diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
index 22b921fb2084f..5f1983791cfae 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
@@ -45,12 +45,6 @@ static cl::opt<bool> WidenLoads(
   cl::ReallyHidden,
   cl::init(false));
 
-static cl::opt<bool> Widen16BitOps(
-    "amdgpu-codegenprepare-widen-16-bit-ops",
-    cl::desc(
-        "Widen uniform 16-bit instructions to 32-bit in AMDGPUCodeGenPrepare"),
-    cl::ReallyHidden, cl::init(false));
-
 static cl::opt<bool>
     BreakLargePHIs("amdgpu-codegenprepare-break-large-phis",
                    cl::desc("Break large PHI nodes for DAGISel"),
@@ -150,18 +144,6 @@ class AMDGPUCodeGenPrepareImpl
 
   bool canBreakPHINode(const PHINode &I);
 
-  /// Copies exact/nsw/nuw flags (if any) from binary operation \p I to
-  /// binary operation \p V.
-  ///
-  /// \returns Binary operation \p V.
-  /// \returns \p T's base element bit width.
-  unsigned getBaseElementBitWidth(const Type *T) const;
-
-  /// \returns Equivalent 32 bit integer type for given type \p T. For example,
-  /// if \p T is i7, then i32 is returned; if \p T is <3 x i12>, then <3 x i32>
-  /// is returned.
-  Type *getI32Ty(IRBuilder<> &B, const Type *T) const;
-
   /// \returns True if binary operation \p I is a signed binary operation, false
   /// otherwise.
   bool isSigned(const BinaryOperator &I) const;
@@ -170,10 +152,6 @@ class AMDGPUCodeGenPrepareImpl
   /// signed 'icmp' operation, false otherwise.
   bool isSigned(const SelectInst &I) const;
 
-  /// \returns True if type \p T needs to be promoted to 32 bit integer type,
-  /// false otherwise.
-  bool needsPromotionToI32(const Type *T) const;
-
   /// Return true if \p T is a legal scalar floating point type.
   bool isLegalFloatingTy(const Type *T) const;
 
@@ -188,52 +166,6 @@ class AMDGPUCodeGenPrepareImpl
            computeKnownFPClass(V, fcSubnormal, CtxI).isKnownNeverSubnormal();
   }
 
-  /// Promotes uniform binary operation \p I to equivalent 32 bit binary
-  /// operation.
-  ///
-  /// \details \p I's base element bit width must be greater than 1 and less
-  /// than or equal 16. Promotion is done by sign or zero extending operands to
-  /// 32 bits, replacing \p I with equivalent 32 bit binary operation, and
-  /// truncating the result of 32 bit binary operation back to \p I's original
-  /// type. Division operation is not promoted.
-  ///
-  /// \returns True if \p I is promoted to equivalent 32 bit binary operation,
-  /// false otherwise.
-  bool promoteUniformOpToI32(BinaryOperator &I) const;
-
-  /// Promotes uniform 'icmp' operation \p I to 32 bit 'icmp' operation.
-  ///
-  /// \details \p I's base element bit width must be greater than 1 and less
-  /// than or equal 16. Promotion is done by sign or zero extending operands to
-  /// 32 bits, and replacing \p I with 32 bit 'icmp' operation.
-  ///
-  /// \returns True.
-  bool promoteUniformOpToI32(ICmpInst &I) const;
-
-  /// Promotes uniform 'select' operation \p I to 32 bit 'select'
-  /// operation.
-  ///
-  /// \details \p I's base element bit width must be greater than 1 and less
-  /// than or equal 16. Promotion is done by sign or zero extending operands to
-  /// 32 bits, replacing \p I with 32 bit 'select' operation, and truncating the
-  /// result of 32 bit 'select' operation back to \p I's original type.
-  ///
-  /// \returns True.
-  bool promoteUniformOpToI32(SelectInst &I) const;
-
-  /// Promotes uniform 'bitreverse' intrinsic \p I to 32 bit 'bitreverse'
-  /// intrinsic.
-  ///
-  /// \details \p I's base element bit width must be greater than 1 and less
-  /// than or equal 16. Promotion is done by zero extending the operand to 32
-  /// bits, replacing \p I with 32 bit 'bitreverse' intrinsic, shifting the
-  /// result of 32 bit 'bitreverse' intrinsic to the right with zero fill (the
-  /// shift amount is 32 minus \p I's base element bit width), and truncating
-  /// the result of the shift operation back to \p I's original type.
-  ///
-  /// \returns True.
-  bool promoteUniformBitreverseToI32(IntrinsicInst &I) const;
-
   /// \returns The minimum number of bits needed to store the value of \Op as an
   /// unsigned integer. Truncating to this size and then zero-extending to
   /// the original will not change the value.
@@ -320,13 +252,11 @@ class AMDGPUCodeGenPrepareImpl
   bool visitInstruction(Instruction &I) { return false; }
   bool visitBinaryOperator(BinaryOperator &I);
   bool visitLoadInst(LoadInst &I);
-  bool visitICmpInst(ICmpInst &I);
   bool visitSelectInst(SelectInst &I);
   bool visitPHINode(PHINode &I);
   bool visitAddrSpaceCastInst(AddrSpaceCastInst &I);
 
   bool visitIntrinsicInst(IntrinsicInst &I);
-  bool visitBitreverseIntrinsicInst(IntrinsicInst &I);
   bool visitFMinLike(IntrinsicInst &I);
   bool visitSqrt(IntrinsicInst &I);
   bool run();
@@ -380,22 +310,6 @@ bool AMDGPUCodeGenPrepareImpl::run() {
   return MadeChange;
 }
 
