[llvm] [VPlan] Move call widening decision to VPlan. (NFCI) (PR #195518)

Florian Hahn via llvm-commits llvm-commits at lists.llvm.org
Wed May 6 02:19:14 PDT 2026


https://github.com/fhahn updated https://github.com/llvm/llvm-project/pull/195518

>From 58907fe363f4dce6c4521c799a852a95957b32e8 Mon Sep 17 00:00:00 2001
From: Florian Hahn <flo at fhahn.com>
Date: Sun, 3 May 2026 12:26:03 +0100
Subject: [PATCH 1/2] [VPlan] Move call widening decision to VPlan. (NFCI)

This patch adds a new makeCallWideningDecisions transform which converts
Call VPInstructions to VPWidenCallRecipe/VPWidenIntrinsicRecipe/VPReplicateRecipe
depending on their costs.

To compute the costs, static helpers are introduced to re-use the
existing VPlan cost model logic:
 * VPWidenIntrinsicRecipe::getIntrinsicCost
 * VPReplicateRecipe::computeScalarCallCost

The cost-model logic is still retained; we assert that the decisions
match to make sure we do not miss any edge cases. The legacy logic will
be removed in a follow-up.
---
 .../Transforms/Vectorize/LoopVectorize.cpp    | 134 ++++---------
 .../Transforms/Vectorize/VPRecipeBuilder.h    |  14 +-
 llvm/lib/Transforms/Vectorize/VPlan.h         |  21 +-
 llvm/lib/Transforms/Vectorize/VPlanHelpers.h  |  19 ++
 .../lib/Transforms/Vectorize/VPlanRecipes.cpp |  79 ++++----
 .../Transforms/Vectorize/VPlanTransforms.cpp  | 186 ++++++++++++++++++
 .../Transforms/Vectorize/VPlanTransforms.h    |   6 +
 .../VPlan/AArch64/call-decisions.ll           |   6 +-
 .../VPlan/vplan-print-after-all.ll            |   1 +
 9 files changed, 321 insertions(+), 145 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 6535f2dc6c72c..a08f58ec89aac 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -1312,6 +1312,12 @@ class LoopVectorizationCostModel {
   /// trivially hoistable.
   bool shouldConsiderInvariant(Value *Op);
 
+  /// Returns true if \p I has been forced to be scalarized at \p VF.
+  bool isForcedScalar(Instruction *I, ElementCount VF) const {
+    auto FS = ForcedScalars.find(VF);
+    return FS != ForcedScalars.end() && FS->second.contains(I);
+  }
+
 private:
   unsigned NumPredStores = 0;
 
@@ -5848,6 +5854,32 @@ uint64_t VPCostContext::getPredBlockCostDivisor(BasicBlock *BB) const {
   return CM.getPredBlockCostDivisor(CostKind, BB);
 }
 
+bool VPCostContext::willBeScalarized(Instruction *I, ElementCount VF) const {
+  return CM.isScalarWithPredication(I, VF) ||
+         CM.isUniformAfterVectorization(I, VF) || CM.isForcedScalar(I, VF) ||
+         (VF.isVector() && CM.isProfitableToScalarize(I, VF));
+}
+
+bool VPCostContext::isMaskRequired(Instruction *I) const {
+  return CM.isMaskRequired(I);
+}
+
+std::optional<VPCostContext::CallWideningKind>
+VPCostContext::getLegacyCallKind(CallInst *CI, ElementCount VF) const {
+  if (VF.isScalar())
+    return std::nullopt;
+  switch (CM.getCallWideningDecision(CI, VF).Kind) {
+  case LoopVectorizationCostModel::CM_Scalarize:
+    return CallWideningKind::Scalarize;
+  case LoopVectorizationCostModel::CM_IntrinsicCall:
+    return CallWideningKind::Intrinsic;
+  case LoopVectorizationCostModel::CM_VectorCall:
+    return CallWideningKind::VectorVariant;
+  default:
+    return std::nullopt;
+  }
+}
+
 InstructionCost
 LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF,
                                           VPCostContext &CostCtx) const {
@@ -6491,90 +6523,6 @@ VPRecipeBuilder::tryToOptimizeInductionTruncate(VPInstruction *VPI,
       Phi, Start, Step, &Plan.getVF(), IndDesc, I, Flags, VPI->getDebugLoc());
 }
 
