  .LBB6_2:
 ; RV64I-NEXT:    lbu a6, 1(a0)
-; RV64I-NEXT:    lbu a7, 0(a0)
-; RV64I-NEXT:    lbu t0, 2(a0)
-; RV64I-NEXT:    lbu t1, 3(a0)
+; RV64I-NEXT:    lbu a7, 2(a0)
+; RV64I-NEXT:    lbu t0, 3(a0)
+; RV64I-NEXT:    lbu t1, 0(a0)
 ; RV64I-NEXT:    slli a6, a6, 8
-; RV64I-NEXT:    or a6, a6, a7
-; RV64I-NEXT:    lbu a7, 4(a0)
-; RV64I-NEXT:    lbu t2, 5(a0)
-; RV64I-NEXT:    slli t0, t0, 16
-; RV64I-NEXT:    slli t1, t1, 24
-; RV64I-NEXT:    or t0, t1, t0
-; RV64I-NEXT:    lbu t1, 6(a0)
+; RV64I-NEXT:    slli a7, a7, 16
+; RV64I-NEXT:    slli t0, t0, 24
+; RV64I-NEXT:    or a6, a6, t1
+; RV64I-NEXT:    or a7, t0, a7
+; RV64I-NEXT:    lbu t0, 5(a0)
+; RV64I-NEXT:    lbu t1, 4(a0)
+; RV64I-NEXT:    lbu t2, 6(a0)
 ; RV64I-NEXT:    lbu a0, 7(a0)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
-; RV64I-NEXT:    slli t1, t1, 16
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
+; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, t0, a6
-; RV64I-NEXT:    not t0, a4
+; RV64I-NEXT:    or a0, a0, t2
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    not a7, a4
 ; RV64I-NEXT:    slli a5, a5, 1
-; RV64I-NEXT:    or a0, a0, a7
+; RV64I-NEXT:    or a0, a0, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a6
 ; RV64I-NEXT:    srl a0, a0, a4
-; RV64I-NEXT:    sll a4, a5, t0
+; RV64I-NEXT:    sll a4, a5, a7
 ; RV64I-NEXT:    or a0, a0, a4
 ; RV64I-NEXT:  .LBB6_3:
 ; RV64I-NEXT:    srai a3, a3, 63
@@ -736,24 +736,24 @@ define void @lshr_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    slli t3, t3, 16
 ; RV32I-NEXT:    slli t4, t4, 24
 ; RV32I-NEXT:    slli t1, t1, 8
-; RV32I-NEXT:    or t3, t4, t3
-; RV32I-NEXT:    or a6, t1, a6
-; RV32I-NEXT:    lbu t1, 0(a1)
-; RV32I-NEXT:    lbu t4, 1(a1)
 ; RV32I-NEXT:    slli t2, t2, 16
 ; RV32I-NEXT:    slli a0, a0, 24
+; RV32I-NEXT:    or t3, t4, t3
+; RV32I-NEXT:    or a6, t1, a6
 ; RV32I-NEXT:    or a0, a0, t2
-; RV32I-NEXT:    lbu t2, 2(a1)
+; RV32I-NEXT:    lbu t1, 1(a1)
+; RV32I-NEXT:    lbu t2, 0(a1)
+; RV32I-NEXT:    lbu t4, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli t4, t4, 8
-; RV32I-NEXT:    or t1, t4, t1
+; RV32I-NEXT:    slli t1, t1, 8
+; RV32I-NEXT:    or t1, t1, t2
 ; RV32I-NEXT:    sw zero, 16(sp)
 ; RV32I-NEXT:    sw zero, 20(sp)
 ; RV32I-NEXT:    sw zero, 24(sp)
 ; RV32I-NEXT:    sw zero, 28(sp)
-; RV32I-NEXT:    slli t2, t2, 16
+; RV32I-NEXT:    slli t4, t4, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, t2
+; RV32I-NEXT:    or a1, a1, t4
 ; RV32I-NEXT:    mv t2, sp
 ; RV32I-NEXT:    or a3, a4, a3
 ; RV32I-NEXT:    or a4, t0, a7
@@ -767,28 +767,28 @@ define void @lshr_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    srli a0, a1, 3
 ; RV32I-NEXT:    andi a3, a1, 31
 ; RV32I-NEXT:    andi a0, a0, 12
-; RV32I-NEXT:    add a0, t2, a0
-; RV32I-NEXT:    lw a4, 0(a0)
-; RV32I-NEXT:    lw a5, 4(a0)
-; RV32I-NEXT:    lw a6, 8(a0)
 ; RV32I-NEXT:    xori a3, a3, 31
+; RV32I-NEXT:    add a0, t2, a0
+; RV32I-NEXT:    lw a4, 4(a0)
+; RV32I-NEXT:    lw a5, 8(a0)
+; RV32I-NEXT:    lw a6, 0(a0)
 ; RV32I-NEXT:    lw a0, 12(a0)
-; RV32I-NEXT:    srl a7, a5, a1
