[llvm-branch-commits] [llvm] [LV] Only create partial reductions when profitable. (PR #181706)

Sander de Smalen via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Thu Mar 12 08:58:00 PDT 2026


https://github.com/sdesmalen-arm updated https://github.com/llvm/llvm-project/pull/181706

>From 3a12a31edd65c0b7633f5c0b624d1442188da113 Mon Sep 17 00:00:00 2001
From: Sander de Smalen <sander.desmalen at arm.com>
Date: Tue, 10 Feb 2026 09:33:11 +0000
Subject: [PATCH 1/4] [LV] Only create partial reductions when profitable.

We want the LV cost-model to make the best possible decision of
VF and whether or not to use partial reductions.  At the moment,
when the LV can use partial reductions for a given VF range, it
assumes those are always preferred. After transforming the plan to
use partial reductions, it then chooses the most profitable VF. It
is possible for a different VF to have been more profitable, if it
wouldn't have chosen to use partial reductions.

This PR changes that, to first decide whether partial reductions
are more profitable for a given chain. If not, then it won't do
the transform.
---
 .../Transforms/Vectorize/VPlanTransforms.cpp  | 166 ++++++++++--------
 .../AArch64/partial-reduce-chained.ll         |  14 +-
 .../LoopVectorize/AArch64/vector-reverse.ll   |   8 +-
 3 files changed, 101 insertions(+), 87 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index dc8631404e8aa..e918cebf19258 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -4339,23 +4339,19 @@ tryToMatchAndCreateExtendedReduction(VPReductionRecipe *Red, VPCostContext &Ctx,
               cast<VPWidenCastRecipe>(VecOp)->computeCost(VF, Ctx);
           InstructionCost RedCost = Red->computeCost(VF, Ctx);
 
-          if (Red->isPartialReduction()) {
-            TargetTransformInfo::PartialReductionExtendKind ExtKind =
-                TargetTransformInfo::getPartialReductionExtendKind(ExtOpc);
-            // FIXME: Move partial reduction creation, costing and clamping
-            // here from LoopVectorize.cpp.
-            ExtRedCost = Ctx.TTI.getPartialReductionCost(
-                Opcode, SrcTy, nullptr, RedTy, VF, ExtKind,
-                llvm::TargetTransformInfo::PR_None, std::nullopt, Ctx.CostKind,
-                RedTy->isFloatingPointTy()
-                    ? std::optional{Red->getFastMathFlags()}
-                    : std::nullopt);
-          } else if (!RedTy->isFloatingPointTy()) {
-            // TTI::getExtendedReductionCost only supports integer types.
-            ExtRedCost = Ctx.TTI.getExtendedReductionCost(
-                Opcode, ExtOpc == Instruction::CastOps::ZExt, RedTy, SrcVecTy,
-                Red->getFastMathFlags(), CostKind);
-          }
+          // For partial reductions, the decision has already been
+          // made at the point of transforming reductions -> partial
+          // reductions for a given plan, based on the cost-model.
+          if (Red->isPartialReduction())
+            return true;
+
+          // TTI::getExtendedReductionCost for in-loop reductions
+          // only supports integer types.
+          if (RedTy->isFloatingPointTy())
+            return false;
+          ExtRedCost = Ctx.TTI.getExtendedReductionCost(
+              Opcode, ExtOpc == Instruction::CastOps::ZExt, RedTy, SrcVecTy,
+              Red->getFastMathFlags(), CostKind);
           return ExtRedCost.isValid() && ExtRedCost < ExtCost + RedCost;
         },
         Range);
@@ -4405,37 +4401,24 @@ tryToMatchAndCreateMulAccumulateReduction(VPReductionRecipe *Red,
               Ext0 ? Ctx.Types.inferScalarType(Ext0->getOperand(0)) : RedTy;
           InstructionCost MulAccCost;
 
-          if (Red->isPartialReduction()) {
-            Type *SrcTy2 =
-                Ext1 ? Ctx.Types.inferScalarType(Ext1->getOperand(0)) : nullptr;
-            // FIXME: Move partial reduction creation, costing and clamping
-            // here from LoopVectorize.cpp.
-            MulAccCost = Ctx.TTI.getPartialReductionCost(
-                Opcode, SrcTy, SrcTy2, RedTy, VF,
-                Ext0 ? TargetTransformInfo::getPartialReductionExtendKind(
-                           Ext0->getOpcode())
-                     : TargetTransformInfo::PR_None,
-                Ext1 ? TargetTransformInfo::getPartialReductionExtendKind(
-                           Ext1->getOpcode())
-                     : TargetTransformInfo::PR_None,
-                Mul->getOpcode(), CostKind,
-                RedTy->isFloatingPointTy()
-                    ? std::optional{Red->getFastMathFlags()}
-                    : std::nullopt);
-          } else {
-            // Only partial reductions support mixed or floating-point extends
-            // at the moment.
-            if (Ext0 && Ext1 &&
-                (Ext0->getOpcode() != Ext1->getOpcode() ||
-                 Ext0->getOpcode() == Instruction::CastOps::FPExt))
-              return false;
+          // For partial reductions, the decision has already been
+          // made at the point of transforming reductions -> partial
+          // reductions for a given plan, based on the cost-model.
+          if (Red->isPartialReduction())
+            return true;
+
+          // Only partial reductions support mixed or floating-point extends
+          // at the moment.
+          if (Ext0 && Ext1 &&
+              (Ext0->getOpcode() != Ext1->getOpcode() ||
+               Ext0->getOpcode() == Instruction::CastOps::FPExt))
+            return false;
 