-unsigned AMDGPUCodeGenPrepareImpl::getBaseElementBitWidth(const Type *T) const {
-  assert(needsPromotionToI32(T) && "T does not need promotion to i32");
-
-  if (T->isIntegerTy())
-    return T->getIntegerBitWidth();
-  return cast<VectorType>(T)->getElementType()->getIntegerBitWidth();
-}
-
-Type *AMDGPUCodeGenPrepareImpl::getI32Ty(IRBuilder<> &B, const Type *T) const {
-  assert(needsPromotionToI32(T) && "T does not need promotion to i32");
-
-  if (T->isIntegerTy())
-    return B.getInt32Ty();
-  return FixedVectorType::get(B.getInt32Ty(), cast<FixedVectorType>(T));
-}
-
 bool AMDGPUCodeGenPrepareImpl::isSigned(const BinaryOperator &I) const {
   return I.getOpcode() == Instruction::AShr ||
       I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::SRem;
@@ -406,59 +320,11 @@ bool AMDGPUCodeGenPrepareImpl::isSigned(const SelectInst &I) const {
          cast<ICmpInst>(I.getOperand(0))->isSigned();
 }
 
-bool AMDGPUCodeGenPrepareImpl::needsPromotionToI32(const Type *T) const {
-  if (!Widen16BitOps)
-    return false;
-
-  const IntegerType *IntTy = dyn_cast<IntegerType>(T);
-  if (IntTy && IntTy->getBitWidth() > 1 && IntTy->getBitWidth() <= 16)
-    return true;
-
-  if (const VectorType *VT = dyn_cast<VectorType>(T)) {
-    // TODO: The set of packed operations is more limited, so may want to
-    // promote some anyway.
-    if (ST.hasVOP3PInsts())
-      return false;
-
-    return needsPromotionToI32(VT->getElementType());
-  }
-
-  return false;
-}
-
 bool AMDGPUCodeGenPrepareImpl::isLegalFloatingTy(const Type *Ty) const {
   return Ty->isFloatTy() || Ty->isDoubleTy() ||
          (Ty->isHalfTy() && ST.has16BitInsts());
 }
 
-// Return true if the op promoted to i32 should have nsw set.
-static bool promotedOpIsNSW(const Instruction &I) {
-  switch (I.getOpcode()) {
-  case Instruction::Shl:
-  case Instruction::Add:
-  case Instruction::Sub:
-    return true;
-  case Instruction::Mul:
-    return I.hasNoUnsignedWrap();
-  default:
-    return false;
-  }
-}
-
-// Return true if the op promoted to i32 should have nuw set.
-static bool promotedOpIsNUW(const Instruction &I) {
-  switch (I.getOpcode()) {
-  case Instruction::Shl:
-  case Instruction::Add:
-  case Instruction::Mul:
-    return true;
-  case Instruction::Sub:
-    return I.hasNoUnsignedWrap();
-  default:
-    return false;
-  }
-}
-
 bool AMDGPUCodeGenPrepareImpl::canWidenScalarExtLoad(LoadInst &I) const {
   Type *Ty = I.getType();
   int TySize = DL.getTypeSizeInBits(Ty);
@@ -467,134 +333,6 @@ bool AMDGPUCodeGenPrepareImpl::canWidenScalarExtLoad(LoadInst &I) const {
   return I.isSimple() && TySize < 32 && Alignment >= 4 && UA.isUniform(&I);
 }
 
-bool AMDGPUCodeGenPrepareImpl::promoteUniformOpToI32(BinaryOperator &I) const {
-  assert(needsPromotionToI32(I.getType()) &&
-         "I does not need promotion to i32");
-
-  if (I.getOpcode() == Instruction::SDiv ||
-      I.getOpcode() == Instruction::UDiv ||
-      I.getOpcode() == Instruction::SRem ||
-      I.getOpcode() == Instruction::URem)
-    return false;
-
-  IRBuilder<> Builder(&I);
-  Builder.SetCurrentDebugLocation(I.getDebugLoc());
-
-  Type *I32Ty = getI32Ty(Builder, I.getType());
-  Value *ExtOp0 = nullptr;
-  Value *ExtOp1 = nullptr;
-  Value *ExtRes = nullptr;
-  Value *TruncRes = nullptr;
-
-  if (isSigned(I)) {
-    ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
-    ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
-  } else {
-    ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
-    ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
-  }
-
-  ExtRes = Builder.CreateBinOp(I.getOpcode(), ExtOp0, ExtOp1);
-  if (Instruction *Inst = dyn_cast<Instruction>(ExtRes)) {
-    if (promotedOpIsNSW(cast<Instruction>(I)))
-      Inst->setHasNoSignedWrap();
-
-    if (promotedOpIsNUW(cast<Instruction>(I)))
-      Inst->setHasNoUnsignedWrap();
-
-    if (const auto *ExactOp = dyn_cast<PossiblyExactOperator>(&I))
-      Inst->setIsExact(ExactOp->isExact());
-  }
-
-  TruncRes = Builder.CreateTrunc(ExtRes, I.getType());
-
-  I.replaceAllUsesWith(TruncRes);
-  I.eraseFromParent();
-
-  return true;
-}
-
-bool AMDGPUCodeGenPrepareImpl::promoteUniformOpToI32(ICmpInst &I) const {
-  assert(needsPromotionToI32(I.getOperand(0)->getType()) &&
-         "I does not need promotion to i32");
-
-  IRBuilder<> Builder(&I);
-  Builder.SetCurrentDebugLocation(I.getDebugLoc());
-
-  Type *I32Ty = getI32Ty(Builder, I.getOperand(0)->getType());
-  Value *ExtOp0 = nullptr;
-  Value *ExtOp1 = nullptr;
-  Value *NewICmp  = nullptr;
-
-  if (I.isSigned()) {
-    ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
-    ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
-  } else {
-    ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
-    ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
-  }
-  NewICmp = Builder.CreateICmp(I.getPredicate(), ExtOp0, ExtOp1);
-
-  I.replaceAllUsesWith(NewICmp);
-  I.eraseFromParent();
-
-  return true;
-}
-
-bool AMDGPUCodeGenPrepareImpl::promoteUniformOpToI32(SelectInst &I) const {
-  assert(needsPromotionToI32(I.getType()) &&
-         "I does not need promotion to i32");
-
-  IRBuilder<> Builder(&I);
-  Builder.SetCurrentDebugLocation(I.getDebugLoc());
-
-  Type *I32Ty = getI32Ty(Builder, I.getType());
-  Value *ExtOp1 = nullptr;
-  Value *ExtOp2 = nullptr;
-  Value *ExtRes = nullptr;
-  Value *TruncRes = nullptr;
-
-  if (isSigned(I)) {
-    ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
-    ExtOp2 = Builder.CreateSExt(I.getOperand(2), I32Ty);
-  } else {
-    ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
-    ExtOp2 = Builder.CreateZExt(I.getOperand(2), I32Ty);
-  }
-  ExtRes = Builder.CreateSelect(I.getOperand(0), ExtOp1, ExtOp2);
-  TruncRes = Builder.CreateTrunc(ExtRes, I.getType());
-
-  I.replaceAllUsesWith(TruncRes);
-  I.eraseFromParent();
-
-  return true;
-}
-
-bool AMDGPUCodeGenPrepareImpl::promoteUniformBitreverseToI32(
-    IntrinsicInst &I) const {
-  assert(I.getIntrinsicID() == Intrinsic::bitreverse &&
-         "I must be bitreverse intrinsic");
-  assert(needsPromotionToI32(I.getType()) &&
-         "I does not need promotion to i32");
-
-  IRBuilder<> Builder(&I);
-  Builder.SetCurrentDebugLocation(I.getDebugLoc());
-
-  Type *I32Ty = getI32Ty(Builder, I.getType());
-  Value *ExtOp = Builder.CreateZExt(I.getOperand(0), I32Ty);
-  Value *ExtRes =
-      Builder.CreateIntrinsic(Intrinsic::bitreverse, {I32Ty}, {ExtOp});
-  Value *LShrOp =
-      Builder.CreateLShr(ExtRes, 32 - getBaseElementBitWidth(I.getType()));
-  Value *TruncRes =
-      Builder.CreateTrunc(LShrOp, I.getType());
-
-  I.replaceAllUsesWith(TruncRes);
-  I.eraseFromParent();
-
-  return true;
-}
-
 unsigned AMDGPUCodeGenPrepareImpl::numBitsUnsigned(Value *Op) const {
   return computeKnownBits(Op, DL, AC).countMaxActiveBits();
 }
@@ -1635,10 +1373,6 @@ bool AMDGPUCodeGenPrepareImpl::visitBinaryOperator(BinaryOperator &I) {
   if (foldBinOpIntoSelect(I))
     return true;
 