-; RV32I-NEXT:    slli t0, a6, 1
-; RV32I-NEXT:    srl a4, a4, a1
-; RV32I-NEXT:    slli a5, a5, 1
+; RV32I-NEXT:    srl a7, a4, a1
+; RV32I-NEXT:    slli t0, a5, 1
 ; RV32I-NEXT:    srl a6, a6, a1
+; RV32I-NEXT:    slli a4, a4, 1
+; RV32I-NEXT:    srl a5, a5, a1
 ; RV32I-NEXT:    slli t1, a0, 1
 ; RV32I-NEXT:    srl a0, a0, a1
 ; RV32I-NEXT:    sll a1, t0, a3
-; RV32I-NEXT:    sll a5, a5, a3
+; RV32I-NEXT:    sll a4, a4, a3
 ; RV32I-NEXT:    sll a3, t1, a3
 ; RV32I-NEXT:    srli t0, a0, 16
 ; RV32I-NEXT:    srli t1, a0, 24
 ; RV32I-NEXT:    srli t2, a0, 8
 ; RV32I-NEXT:    or a1, a7, a1
-; RV32I-NEXT:    or a4, a4, a5
-; RV32I-NEXT:    or a3, a6, a3
+; RV32I-NEXT:    or a4, a6, a4
+; RV32I-NEXT:    or a3, a5, a3
 ; RV32I-NEXT:    sb a0, 12(a2)
 ; RV32I-NEXT:    sb t2, 13(a2)
 ; RV32I-NEXT:    sb t0, 14(a2)
@@ -847,24 +847,24 @@ define void @shl_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli t2, t2, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or t1, t2, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 4(a1)
-; RV64I-NEXT:    lbu t2, 5(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t3, t3, 24
-; RV64I-NEXT:    or t0, t3, t0
+; RV64I-NEXT:    or t1, t2, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t3, t0
+; RV64I-NEXT:    lbu t0, 5(a1)
+; RV64I-NEXT:    lbu t2, 4(a1)
 ; RV64I-NEXT:    lbu t3, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t2
 ; RV64I-NEXT:    slli t3, t3, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t3
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a4, t1, a5
-; RV64I-NEXT:    or a6, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a1, a1, 32
 ; RV64I-NEXT:    or a5, a4, a3
@@ -877,31 +877,31 @@ define void @shl_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    j .LBB7_3
 ; RV64I-NEXT:  .LBB7_2:
 ; RV64I-NEXT:    lbu a6, 9(a0)
-; RV64I-NEXT:    lbu a7, 8(a0)
-; RV64I-NEXT:    lbu t0, 10(a0)
-; RV64I-NEXT:    lbu t1, 11(a0)
+; RV64I-NEXT:    lbu a7, 10(a0)
+; RV64I-NEXT:    lbu t0, 11(a0)
+; RV64I-NEXT:    lbu t1, 8(a0)
 ; RV64I-NEXT:    slli a6, a6, 8
-; RV64I-NEXT:    or a6, a6, a7
-; RV64I-NEXT:    lbu a7, 12(a0)
-; RV64I-NEXT:    lbu t2, 13(a0)
-; RV64I-NEXT:    slli t0, t0, 16
-; RV64I-NEXT:    slli t1, t1, 24
-; RV64I-NEXT:    or t0, t1, t0
-; RV64I-NEXT:    lbu t1, 14(a0)
+; RV64I-NEXT:    slli a7, a7, 16
+; RV64I-NEXT:    slli t0, t0, 24
+; RV64I-NEXT:    or a6, a6, t1
+; RV64I-NEXT:    or a7, t0, a7
+; RV64I-NEXT:    lbu t0, 13(a0)
+; RV64I-NEXT:    lbu t1, 12(a0)
+; RV64I-NEXT:    lbu t2, 14(a0)
 ; RV64I-NEXT:    lbu a0, 15(a0)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
-; RV64I-NEXT:    slli t1, t1, 16
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t1
+; RV64I-NEXT:    slli t2, t2, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, t1
-; RV64I-NEXT:    or a6, t0, a6
-; RV64I-NEXT:    not t0, a4
+; RV64I-NEXT:    or a0, a0, t2
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    not a7, a4
 ; RV64I-NEXT:    srli a5, a5, 1
-; RV64I-NEXT:    or a0, a0, a7
+; RV64I-NEXT:    or a0, a0, t0
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a6
 ; RV64I-NEXT:    sll a0, a0, a4