-VPSingleDefRecipe *VPRecipeBuilder::tryToWidenCall(VPInstruction *VPI,
-                                                   VFRange &Range) {
-  CallInst *CI = cast<CallInst>(VPI->getUnderlyingInstr());
-  bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange(
-      [this, CI](ElementCount VF) {
-        return CM.isScalarWithPredication(CI, VF);
-      },
-      Range);
-
-  if (IsPredicated)
-    return nullptr;
-
-  Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
-  if (VPCostContext::isFreeScalarIntrinsic(ID))
-    return nullptr;
-
-  SmallVector<VPValue *, 4> Ops(VPI->op_begin(),
-                                VPI->op_begin() + CI->arg_size());
-
-  // Is it beneficial to perform intrinsic call compared to lib call?
-  bool ShouldUseVectorIntrinsic =
-      ID && LoopVectorizationPlanner::getDecisionAndClampRange(
-                [&](ElementCount VF) -> bool {
-                  return CM.getCallWideningDecision(CI, VF).Kind ==
-                         LoopVectorizationCostModel::CM_IntrinsicCall;
-                },
-                Range);
-  if (ShouldUseVectorIntrinsic)
-    return new VPWidenIntrinsicRecipe(*CI, ID, Ops, CI->getType(), *VPI, *VPI,
-                                      VPI->getDebugLoc());
-
-  Function *Variant = nullptr;
-  std::optional<unsigned> MaskPos;
-  // Is better to call a vectorized version of the function than to to scalarize
-  // the call?
-  auto ShouldUseVectorCall = LoopVectorizationPlanner::getDecisionAndClampRange(
-      [&](ElementCount VF) -> bool {
-        // The following case may be scalarized depending on the VF.
-        // The flag shows whether we can use a usual Call for vectorized
-        // version of the instruction.
-
-        // If we've found a variant at a previous VF, then stop looking. A
-        // vectorized variant of a function expects input in a certain shape
-        // -- basically the number of input registers, the number of lanes
-        // per register, and whether there's a mask required.
-        // We store a pointer to the variant in the VPWidenCallRecipe, so
-        // once we have an appropriate variant it's only valid for that VF.
-        // This will force a different vplan to be generated for each VF that
-        // finds a valid variant.
-        if (Variant)
-          return false;
-        LoopVectorizationCostModel::CallWideningDecision Decision =
-            CM.getCallWideningDecision(CI, VF);
-        if (Decision.Kind == LoopVectorizationCostModel::CM_VectorCall) {
-          Variant = Decision.Variant;
-          MaskPos = Decision.MaskPos;
-          return true;
-        }
-
-        return false;
-      },
-      Range);
-  if (ShouldUseVectorCall) {
-    if (MaskPos.has_value()) {
-      // We have 2 cases that would require a mask:
-      //   1) The call needs to be predicated, either due to a conditional
-      //      in the scalar loop or use of an active lane mask with
-      //      tail-folding, and we use the appropriate mask for the block.
-      //   2) No mask is required for the call instruction, but the only
-      //      available vector variant at this VF requires a mask, so we
-      //      synthesize an all-true mask.
-      VPValue *Mask = VPI->isMasked() ? VPI->getMask() : Plan.getTrue();
-
-      Ops.insert(Ops.begin() + *MaskPos, Mask);
-    }
-
-    Ops.push_back(VPI->getOperand(VPI->getNumOperandsWithoutMask() - 1));
-    return new VPWidenCallRecipe(CI, Variant, Ops, *VPI, *VPI,
-                                 VPI->getDebugLoc());
-  }
-
-  return nullptr;
-}
-
 bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const {
   assert((!isa<UncondBrInst, CondBrInst, PHINode, LoadInst, StoreInst>(I)) &&
          "Instruction should have been handled earlier");
@@ -6760,10 +6708,12 @@ VPRecipeBuilder::tryToCreateWidenNonPhiRecipe(VPSingleDefRecipe *R,
                                               VFRange &Range) {
   assert(!R->isPhi() && "phis must be handled earlier");
   // First, check for specific widening recipes that deal with optimizing
-  // truncates, calls and memory operations.
+  // truncates and memory operations
+  auto *VPI = cast<VPInstruction>(R);
+  assert(VPI->getOpcode() != Instruction::Call &&
+         "Call should have been handled by makeCallWideningDecisions");
 
   VPRecipeBase *Recipe;
-  auto *VPI = cast<VPInstruction>(R);
   if (VPI->getOpcode() == Instruction::Trunc &&
       (Recipe = tryToOptimizeInductionTruncate(VPI, Range)))
     return Recipe;
@@ -6773,9 +6723,6 @@ VPRecipeBuilder::tryToCreateWidenNonPhiRecipe(VPSingleDefRecipe *R,
           [&](ElementCount VF) { return VF.isScalar(); }, Range))
     return nullptr;
 
-  if (VPI->getOpcode() == Instruction::Call)
-    return tryToWidenCall(VPI, Range);
-
   Instruction *Instr = R->getUnderlyingInstr();
   assert(!is_contained({Instruction::Load, Instruction::Store},
                        VPI->getOpcode()) &&
@@ -6972,7 +6919,7 @@ LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VPlanPtr Plan,
   // Construct wide recipes and apply predication for original scalar
   // VPInstructions in the loop.
   // ---------------------------------------------------------------------------
-  VPRecipeBuilder RecipeBuilder(*Plan, TLI, Legal, CM, Builder);
+  VPRecipeBuilder RecipeBuilder(*Plan, Legal, CM, Builder);
 
   // Scan the body of the loop in a topological order to visit each basic block
   // after having visited its predecessor basic blocks.
@@ -6991,6 +6938,9 @@ LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VPlanPtr Plan,
 
   RUN_VPLAN_PASS(VPlanTransforms::makeScalarizationDecisions, *Plan, Range);
 
+  RUN_VPLAN_PASS_NO_VERIFY(VPlanTransforms::makeCallWideningDecisions, *Plan,
+                           Range, RecipeBuilder, CostCtx);
+
   // Now process all other blocks and instructions.
   for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
     // Convert input VPInstructions to widened recipes.
@@ -7000,8 +6950,8 @@ LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VPlanPtr Plan,
       // transformed.
       if (isa<VPWidenCanonicalIVRecipe, VPBlendRecipe, VPReductionRecipe,
               VPReplicateRecipe, VPWidenLoadRecipe, VPWidenStoreRecipe,
-              VPVectorPointerRecipe, VPVectorEndPointerRecipe,
-              VPHistogramRecipe>(&R))
+              VPWidenCallRecipe, VPWidenIntrinsicRecipe, VPVectorPointerRecipe,
+              VPVectorEndPointerRecipe, VPHistogramRecipe>(&R))
         continue;
       auto *VPI = cast<VPInstruction>(&R);
       if (!VPI->getUnderlyingValue())
diff --git a/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h b/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h
index a84c77d614673..aff84cdbd0cf7 100644
--- a/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h
+++ b/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h
@@ -17,7 +17,6 @@ namespace llvm {
 
 class LoopVectorizationLegality;
 class LoopVectorizationCostModel;
-class TargetLibraryInfo;
 struct HistogramInfo;
 struct VFRange;
 
@@ -26,9 +25,6 @@ class VPRecipeBuilder {
   /// The VPlan new recipes are added to.
   VPlan &Plan;
 
-  /// Target Library Info.
-  const TargetLibraryInfo *TLI;
-
   /// The legality analysis.
   LoopVectorizationLegality *Legal;
 
@@ -47,21 +43,15 @@ class VPRecipeBuilder {
   VPWidenIntOrFpInductionRecipe *
   tryToOptimizeInductionTruncate(VPInstruction *VPI, VFRange &Range);
 
