[VPlan] Add new VPScalarCastRecipe, use for IV & step trunc.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -2297,6 +2297,7 @@ emitTransformedIndex(IRBuilderBase &B, Value *Index, Value *StartValue,
                            ? B.CreateSExtOrTrunc(Index, StepTy)
                            : B.CreateCast(Instruction::SIToFP, Index, StepTy);
   if (CastedIndex != Index) {
+    assert(!isa<SExtInst>(CastedIndex));
     CastedIndex->setName(CastedIndex->getName() + ".cast");
     Index = CastedIndex;
   }

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -859,6 +859,7 @@ class VPSingleDefRecipe : public VPRecipeBase, public VPValue {
     case VPRecipeBase::VPWidenIntOrFpInductionSC:
     case VPRecipeBase::VPWidenPointerInductionSC:
     case VPRecipeBase::VPReductionPHISC:
+    case VPRecipeBase::VPScalarCastSC:
       return true;
     case VPRecipeBase::VPInterleaveSC:
     case VPRecipeBase::VPBranchOnMaskSC:
@@ -1338,6 +1339,32 @@ class VPWidenCastRecipe : public VPRecipeWithIRFlags {
   Type *getResultType() const { return ResultTy; }
 };
 
+/// VPScalarCastRecipe is a recipe to create scalar cast instructions.
+class VPScalarCastRecipe : public VPSingleDefRecipe {
+  Instruction::CastOps Opcode;
+
+  /// Result type for the cast.
+  Type *ResultTy;
+
+  Value *generate(VPTransformState &State, unsigned Part);
+
+public:
+  VPScalarCastRecipe(Instruction::CastOps Opcode, VPValue *Op, Type *ResultTy)
+      : VPSingleDefRecipe(VPDef::VPScalarCastSC, {Op}), Opcode(Opcode),
+        ResultTy(ResultTy) {}
+
+  ~VPScalarCastRecipe() override = default;
+
+  VP_CLASSOF_IMPL(VPDef::VPScalarCastSC)
+
+  void execute(VPTransformState &State) override;
+
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+  void print(raw_ostream &O, const Twine &Indent,
+             VPSlotTracker &SlotTracker) const override;
+#endif
+};
+
 /// A recipe for widening Call instructions.
 class VPWidenCallRecipe : public VPSingleDefRecipe {
   /// ID of the vector intrinsic to call when widening the call. If set the

llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp

@@ -230,7 +230,11 @@ Type *VPTypeAnalysis::inferScalarType(const VPValue *V) {
             return V->getUnderlyingValue()->getType();
           })
           .Case<VPWidenCastRecipe>(
-              [](const VPWidenCastRecipe *R) { return R->getResultType(); });
+              [](const VPWidenCastRecipe *R) { return R->getResultType(); })
+          .Case<VPExpandSCEVRecipe>([](const VPExpandSCEVRecipe *R) {
+            return R->getSCEV()->getType();
+          });
+
   assert(ResultTy && "could not infer type for the given VPValue");
   CachedTypes[V] = ResultTy;
   return ResultTy;

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -117,6 +117,7 @@ bool VPRecipeBase::mayHaveSideEffects() const {
   switch (getVPDefID()) {
   case VPDerivedIVSC:
   case VPPredInstPHISC:
+  case VPScalarCastSC:
     return false;
   case VPInstructionSC:
     switch (cast<VPInstruction>(this)->getOpcode()) {
@@ -1117,13 +1118,7 @@ void VPScalarIVStepsRecipe::execute(VPTransformState &State) {
 
   // Ensure step has the same type as that of scalar IV.
   Type *BaseIVTy = BaseIV->getType()->getScalarType();
-  if (BaseIVTy != Step->getType()) {
-    // TODO: Also use VPDerivedIVRecipe when only the step needs truncating, to
-    // avoid separate truncate here.
-    assert(Step->getType()->isIntegerTy() &&
-           "Truncation requires an integer step");
-    Step = State.Builder.CreateTrunc(Step, BaseIVTy);
-  }
+  assert(BaseIVTy == Step->getType() && "Types of BaseIV and Step must match!");
 