-; RV64I-NEXT:    srl a4, a5, t0
+; RV64I-NEXT:    srl a4, a5, a7
 ; RV64I-NEXT:    or a0, a0, a4
 ; RV64I-NEXT:  .LBB7_3:
 ; RV64I-NEXT:    srai a3, a3, 63
@@ -972,24 +972,24 @@ define void @shl_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    slli t3, t3, 16
 ; RV32I-NEXT:    slli t4, t4, 24
 ; RV32I-NEXT:    slli t1, t1, 8
-; RV32I-NEXT:    or t3, t4, t3
-; RV32I-NEXT:    or a6, t1, a6
-; RV32I-NEXT:    lbu t1, 0(a1)
-; RV32I-NEXT:    lbu t4, 1(a1)
 ; RV32I-NEXT:    slli t2, t2, 16
 ; RV32I-NEXT:    slli a0, a0, 24
+; RV32I-NEXT:    or t3, t4, t3
+; RV32I-NEXT:    or a6, t1, a6
 ; RV32I-NEXT:    or a0, a0, t2
-; RV32I-NEXT:    lbu t2, 2(a1)
+; RV32I-NEXT:    lbu t1, 1(a1)
+; RV32I-NEXT:    lbu t2, 0(a1)
+; RV32I-NEXT:    lbu t4, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli t4, t4, 8
-; RV32I-NEXT:    or t1, t4, t1
+; RV32I-NEXT:    slli t1, t1, 8
+; RV32I-NEXT:    or t1, t1, t2
 ; RV32I-NEXT:    sw zero, 0(sp)
 ; RV32I-NEXT:    sw zero, 4(sp)
 ; RV32I-NEXT:    sw zero, 8(sp)
 ; RV32I-NEXT:    sw zero, 12(sp)
-; RV32I-NEXT:    slli t2, t2, 16
+; RV32I-NEXT:    slli t4, t4, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, t2
+; RV32I-NEXT:    or a1, a1, t4
 ; RV32I-NEXT:    addi t2, sp, 16
 ; RV32I-NEXT:    or a3, a4, a3
 ; RV32I-NEXT:    or a4, t0, a7
@@ -1083,24 +1083,24 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli t2, t2, 24
 ; RV64I-NEXT:    slli a7, a7, 8
-; RV64I-NEXT:    or t1, t2, t1
-; RV64I-NEXT:    or a6, a7, a6
-; RV64I-NEXT:    lbu a7, 4(a1)
-; RV64I-NEXT:    lbu t2, 5(a1)
 ; RV64I-NEXT:    slli t0, t0, 16
 ; RV64I-NEXT:    slli t3, t3, 24
-; RV64I-NEXT:    or t0, t3, t0
+; RV64I-NEXT:    or t1, t2, t1
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a7, t3, t0
+; RV64I-NEXT:    lbu t0, 5(a1)
+; RV64I-NEXT:    lbu t2, 4(a1)
 ; RV64I-NEXT:    lbu t3, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli t2, t2, 8
-; RV64I-NEXT:    or a7, t2, a7
+; RV64I-NEXT:    slli t0, t0, 8
+; RV64I-NEXT:    or t0, t0, t2
 ; RV64I-NEXT:    slli t3, t3, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, t3
 ; RV64I-NEXT:    or a3, a4, a3
 ; RV64I-NEXT:    or a5, t1, a5
-; RV64I-NEXT:    or a6, t0, a6
-; RV64I-NEXT:    or a1, a1, a7
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    or a1, a1, t0
 ; RV64I-NEXT:    slli a4, a5, 32
 ; RV64I-NEXT:    slli a1, a1, 32
 ; RV64I-NEXT:    or a4, a4, a3
@@ -1115,31 +1115,31 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    j .LBB8_3
 ; RV64I-NEXT:  .LBB8_2:
 ; RV64I-NEXT:    lbu a5, 1(a0)
-; RV64I-NEXT:    lbu a6, 0(a0)
-; RV64I-NEXT:    lbu a7, 2(a0)
-; RV64I-NEXT:    lbu t0, 3(a0)
+; RV64I-NEXT:    lbu a6, 2(a0)
+; RV64I-NEXT:    lbu a7, 3(a0)
+; RV64I-NEXT:    lbu t0, 0(a0)
 ; RV64I-NEXT:    slli a5, a5, 8
-; RV64I-NEXT:    or a5, a5, a6
-; RV64I-NEXT:    lbu a6, 4(a0)
-; RV64I-NEXT:    lbu t1, 5(a0)
-; RV64I-NEXT:    slli a7, a7, 16
-; RV64I-NEXT:    slli t0, t0, 24
-; RV64I-NEXT:    or a7, t0, a7
-; RV64I-NEXT:    lbu t0, 6(a0)
+; RV64I-NEXT:    slli a6, a6, 16
+; RV64I-NEXT:    slli a7, a7, 24
+; RV64I-NEXT:    or a5, a5, t0