-            bool IsZExt =
-                !Ext0 || Ext0->getOpcode() == Instruction::CastOps::ZExt;
-            auto *SrcVecTy = cast<VectorType>(toVectorTy(SrcTy, VF));
-            MulAccCost = Ctx.TTI.getMulAccReductionCost(IsZExt, Opcode, RedTy,
-                                                        SrcVecTy, CostKind);
-          }
+          bool IsZExt =
+              !Ext0 || Ext0->getOpcode() == Instruction::CastOps::ZExt;
+          auto *SrcVecTy = cast<VectorType>(toVectorTy(SrcTy, VF));
+          MulAccCost = Ctx.TTI.getMulAccReductionCost(IsZExt, Opcode, RedTy,
+                                                      SrcVecTy, CostKind);
 
           InstructionCost MulCost = Mul->computeCost(VF, Ctx);
           InstructionCost RedCost = Red->computeCost(VF, Ctx);
@@ -6010,12 +5993,11 @@ static void transformToPartialReduction(const VPPartialReductionChain &Chain,
       [&NewResult](VPUser &U, unsigned Idx) { return &U != NewResult; });
 }
 
-/// Check if a partial reduction chain is is supported by the target (i.e. does
-/// not have an invalid cost) for the given VF range. Clamps the range and
-/// returns true if profitable for any VF.
-static bool isValidPartialReduction(const VPPartialReductionChain &Chain,
-                                    Type *PhiType, VPCostContext &CostCtx,
-                                    VFRange &Range) {
+/// Returns the cost of a link in a partial-reduction chain for a given VF.
+static InstructionCost
+getPartialReductionLinkCost(VPCostContext &CostCtx,
+                            const VPPartialReductionChain &Chain,
+                            ElementCount VF) {
   auto GetExtInfo = [&CostCtx](VPWidenCastRecipe *Ext)
       -> std::pair<Type *, TargetTransformInfo::PartialReductionExtendKind> {
     if (!Ext)
@@ -6025,6 +6007,7 @@ static bool isValidPartialReduction(const VPPartialReductionChain &Chain,
         static_cast<Instruction::CastOps>(Ext->getOpcode()));
     return {ExtOpType, ExtKind};
   };
+
   ExtendedReductionOperand ExtendedOp = Chain.ExtendedOp;
   VPWidenCastRecipe *ExtendA = ExtendedOp.CastRecipes[0];
   VPWidenCastRecipe *ExtendB = ExtendedOp.CastRecipes[1];
@@ -6034,38 +6017,35 @@ static bool isValidPartialReduction(const VPPartialReductionChain &Chain,
   std::tie(ExtOpTypeA, ExtKindA) = GetExtInfo(ExtendA);
   std::tie(ExtOpTypeB, ExtKindB) = GetExtInfo(ExtendB);
 
+  std::optional<unsigned> BinOpc;
+  if (ExtendedOp.BinOp && ExtendedOp.BinOp != Chain.ReductionBinOp)
+    BinOpc = ExtendedOp.BinOp->getOpcode();
+
   // If ExtendB is nullptr but there's a separate BinOp, the second operand
   // was a constant that can use the same extend kind as the first.
-  if (!ExtendB && ExtendedOp.BinOp &&
-      ExtendedOp.BinOp != Chain.ReductionBinOp) {
+  if (!ExtendB && BinOpc) {
     const APInt *Const = nullptr;
     for (VPValue *Op : ExtendedOp.BinOp->operands()) {
       if (match(Op, m_APInt(Const)))
         break;
     }
     if (!Const || !canConstantBeExtended(Const, ExtOpTypeA, ExtKindA))
-      return false;
+      return InstructionCost::getInvalid();
     ExtOpTypeB = ExtOpTypeA;
     ExtKindB = ExtKindA;
   }
 