   // We build scalar steps for both integer and floating-point induction
   // variables. Here, we determine the kind of arithmetic we will perform.
@@ -1469,6 +1464,47 @@ void VPReplicateRecipe::print(raw_ostream &O, const Twine &Indent,
 }
 #endif
 
+Value *VPScalarCastRecipe ::generate(VPTransformState &State, unsigned Part) {
+  assert(vputils::onlyFirstLaneUsed(this) &&
+         "Codegen only implemented for first lane.");
+  switch (Opcode) {
+  case Instruction::SExt:
+  case Instruction::ZExt:
+  case Instruction::Trunc: {
+    Value *Op = State.get(getOperand(0), VPIteration(Part, 0));
+    return State.Builder.CreateCast(Instruction::CastOps(Opcode), Op, ResultTy);
+  }
+  default:
+    llvm_unreachable("opcode not implemented yet");
+  }
+}
+
+void VPScalarCastRecipe ::execute(VPTransformState &State) {
+  bool UniformAcrossUFs = all_of(operands(), [](VPValue *Op) {
+    return Op->isDefinedOutsideVectorRegions();
+  });
+  for (unsigned Part = 0; Part != State.UF; ++Part) {
+    Value *Res;
+    // Only generate a single instance, if the recipe is uniform across all UFs.
+    if (Part > 0 && UniformAcrossUFs)
+      Res = State.get(this, VPIteration(0, 0));
+    else
+      Res = generate(State, Part);
+    State.set(this, Res, VPIteration(Part, 0));
+  }
+}
+
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+void VPScalarCastRecipe ::print(raw_ostream &O, const Twine &Indent,
+                                VPSlotTracker &SlotTracker) const {
+  O << Indent << "SCALAR-CAST ";
+  printAsOperand(O, SlotTracker);
+  O << " = " << Instruction::getOpcodeName(Opcode) << " ";
+  printOperands(O, SlotTracker);
+  O << " to " << *ResultTy;
+}
+#endif
+
 void VPBranchOnMaskRecipe::execute(VPTransformState &State) {
   assert(State.Instance && "Branch on Mask works only on single instance.");
 

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -504,6 +504,15 @@ static VPValue *createScalarIVSteps(VPlan &Plan, const InductionDescriptor &ID,
     HeaderVPBB->insert(BaseIV->getDefiningRecipe(), IP);
   }
 
+  VPTypeAnalysis TypeInfo(SE.getContext());
+  if (TypeInfo.inferScalarType(BaseIV) != TypeInfo.inferScalarType(Step)) {
+    Step = new VPScalarCastRecipe(Instruction::Trunc, Step,
+                                  TypeInfo.inferScalarType(BaseIV));
+    auto *VecPreheader =
+        cast<VPBasicBlock>(Plan.getVectorLoopRegion()->getSinglePredecessor());
+    VecPreheader->appendRecipe(Step->getDefiningRecipe());
+  }
+
   VPScalarIVStepsRecipe *Steps = new VPScalarIVStepsRecipe(ID, BaseIV, Step);
   HeaderVPBB->insert(Steps, IP);
   return Steps;

llvm/lib/Transforms/Vectorize/VPlanValue.h

@@ -350,6 +350,7 @@ class VPDef {
     VPInterleaveSC,
     VPReductionSC,
     VPReplicateSC,
+    VPScalarCastSC,
     VPScalarIVStepsSC,
     VPVectorPointerSC,
     VPWidenCallSC,