-  if (ST.has16BitInsts() && needsPromotionToI32(I.getType()) &&
-      UA.isUniform(&I) && promoteUniformOpToI32(I))
-    return true;
-
   if (UseMul24Intrin && replaceMulWithMul24(I))
     return true;
   if (tryNarrowMathIfNoOverflow(&I, ST.getTargetLowering(),
@@ -1770,16 +1504,6 @@ bool AMDGPUCodeGenPrepareImpl::visitLoadInst(LoadInst &I) {
   return false;
 }
 
-bool AMDGPUCodeGenPrepareImpl::visitICmpInst(ICmpInst &I) {
-  bool Changed = false;
-
-  if (ST.has16BitInsts() && needsPromotionToI32(I.getOperand(0)->getType()) &&
-      UA.isUniform(&I))
-    Changed |= promoteUniformOpToI32(I);
-
-  return Changed;
-}
-
 bool AMDGPUCodeGenPrepareImpl::visitSelectInst(SelectInst &I) {
   Value *Cond = I.getCondition();
   Value *TrueVal = I.getTrueValue();
@@ -1787,12 +1511,6 @@ bool AMDGPUCodeGenPrepareImpl::visitSelectInst(SelectInst &I) {
   Value *CmpVal;
   CmpPredicate Pred;
 
-  if (ST.has16BitInsts() && needsPromotionToI32(I.getType())) {
-    if (UA.isUniform(&I))
-      return promoteUniformOpToI32(I);
-    return false;
-  }
-
   // Match fract pattern with nan check.
   if (!match(Cond, m_FCmp(Pred, m_Value(CmpVal), m_NonNaN())))
     return false;
@@ -2196,8 +1914,6 @@ bool AMDGPUCodeGenPrepareImpl::visitAddrSpaceCastInst(AddrSpaceCastInst &I) {
 
 bool AMDGPUCodeGenPrepareImpl::visitIntrinsicInst(IntrinsicInst &I) {
   switch (I.getIntrinsicID()) {
-  case Intrinsic::bitreverse:
-    return visitBitreverseIntrinsicInst(I);
   case Intrinsic::minnum:
   case Intrinsic::minimumnum:
   case Intrinsic::minimum:
@@ -2209,16 +1925,6 @@ bool AMDGPUCodeGenPrepareImpl::visitIntrinsicInst(IntrinsicInst &I) {
   }
 }
 
-bool AMDGPUCodeGenPrepareImpl::visitBitreverseIntrinsicInst(IntrinsicInst &I) {
-  bool Changed = false;
-
-  if (ST.has16BitInsts() && needsPromotionToI32(I.getType()) &&
-      UA.isUniform(&I))
-    Changed |= promoteUniformBitreverseToI32(I);
-
-  return Changed;
-}
-
 /// Match non-nan fract pattern.
 ///   minnum(fsub(x, floor(x)), nextafter(1.0, -1.0))
 ///   minimumnum(fsub(x, floor(x)), nextafter(1.0, -1.0))
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/andn2.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/andn2.ll
index 32e461ba09f06..e1ef3f9be0a5d 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/andn2.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/andn2.ll
@@ -1,8 +1,8 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -global-isel -amdgpu-codegenprepare-widen-16-bit-ops=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti < %s | FileCheck -check-prefixes=GCN,GFX6 %s
-; RUN: llc -global-isel -amdgpu-codegenprepare-widen-16-bit-ops=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
-; RUN: llc -global-isel -amdgpu-codegenprepare-widen-16-bit-ops=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX10 %s
-; RUN: llc -global-isel -amdgpu-codegenprepare-widen-16-bit-ops=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti < %s | FileCheck -check-prefixes=GCN,GFX6 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX10 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11 %s
 