+; RV64I-NEXT:    or a6, a7, a6
+; RV64I-NEXT:    lbu a7, 5(a0)
+; RV64I-NEXT:    lbu t0, 4(a0)
+; RV64I-NEXT:    lbu t1, 6(a0)
 ; RV64I-NEXT:    lbu a0, 7(a0)
-; RV64I-NEXT:    slli t1, t1, 8
-; RV64I-NEXT:    or a6, t1, a6
-; RV64I-NEXT:    slli t0, t0, 16
+; RV64I-NEXT:    slli a7, a7, 8
+; RV64I-NEXT:    or a7, a7, t0
+; RV64I-NEXT:    slli t1, t1, 16
 ; RV64I-NEXT:    slli a0, a0, 24
-; RV64I-NEXT:    or a0, a0, t0
-; RV64I-NEXT:    or a5, a7, a5
-; RV64I-NEXT:    not a7, a3
+; RV64I-NEXT:    or a0, a0, t1
+; RV64I-NEXT:    or a5, a6, a5
+; RV64I-NEXT:    not a6, a3
 ; RV64I-NEXT:    slli a4, a4, 1
-; RV64I-NEXT:    or a0, a0, a6
+; RV64I-NEXT:    or a0, a0, a7
 ; RV64I-NEXT:    slli a0, a0, 32
 ; RV64I-NEXT:    or a0, a0, a5
 ; RV64I-NEXT:    srl a0, a0, a3
-; RV64I-NEXT:    sll a3, a4, a7
+; RV64I-NEXT:    sll a3, a4, a6
 ; RV64I-NEXT:    or a0, a0, a3
 ; RV64I-NEXT:  .LBB8_3:
 ; RV64I-NEXT:    srli a3, a1, 56
@@ -1206,29 +1206,29 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    slli t3, t3, 8
 ; RV32I-NEXT:    slli t4, t4, 16
 ; RV32I-NEXT:    slli t5, t5, 24
+; RV32I-NEXT:    slli t1, t1, 8
 ; RV32I-NEXT:    or a4, t3, a4
 ; RV32I-NEXT:    or t3, t5, t4
-; RV32I-NEXT:    lbu t4, 0(a1)
-; RV32I-NEXT:    lbu t5, 1(a1)
-; RV32I-NEXT:    slli t1, t1, 8
 ; RV32I-NEXT:    or t0, t1, t0
-; RV32I-NEXT:    lbu t1, 2(a1)
+; RV32I-NEXT:    lbu t1, 1(a1)
+; RV32I-NEXT:    lbu t4, 0(a1)
+; RV32I-NEXT:    lbu t5, 2(a1)
 ; RV32I-NEXT:    lbu a1, 3(a1)
-; RV32I-NEXT:    slli t5, t5, 8
-; RV32I-NEXT:    or t4, t5, t4
-; RV32I-NEXT:    slli t1, t1, 16
+; RV32I-NEXT:    slli t1, t1, 8
+; RV32I-NEXT:    or t1, t1, t4
+; RV32I-NEXT:    slli t5, t5, 16
 ; RV32I-NEXT:    slli a1, a1, 24
-; RV32I-NEXT:    or a1, a1, t1
+; RV32I-NEXT:    or a1, a1, t5
 ; RV32I-NEXT:    or a3, a5, a3
 ; RV32I-NEXT:    mv a5, sp
 ; RV32I-NEXT:    slli t2, t2, 16
 ; RV32I-NEXT:    slli a0, a0, 24
-; RV32I-NEXT:    or t1, a0, t2
+; RV32I-NEXT:    or t2, a0, t2
 ; RV32I-NEXT:    srai a0, a0, 31
 ; RV32I-NEXT:    or a6, a7, a6
 ; RV32I-NEXT:    or a4, t3, a4
-; RV32I-NEXT:    or a7, t1, t0
-; RV32I-NEXT:    or a1, a1, t4
+; RV32I-NEXT:    or a7, t2, t0
+; RV32I-NEXT:    or a1, a1, t1
 ; RV32I-NEXT:    sw a0, 16(sp)
 ; RV32I-NEXT:    sw a0, 20(sp)
 ; RV32I-NEXT:    sw a0, 24(sp)
@@ -1240,28 +1240,28 @@ define void @ashr_16bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV32I-NEXT:    srli a0, a1, 3
 ; RV32I-NEXT:    andi a3, a1, 31
 ; RV32I-NEXT:    andi a0, a0, 12
-; RV32I-NEXT:    add a0, a5, a0
-; RV32I-NEXT:    lw a4, 0(a0)
-; RV32I-NEXT:    lw a5, 4(a0)
-; RV32I-NEXT:    lw a6, 8(a0)
 ; RV32I-NEXT:    xori a3, a3, 31
+; RV32I-NEXT:    add a0, a5, a0
+; RV32I-NEXT:    lw a4, 4(a0)
+; RV32I-NEXT:    lw a5, 8(a0)
+; RV32I-NEXT:    lw a6, 0(a0)
 ; RV32I-NEXT:    lw a0, 12(a0)
-; RV32I-NEXT:    srl a7, a5, a1
-; RV32I-NEXT:    slli t0, a6, 1
-; RV32I-NEXT:    srl a4, a4, a1
-; RV32I-NEXT:    slli a5, a5, 1
+; RV32I-NEXT:    srl a7, a4, a1
+; RV32I-NEXT:    slli t0, a5, 1
 ; RV32I-NEXT:    srl a6, a6, a1
+; RV32I-NEXT:    slli a4, a4, 1
+; RV32I-NEXT:    srl a5, a5, a1
 ; RV32I-NEXT:    slli t1, a0, 1