llvm/test/Transforms/LoopVectorize/cast-induction.ll

@@ -83,12 +83,14 @@ define void @cast_variable_step(i64 %step) {
 ; VF4: middle.block:
 ;
 ; IC2-LABEL: @cast_variable_step(
+; IC2:   [[TRUNC_STEP:%.+]] = trunc i64 %step to i32
+; IC2:   br label %vector.body
+
 ; IC2-LABEL: vector.body:
 ; IC2-NEXT:   [[INDEX:%.+]] = phi i64 [ 0, %vector.ph ]
 ; IC2-NEXT:   [[MUL:%.+]] = mul i64 %index, %step
 ; IC2-NEXT:   [[OFFSET_IDX:%.+]] = add i64 10, [[MUL]]
 ; IC2-NEXT:   [[TRUNC_OFF:%.+]] = trunc i64 [[OFFSET_IDX]] to i32
-; IC2-NEXT:   [[TRUNC_STEP:%.+]] = trunc i64 %step to i32
 ; IC2-NEXT:   [[STEP0:%.+]] = mul i32 0, [[TRUNC_STEP]]
 ; IC2-NEXT:   [[T0:%.+]] = add i32 [[TRUNC_OFF]], [[STEP0]]
 ; IC2-NEXT:   [[STEP1:%.+]] = mul i32 1, [[TRUNC_STEP]]

llvm/test/Transforms/LoopVectorize/interleave-and-scalarize-only.ll

@@ -184,14 +184,15 @@ exit:
 ; DBG-NEXT: No successors
 ; DBG-EMPTY:
 ; DBG-NEXT: vector.ph:
+; DBG-NEXT:   SCALAR-CAST vp<[[CAST:%.+]]> = trunc ir<1> to i32
 ; DBG-NEXT: Successor(s): vector loop
 ; DBG-EMPTY:
 ; DBG-NEXT: <x1> vector loop: {
 ; DBG-NEXT:   vector.body:
 ; DBG-NEXT:     EMIT vp<[[CAN_IV:%.+]]> = CANONICAL-INDUCTION
 ; DBG-NEXT:     FIRST-ORDER-RECURRENCE-PHI ir<%for> = phi ir<0>, vp<[[SCALAR_STEPS:.+]]>
 ; DBG-NEXT:     vp<[[DERIVED_IV:%.+]]> = DERIVED-IV ir<0> + vp<[[CAN_IV]]> * ir<1> (truncated to i32)
-; DBG-NEXT:     vp<[[SCALAR_STEPS]]> = SCALAR-STEPS vp<[[DERIVED_IV]]>, ir<1>
+; DBG-NEXT:     vp<[[SCALAR_STEPS]]> = SCALAR-STEPS vp<[[DERIVED_IV]]>, vp<[[CAST]]>
 ; DBG-NEXT:     EMIT vp<[[SPLICE:%.+]]> = first-order splice ir<%for>, vp<[[SCALAR_STEPS]]>
 ; DBG-NEXT:     CLONE store vp<[[SPLICE]]>, ir<%dst>
 ; DBG-NEXT:     EMIT vp<[[IV_INC:%.+]]> = add nuw vp<[[CAN_IV]]>, vp<[[VFxUF]]>

llvm/test/Transforms/LoopVectorize/pr46525-expander-insertpoint.ll

@@ -43,7 +43,7 @@ define void @test(i16 %x, i64 %y, ptr %ptr) {
 ; CHECK-NEXT:    [[V3:%.*]] = add i8 [[V2]], 1
 ; CHECK-NEXT:    [[CMP15:%.*]] = icmp slt i8 [[V3]], 5
 ; CHECK-NEXT:    [[IV_NEXT]] = add i64 [[IV]], [[INC]]
-; CHECK-NEXT:    br i1 [[CMP15]], label [[LOOP]], label [[LOOP_EXIT]], !llvm.loop [[LOOP2:![0-9]+]]
+; CHECK-NEXT:    br i1 [[CMP15]], label [[LOOP]], label [[LOOP_EXIT]], !llvm.loop [[LOOP3:![0-9]+]]
 ; CHECK:       loop.exit:
 ; CHECK-NEXT:    [[DIV_1:%.*]] = udiv i64 [[Y]], [[ADD]]
 ; CHECK-NEXT:    [[V1:%.*]] = add i64 [[DIV_1]], 1