 define amdgpu_ps i32 @s_andn2_i32(i32 inreg %src0, i32 inreg %src1) {
 ; GCN-LABEL: s_andn2_i32:
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/orn2.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/orn2.ll
index 12b37c386c140..afabc7b62386f 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/orn2.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/orn2.ll
@@ -1,8 +1,8 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -global-isel -amdgpu-codegenprepare-widen-16-bit-ops=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti < %s | FileCheck -check-prefixes=GCN,GFX6 %s
-; RUN: llc -global-isel -amdgpu-codegenprepare-widen-16-bit-ops=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
-; RUN: llc -global-isel -amdgpu-codegenprepare-widen-16-bit-ops=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX10 %s
-; RUN: llc -global-isel -amdgpu-codegenprepare-widen-16-bit-ops=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti < %s | FileCheck -check-prefixes=GCN,GFX6 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX10 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11 %s
 
 define amdgpu_ps i32 @s_orn2_i32(i32 inreg %src0, i32 inreg %src1) {
 ; GCN-LABEL: s_orn2_i32:
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-i16-to-i32.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-i16-to-i32.ll
deleted file mode 100644
index 8945708e0f0ca..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-i16-to-i32.ll
+++ /dev/null
@@ -1,2853 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt -S -amdgpu-codegenprepare-widen-16-bit-ops -mtriple=amdgcn-- -amdgpu-codegenprepare %s | FileCheck -check-prefix=SI %s
-; RUN: opt -S -amdgpu-codegenprepare-widen-16-bit-ops -mtriple=amdgcn-- -mcpu=tonga -amdgpu-codegenprepare %s | FileCheck -check-prefix=VI %s
-
-define amdgpu_kernel void @add_i3(i3 %a, i3 %b) {
-; SI-LABEL: @add_i3(
-; SI-NEXT:    [[R:%.*]] = add i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = add i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nsw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @add_nsw_i3(
-; SI-NEXT:    [[R:%.*]] = add nsw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nsw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = add nsw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nuw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @add_nuw_i3(
-; SI-NEXT:    [[R:%.*]] = add nuw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nuw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = add nuw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nuw_nsw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @add_nuw_nsw_i3(
-; SI-NEXT:    [[R:%.*]] = add nuw nsw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nuw_nsw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = add nuw nsw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_i3(i3 %a, i3 %b) {
-; SI-LABEL: @sub_i3(
-; SI-NEXT:    [[R:%.*]] = sub i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = sub nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = sub i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nsw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @sub_nsw_i3(
-; SI-NEXT:    [[R:%.*]] = sub nsw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nsw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = sub nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = sub nsw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nuw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @sub_nuw_i3(
-; SI-NEXT:    [[R:%.*]] = sub nuw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nuw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = sub nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = sub nuw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nuw_nsw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @sub_nuw_nsw_i3(
-; SI-NEXT:    [[R:%.*]] = sub nuw nsw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nuw_nsw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = sub nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = sub nuw nsw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_i3(i3 %a, i3 %b) {
-; SI-LABEL: @mul_i3(
-; SI-NEXT:    [[R:%.*]] = mul i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = mul i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nsw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @mul_nsw_i3(
-; SI-NEXT:    [[R:%.*]] = mul nsw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nsw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = mul nsw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nuw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @mul_nuw_i3(
-; SI-NEXT:    [[R:%.*]] = mul nuw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nuw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = mul nuw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nuw_nsw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @mul_nuw_nsw_i3(
-; SI-NEXT:    [[R:%.*]] = mul nuw nsw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nuw_nsw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = mul nuw nsw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_i3(i3 %a, i3 %b) {
-; SI-LABEL: @shl_i3(
-; SI-NEXT:    [[R:%.*]] = shl i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = shl i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nsw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @shl_nsw_i3(
-; SI-NEXT:    [[R:%.*]] = shl nsw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nsw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = shl nsw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nuw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @shl_nuw_i3(
-; SI-NEXT:    [[R:%.*]] = shl nuw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nuw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = shl nuw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nuw_nsw_i3(i3 %a, i3 %b) {
-; SI-LABEL: @shl_nuw_nsw_i3(
-; SI-NEXT:    [[R:%.*]] = shl nuw nsw i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nuw_nsw_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = shl nuw nsw i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @lshr_i3(i3 %a, i3 %b) {
-; SI-LABEL: @lshr_i3(
-; SI-NEXT:    [[R:%.*]] = lshr i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @lshr_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = lshr i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = lshr i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @lshr_exact_i3(i3 %a, i3 %b) {
-; SI-LABEL: @lshr_exact_i3(
-; SI-NEXT:    [[R:%.*]] = lshr exact i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @lshr_exact_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = lshr exact i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = lshr exact i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @ashr_i3(i3 %a, i3 %b) {
-; SI-LABEL: @ashr_i3(
-; SI-NEXT:    [[R:%.*]] = ashr i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @ashr_i3(
-; VI-NEXT:    [[TMP1:%.*]] = sext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = ashr i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = ashr i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @ashr_exact_i3(i3 %a, i3 %b) {
-; SI-LABEL: @ashr_exact_i3(
-; SI-NEXT:    [[R:%.*]] = ashr exact i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @ashr_exact_i3(
-; VI-NEXT:    [[TMP1:%.*]] = sext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = ashr exact i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = ashr exact i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @and_i3(i3 %a, i3 %b) {
-; SI-LABEL: @and_i3(
-; SI-NEXT:    [[R:%.*]] = and i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @and_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = and i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = and i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @or_i3(i3 %a, i3 %b) {
-; SI-LABEL: @or_i3(
-; SI-NEXT:    [[R:%.*]] = or i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @or_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = or i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = or i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @xor_i3(i3 %a, i3 %b) {
-; SI-LABEL: @xor_i3(
-; SI-NEXT:    [[R:%.*]] = xor i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i3 [[R]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @xor_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = xor i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %r = xor i3 %a, %b
-  store volatile i3 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_eq_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_eq_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp eq i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_eq_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp eq i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp eq i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ne_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_ne_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp ne i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ne_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ne i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ugt_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_ugt_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp ugt i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ugt_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp ugt i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ugt i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_uge_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_uge_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp uge i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_uge_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp uge i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp uge i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ult_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_ult_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp ult i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ult_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp ult i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ult i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ule_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_ule_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp ule i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ule_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp ule i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ule i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sgt_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_sgt_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp sgt i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sgt_i3(