 ; RV32I-NEXT:    sra a0, a0, a1
 ; RV32I-NEXT:    sll a1, t0, a3
-; RV32I-NEXT:    sll a5, a5, a3
+; RV32I-NEXT:    sll a4, a4, a3
 ; RV32I-NEXT:    sll a3, t1, a3
 ; RV32I-NEXT:    srli t0, a0, 16
 ; RV32I-NEXT:    srli t1, a0, 24
 ; RV32I-NEXT:    srli t2, a0, 8
 ; RV32I-NEXT:    or a1, a7, a1
-; RV32I-NEXT:    or a4, a4, a5
-; RV32I-NEXT:    or a3, a6, a3
+; RV32I-NEXT:    or a4, a6, a4
+; RV32I-NEXT:    or a3, a5, a3
 ; RV32I-NEXT:    sb a0, 12(a2)
 ; RV32I-NEXT:    sb t2, 13(a2)
 ; RV32I-NEXT:    sb t0, 14(a2)
@@ -1388,17 +1388,17 @@ define void @lshr_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli s5, s5, 16
 ; RV64I-NEXT:    slli s6, s6, 24
 ; RV64I-NEXT:    slli s3, s3, 8
-; RV64I-NEXT:    or s5, s6, s5
-; RV64I-NEXT:    or s2, s3, s2
-; RV64I-NEXT:    lbu s3, 4(a1)
-; RV64I-NEXT:    lbu s6, 5(a1)
 ; RV64I-NEXT:    slli s4, s4, 16
 ; RV64I-NEXT:    slli s7, s7, 24
-; RV64I-NEXT:    or s4, s7, s4
+; RV64I-NEXT:    or s5, s6, s5
+; RV64I-NEXT:    or s2, s3, s2
+; RV64I-NEXT:    or s3, s7, s4
+; RV64I-NEXT:    lbu s4, 5(a1)
+; RV64I-NEXT:    lbu s6, 4(a1)
 ; RV64I-NEXT:    lbu s7, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli s6, s6, 8
-; RV64I-NEXT:    or s3, s6, s3
+; RV64I-NEXT:    slli s4, s4, 8
+; RV64I-NEXT:    or s4, s4, s6
 ; RV64I-NEXT:    slli s7, s7, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, s7
@@ -1415,8 +1415,8 @@ define void @lshr_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    or a0, a0, t5
 ; RV64I-NEXT:    or t1, s0, t6
 ; RV64I-NEXT:    or t2, s5, s1
-; RV64I-NEXT:    or t3, s4, s2
-; RV64I-NEXT:    or a1, a1, s3
+; RV64I-NEXT:    or t3, s3, s2
+; RV64I-NEXT:    or a1, a1, s4
 ; RV64I-NEXT:    slli a3, a3, 32
 ; RV64I-NEXT:    slli a7, a7, 32
 ; RV64I-NEXT:    slli a0, a0, 32
@@ -1434,23 +1434,23 @@ define void @lshr_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    srli a0, a1, 3
 ; RV64I-NEXT:    andi a3, a1, 63
 ; RV64I-NEXT:    andi a0, a0, 24
-; RV64I-NEXT:    add a0, a6, a0
-; RV64I-NEXT:    ld a4, 0(a0)
-; RV64I-NEXT:    ld a5, 8(a0)
-; RV64I-NEXT:    ld a6, 16(a0)
 ; RV64I-NEXT:    xori a3, a3, 63
+; RV64I-NEXT:    add a0, a6, a0
+; RV64I-NEXT:    ld a4, 8(a0)
+; RV64I-NEXT:    ld a5, 16(a0)
+; RV64I-NEXT:    ld a6, 0(a0)
 ; RV64I-NEXT:    ld a0, 24(a0)
-; RV64I-NEXT:    srl a7, a5, a1
-; RV64I-NEXT:    slli t0, a6, 1
-; RV64I-NEXT:    srl a4, a4, a1
-; RV64I-NEXT:    slli a5, a5, 1
+; RV64I-NEXT:    srl a7, a4, a1
+; RV64I-NEXT:    slli t0, a5, 1
 ; RV64I-NEXT:    srl a6, a6, a1
+; RV64I-NEXT:    slli a4, a4, 1
+; RV64I-NEXT:    srl a5, a5, a1
 ; RV64I-NEXT:    slli t1, a0, 1
 ; RV64I-NEXT:    srl t2, a0, a1
 ; RV64I-NEXT:    sll a0, t0, a3
-; RV64I-NEXT:    sll a1, a5, a3
+; RV64I-NEXT:    sll a1, a4, a3
 ; RV64I-NEXT:    sll a3, t1, a3
-; RV64I-NEXT:    srli a5, t2, 56
+; RV64I-NEXT:    srli a4, t2, 56
 ; RV64I-NEXT:    srli t0, t2, 48
 ; RV64I-NEXT:    srli t1, t2, 40
 ; RV64I-NEXT:    srli t3, t2, 32
@@ -1458,12 +1458,12 @@ define void @lshr_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    srli t5, t2, 16
 ; RV64I-NEXT:    srli t6, t2, 8
 ; RV64I-NEXT:    or a0, a7, a0