-  /// Handle call instructions. If \p VPI can be widened for \p Range.Start,
-  /// return a new VPWidenCallRecipe or VPWidenIntrinsicRecipe. Range.End may be
-  /// decreased to ensure same decision from \p Range.Start to \p Range.End.
-  VPSingleDefRecipe *tryToWidenCall(VPInstruction *VPI, VFRange &Range);
-
   /// Check if \p VPI has an opcode that can be widened and return a
   /// VPWidenRecipe if it can. The function should only be called if the
   /// cost-model indicates that widening should be performed.
   VPWidenRecipe *tryToWiden(VPInstruction *VPI);
 
 public:
-  VPRecipeBuilder(VPlan &Plan, const TargetLibraryInfo *TLI,
-                  LoopVectorizationLegality *Legal,
+  VPRecipeBuilder(VPlan &Plan, LoopVectorizationLegality *Legal,
                   LoopVectorizationCostModel &CM, VPBuilder &Builder)
-      : Plan(Plan), TLI(TLI), Legal(Legal), CM(CM), Builder(Builder) {}
+      : Plan(Plan), Legal(Legal), CM(CM), Builder(Builder) {}
 
   /// Create and return a widened recipe for a non-phi recipe \p R if one can be
   /// created within the given VF \p Range.
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 6a1ea6b3439bf..f42692847dde2 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -70,10 +70,6 @@ class LoopVectorizationCostModel;
 
 struct VPCostContext;
 
-namespace Intrinsic {
-typedef unsigned ID;
-}
-
 using VPlanPtr = std::unique_ptr<VPlan>;
 
 /// \enum UncountableExitStyle
@@ -1951,6 +1947,12 @@ class VPWidenIntrinsicRecipe : public VPRecipeWithIRFlags, public VPIRMetadata {
   /// Produce a widened version of the vector intrinsic.
   LLVM_ABI_FOR_TEST void execute(VPTransformState &State) override;
 
+  /// Compute the cost of a vector intrinsic with \p ID and \p Operands.
+  static InstructionCost
+  computeIntrinsicCost(Intrinsic::ID ID, ArrayRef<const VPValue *> Operands,
+                       const VPRecipeWithIRFlags &R, ElementCount VF,
+                       VPCostContext &Ctx);
+
   /// Return the cost of this vector intrinsic.
   LLVM_ABI_FOR_TEST InstructionCost
   computeCost(ElementCount VF, VPCostContext &Ctx) const override;
@@ -2021,6 +2023,10 @@ class LLVM_ABI_FOR_TEST VPWidenCallRecipe : public VPRecipeWithIRFlags,
   InstructionCost computeCost(ElementCount VF,
                               VPCostContext &Ctx) const override;
 
+  /// Return the cost of widening a call using the vector function \p Variant.
+  static InstructionCost computeVectorCallCost(Function *Variant,
+                                               VPCostContext &Ctx);
+
   Function *getCalledScalarFunction() const {
     return cast<Function>(getOperand(getNumOperands() - 1)->getLiveInIRValue());
   }
@@ -3232,6 +3238,13 @@ class LLVM_ABI_FOR_TEST VPReplicateRecipe : public VPRecipeWithIRFlags,
   InstructionCost computeCost(ElementCount VF,
                               VPCostContext &Ctx) const override;
 
+  /// Return the cost of scalarizing a call to \p CalledFn with argument
+  /// operands \p ArgOps for a given \p VF.
+  static InstructionCost
+  computeScalarCallCost(Function *CalledFn, Type *ResultTy,
+                        ArrayRef<const VPValue *> ArgOps, bool IsSingleScalar,
+                        ElementCount VF, VPCostContext &Ctx);
+
   bool isSingleScalar() const { return IsSingleScalar; }
 
   bool isPredicated() const { return IsPredicated; }
diff --git a/llvm/lib/Transforms/Vectorize/VPlanHelpers.h b/llvm/lib/Transforms/Vectorize/VPlanHelpers.h
index 62d7782e097dd..ff4b2d7f32964 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanHelpers.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanHelpers.h
@@ -30,6 +30,7 @@ namespace llvm {
 
 class AssumptionCache;
 class BasicBlock;
+class CallInst;
 class DominatorTree;
 class InnerLoopVectorizer;
 class IRBuilderBase;
@@ -41,6 +42,10 @@ class VPRegionBlock;
 class VPlan;
 class Value;
 
+namespace Intrinsic {
+typedef unsigned ID;
+}
+
 /// Returns a calculation for the total number of elements for a given \p VF.
 /// For fixed width vectors this value is a constant, whereas for scalable
 /// vectors it is an expression determined at runtime.
@@ -324,6 +329,9 @@ struct VPTransformState {
 
 /// Struct to hold various analysis needed for cost computations.
 struct VPCostContext {
+  /// Choice for how to widen a call at a given VF.
+  enum class CallWideningKind { Scalarize, Intrinsic, VectorVariant };
+
   const TargetTransformInfo &TTI;
   const TargetLibraryInfo &TLI;
   VPTypeAnalysis Types;
@@ -356,6 +364,17 @@ struct VPCostContext {
   /// Forwards to LoopVectorizationCostModel::getPredBlockCostDivisor.
   uint64_t getPredBlockCostDivisor(BasicBlock *BB) const;
 
+  /// Returns true if \p I is known to be scalarized at \p VF.
+  bool willBeScalarized(Instruction *I, ElementCount VF) const;
+
+  /// Forwards to LoopVectorizationCostModel::isMaskRequired.
+  bool isMaskRequired(Instruction *I) const;
+
+  /// Returns the legacy call widening decision for \p CI at \p VF, or
+  /// std::nullopt if none was recorded. Used only in asserts.
+  std::optional<CallWideningKind> getLegacyCallKind(CallInst *CI,
+                                                    ElementCount VF) const;
+
   /// Returns the OperandInfo for \p V, if it is a live-in.
   TargetTransformInfo::OperandValueInfo getOperandInfo(VPValue *V) const;
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index d04b5edcfc212..3ed0efc7fc226 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -1878,6 +1878,11 @@ void VPWidenCallRecipe::execute(VPTransformState &State) {
 
 InstructionCost VPWidenCallRecipe::computeCost(ElementCount VF,
                                                VPCostContext &Ctx) const {
+  return computeVectorCallCost(Variant, Ctx);
+}
+
+InstructionCost VPWidenCallRecipe::computeVectorCallCost(Function *Variant,
+                                                         VPCostContext &Ctx) {
   return Ctx.TTI.getCallInstrCost(nullptr, Variant->getReturnType(),
                                   Variant->getFunctionType()->params(),
                                   Ctx.CostKind);
@@ -1963,12 +1968,9 @@ void VPWidenIntrinsicRecipe::execute(VPTransformState &State) {
     State.set(this, V);
 }
 