-; VI-NEXT:    [[TMP1:%.*]] = sext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp sgt i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = sext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sgt i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sge_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_sge_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp sge i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sge_i3(
-; VI-NEXT:    [[TMP1:%.*]] = sext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp sge i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = sext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sge i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_slt_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_slt_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp slt i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_slt_i3(
-; VI-NEXT:    [[TMP1:%.*]] = sext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp slt i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = sext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp slt i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sle_i3(i3 %a, i3 %b) {
-; SI-LABEL: @select_sle_i3(
-; SI-NEXT:    [[CMP:%.*]] = icmp sle i3 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i3 [[A]], i3 [[B]]
-; SI-NEXT:    store volatile i3 [[SEL]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sle_i3(
-; VI-NEXT:    [[TMP1:%.*]] = sext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i3 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp sle i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext i3 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = sext i3 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i3
-; VI-NEXT:    store volatile i3 [[TMP7]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sle i3 %a, %b
-  %sel = select i1 %cmp, i3 %a, i3 %b
-  store volatile i3 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-declare i3 @llvm.bitreverse.i3(i3)
-define amdgpu_kernel void @bitreverse_i3(i3 %a) {
-; SI-LABEL: @bitreverse_i3(
-; SI-NEXT:    [[BREV:%.*]] = call i3 @llvm.bitreverse.i3(i3 [[A:%.*]])
-; SI-NEXT:    store volatile i3 [[BREV]], ptr addrspace(1) poison, align 1
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @bitreverse_i3(
-; VI-NEXT:    [[TMP1:%.*]] = zext i3 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = call i32 @llvm.bitreverse.i32(i32 [[TMP1]])
-; VI-NEXT:    [[TMP3:%.*]] = lshr i32 [[TMP2]], 29
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i3
-; VI-NEXT:    store volatile i3 [[TMP4]], ptr addrspace(1) poison, align 1
-; VI-NEXT:    ret void
-;
-  %brev = call i3 @llvm.bitreverse.i3(i3 %a)
-  store volatile i3 %brev, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_i16(i16 %a, i16 %b) {
-; SI-LABEL: @add_i16(
-; SI-NEXT:    [[R:%.*]] = add i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = add i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @constant_add_i16() {
-; SI-LABEL: @constant_add_i16(
-; SI-NEXT:    [[R:%.*]] = add i16 1, 2
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @constant_add_i16(
-; VI-NEXT:    store volatile i16 3, ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = add i16 1, 2
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @constant_add_nsw_i16() {
-; SI-LABEL: @constant_add_nsw_i16(
-; SI-NEXT:    [[R:%.*]] = add nsw i16 1, 2
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @constant_add_nsw_i16(
-; VI-NEXT:    store volatile i16 3, ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = add nsw i16 1, 2
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @constant_add_nuw_i16() {
-; SI-LABEL: @constant_add_nuw_i16(
-; SI-NEXT:    [[R:%.*]] = add nsw i16 1, 2
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @constant_add_nuw_i16(
-; VI-NEXT:    store volatile i16 3, ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = add nsw i16 1, 2
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nsw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @add_nsw_i16(
-; SI-NEXT:    [[R:%.*]] = add nsw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nsw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = add nsw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nuw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @add_nuw_i16(
-; SI-NEXT:    [[R:%.*]] = add nuw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nuw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = add nuw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nuw_nsw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @add_nuw_nsw_i16(
-; SI-NEXT:    [[R:%.*]] = add nuw nsw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nuw_nsw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = add nuw nsw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_i16(i16 %a, i16 %b) {
-; SI-LABEL: @sub_i16(
-; SI-NEXT:    [[R:%.*]] = sub i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = sub nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = sub i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nsw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @sub_nsw_i16(
-; SI-NEXT:    [[R:%.*]] = sub nsw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nsw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = sub nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = sub nsw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nuw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @sub_nuw_i16(
-; SI-NEXT:    [[R:%.*]] = sub nuw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nuw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = sub nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = sub nuw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nuw_nsw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @sub_nuw_nsw_i16(
-; SI-NEXT:    [[R:%.*]] = sub nuw nsw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nuw_nsw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = sub nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = sub nuw nsw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_i16(i16 %a, i16 %b) {
-; SI-LABEL: @mul_i16(
-; SI-NEXT:    [[R:%.*]] = mul i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = mul i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nsw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @mul_nsw_i16(
-; SI-NEXT:    [[R:%.*]] = mul nsw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nsw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = mul nsw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nuw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @mul_nuw_i16(
-; SI-NEXT:    [[R:%.*]] = mul nuw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nuw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = mul nuw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nuw_nsw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @mul_nuw_nsw_i16(
-; SI-NEXT:    [[R:%.*]] = mul nuw nsw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nuw_nsw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = mul nuw nsw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_i16(i16 %a, i16 %b) {
-; SI-LABEL: @shl_i16(
-; SI-NEXT:    [[R:%.*]] = shl i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = shl i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nsw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @shl_nsw_i16(
-; SI-NEXT:    [[R:%.*]] = shl nsw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nsw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = shl nsw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nuw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @shl_nuw_i16(
-; SI-NEXT:    [[R:%.*]] = shl nuw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nuw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = shl nuw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nuw_nsw_i16(i16 %a, i16 %b) {
-; SI-LABEL: @shl_nuw_nsw_i16(
-; SI-NEXT:    [[R:%.*]] = shl nuw nsw i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nuw_nsw_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = shl nuw nsw i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @lshr_i16(i16 %a, i16 %b) {
-; SI-LABEL: @lshr_i16(
-; SI-NEXT:    [[R:%.*]] = lshr i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @lshr_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = lshr i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = lshr i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @lshr_exact_i16(i16 %a, i16 %b) {
-; SI-LABEL: @lshr_exact_i16(
-; SI-NEXT:    [[R:%.*]] = lshr exact i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @lshr_exact_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = lshr exact i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = lshr exact i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @ashr_i16(i16 %a, i16 %b) {
-; SI-LABEL: @ashr_i16(
-; SI-NEXT:    [[R:%.*]] = ashr i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @ashr_i16(
-; VI-NEXT:    [[TMP1:%.*]] = sext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = ashr i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = ashr i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @ashr_exact_i16(i16 %a, i16 %b) {
-; SI-LABEL: @ashr_exact_i16(
-; SI-NEXT:    [[R:%.*]] = ashr exact i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @ashr_exact_i16(
-; VI-NEXT:    [[TMP1:%.*]] = sext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = ashr exact i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = ashr exact i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @constant_lshr_exact_i16(i16 %a, i16 %b) {
-; SI-LABEL: @constant_lshr_exact_i16(
-; SI-NEXT:    [[R:%.*]] = lshr exact i16 4, 1
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @constant_lshr_exact_i16(