-; RV64I-NEXT:    or a1, a4, a1
-; RV64I-NEXT:    or a3, a6, a3
+; RV64I-NEXT:    or a1, a6, a1
+; RV64I-NEXT:    or a3, a5, a3
 ; RV64I-NEXT:    sb t3, 28(a2)
 ; RV64I-NEXT:    sb t1, 29(a2)
 ; RV64I-NEXT:    sb t0, 30(a2)
-; RV64I-NEXT:    sb a5, 31(a2)
+; RV64I-NEXT:    sb a4, 31(a2)
 ; RV64I-NEXT:    sb t2, 24(a2)
 ; RV64I-NEXT:    sb t6, 25(a2)
 ; RV64I-NEXT:    sb t5, 26(a2)
@@ -1864,17 +1864,17 @@ define void @shl_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli s5, s5, 16
 ; RV64I-NEXT:    slli s6, s6, 24
 ; RV64I-NEXT:    slli s3, s3, 8
-; RV64I-NEXT:    or s5, s6, s5
-; RV64I-NEXT:    or s2, s3, s2
-; RV64I-NEXT:    lbu s3, 4(a1)
-; RV64I-NEXT:    lbu s6, 5(a1)
 ; RV64I-NEXT:    slli s4, s4, 16
 ; RV64I-NEXT:    slli s7, s7, 24
-; RV64I-NEXT:    or s4, s7, s4
+; RV64I-NEXT:    or s5, s6, s5
+; RV64I-NEXT:    or s2, s3, s2
+; RV64I-NEXT:    or s3, s7, s4
+; RV64I-NEXT:    lbu s4, 5(a1)
+; RV64I-NEXT:    lbu s6, 4(a1)
 ; RV64I-NEXT:    lbu s7, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli s6, s6, 8
-; RV64I-NEXT:    or s3, s6, s3
+; RV64I-NEXT:    slli s4, s4, 8
+; RV64I-NEXT:    or s4, s4, s6
 ; RV64I-NEXT:    slli s7, s7, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, s7
@@ -1891,8 +1891,8 @@ define void @shl_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    or a0, a0, t5
 ; RV64I-NEXT:    or t1, s0, t6
 ; RV64I-NEXT:    or t2, s5, s1
-; RV64I-NEXT:    or t3, s4, s2
-; RV64I-NEXT:    or a1, a1, s3
+; RV64I-NEXT:    or t3, s3, s2
+; RV64I-NEXT:    or a1, a1, s4
 ; RV64I-NEXT:    slli a3, a3, 32
 ; RV64I-NEXT:    slli a7, a7, 32
 ; RV64I-NEXT:    slli a0, a0, 32
@@ -2340,17 +2340,17 @@ define void @ashr_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    slli s5, s5, 16
 ; RV64I-NEXT:    slli s6, s6, 24
 ; RV64I-NEXT:    slli s3, s3, 8
-; RV64I-NEXT:    or s5, s6, s5
-; RV64I-NEXT:    or s2, s3, s2
-; RV64I-NEXT:    lbu s3, 4(a1)
-; RV64I-NEXT:    lbu s6, 5(a1)
 ; RV64I-NEXT:    slli s4, s4, 16
 ; RV64I-NEXT:    slli s7, s7, 24
-; RV64I-NEXT:    or s4, s7, s4
+; RV64I-NEXT:    or s5, s6, s5
+; RV64I-NEXT:    or s2, s3, s2
+; RV64I-NEXT:    or s3, s7, s4
+; RV64I-NEXT:    lbu s4, 5(a1)
+; RV64I-NEXT:    lbu s6, 4(a1)
 ; RV64I-NEXT:    lbu s7, 6(a1)
 ; RV64I-NEXT:    lbu a1, 7(a1)
-; RV64I-NEXT:    slli s6, s6, 8
-; RV64I-NEXT:    or s3, s6, s3
+; RV64I-NEXT:    slli s4, s4, 8
+; RV64I-NEXT:    or s4, s4, s6
 ; RV64I-NEXT:    slli s7, s7, 16
 ; RV64I-NEXT:    slli a1, a1, 24
 ; RV64I-NEXT:    or a1, a1, s7
@@ -2363,8 +2363,8 @@ define void @ashr_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    or a0, a0, t5
 ; RV64I-NEXT:    or t0, s0, t6
 ; RV64I-NEXT:    or t1, s5, s1
-; RV64I-NEXT:    or t2, s4, s2
-; RV64I-NEXT:    or a1, a1, s3
+; RV64I-NEXT:    or t2, s3, s2
+; RV64I-NEXT:    or a1, a1, s4
 ; RV64I-NEXT:    slli a4, a4, 32
 ; RV64I-NEXT:    slli a6, a6, 32
 ; RV64I-NEXT:    slli a0, a0, 32
@@ -2387,23 +2387,23 @@ define void @ashr_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    srli a0, a1, 3
 ; RV64I-NEXT:    andi a3, a1, 63
 ; RV64I-NEXT:    andi a0, a0, 24
-; RV64I-NEXT:    add a0, s6, a0