-  std::optional<unsigned> BinOpc;
-  if (ExtendedOp.BinOp && ExtendedOp.BinOp != Chain.ReductionBinOp)
-    BinOpc = ExtendedOp.BinOp->getOpcode();
+  Type *RdxType = CostCtx.Types.inferScalarType(Chain.ReductionBinOp);
+  std::optional<llvm::FastMathFlags> Flags;
+  if (RdxType->isFloatingPointTy())
+    Flags = Chain.ReductionBinOp->getFastMathFlags();
 
-  VPWidenRecipe *WidenRecipe = Chain.ReductionBinOp;
-  return LoopVectorizationPlanner::getDecisionAndClampRange(
-      [&](ElementCount VF) {
-        return CostCtx.TTI
-            .getPartialReductionCost(
-                WidenRecipe->getOpcode(), ExtOpTypeA, ExtOpTypeB, PhiType, VF,
-                ExtKindA, ExtKindB, BinOpc, CostCtx.CostKind,
-                PhiType->isFloatingPointTy()
-                    ? std::optional{WidenRecipe->getFastMathFlags()}
-                    : std::nullopt)
-            .isValid();
-      },
-      Range);
+  unsigned Opcode = Chain.RK == RecurKind::Sub
+                        ? (unsigned)Instruction::Add
+                        : Chain.ReductionBinOp->getOpcode();
+  return CostCtx.TTI.getPartialReductionCost(Opcode, ExtOpTypeA, ExtOpTypeB,
+                                             RdxType, VF, ExtKindA, ExtKindB,
+                                             BinOpc, CostCtx.CostKind, Flags);
 }
 
 /// Checks if \p Op (which is an operand of \p UpdateR) is an extended reduction
@@ -6114,8 +6094,7 @@ matchExtendedReductionOperand(VPWidenRecipe *UpdateR, VPValue *Op) {
     assert(Operands.size() <= 2 && "expected at most 2 operands");
 
     for (const auto &[I, OpVal] : enumerate(Operands)) {
-      // Allow constant as second operand - validation happens in
-      // isValidPartialReduction.
+      // Allow constant as second operand - validation happens later.
       const APInt *Unused;
       if (I > 0 && CastRecipes[0] && match(OpVal, m_APInt(Unused)))
         continue;
@@ -6216,10 +6195,17 @@ getScaledReductions(VPReductionPHIRecipe *RedPhiR, VPCostContext &CostCtx,
     if (!PHISize.hasKnownScalarFactor(ExtSrcSize))
       return std::nullopt;
 
+    /// Check if a partial reduction chain is supported by the target (i.e.
+    /// does not have an invalid cost) for the given VF range. Clamps the range
+    /// and returns true if feasible for any VF.
     VPPartialReductionChain Chain(
         {UpdateR, *ExtendedOp,
          static_cast<unsigned>(PHISize.getKnownScalarFactor(ExtSrcSize)), RK});
-    if (!isValidPartialReduction(Chain, PhiType, CostCtx, Range))
+    if (!LoopVectorizationPlanner::getDecisionAndClampRange(
+            [&](ElementCount VF) {
+              return getPartialReductionLinkCost(CostCtx, Chain, VF).isValid();
+            },
+            Range))
       return std::nullopt;
 
     Chains.push_back(Chain);
@@ -6273,6 +6259,28 @@ void VPlanTransforms::createPartialReductions(VPlan &Plan,
     });
   };
 
+  auto IsProfitablePartialReductionChainForVF =
+      [&](ArrayRef<VPPartialReductionChain> Chain, ElementCount VF) -> bool {
+    InstructionCost PartialCost = 0, RegularCost = 0;
+
+    // The chain is a profitable partial reduction chain if
+    // the cost of handling the entire chain is cheaper when
+    // using partial reductions than when handling the entire
+    // chain using regular reductions.
+    for (const VPPartialReductionChain &Link : Chain) {
+      ExtendedReductionOperand ExtendedOp = Link.ExtendedOp;
+      PartialCost += getPartialReductionLinkCost(CostCtx, Link, VF);
+      RegularCost += Link.ReductionBinOp->computeCost(VF, CostCtx);
+      if (ExtendedOp.BinOp && ExtendedOp.BinOp != Link.ReductionBinOp)
+        RegularCost += ExtendedOp.BinOp->computeCost(VF, CostCtx);
+      if (ExtendedOp.CastRecipes[0])
+        RegularCost += ExtendedOp.CastRecipes[0]->computeCost(VF, CostCtx);
+      if (ExtendedOp.CastRecipes[1])
+        RegularCost += ExtendedOp.CastRecipes[1]->computeCost(VF, CostCtx);
+    }
+    return PartialCost.isValid() && PartialCost <= RegularCost;
+  };
+
   // Validate chains: check that extends are only used by partial reductions,
   // and that reduction bin ops are only used by other partial reductions with
   // matching scale factors, are outside the loop region or the select
@@ -6311,6 +6319,14 @@ void VPlanTransforms::createPartialReductions(VPlan &Plan,
         }
       }
     }
+
+    // Clear the chain if it is not profitable.
+    if (!LoopVectorizationPlanner::getDecisionAndClampRange(
+            [&](ElementCount VF) {
+              return IsProfitablePartialReductionChainForVF(Chains, VF);
+            },
+            Range))
+      Chains.clear();
   }
 