-/// Compute the cost for the intrinsic \p ID with \p Operands, produced by \p R.
-static InstructionCost getCostForIntrinsics(Intrinsic::ID ID,
-                                            ArrayRef<const VPValue *> Operands,
-                                            const VPRecipeWithIRFlags &R,
-                                            ElementCount VF,
-                                            VPCostContext &Ctx) {
+InstructionCost VPWidenIntrinsicRecipe::computeIntrinsicCost(
+    Intrinsic::ID ID, ArrayRef<const VPValue *> Operands,
+    const VPRecipeWithIRFlags &R, ElementCount VF, VPCostContext &Ctx) {
   Type *ScalarRetTy = Ctx.Types.inferScalarType(&R);
   // Skip the reverse operation cost for the mask.
   // FIXME: Remove this once redundant mask reverse operations can be eliminated
@@ -2012,7 +2014,7 @@ static InstructionCost getCostForIntrinsics(Intrinsic::ID ID,
 InstructionCost VPWidenIntrinsicRecipe::computeCost(ElementCount VF,
                                                     VPCostContext &Ctx) const {
   SmallVector<const VPValue *> ArgOps(operands());
-  return getCostForIntrinsics(VectorIntrinsicID, ArgOps, *this, VF, Ctx);
+  return computeIntrinsicCost(VectorIntrinsicID, ArgOps, *this, VF, Ctx);
 }
 
 StringRef VPWidenIntrinsicRecipe::getIntrinsicName() const {
@@ -3436,34 +3438,10 @@ InstructionCost VPReplicateRecipe::computeCost(ElementCount VF,
   case Instruction::Call: {
     auto *CalledFn =
         cast<Function>(getOperand(getNumOperands() - 1)->getLiveInIRValue());
-
-    SmallVector<const VPValue *> ArgOps(drop_end(operands()));
-    SmallVector<Type *, 4> Tys;
-    for (const VPValue *ArgOp : ArgOps)
-      Tys.push_back(Ctx.Types.inferScalarType(ArgOp));
-
-    if (CalledFn->isIntrinsic() &&
-        VPCostContext::isFreeScalarIntrinsic(CalledFn->getIntrinsicID())) {
-      assert(getCostForIntrinsics(CalledFn->getIntrinsicID(), ArgOps, *this,
-                                  ElementCount::getFixed(1), Ctx) == 0 &&
-             "scalarizing intrinsic should be free");
-      return InstructionCost(0);
-    }
-
     Type *ResultTy = Ctx.Types.inferScalarType(this);
-    InstructionCost ScalarCallCost =
-        Ctx.TTI.getCallInstrCost(CalledFn, ResultTy, Tys, Ctx.CostKind);
-    if (isSingleScalar()) {
-      if (CalledFn->isIntrinsic())
-        ScalarCallCost = std::min(
-            ScalarCallCost,
-            getCostForIntrinsics(CalledFn->getIntrinsicID(), ArgOps, *this,
-                                 ElementCount::getFixed(1), Ctx));
-      return ScalarCallCost;
-    }
-
-    return ScalarCallCost * VF.getFixedValue() +
-           Ctx.getScalarizationOverhead(ResultTy, ArgOps, VF);
+    SmallVector<const VPValue *> ArgOps(drop_end(operands()));
+    return computeScalarCallCost(CalledFn, ResultTy, ArgOps, isSingleScalar(),
+                                 VF, Ctx);
   }
   case Instruction::Add:
   case Instruction::Sub:
@@ -3644,6 +3622,39 @@ InstructionCost VPReplicateRecipe::computeCost(ElementCount VF,
   return Ctx.getLegacyCost(UI, VF);
 }
 