-; VI-NEXT:    store volatile i16 2, ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = lshr exact i16 4, 1
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @and_i16(i16 %a, i16 %b) {
-; SI-LABEL: @and_i16(
-; SI-NEXT:    [[R:%.*]] = and i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @and_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = and i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = and i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @or_i16(i16 %a, i16 %b) {
-; SI-LABEL: @or_i16(
-; SI-NEXT:    [[R:%.*]] = or i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @or_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = or i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = or i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @xor_i16(i16 %a, i16 %b) {
-; SI-LABEL: @xor_i16(
-; SI-NEXT:    [[R:%.*]] = xor i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile i16 [[R]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @xor_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = xor i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %r = xor i16 %a, %b
-  store volatile i16 %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_eq_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_eq_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp eq i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_eq_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp eq i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp eq i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ne_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_ne_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp ne i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ne_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ne i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ugt_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_ugt_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp ugt i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ugt_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp ugt i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ugt i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_uge_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_uge_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp uge i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_uge_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp uge i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp uge i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ult_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_ult_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp ult i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ult_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp ult i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ult i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ule_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_ule_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp ule i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ule_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = zext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp ule i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = zext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ule i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sgt_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_sgt_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp sgt i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sgt_i16(
-; VI-NEXT:    [[TMP1:%.*]] = sext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp sgt i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = sext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sgt i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sge_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_sge_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp sge i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sge_i16(
-; VI-NEXT:    [[TMP1:%.*]] = sext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp sge i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = sext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sge i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_slt_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_slt_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp slt i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_slt_i16(
-; VI-NEXT:    [[TMP1:%.*]] = sext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp slt i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = sext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp slt i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sle_i16(i16 %a, i16 %b) {
-; SI-LABEL: @select_sle_i16(
-; SI-NEXT:    [[CMP:%.*]] = icmp sle i16 [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i16 [[A]], i16 [[B]]
-; SI-NEXT:    store volatile i16 [[SEL]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sle_i16(
-; VI-NEXT:    [[TMP1:%.*]] = sext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = sext i16 [[B:%.*]] to i32
-; VI-NEXT:    [[TMP3:%.*]] = icmp sle i32 [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext i16 [[A]] to i32
-; VI-NEXT:    [[TMP5:%.*]] = sext i16 [[B]] to i32
-; VI-NEXT:    [[TMP6:%.*]] = select i1 [[TMP3]], i32 [[TMP4]], i32 [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
-; VI-NEXT:    store volatile i16 [[TMP7]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sle i16 %a, %b
-  %sel = select i1 %cmp, i16 %a, i16 %b
-  store volatile i16 %sel, ptr addrspace(1) poison
-  ret void
-}
-
-declare i16 @llvm.bitreverse.i16(i16)
-
-define amdgpu_kernel void @bitreverse_i16(i16 %a) {
-; SI-LABEL: @bitreverse_i16(
-; SI-NEXT:    [[BREV:%.*]] = call i16 @llvm.bitreverse.i16(i16 [[A:%.*]])
-; SI-NEXT:    store volatile i16 [[BREV]], ptr addrspace(1) poison, align 2
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @bitreverse_i16(
-; VI-NEXT:    [[TMP1:%.*]] = zext i16 [[A:%.*]] to i32
-; VI-NEXT:    [[TMP2:%.*]] = call i32 @llvm.bitreverse.i32(i32 [[TMP1]])
-; VI-NEXT:    [[TMP3:%.*]] = lshr i32 [[TMP2]], 16
-; VI-NEXT:    [[TMP4:%.*]] = trunc i32 [[TMP3]] to i16
-; VI-NEXT:    store volatile i16 [[TMP4]], ptr addrspace(1) poison, align 2
-; VI-NEXT:    ret void
-;
-  %brev = call i16 @llvm.bitreverse.i16(i16 %a)
-  store volatile i16 %brev, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @add_3xi15(
-; SI-NEXT:    [[R:%.*]] = add <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = add <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nsw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @add_nsw_3xi15(
-; SI-NEXT:    [[R:%.*]] = add nsw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nsw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = add nsw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nuw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @add_nuw_3xi15(
-; SI-NEXT:    [[R:%.*]] = add nuw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nuw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = add nuw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nuw_nsw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @add_nuw_nsw_3xi15(
-; SI-NEXT:    [[R:%.*]] = add nuw nsw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nuw_nsw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = add nuw nsw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @sub_3xi15(
-; SI-NEXT:    [[R:%.*]] = sub <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = sub nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = sub <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nsw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @sub_nsw_3xi15(
-; SI-NEXT:    [[R:%.*]] = sub nsw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nsw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = sub nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = sub nsw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nuw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @sub_nuw_3xi15(
-; SI-NEXT:    [[R:%.*]] = sub nuw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nuw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = sub nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = sub nuw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nuw_nsw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @sub_nuw_nsw_3xi15(
-; SI-NEXT:    [[R:%.*]] = sub nuw nsw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nuw_nsw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = sub nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = sub nuw nsw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @mul_3xi15(
-; SI-NEXT:    [[R:%.*]] = mul <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = mul <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nsw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @mul_nsw_3xi15(
-; SI-NEXT:    [[R:%.*]] = mul nsw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nsw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = mul nsw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nuw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @mul_nuw_3xi15(
-; SI-NEXT:    [[R:%.*]] = mul nuw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nuw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = mul nuw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nuw_nsw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @mul_nuw_nsw_3xi15(
-; SI-NEXT:    [[R:%.*]] = mul nuw nsw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nuw_nsw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = mul nuw nsw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @shl_3xi15(
-; SI-NEXT:    [[R:%.*]] = shl <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = shl <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nsw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @shl_nsw_3xi15(
-; SI-NEXT:    [[R:%.*]] = shl nsw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nsw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = shl nsw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nuw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @shl_nuw_3xi15(