   for (auto &[Phi, Chains] : ChainsByPhi)
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-chained.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-chained.ll
index 15e0220b71d61..98d3ee86cf276 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-chained.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-chained.ll
@@ -893,7 +893,7 @@ define i32 @chained_partial_reduce_sub_add_sub(ptr %a, ptr %b, ptr %c, i32 %N) #
 ; CHECK-SVE-MAXBW-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK-SVE-MAXBW:       vector.body:
 ; CHECK-SVE-MAXBW-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-SVE-MAXBW-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 2 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE4:%.*]], [[VECTOR_BODY]] ]
+; CHECK-SVE-MAXBW-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 8 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-SVE-MAXBW-NEXT:    [[TMP7:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 [[INDEX]]
 ; CHECK-SVE-MAXBW-NEXT:    [[TMP8:%.*]] = getelementptr inbounds nuw i8, ptr [[B]], i64 [[INDEX]]
 ; CHECK-SVE-MAXBW-NEXT:    [[TMP9:%.*]] = getelementptr inbounds nuw i8, ptr [[C]], i64 [[INDEX]]
@@ -902,20 +902,18 @@ define i32 @chained_partial_reduce_sub_add_sub(ptr %a, ptr %b, ptr %c, i32 %N) #
 ; CHECK-SVE-MAXBW-NEXT:    [[WIDE_LOAD2:%.*]] = load <vscale x 8 x i8>, ptr [[TMP9]], align 1
 ; CHECK-SVE-MAXBW-NEXT:    [[TMP13:%.*]] = sext <vscale x 8 x i8> [[WIDE_LOAD]] to <vscale x 8 x i32>
 ; CHECK-SVE-MAXBW-NEXT:    [[TMP14:%.*]] = sext <vscale x 8 x i8> [[WIDE_LOAD1]] to <vscale x 8 x i32>
-; CHECK-SVE-MAXBW-NEXT:    [[TMP16:%.*]] = mul nsw <vscale x 8 x i32> [[TMP13]], [[TMP14]]
-; CHECK-SVE-MAXBW-NEXT:    [[TMP17:%.*]] = sub nsw <vscale x 8 x i32> zeroinitializer, [[TMP16]]
-; CHECK-SVE-MAXBW-NEXT:    [[PARTIAL_REDUCE:%.*]] = call <vscale x 2 x i32> @llvm.vector.partial.reduce.add.nxv2i32.nxv8i32(<vscale x 2 x i32> [[VEC_PHI]], <vscale x 8 x i32> [[TMP17]])
 ; CHECK-SVE-MAXBW-NEXT:    [[TMP12:%.*]] = sext <vscale x 8 x i8> [[WIDE_LOAD2]] to <vscale x 8 x i32>
+; CHECK-SVE-MAXBW-NEXT:    [[TMP10:%.*]] = mul nsw <vscale x 8 x i32> [[TMP13]], [[TMP14]]
+; CHECK-SVE-MAXBW-NEXT:    [[TMP11:%.*]] = sub <vscale x 8 x i32> [[VEC_PHI]], [[TMP10]]
 ; CHECK-SVE-MAXBW-NEXT:    [[TMP18:%.*]] = mul nsw <vscale x 8 x i32> [[TMP13]], [[TMP12]]
-; CHECK-SVE-MAXBW-NEXT:    [[PARTIAL_REDUCE3:%.*]] = call <vscale x 2 x i32> @llvm.vector.partial.reduce.add.nxv2i32.nxv8i32(<vscale x 2 x i32> [[PARTIAL_REDUCE]], <vscale x 8 x i32> [[TMP18]])
+; CHECK-SVE-MAXBW-NEXT:    [[TMP16:%.*]] = add <vscale x 8 x i32> [[TMP11]], [[TMP18]]
 ; CHECK-SVE-MAXBW-NEXT:    [[TMP19:%.*]] = mul nsw <vscale x 8 x i32> [[TMP14]], [[TMP12]]
-; CHECK-SVE-MAXBW-NEXT:    [[TMP20:%.*]] = sub <vscale x 8 x i32> zeroinitializer, [[TMP19]]
-; CHECK-SVE-MAXBW-NEXT:    [[PARTIAL_REDUCE4]] = call <vscale x 2 x i32> @llvm.vector.partial.reduce.add.nxv2i32.nxv8i32(<vscale x 2 x i32> [[PARTIAL_REDUCE3]], <vscale x 8 x i32> [[TMP20]])
+; CHECK-SVE-MAXBW-NEXT:    [[TMP15]] = sub <vscale x 8 x i32> [[TMP16]], [[TMP19]]
 ; CHECK-SVE-MAXBW-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]
 ; CHECK-SVE-MAXBW-NEXT:    [[TMP22:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-SVE-MAXBW-NEXT:    br i1 [[TMP22]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]]
 ; CHECK-SVE-MAXBW:       middle.block:
-; CHECK-SVE-MAXBW-NEXT:    [[TMP23:%.*]] = call i32 @llvm.vector.reduce.add.nxv2i32(<vscale x 2 x i32> [[PARTIAL_REDUCE4]])