+InstructionCost VPReplicateRecipe::computeScalarCallCost(
+    Function *CalledFn, Type *ResultTy, ArrayRef<const VPValue *> ArgOps,
+    bool IsSingleScalar, ElementCount VF, VPCostContext &Ctx) {
+  SmallVector<Type *, 4> Tys = map_to_vector<4>(
+      ArgOps, [&](const VPValue *Op) { return Ctx.Types.inferScalarType(Op); });
+
+  Intrinsic::ID IntrinID = CalledFn->getIntrinsicID();
+  auto GetIntrinsicCost = [&] {
+    return Ctx.TTI.getIntrinsicInstrCost(
+        IntrinsicCostAttributes(IntrinID, ResultTy, Tys), Ctx.CostKind);
+  };
+
+  if (IntrinID && VPCostContext::isFreeScalarIntrinsic(IntrinID)) {
+    assert(GetIntrinsicCost() == 0 && "scalarizing intrinsic should be free");
+    return InstructionCost(0);
+  }
+
+  InstructionCost ScalarCallCost =
+      Ctx.TTI.getCallInstrCost(CalledFn, ResultTy, Tys, Ctx.CostKind);
+  if (IsSingleScalar) {
+    if (IntrinID)
+      ScalarCallCost = std::min(ScalarCallCost, GetIntrinsicCost());
+    return ScalarCallCost;
+  }
+
+  // Scalarization overhead is undefined for scalable VFs.
+  if (VF.isScalable())
+    return InstructionCost::getInvalid();
+
+  return ScalarCallCost * VF.getFixedValue() +
+         Ctx.getScalarizationOverhead(ResultTy, ArgOps, VF);
+}
+
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 void VPReplicateRecipe::printRecipe(raw_ostream &O, const Twine &Indent,
                                     VPSlotTracker &SlotTracker) const {
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 3dadd9253ad57..ad2acdb956df0 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -6552,3 +6552,189 @@ void VPlanTransforms::makeScalarizationDecisions(VPlan &Plan, VFRange &Range) {
     }
   }
 }
+
+/// Returns true if \p Info's parameter kinds are compatible with \p Args.
+static bool areVFParamsOk(const VFInfo &Info, ArrayRef<VPValue *> Args,
+                          PredicatedScalarEvolution &PSE, const Loop *L) {
+  return all_of(Info.Shape.Parameters, [&](VFParameter Param) {
+    switch (Param.ParamKind) {
+    case VFParamKind::Vector:
+    case VFParamKind::GlobalPredicate:
+      return true;
+    case VFParamKind::OMP_Uniform:
+      return PSE.getSE()->isLoopInvariant(
+          vputils::getSCEVExprForVPValue(Args[Param.ParamPos], PSE, L), L);
+    case VFParamKind::OMP_Linear:
+      return match(vputils::getSCEVExprForVPValue(Args[Param.ParamPos], PSE, L),
+                   m_scev_AffineAddRec(
+                       m_SCEV(), m_scev_SpecificSInt(Param.LinearStepOrPos),
+                       m_SpecificLoop(L)));
+    default:
+      return false;
+    }
+  });
+}
+
+/// Find a vector variant of \p CI for \p VF, respecting \p MaskRequired.
+/// Returns the variant function and the position of its mask parameter
+/// (if any), or {nullptr, std::nullopt}.
+static std::pair<Function *, std::optional<unsigned>>
+findVectorVariant(CallInst *CI, ArrayRef<VPValue *> Args, ElementCount VF,
+                  bool MaskRequired, PredicatedScalarEvolution &PSE,
+                  const Loop *L) {
+  if (CI->isNoBuiltin())
+    return {nullptr, std::nullopt};
+  auto Mappings = VFDatabase::getMappings(*CI);
+  const auto *It = find_if(Mappings, [&](const VFInfo &Info) {
+    return Info.Shape.VF == VF && (!MaskRequired || Info.isMasked()) &&
+           areVFParamsOk(Info, Args, PSE, L);
+  });
+  if (It == Mappings.end())
+    return {nullptr, std::nullopt};
+  if (Function *VecFunc = CI->getModule()->getFunction(It->VectorName))
+    return {VecFunc, It->getParamIndexForOptionalMask()};
+  return {nullptr, std::nullopt};
+}
+
+namespace {
+/// The outcome of choosing how to widen a call at a given VF.
+struct CallWideningDecision {
+  using KindTy = VPCostContext::CallWideningKind;
+  KindTy Kind = KindTy::Scalarize;
+  /// Set when Kind == VectorVariant.
+  Function *Variant = nullptr;
+  /// Position of the mask parameter for \p Variant, if any.
+  std::optional<unsigned> MaskPos;
+};
+} // namespace
+
+/// Pick the cheapest widening for the call \p VPI at \p VF among scalarization,
+/// vector intrinsic, and vector library variant.
+static CallWideningDecision decideCallWidening(VPInstruction &VPI,
+                                               ArrayRef<VPValue *> Ops,
+                                               ElementCount VF,
+                                               VPCostContext &CostCtx) {
+  auto *CI = cast<CallInst>(VPI.getUnderlyingInstr());
+  auto *CalledFn = cast<Function>(
+      VPI.getOperand(VPI.getNumOperandsWithoutMask() - 1)->getLiveInIRValue());
+  Type *ResultTy = CostCtx.Types.inferScalarType(&VPI);
+  Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, &CostCtx.TLI);
+  bool MaskRequired = CostCtx.isMaskRequired(CI);
+
+  // Pseudo intrinsics (assume, lifetime, ...) are always scalarized.
+  if (ID && VPCostContext::isFreeScalarIntrinsic(ID))
+    return {};
+
+  InstructionCost ScalarCost = VPReplicateRecipe::computeScalarCallCost(
+      CalledFn, ResultTy, Ops,
+      /*IsSingleScalar=*/false, VF, CostCtx);
+
+  auto [VecFunc, MaskPos] =
+      findVectorVariant(CI, Ops, VF, MaskRequired, CostCtx.PSE, CostCtx.L);
+  InstructionCost VecCallCost = InstructionCost::getInvalid();
+  if (VecFunc)
+    VecCallCost = VPWidenCallRecipe::computeVectorCallCost(VecFunc, CostCtx);
+
+  // Prefer the intrinsic if it is at least as cheap as scalarizing and any
+  // available vector variant.
+  if (ID) {
+    InstructionCost IntrinsicCost =
+        VPWidenIntrinsicRecipe::computeIntrinsicCost(ID, Ops, VPI, VF, CostCtx);
+    if (IntrinsicCost.isValid() && ScalarCost >= IntrinsicCost &&
+        (!VecFunc || VecCallCost >= IntrinsicCost))
+      return {CallWideningDecision::KindTy::Intrinsic, nullptr, std::nullopt};
+  }
+
+  // Otherwise, use a vector library variant when it beats scalarizing.
+  if (VecFunc && ScalarCost >= VecCallCost)
+    return {CallWideningDecision::KindTy::VectorVariant, VecFunc, MaskPos};
+
+  return {};
+}
+
+void VPlanTransforms::makeCallWideningDecisions(VPlan &Plan, VFRange &Range,
+                                                VPRecipeBuilder &RecipeBuilder,
+                                                VPCostContext &CostCtx) {
+  bool IsScalarVPlan = LoopVectorizationPlanner::getDecisionAndClampRange(
+      [](ElementCount VF) { return VF.isScalar(); }, Range);
+
+  SmallVector<VPInstruction *, 8> ToErase;
+  for (VPBasicBlock *VPBB :
+       VPBlockUtils::blocksOnly<VPBasicBlock>(vp_depth_first_shallow(
+           Plan.getVectorLoopRegion()->getEntryBasicBlock()))) {
+    for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
+      auto *VPI = dyn_cast<VPInstruction>(&R);
+      if (!VPI || !VPI->getUnderlyingValue() ||
+          VPI->getOpcode() != Instruction::Call)
+        continue;
+
+      // Scalar VPlans and known-scalarized calls fall through to replication.
+      auto *CI = cast<CallInst>(VPI->getUnderlyingInstr());
+      bool KeepScalar =
+          IsScalarVPlan ||
+          LoopVectorizationPlanner::getDecisionAndClampRange(
+              [&](ElementCount VF) { return CostCtx.willBeScalarized(CI, VF); },
+              Range);
+
+      VPSingleDefRecipe *Recipe = nullptr;
+      CallWideningDecision Decision;
+      if (!KeepScalar) {
+        SmallVector<VPValue *, 4> Ops(VPI->op_begin(),
+                                      VPI->op_begin() + CI->arg_size());
+
+        // Pick the cheapest widening at Range.Start, then clamp the range.
+        Decision = decideCallWidening(*VPI, Ops, Range.Start, CostCtx);
+        LoopVectorizationPlanner::getDecisionAndClampRange(
+            [&](ElementCount VF) {
+              CallWideningDecision D =
+                  decideCallWidening(*VPI, Ops, VF, CostCtx);
+              return D.Kind == Decision.Kind && D.Variant == Decision.Variant;
+            },
+            Range);
+
+        switch (Decision.Kind) {
+        case CallWideningDecision::KindTy::Intrinsic: {
+          Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, &CostCtx.TLI);
+          Type *ResultTy = CostCtx.Types.inferScalarType(VPI);
+          Recipe = new VPWidenIntrinsicRecipe(*CI, ID, Ops, ResultTy, *VPI,
+                                              *VPI, VPI->getDebugLoc());
+          break;
+        }
+        case CallWideningDecision::KindTy::VectorVariant: {
+          if (Decision.MaskPos) {
+            VPValue *Mask = VPI->isMasked() ? VPI->getMask() : Plan.getTrue();
+            Ops.insert(Ops.begin() + *Decision.MaskPos, Mask);
+          }
+          Ops.push_back(VPI->getOperand(VPI->getNumOperandsWithoutMask() - 1));
+          Recipe =
+              new VPWidenCallRecipe(VPI->getUnderlyingValue(), Decision.Variant,
+                                    Ops, *VPI, *VPI, VPI->getDebugLoc());
+          break;
+        }
+        case CallWideningDecision::KindTy::Scalarize:
+          break;
+        }
+      }
+
+      if (!Recipe)
+        Recipe = RecipeBuilder.handleReplication(VPI, Range);
+
+      assert(all_of(Range,
+                    [&](ElementCount VF) {
+                      Intrinsic::ID IID =
+                          getVectorIntrinsicIDForCall(CI, &CostCtx.TLI);
+                      if (IID && VPCostContext::isFreeScalarIntrinsic(IID))
+                        return true;
+                      auto Legacy = CostCtx.getLegacyCallKind(CI, VF);
+                      return !Legacy || *Legacy == Decision.Kind;
+                    }) &&
+             "VPlan call widening decision must match legacy decision");
+
+      Recipe->insertBefore(VPI);
+      VPI->replaceAllUsesWith(Recipe->getVPSingleValue());
+      ToErase.push_back(VPI);
+    }
+  }
+  for (VPInstruction *VPI : ToErase)
+    VPI->eraseFromParent();
+}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
index d0ce6ce71d80c..8ef6a89e89085 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -544,6 +544,12 @@ struct VPlanTransforms {
   /// made by the legacy CM. Only transforms "usesFirstLaneOnly` def-use chains
   /// enabled by prior widening of consecutive memory operations for now.
   static void makeScalarizationDecisions(VPlan &Plan, VFRange &Range);
+
+  /// Convert call VPInstructions in \p Plan into widened call, vector
+  /// intrinsic or replicate recipes based on a cost comparison via \p CostCtx.
+  static void makeCallWideningDecisions(VPlan &Plan, VFRange &Range,
+                                        VPRecipeBuilder &RecipeBuilder,
+                                        VPCostContext &CostCtx);
 };
 