-; RV64I-NEXT:    ld a4, 0(a0)
-; RV64I-NEXT:    ld a5, 8(a0)
-; RV64I-NEXT:    ld a6, 16(a0)
 ; RV64I-NEXT:    xori a3, a3, 63
+; RV64I-NEXT:    add a0, s6, a0
+; RV64I-NEXT:    ld a4, 8(a0)
+; RV64I-NEXT:    ld a5, 16(a0)
+; RV64I-NEXT:    ld a6, 0(a0)
 ; RV64I-NEXT:    ld a0, 24(a0)
-; RV64I-NEXT:    srl a7, a5, a1
-; RV64I-NEXT:    slli t0, a6, 1
-; RV64I-NEXT:    srl a4, a4, a1
-; RV64I-NEXT:    slli a5, a5, 1
+; RV64I-NEXT:    srl a7, a4, a1
+; RV64I-NEXT:    slli t0, a5, 1
 ; RV64I-NEXT:    srl a6, a6, a1
+; RV64I-NEXT:    slli a4, a4, 1
+; RV64I-NEXT:    srl a5, a5, a1
 ; RV64I-NEXT:    slli t1, a0, 1
 ; RV64I-NEXT:    sra t2, a0, a1
 ; RV64I-NEXT:    sll a0, t0, a3
-; RV64I-NEXT:    sll a1, a5, a3
+; RV64I-NEXT:    sll a1, a4, a3
 ; RV64I-NEXT:    sll a3, t1, a3
-; RV64I-NEXT:    srli a5, t2, 56
+; RV64I-NEXT:    srli a4, t2, 56
 ; RV64I-NEXT:    srli t0, t2, 48
 ; RV64I-NEXT:    srli t1, t2, 40
 ; RV64I-NEXT:    srli t3, t2, 32
@@ -2411,12 +2411,12 @@ define void @ashr_32bytes(ptr %src.ptr, ptr %bitOff.ptr, ptr %dst) nounwind {
 ; RV64I-NEXT:    srli t5, t2, 16
 ; RV64I-NEXT:    srli t6, t2, 8
 ; RV64I-NEXT:    or a0, a7, a0
-; RV64I-NEXT:    or a1, a4, a1
-; RV64I-NEXT:    or a3, a6, a3
+; RV64I-NEXT:    or a1, a6, a1
+; RV64I-NEXT:    or a3, a5, a3
 ; RV64I-NEXT:    sb t3, 28(a2)
 ; RV64I-NEXT:    sb t1, 29(a2)
 ; RV64I-NEXT:    sb t0, 30(a2)
-; RV64I-NEXT:    sb a5, 31(a2)
+; RV64I-NEXT:    sb a4, 31(a2)
 ; RV64I-NEXT:    sb t2, 24(a2)
 ; RV64I-NEXT:    sb t6, 25(a2)
 ; RV64I-NEXT:    sb t5, 26(a2)
diff --git a/llvm/test/CodeGen/RISCV/xtheadmempair.ll b/llvm/test/CodeGen/RISCV/xtheadmempair.ll
index 3525c40026064..4df61dad7d039 100644
--- a/llvm/test/CodeGen/RISCV/xtheadmempair.ll
+++ b/llvm/test/CodeGen/RISCV/xtheadmempair.ll
@@ -57,14 +57,14 @@ define i64 @lwud(ptr %a) {
 define i64 @ldd(ptr %a) {
 ; RV32XTHEADMEMPAIR-LABEL: ldd:
 ; RV32XTHEADMEMPAIR:       # %bb.0:
-; RV32XTHEADMEMPAIR-NEXT:    lw a1, 44(a0)
-; RV32XTHEADMEMPAIR-NEXT:    lw a2, 32(a0)
-; RV32XTHEADMEMPAIR-NEXT:    lw a3, 36(a0)
+; RV32XTHEADMEMPAIR-NEXT:    lw a1, 32(a0)
+; RV32XTHEADMEMPAIR-NEXT:    lw a2, 36(a0)
+; RV32XTHEADMEMPAIR-NEXT:    lw a3, 44(a0)
 ; RV32XTHEADMEMPAIR-NEXT:    lw a0, 40(a0)
-; RV32XTHEADMEMPAIR-NEXT:    add a1, a3, a1
-; RV32XTHEADMEMPAIR-NEXT:    add a0, a2, a0
-; RV32XTHEADMEMPAIR-NEXT:    sltu a2, a0, a2
-; RV32XTHEADMEMPAIR-NEXT:    add a1, a1, a2
+; RV32XTHEADMEMPAIR-NEXT:    add a2, a2, a3
+; RV32XTHEADMEMPAIR-NEXT:    add a0, a1, a0
+; RV32XTHEADMEMPAIR-NEXT:    sltu a1, a0, a1
+; RV32XTHEADMEMPAIR-NEXT:    add a1, a2, a1
 ; RV32XTHEADMEMPAIR-NEXT:    ret
 ;
 ; RV64XTHEADMEMPAIR-LABEL: ldd:
diff --git a/llvm/test/CodeGen/RISCV/zilsd.ll b/llvm/test/CodeGen/RISCV/zilsd.ll
index 6a4578bb02d8d..09b065a8e9ddd 100644
--- a/llvm/test/CodeGen/RISCV/zilsd.ll
+++ b/llvm/test/CodeGen/RISCV/zilsd.ll
@@ -36,25 +36,25 @@ define i64 @load_unaligned(ptr %p) {
 ; SLOW-LABEL: load_unaligned:
 ; SLOW:       # %bb.0:
 ; SLOW-NEXT:    lbu a1, 1(a0)
-; SLOW-NEXT:    lbu a2, 0(a0)
-; SLOW-NEXT:    lbu a3, 2(a0)
-; SLOW-NEXT:    lbu a4, 3(a0)