+; CHECK-SVE-MAXBW-NEXT:    [[TMP17:%.*]] = call i32 @llvm.vector.reduce.add.nxv8i32(<vscale x 8 x i32> [[TMP15]])
 ; CHECK-SVE-MAXBW-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
 ; CHECK-SVE-MAXBW-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
 ; CHECK-SVE-MAXBW:       scalar.ph:
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/vector-reverse.ll b/llvm/test/Transforms/LoopVectorize/AArch64/vector-reverse.ll
index 114ed5b0a4e5e..ede64f09a2aec 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/vector-reverse.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/vector-reverse.ll
@@ -203,17 +203,17 @@ define i32 @reverse_store_with_partial_reduction(ptr noalias %dst, ptr noalias %
 ; CHECK-NEXT:    [[N_MOD_VF10:%.*]] = urem i64 [[TMP0]], 4
 ; CHECK-NEXT:    [[N_VEC11:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF10]]
 ; CHECK-NEXT:    [[TMP32:%.*]] = sub i64 [[N]], [[N_VEC11]]
-; CHECK-NEXT:    [[TMP33:%.*]] = insertelement <2 x i32> zeroinitializer, i32 [[BC_MERGE_RDX]], i32 0
+; CHECK-NEXT:    [[TMP26:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[BC_MERGE_RDX]], i32 0
 ; CHECK-NEXT:    br label %[[VEC_EPILOG_VECTOR_BODY:.*]]
 ; CHECK:       [[VEC_EPILOG_VECTOR_BODY]]:
 ; CHECK-NEXT:    [[INDEX12:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], %[[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT18:%.*]], %[[VEC_EPILOG_VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI13:%.*]] = phi <2 x i32> [ [[TMP33]], %[[VEC_EPILOG_PH]] ], [ [[PARTIAL_REDUCE16:%.*]], %[[VEC_EPILOG_VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI13:%.*]] = phi <4 x i32> [ [[TMP26]], %[[VEC_EPILOG_PH]] ], [ [[TMP33:%.*]], %[[VEC_EPILOG_VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = sub i64 [[N]], [[INDEX12]]
 ; CHECK-NEXT:    [[TMP34:%.*]] = load i16, ptr [[SRC]], align 2
 ; CHECK-NEXT:    [[BROADCAST_SPLATINSERT14:%.*]] = insertelement <4 x i16> poison, i16 [[TMP34]], i64 0
 ; CHECK-NEXT:    [[BROADCAST_SPLAT15:%.*]] = shufflevector <4 x i16> [[BROADCAST_SPLATINSERT14]], <4 x i16> poison, <4 x i32> zeroinitializer
 ; CHECK-NEXT:    [[TMP35:%.*]] = sext <4 x i16> [[BROADCAST_SPLAT15]] to <4 x i32>
-; CHECK-NEXT:    [[PARTIAL_REDUCE16]] = call <2 x i32> @llvm.vector.partial.reduce.add.v2i32.v4i32(<2 x i32> [[VEC_PHI13]], <4 x i32> [[TMP35]])
+; CHECK-NEXT:    [[TMP33]] = add <4 x i32> [[VEC_PHI13]], [[TMP35]]
 ; CHECK-NEXT:    [[TMP36:%.*]] = getelementptr i16, ptr [[DST]], i64 [[OFFSET_IDX]]
 ; CHECK-NEXT:    [[TMP38:%.*]] = getelementptr i16, ptr [[TMP36]], i64 -3
 ; CHECK-NEXT:    [[REVERSE17:%.*]] = shufflevector <4 x i16> [[BROADCAST_SPLAT15]], <4 x i16> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
@@ -222,7 +222,7 @@ define i32 @reverse_store_with_partial_reduction(ptr noalias %dst, ptr noalias %
 ; CHECK-NEXT:    [[TMP39:%.*]] = icmp eq i64 [[INDEX_NEXT18]], [[N_VEC11]]
 ; CHECK-NEXT:    br i1 [[TMP39]], label %[[VEC_EPILOG_MIDDLE_BLOCK:.*]], label %[[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
 ; CHECK:       [[VEC_EPILOG_MIDDLE_BLOCK]]:
-; CHECK-NEXT:    [[TMP40:%.*]] = call i32 @llvm.vector.reduce.add.v2i32(<2 x i32> [[PARTIAL_REDUCE16]])
+; CHECK-NEXT:    [[TMP37:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP33]])
 ; CHECK-NEXT:    [[CMP_N19:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC11]]
 ; CHECK-NEXT:    br i1 [[CMP_N19]], [[EXIT]], label %[[VEC_EPILOG_SCALAR_PH]]
 ; CHECK:       [[VEC_EPILOG_SCALAR_PH]]:

>From ad51459fcb791e1b707bc2571a7169426d80893f Mon Sep 17 00:00:00 2001
From: Sander de Smalen <sander.desmalen at arm.com>
Date: Thu, 12 Mar 2026 13:41:14 +0000
Subject: [PATCH 2/4] Address suggestions

---
 .../Transforms/Vectorize/VPlanTransforms.cpp  | 59 +++++++++++--------
 1 file changed, 33 insertions(+), 26 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index e918cebf19258..226354c2bf486 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -4326,6 +4326,12 @@ tryToMatchAndCreateExtendedReduction(VPReductionRecipe *Red, VPCostContext &Ctx,
   Type *RedTy = Ctx.Types.inferScalarType(Red);
   VPValue *VecOp = Red->getVecOp();
 
+  // For partial reductions, the decision has already been
+  // made at the point of transforming reductions -> partial
+  // reductions for a given plan, based on the cost-model.
+  if (Red->isPartialReduction())
+    return new VPExpressionRecipe(cast<VPWidenCastRecipe>(VecOp), Red);
+
   // Clamp the range if using extended-reduction is profitable.
   auto IsExtendedRedValidAndClampRange =
       [&](unsigned Opcode, Instruction::CastOps ExtOpc, Type *SrcTy) -> bool {
@@ -4339,12 +4345,6 @@ tryToMatchAndCreateExtendedReduction(VPReductionRecipe *Red, VPCostContext &Ctx,
               cast<VPWidenCastRecipe>(VecOp)->computeCost(VF, Ctx);
           InstructionCost RedCost = Red->computeCost(VF, Ctx);
 
-          // For partial reductions, the decision has already been
-          // made at the point of transforming reductions -> partial
-          // reductions for a given plan, based on the cost-model.
-          if (Red->isPartialReduction())
-            return true;
-
           // TTI::getExtendedReductionCost for in-loop reductions
           // only supports integer types.
           if (RedTy->isFloatingPointTy())
@@ -4396,17 +4396,17 @@ tryToMatchAndCreateMulAccumulateReduction(VPReductionRecipe *Red,
           VPWidenCastRecipe *OuterExt) -> bool {
     return LoopVectorizationPlanner::getDecisionAndClampRange(
         [&](ElementCount VF) {
-          TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
-          Type *SrcTy =
-              Ext0 ? Ctx.Types.inferScalarType(Ext0->getOperand(0)) : RedTy;
-          InstructionCost MulAccCost;
-
           // For partial reductions, the decision has already been
           // made at the point of transforming reductions -> partial
           // reductions for a given plan, based on the cost-model.
           if (Red->isPartialReduction())
             return true;
 
+          TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
+          Type *SrcTy =
+              Ext0 ? Ctx.Types.inferScalarType(Ext0->getOperand(0)) : RedTy;
+          InstructionCost MulAccCost;
+
           // Only partial reductions support mixed or floating-point extends
           // at the moment.
           if (Ext0 && Ext1 &&
@@ -5830,6 +5830,8 @@ struct VPPartialReductionChain {
   /// This allows distinguishing between Sub and AddWithSub recurrences,
   /// when the ReductionBinOp is a Instruction::Sub.
   RecurKind RK;
+  /// The cost of the link in the reduction chain.
+  DenseMap<ElementCount, InstructionCost> PartialReductionCost;
 };
 
 static VPSingleDefRecipe *
@@ -5996,7 +5998,7 @@ static void transformToPartialReduction(const VPPartialReductionChain &Chain,
 /// Returns the cost of a link in a partial-reduction chain for a given VF.
 static InstructionCost
 getPartialReductionLinkCost(VPCostContext &CostCtx,
-                            const VPPartialReductionChain &Chain,
+                            const VPPartialReductionChain &Link,
                             ElementCount VF) {
   auto GetExtInfo = [&CostCtx](VPWidenCastRecipe *Ext)
       -> std::pair<Type *, TargetTransformInfo::PartialReductionExtendKind> {
@@ -6008,7 +6010,7 @@ getPartialReductionLinkCost(VPCostContext &CostCtx,
     return {ExtOpType, ExtKind};
   };
 