-; SI-NEXT:    [[R:%.*]] = shl nuw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nuw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = shl nuw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nuw_nsw_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @shl_nuw_nsw_3xi15(
-; SI-NEXT:    [[R:%.*]] = shl nuw nsw <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nuw_nsw_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = shl nuw nsw <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @lshr_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @lshr_3xi15(
-; SI-NEXT:    [[R:%.*]] = lshr <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @lshr_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = lshr <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = lshr <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @lshr_exact_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @lshr_exact_3xi15(
-; SI-NEXT:    [[R:%.*]] = lshr exact <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @lshr_exact_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = lshr exact <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = lshr exact <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @ashr_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @ashr_3xi15(
-; SI-NEXT:    [[R:%.*]] = ashr <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @ashr_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = ashr <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = ashr <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @ashr_exact_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @ashr_exact_3xi15(
-; SI-NEXT:    [[R:%.*]] = ashr exact <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @ashr_exact_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = ashr exact <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = ashr exact <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @and_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @and_3xi15(
-; SI-NEXT:    [[R:%.*]] = and <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @and_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = and <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = and <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @or_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @or_3xi15(
-; SI-NEXT:    [[R:%.*]] = or <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @or_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = or <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = or <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @xor_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @xor_3xi15(
-; SI-NEXT:    [[R:%.*]] = xor <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i15> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @xor_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = xor <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = xor <3 x i15> %a, %b
-  store volatile <3 x i15> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_eq_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_eq_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp eq <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_eq_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp eq <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp eq <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ne_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_ne_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp ne <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ne_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp ne <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ne <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ugt_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_ugt_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp ugt <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ugt_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp ugt <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ugt <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_uge_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_uge_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp uge <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_uge_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp uge <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp uge <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ult_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_ult_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp ult <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ult_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp ult <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ult <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ule_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_ule_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp ule <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ule_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp ule <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ule <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sgt_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_sgt_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp sgt <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sgt_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp sgt <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = sext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sgt <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sge_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_sge_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp sge <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sge_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp sge <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = sext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sge <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_slt_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_slt_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp slt <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_slt_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp slt <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = sext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp slt <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sle_3xi15(<3 x i15> %a, <3 x i15> %b) {
-; SI-LABEL: @select_sle_3xi15(
-; SI-NEXT:    [[CMP:%.*]] = icmp sle <3 x i15> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i15> [[A]], <3 x i15> [[B]]
-; SI-NEXT:    store volatile <3 x i15> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sle_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i15> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp sle <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext <3 x i15> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = sext <3 x i15> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sle <3 x i15> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i15> %a, <3 x i15> %b
-  store volatile <3 x i15> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-declare <3 x i15> @llvm.bitreverse.v3i15(<3 x i15>)
-define amdgpu_kernel void @bitreverse_3xi15(<3 x i15> %a) {
-; SI-LABEL: @bitreverse_3xi15(
-; SI-NEXT:    [[BREV:%.*]] = call <3 x i15> @llvm.bitreverse.v3i15(<3 x i15> [[A:%.*]])
-; SI-NEXT:    store volatile <3 x i15> [[BREV]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @bitreverse_3xi15(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i15> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = call <3 x i32> @llvm.bitreverse.v3i32(<3 x i32> [[TMP1]])
-; VI-NEXT:    [[TMP3:%.*]] = lshr <3 x i32> [[TMP2]], splat (i32 17)
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i15>
-; VI-NEXT:    store volatile <3 x i15> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %brev = call <3 x i15> @llvm.bitreverse.v3i15(<3 x i15> %a)
-  store volatile <3 x i15> %brev, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @add_3xi16(
-; SI-NEXT:    [[R:%.*]] = add <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = add <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nsw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @add_nsw_3xi16(
-; SI-NEXT:    [[R:%.*]] = add nsw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nsw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = add nsw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nuw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @add_nuw_3xi16(
-; SI-NEXT:    [[R:%.*]] = add nuw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nuw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = add nuw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @add_nuw_nsw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @add_nuw_nsw_3xi16(
-; SI-NEXT:    [[R:%.*]] = add nuw nsw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @add_nuw_nsw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = add nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = add nuw nsw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @sub_3xi16(
-; SI-NEXT:    [[R:%.*]] = sub <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = sub nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = sub <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nsw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @sub_nsw_3xi16(
-; SI-NEXT:    [[R:%.*]] = sub nsw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nsw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = sub nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = sub nsw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nuw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @sub_nuw_3xi16(
-; SI-NEXT:    [[R:%.*]] = sub nuw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nuw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = sub nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = sub nuw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @sub_nuw_nsw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @sub_nuw_nsw_3xi16(
-; SI-NEXT:    [[R:%.*]] = sub nuw nsw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @sub_nuw_nsw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = sub nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = sub nuw nsw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @mul_3xi16(