+; SLOW-NEXT:    lbu a2, 2(a0)
+; SLOW-NEXT:    lbu a3, 3(a0)
+; SLOW-NEXT:    lbu a4, 0(a0)
 ; SLOW-NEXT:    slli a1, a1, 8
-; SLOW-NEXT:    or a1, a1, a2
-; SLOW-NEXT:    lbu a2, 4(a0)
-; SLOW-NEXT:    lbu a5, 5(a0)
-; SLOW-NEXT:    slli a3, a3, 16
-; SLOW-NEXT:    slli a4, a4, 24
-; SLOW-NEXT:    or a3, a4, a3
-; SLOW-NEXT:    lbu a4, 6(a0)
+; SLOW-NEXT:    slli a2, a2, 16
+; SLOW-NEXT:    slli a3, a3, 24
+; SLOW-NEXT:    or a1, a1, a4
+; SLOW-NEXT:    or a2, a3, a2
+; SLOW-NEXT:    lbu a3, 5(a0)
+; SLOW-NEXT:    lbu a4, 4(a0)
+; SLOW-NEXT:    lbu a5, 6(a0)
 ; SLOW-NEXT:    lbu a0, 7(a0)
-; SLOW-NEXT:    slli a5, a5, 8
-; SLOW-NEXT:    or a2, a5, a2
-; SLOW-NEXT:    slli a4, a4, 16
+; SLOW-NEXT:    slli a3, a3, 8
+; SLOW-NEXT:    or a3, a3, a4
+; SLOW-NEXT:    slli a5, a5, 16
 ; SLOW-NEXT:    slli a0, a0, 24
-; SLOW-NEXT:    or a4, a0, a4
-; SLOW-NEXT:    or a0, a3, a1
-; SLOW-NEXT:    or a1, a4, a2
+; SLOW-NEXT:    or a5, a0, a5
+; SLOW-NEXT:    or a0, a2, a1
+; SLOW-NEXT:    or a1, a5, a3
 ; SLOW-NEXT:    ret
 ;
 ; FAST-LABEL: load_unaligned:
diff --git a/llvm/test/Transforms/LoopStrengthReduce/X86/ivchain-X86.ll b/llvm/test/Transforms/LoopStrengthReduce/X86/ivchain-X86.ll
index 93203ef6e17f5..c12d8135e5eba 100644
--- a/llvm/test/Transforms/LoopStrengthReduce/X86/ivchain-X86.ll
+++ b/llvm/test/Transforms/LoopStrengthReduce/X86/ivchain-X86.ll
@@ -25,20 +25,18 @@ define i32 @simple(ptr %a, ptr %b, i32 %x) nounwind {
 ; X64-NEXT:    # =>This Inner Loop Header: Depth=1
 ; X64-NEXT:    addl (%rdi), %eax
 ; X64-NEXT:    addl (%rdi,%rcx), %eax
-; X64-NEXT:    leaq (%rdi,%rcx), %r8
-; X64-NEXT:    addl (%rcx,%r8), %eax
-; X64-NEXT:    addq %rcx, %r8
-; X64-NEXT:    addl (%rcx,%r8), %eax
-; X64-NEXT:    addq %rdx, %r8
-; X64-NEXT:    movq %r8, %rdi
-; X64-NEXT:    cmpq %rsi, %r8
+; X64-NEXT:    addq %rcx, %rdi
+; X64-NEXT:    addl (%rcx,%rdi), %eax
+; X64-NEXT:    addq %rcx, %rdi
+; X64-NEXT:    addl (%rcx,%rdi), %eax
+; X64-NEXT:    addq %rdx, %rdi
+; X64-NEXT:    cmpq %rsi, %rdi
 ; X64-NEXT:    jne .LBB0_1
 ; X64-NEXT:  # %bb.2: # %exit
 ; X64-NEXT:    retq
 ;
 ; X32-LABEL: simple:
 ; X32:       # %bb.0: # %entry
-; X32-NEXT:    pushl %ebx
 ; X32-NEXT:    pushl %edi
 ; X32-NEXT:    pushl %esi
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
@@ -52,18 +50,16 @@ define i32 @simple(ptr %a, ptr %b, i32 %x) nounwind {
 ; X32-NEXT:    # =>This Inner Loop Header: Depth=1
 ; X32-NEXT:    addl (%edi), %eax
 ; X32-NEXT:    addl (%edi,%edx), %eax
-; X32-NEXT:    leal (%edi,%edx), %ebx
-; X32-NEXT:    addl (%edx,%ebx), %eax
-; X32-NEXT:    addl %edx, %ebx
-; X32-NEXT:    addl (%edx,%ebx), %eax
-; X32-NEXT:    addl %esi, %ebx
-; X32-NEXT:    movl %ebx, %edi
-; X32-NEXT:    cmpl %ecx, %ebx
+; X32-NEXT:    addl %edx, %edi
+; X32-NEXT:    addl (%edx,%edi), %eax
+; X32-NEXT:    addl %edx, %edi
+; X32-NEXT:    addl (%edx,%edi), %eax
+; X32-NEXT:    addl %esi, %edi
+; X32-NEXT:    cmpl %ecx, %edi
 ; X32-NEXT:    jne .LBB0_1
 ; X32-NEXT:  # %bb.2: # %exit
 ; X32-NEXT:    popl %esi
 ; X32-NEXT:    popl %edi
-; X32-NEXT:    popl %ebx
 ; X32-NEXT:    retl
 entry:
   br label %loop