-  ExtendedReductionOperand ExtendedOp = Chain.ExtendedOp;
+  ExtendedReductionOperand ExtendedOp = Link.ExtendedOp;
   VPWidenCastRecipe *ExtendA = ExtendedOp.CastRecipes[0];
   VPWidenCastRecipe *ExtendB = ExtendedOp.CastRecipes[1];
 
@@ -6018,7 +6020,7 @@ getPartialReductionLinkCost(VPCostContext &CostCtx,
   std::tie(ExtOpTypeB, ExtKindB) = GetExtInfo(ExtendB);
 
   std::optional<unsigned> BinOpc;
-  if (ExtendedOp.BinOp && ExtendedOp.BinOp != Chain.ReductionBinOp)
+  if (ExtendedOp.BinOp && ExtendedOp.BinOp != Link.ReductionBinOp)
     BinOpc = ExtendedOp.BinOp->getOpcode();
 
   // If ExtendB is nullptr but there's a separate BinOp, the second operand
@@ -6035,14 +6037,14 @@ getPartialReductionLinkCost(VPCostContext &CostCtx,
     ExtKindB = ExtKindA;
   }
 
-  Type *RdxType = CostCtx.Types.inferScalarType(Chain.ReductionBinOp);
+  Type *RdxType = CostCtx.Types.inferScalarType(Link.ReductionBinOp);
   std::optional<llvm::FastMathFlags> Flags;
   if (RdxType->isFloatingPointTy())
-    Flags = Chain.ReductionBinOp->getFastMathFlags();
+    Flags = Link.ReductionBinOp->getFastMathFlags();
 
-  unsigned Opcode = Chain.RK == RecurKind::Sub
+  unsigned Opcode = Link.RK == RecurKind::Sub
                         ? (unsigned)Instruction::Add
-                        : Chain.ReductionBinOp->getOpcode();
+                        : Link.ReductionBinOp->getOpcode();
   return CostCtx.TTI.getPartialReductionCost(Opcode, ExtOpTypeA, ExtOpTypeB,
                                              RdxType, VF, ExtKindA, ExtKindB,
                                              BinOpc, CostCtx.CostKind, Flags);
@@ -6163,7 +6165,7 @@ getScaledReductions(VPReductionPHIRecipe *RedPhiR, VPCostContext &CostCtx,
   VPValue *ExitValue = RdxResult->getOperand(0);
   match(ExitValue, m_Select(m_VPValue(), m_VPValue(ExitValue), m_VPValue()));
 
-  SmallVector<VPPartialReductionChain> Chains;
+  SmallVector<VPPartialReductionChain> Chain;
   RecurKind RK = RedPhiR->getRecurrenceKind();
   Type *PhiType = CostCtx.Types.inferScalarType(RedPhiR);
   TypeSize PHISize = PhiType->getPrimitiveSizeInBits();
@@ -6198,24 +6200,27 @@ getScaledReductions(VPReductionPHIRecipe *RedPhiR, VPCostContext &CostCtx,
     /// Check if a partial reduction chain is supported by the target (i.e.
     /// does not have an invalid cost) for the given VF range. Clamps the range
     /// and returns true if feasible for any VF.
-    VPPartialReductionChain Chain(
+    VPPartialReductionChain Link(
         {UpdateR, *ExtendedOp,
          static_cast<unsigned>(PHISize.getKnownScalarFactor(ExtSrcSize)), RK});
     if (!LoopVectorizationPlanner::getDecisionAndClampRange(
             [&](ElementCount VF) {
-              return getPartialReductionLinkCost(CostCtx, Chain, VF).isValid();
+              InstructionCost Cost =
+                  getPartialReductionLinkCost(CostCtx, Link, VF);
+              Link.PartialReductionCost[VF] = Cost;
+              return Cost.isValid();
             },
             Range))
       return std::nullopt;
 
-    Chains.push_back(Chain);
+    Chain.push_back(Link);
     CurrentValue = PrevValue;
   }
 
   // The chains were collected by traversing backwards from the exit value.
   // Reverse the chains so they are in program order.
-  std::reverse(Chains.begin(), Chains.end());
-  return Chains;
+  std::reverse(Chain.begin(), Chain.end());
+  return Chain;
 }
 } // namespace
 