 } // namespace llvm
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/AArch64/call-decisions.ll b/llvm/test/Transforms/LoopVectorize/VPlan/AArch64/call-decisions.ll
index 50c605d8aea62..b8054a3ff472a 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/AArch64/call-decisions.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/AArch64/call-decisions.ll
@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --filter-out-after "middle.block:" --version 6
-; -RUN: opt -mattr=+neon -vector-library=LIBMVEC -passes=inject-tli-mappings,loop-vectorize -force-vector-width=2 -force-vector-interleave=1 -S -disable-output -vplan-print-after=makeMemOpWideningDecisions %s 2>&1 | FileCheck %s
+; -RUN: opt -mattr=+neon -vector-library=LIBMVEC -passes=inject-tli-mappings,loop-vectorize -force-vector-width=2 -force-vector-interleave=1 -S -disable-output -vplan-print-after=makeCallWideningDecisions %s 2>&1 | FileCheck %s
 
 target triple = "arm64-apple-macosx"
 
@@ -27,7 +27,7 @@ define void @pick_wide_call(ptr noalias %in.ptr, ptr noalias %out.ptr) {
 ; CHECK-NEXT:      EMIT ir<%in.gep> = getelementptr inbounds ir<%in.ptr>, ir<%iv>
 ; CHECK-NEXT:      vp<[[VP4:%[0-9]+]]> = vector-pointer inbounds ir<%in.gep>
 ; CHECK-NEXT:      WIDEN ir<%in> = load vp<[[VP4]]>
-; CHECK-NEXT:      EMIT ir<%call> = call ir<%in>, ir<@acos>
+; CHECK-NEXT:      WIDEN-CALL ir<%call> = call @acos(ir<%in>) (using library function: _ZGVnN2v_acos)
 ; CHECK-NEXT:      EMIT ir<%out.gep> = getelementptr inbounds ir<%out.ptr>, ir<%iv>
 ; CHECK-NEXT:      vp<[[VP5:%[0-9]+]]> = vector-pointer inbounds ir<%out.gep>
 ; CHECK-NEXT:      WIDEN store vp<[[VP5]]>, ir<%call>
@@ -79,7 +79,7 @@ define void @pick_intrinsic_call(ptr noalias %in.ptr, ptr noalias %out.ptr) {
 ; CHECK-NEXT:      EMIT ir<%in.gep> = getelementptr inbounds ir<%in.ptr>, ir<%iv>
 ; CHECK-NEXT:      vp<[[VP4:%[0-9]+]]> = vector-pointer inbounds ir<%in.gep>
 ; CHECK-NEXT:      WIDEN ir<%in> = load vp<[[VP4]]>
-; CHECK-NEXT:       EMIT ir<%call> = call ir<100>, ir<%in>, ir<@llvm.umin.i64>
+; CHECK-NEXT:      WIDEN-INTRINSIC ir<%call> = call llvm.umin(ir<100>, ir<%in>)
 ; CHECK-NEXT:      EMIT ir<%out.gep> = getelementptr inbounds ir<%out.ptr>, ir<%iv>
 ; CHECK-NEXT:      vp<[[VP5:%[0-9]+]]> = vector-pointer inbounds ir<%out.gep>
 ; CHECK-NEXT:      WIDEN store vp<[[VP5]]>, ir<%call>
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/vplan-print-after-all.ll b/llvm/test/Transforms/LoopVectorize/VPlan/vplan-print-after-all.ll
index 1fb23a06401d4..9babf327060ee 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/vplan-print-after-all.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/vplan-print-after-all.ll
@@ -14,6 +14,7 @@
 ; CHECK: VPlan for loop in 'foo' after VPlanTransforms::createInLoopReductionRecipes
 ; CHECK: VPlan for loop in 'foo' after VPlanTransforms::makeMemOpWideningDecisions
 ; CHECK: VPlan for loop in 'foo' after VPlanTransforms::makeScalarizationDecisions
+; CHECK: VPlan for loop in 'foo' after VPlanTransforms::makeCallWideningDecisions
 ; CHECK: VPlan for loop in 'foo' after VPlanTransforms::adjustFirstOrderRecurrenceMiddleUsers
 ; CHECK: VPlan for loop in 'foo' after VPlanTransforms::clearReductionWrapFlags
 ; CHECK: VPlan for loop in 'foo' after VPlanTransforms::optimizeFindIVReductions