-; SI-NEXT:    [[R:%.*]] = mul <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = mul <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nsw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @mul_nsw_3xi16(
-; SI-NEXT:    [[R:%.*]] = mul nsw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nsw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = mul nsw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nuw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @mul_nuw_3xi16(
-; SI-NEXT:    [[R:%.*]] = mul nuw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nuw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = mul nuw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @mul_nuw_nsw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @mul_nuw_nsw_3xi16(
-; SI-NEXT:    [[R:%.*]] = mul nuw nsw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @mul_nuw_nsw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = mul nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = mul nuw nsw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @shl_3xi16(
-; SI-NEXT:    [[R:%.*]] = shl <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = shl <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nsw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @shl_nsw_3xi16(
-; SI-NEXT:    [[R:%.*]] = shl nsw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nsw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = shl nsw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nuw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @shl_nuw_3xi16(
-; SI-NEXT:    [[R:%.*]] = shl nuw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nuw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = shl nuw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @shl_nuw_nsw_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @shl_nuw_nsw_3xi16(
-; SI-NEXT:    [[R:%.*]] = shl nuw nsw <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @shl_nuw_nsw_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = shl nuw nsw <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = shl nuw nsw <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @lshr_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @lshr_3xi16(
-; SI-NEXT:    [[R:%.*]] = lshr <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @lshr_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = lshr <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = lshr <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @lshr_exact_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @lshr_exact_3xi16(
-; SI-NEXT:    [[R:%.*]] = lshr exact <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @lshr_exact_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = lshr exact <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = lshr exact <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @ashr_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @ashr_3xi16(
-; SI-NEXT:    [[R:%.*]] = ashr <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @ashr_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = ashr <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = ashr <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @ashr_exact_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @ashr_exact_3xi16(
-; SI-NEXT:    [[R:%.*]] = ashr exact <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @ashr_exact_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = ashr exact <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = ashr exact <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @and_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @and_3xi16(
-; SI-NEXT:    [[R:%.*]] = and <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @and_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = and <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = and <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @or_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @or_3xi16(
-; SI-NEXT:    [[R:%.*]] = or <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @or_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = or <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = or <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @xor_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @xor_3xi16(
-; SI-NEXT:    [[R:%.*]] = xor <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    store volatile <3 x i16> [[R]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @xor_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = xor <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %r = xor <3 x i16> %a, %b
-  store volatile <3 x i16> %r, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_eq_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_eq_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp eq <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_eq_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp eq <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp eq <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ne_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_ne_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp ne <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ne_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp ne <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ne <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ugt_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_ugt_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp ugt <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ugt_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp ugt <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ugt <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_uge_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_uge_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp uge <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_uge_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp uge <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp uge <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ult_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_ult_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp ult <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ult_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp ult <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ult <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_ule_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_ule_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp ule <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_ule_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = zext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp ule <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = zext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = zext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp ule <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sgt_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_sgt_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp sgt <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sgt_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp sgt <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = sext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sgt <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sge_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_sge_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp sge <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sge_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp sge <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = sext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sge <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_slt_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_slt_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp slt <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_slt_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp slt <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = sext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp slt <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-define amdgpu_kernel void @select_sle_3xi16(<3 x i16> %a, <3 x i16> %b) {
-; SI-LABEL: @select_sle_3xi16(
-; SI-NEXT:    [[CMP:%.*]] = icmp sle <3 x i16> [[A:%.*]], [[B:%.*]]
-; SI-NEXT:    [[SEL:%.*]] = select <3 x i1> [[CMP]], <3 x i16> [[A]], <3 x i16> [[B]]
-; SI-NEXT:    store volatile <3 x i16> [[SEL]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @select_sle_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = sext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = sext <3 x i16> [[B:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP3:%.*]] = icmp sle <3 x i32> [[TMP1]], [[TMP2]]
-; VI-NEXT:    [[TMP4:%.*]] = sext <3 x i16> [[A]] to <3 x i32>
-; VI-NEXT:    [[TMP5:%.*]] = sext <3 x i16> [[B]] to <3 x i32>
-; VI-NEXT:    [[TMP6:%.*]] = select <3 x i1> [[TMP3]], <3 x i32> [[TMP4]], <3 x i32> [[TMP5]]
-; VI-NEXT:    [[TMP7:%.*]] = trunc <3 x i32> [[TMP6]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP7]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %cmp = icmp sle <3 x i16> %a, %b
-  %sel = select <3 x i1> %cmp, <3 x i16> %a, <3 x i16> %b
-  store volatile <3 x i16> %sel, ptr addrspace(1) poison
-  ret void
-}
-
-declare <3 x i16> @llvm.bitreverse.v3i16(<3 x i16>)
-
-define amdgpu_kernel void @bitreverse_3xi16(<3 x i16> %a) {
-; SI-LABEL: @bitreverse_3xi16(
-; SI-NEXT:    [[BREV:%.*]] = call <3 x i16> @llvm.bitreverse.v3i16(<3 x i16> [[A:%.*]])
-; SI-NEXT:    store volatile <3 x i16> [[BREV]], ptr addrspace(1) poison, align 8
-; SI-NEXT:    ret void
-;
-; VI-LABEL: @bitreverse_3xi16(
-; VI-NEXT:    [[TMP1:%.*]] = zext <3 x i16> [[A:%.*]] to <3 x i32>
-; VI-NEXT:    [[TMP2:%.*]] = call <3 x i32> @llvm.bitreverse.v3i32(<3 x i32> [[TMP1]])
-; VI-NEXT:    [[TMP3:%.*]] = lshr <3 x i32> [[TMP2]], splat (i32 16)
-; VI-NEXT:    [[TMP4:%.*]] = trunc <3 x i32> [[TMP3]] to <3 x i16>
-; VI-NEXT:    store volatile <3 x i16> [[TMP4]], ptr addrspace(1) poison, align 8
-; VI-NEXT:    ret void
-;
-  %brev = call <3 x i16> @llvm.bitreverse.v3i16(<3 x i16> %a)
-  store volatile <3 x i16> %brev, ptr addrspace(1) poison
-  ret void
-}