>From b50eb65bf4b6dd208457bf4de330c142cc803df6 Mon Sep 17 00:00:00 2001
From: Florian Hahn <flo at fhahn.com>
Date: Wed, 6 May 2026 10:18:05 +0100
Subject: [PATCH 2/2] !fixu address latest comments, thanks

---
 .../lib/Transforms/Vectorize/VPlanRecipes.cpp |  7 +-
 .../Transforms/Vectorize/VPlanTransforms.cpp  | 88 ++++++++-----------
 .../VPlan/AArch64/call-decisions.ll           | 10 +--
 3 files changed, 48 insertions(+), 57 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 3ed0efc7fc226..a858a758a74b6 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -1878,6 +1878,8 @@ void VPWidenCallRecipe::execute(VPTransformState &State) {
 
 InstructionCost VPWidenCallRecipe::computeCost(ElementCount VF,
                                                VPCostContext &Ctx) const {
+  assert(getVectorizedTypeVF(Variant->getReturnType()) == VF &&
+         "Variant return type must match VF");
   return computeVectorCallCost(Variant, Ctx);
 }
 
@@ -3630,6 +3632,8 @@ InstructionCost VPReplicateRecipe::computeScalarCallCost(
 
   Intrinsic::ID IntrinID = CalledFn->getIntrinsicID();
   auto GetIntrinsicCost = [&] {
+    if (!IntrinID)
+      return InstructionCost::getInvalid();
     return Ctx.TTI.getIntrinsicInstrCost(
         IntrinsicCostAttributes(IntrinID, ResultTy, Tys), Ctx.CostKind);
   };
@@ -3642,8 +3646,7 @@ InstructionCost VPReplicateRecipe::computeScalarCallCost(
   InstructionCost ScalarCallCost =
       Ctx.TTI.getCallInstrCost(CalledFn, ResultTy, Tys, Ctx.CostKind);
   if (IsSingleScalar) {
-    if (IntrinID)
-      ScalarCallCost = std::min(ScalarCallCost, GetIntrinsicCost());
+    ScalarCallCost = std::min(ScalarCallCost, GetIntrinsicCost());
     return ScalarCallCost;
   }
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index ad2acdb956df0..8f08cfdda2887 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -6615,6 +6615,11 @@ static CallWideningDecision decideCallWidening(VPInstruction &VPI,
                                                ElementCount VF,
                                                VPCostContext &CostCtx) {
   auto *CI = cast<CallInst>(VPI.getUnderlyingInstr());
+
+  // Scalar VFs and calls forced or known to scalarize always replicate.
+  if (VF.isScalar() || CostCtx.willBeScalarized(CI, VF))
+    return {};
+
   auto *CalledFn = cast<Function>(
       VPI.getOperand(VPI.getNumOperandsWithoutMask() - 1)->getLiveInIRValue());
   Type *ResultTy = CostCtx.Types.inferScalarType(&VPI);
@@ -6655,69 +6660,52 @@ static CallWideningDecision decideCallWidening(VPInstruction &VPI,
 void VPlanTransforms::makeCallWideningDecisions(VPlan &Plan, VFRange &Range,
                                                 VPRecipeBuilder &RecipeBuilder,
                                                 VPCostContext &CostCtx) {
-  bool IsScalarVPlan = LoopVectorizationPlanner::getDecisionAndClampRange(
-      [](ElementCount VF) { return VF.isScalar(); }, Range);
-
   SmallVector<VPInstruction *, 8> ToErase;
-  for (VPBasicBlock *VPBB :
-       VPBlockUtils::blocksOnly<VPBasicBlock>(vp_depth_first_shallow(
-           Plan.getVectorLoopRegion()->getEntryBasicBlock()))) {
+  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
+           vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) {
     for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
       auto *VPI = dyn_cast<VPInstruction>(&R);
       if (!VPI || !VPI->getUnderlyingValue() ||
           VPI->getOpcode() != Instruction::Call)
         continue;
 
-      // Scalar VPlans and known-scalarized calls fall through to replication.
       auto *CI = cast<CallInst>(VPI->getUnderlyingInstr());
-      bool KeepScalar =
-          IsScalarVPlan ||
-          LoopVectorizationPlanner::getDecisionAndClampRange(
-              [&](ElementCount VF) { return CostCtx.willBeScalarized(CI, VF); },
-              Range);
+      SmallVector<VPValue *, 4> Ops(VPI->op_begin(),
+                                    VPI->op_begin() + CI->arg_size());
+
+      CallWideningDecision Decision =
+          decideCallWidening(*VPI, Ops, Range.Start, CostCtx);
+      LoopVectorizationPlanner::getDecisionAndClampRange(
+          [&](ElementCount VF) {
+            CallWideningDecision D = decideCallWidening(*VPI, Ops, VF, CostCtx);
+            return D.Kind == Decision.Kind && D.Variant == Decision.Variant;
+          },
+          Range);
 
       VPSingleDefRecipe *Recipe = nullptr;
-      CallWideningDecision Decision;
-      if (!KeepScalar) {
-        SmallVector<VPValue *, 4> Ops(VPI->op_begin(),
-                                      VPI->op_begin() + CI->arg_size());
-
-        // Pick the cheapest widening at Range.Start, then clamp the range.
-        Decision = decideCallWidening(*VPI, Ops, Range.Start, CostCtx);
-        LoopVectorizationPlanner::getDecisionAndClampRange(
-            [&](ElementCount VF) {
-              CallWideningDecision D =
-                  decideCallWidening(*VPI, Ops, VF, CostCtx);
-              return D.Kind == Decision.Kind && D.Variant == Decision.Variant;
-            },
-            Range);
-
-        switch (Decision.Kind) {
-        case CallWideningDecision::KindTy::Intrinsic: {
-          Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, &CostCtx.TLI);
-          Type *ResultTy = CostCtx.Types.inferScalarType(VPI);
-          Recipe = new VPWidenIntrinsicRecipe(*CI, ID, Ops, ResultTy, *VPI,
-                                              *VPI, VPI->getDebugLoc());
-          break;
-        }
-        case CallWideningDecision::KindTy::VectorVariant: {
-          if (Decision.MaskPos) {
-            VPValue *Mask = VPI->isMasked() ? VPI->getMask() : Plan.getTrue();
-            Ops.insert(Ops.begin() + *Decision.MaskPos, Mask);
-          }
-          Ops.push_back(VPI->getOperand(VPI->getNumOperandsWithoutMask() - 1));
-          Recipe =
-              new VPWidenCallRecipe(VPI->getUnderlyingValue(), Decision.Variant,
-                                    Ops, *VPI, *VPI, VPI->getDebugLoc());
-          break;
-        }
-        case CallWideningDecision::KindTy::Scalarize:
-          break;
+      switch (Decision.Kind) {
+      case CallWideningDecision::KindTy::Intrinsic: {
+        Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, &CostCtx.TLI);
+        Type *ResultTy = CostCtx.Types.inferScalarType(VPI);
+        Recipe = new VPWidenIntrinsicRecipe(*CI, ID, Ops, ResultTy, *VPI, *VPI,
+                                            VPI->getDebugLoc());
+        break;
+      }
+      case CallWideningDecision::KindTy::VectorVariant: {
+        if (Decision.MaskPos) {
+          VPValue *Mask = VPI->isMasked() ? VPI->getMask() : Plan.getTrue();
+          Ops.insert(Ops.begin() + *Decision.MaskPos, Mask);
         }
+        Ops.push_back(VPI->getOperand(VPI->getNumOperandsWithoutMask() - 1));
+        Recipe =
+            new VPWidenCallRecipe(VPI->getUnderlyingValue(), Decision.Variant,
+                                  Ops, *VPI, *VPI, VPI->getDebugLoc());
+        break;
       }
-
-      if (!Recipe)
+      case CallWideningDecision::KindTy::Scalarize:
         Recipe = RecipeBuilder.handleReplication(VPI, Range);
+        break;
+      }
 
       assert(all_of(Range,
                     [&](ElementCount VF) {
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/AArch64/call-decisions.ll b/llvm/test/Transforms/LoopVectorize/VPlan/AArch64/call-decisions.ll
index b8054a3ff472a..74156dd8ac8ae 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/AArch64/call-decisions.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/AArch64/call-decisions.ll
@@ -24,11 +24,11 @@ define void @pick_wide_call(ptr noalias %in.ptr, ptr noalias %out.ptr) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0]]>
-; CHECK-NEXT:      EMIT ir<%in.gep> = getelementptr inbounds ir<%in.ptr>, ir<%iv>
+; CHECK-NEXT:      CLONE ir<%in.gep> = getelementptr inbounds ir<%in.ptr>, ir<%iv>
 ; CHECK-NEXT:      vp<[[VP4:%[0-9]+]]> = vector-pointer inbounds ir<%in.gep>
 ; CHECK-NEXT:      WIDEN ir<%in> = load vp<[[VP4]]>
-; CHECK-NEXT:      WIDEN-CALL ir<%call> = call @acos(ir<%in>) (using library function: _ZGVnN2v_acos)
-; CHECK-NEXT:      EMIT ir<%out.gep> = getelementptr inbounds ir<%out.ptr>, ir<%iv>
+; CHECK-NEXT:      WIDEN-CALL ir<%call> = call  @acos(ir<%in>) (using library function: _ZGVnN2v_acos)
+; CHECK-NEXT:      CLONE ir<%out.gep> = getelementptr inbounds ir<%out.ptr>, ir<%iv>
 ; CHECK-NEXT:      vp<[[VP5:%[0-9]+]]> = vector-pointer inbounds ir<%out.gep>
 ; CHECK-NEXT:      WIDEN store vp<[[VP5]]>, ir<%call>
 ; CHECK-NEXT:      EMIT vp<%index.next> = add nuw vp<[[VP3]]>, vp<[[VP1]]>
@@ -76,11 +76,11 @@ define void @pick_intrinsic_call(ptr noalias %in.ptr, ptr noalias %out.ptr) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0]]>
-; CHECK-NEXT:      EMIT ir<%in.gep> = getelementptr inbounds ir<%in.ptr>, ir<%iv>
+; CHECK-NEXT:      CLONE ir<%in.gep> = getelementptr inbounds ir<%in.ptr>, ir<%iv>
 ; CHECK-NEXT:      vp<[[VP4:%[0-9]+]]> = vector-pointer inbounds ir<%in.gep>
 ; CHECK-NEXT:      WIDEN ir<%in> = load vp<[[VP4]]>
 ; CHECK-NEXT:      WIDEN-INTRINSIC ir<%call> = call llvm.umin(ir<100>, ir<%in>)
-; CHECK-NEXT:      EMIT ir<%out.gep> = getelementptr inbounds ir<%out.ptr>, ir<%iv>
+; CHECK-NEXT:      CLONE ir<%out.gep> = getelementptr inbounds ir<%out.ptr>, ir<%iv>
 ; CHECK-NEXT:      vp<[[VP5:%[0-9]+]]> = vector-pointer inbounds ir<%out.gep>
 ; CHECK-NEXT:      WIDEN store vp<[[VP5]]>, ir<%call>
 ; CHECK-NEXT:      EMIT vp<%index.next> = add nuw vp<[[VP3]]>, vp<[[VP1]]>



More information about the llvm-commits mailing list