@@ -6269,7 +6274,9 @@ void VPlanTransforms::createPartialReductions(VPlan &Plan,
     // chain using regular reductions.
     for (const VPPartialReductionChain &Link : Chain) {
       ExtendedReductionOperand ExtendedOp = Link.ExtendedOp;
-      PartialCost += getPartialReductionLinkCost(CostCtx, Link, VF);
+      assert(Link.PartialReductionCost.contains(VF) &&
+             "Expected cost to have been calculated for VF");
+      PartialCost += Link.PartialReductionCost.lookup(VF);
       RegularCost += Link.ReductionBinOp->computeCost(VF, CostCtx);
       if (ExtendedOp.BinOp && ExtendedOp.BinOp != Link.ReductionBinOp)
         RegularCost += ExtendedOp.BinOp->computeCost(VF, CostCtx);
@@ -6322,7 +6329,7 @@ void VPlanTransforms::createPartialReductions(VPlan &Plan,
 
     // Clear the chain if it is not profitable.
     if (!LoopVectorizationPlanner::getDecisionAndClampRange(
-            [&](ElementCount VF) {
+            [&, &Chains = Chains](ElementCount VF) {
               return IsProfitablePartialReductionChainForVF(Chains, VF);
             },
             Range))

>From 48260624239e394d25a53d57fbd42dc125fa311a Mon Sep 17 00:00:00 2001
From: Sander de Smalen <sander.desmalen at arm.com>
Date: Thu, 12 Mar 2026 15:33:35 +0000
Subject: [PATCH 3/4] Address suggestion

---
 .../Transforms/Vectorize/VPlanTransforms.cpp  | 20 +++++--------------
 1 file changed, 5 insertions(+), 15 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 226354c2bf486..25733e994531b 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -5830,8 +5830,6 @@ struct VPPartialReductionChain {
   /// This allows distinguishing between Sub and AddWithSub recurrences,
   /// when the ReductionBinOp is a Instruction::Sub.
   RecurKind RK;
-  /// The cost of the link in the reduction chain.
-  DenseMap<ElementCount, InstructionCost> PartialReductionCost;
 };
 
 static VPSingleDefRecipe *
@@ -6203,16 +6201,6 @@ getScaledReductions(VPReductionPHIRecipe *RedPhiR, VPCostContext &CostCtx,
     VPPartialReductionChain Link(
         {UpdateR, *ExtendedOp,
          static_cast<unsigned>(PHISize.getKnownScalarFactor(ExtSrcSize)), RK});
-    if (!LoopVectorizationPlanner::getDecisionAndClampRange(
-            [&](ElementCount VF) {
-              InstructionCost Cost =
-                  getPartialReductionLinkCost(CostCtx, Link, VF);
-              Link.PartialReductionCost[VF] = Cost;
-              return Cost.isValid();
-            },
-            Range))
-      return std::nullopt;
-
     Chain.push_back(Link);
     CurrentValue = PrevValue;
   }
@@ -6274,9 +6262,11 @@ void VPlanTransforms::createPartialReductions(VPlan &Plan,
     // chain using regular reductions.
     for (const VPPartialReductionChain &Link : Chain) {
       ExtendedReductionOperand ExtendedOp = Link.ExtendedOp;
-      assert(Link.PartialReductionCost.contains(VF) &&
-             "Expected cost to have been calculated for VF");
-      PartialCost += Link.PartialReductionCost.lookup(VF);
+      InstructionCost LinkCost = getPartialReductionLinkCost(CostCtx, Link, VF);
+      if (!LinkCost.isValid())
+        return false;
+
+      PartialCost += LinkCost;
       RegularCost += Link.ReductionBinOp->computeCost(VF, CostCtx);
       if (ExtendedOp.BinOp && ExtendedOp.BinOp != Link.ReductionBinOp)
         RegularCost += ExtendedOp.BinOp->computeCost(VF, CostCtx);

>From fd9f09df4f99a0b11086c7ca23f2654722e520ca Mon Sep 17 00:00:00 2001
From: Sander de Smalen <sander.desmalen at arm.com>
Date: Thu, 12 Mar 2026 15:53:21 +0000
Subject: [PATCH 4/4] Address nit

---
 llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 25733e994531b..c2f0d9d3ef16e 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -6270,10 +6270,9 @@ void VPlanTransforms::createPartialReductions(VPlan &Plan,
       RegularCost += Link.ReductionBinOp->computeCost(VF, CostCtx);
       if (ExtendedOp.BinOp && ExtendedOp.BinOp != Link.ReductionBinOp)
         RegularCost += ExtendedOp.BinOp->computeCost(VF, CostCtx);
-      if (ExtendedOp.CastRecipes[0])
-        RegularCost += ExtendedOp.CastRecipes[0]->computeCost(VF, CostCtx);
-      if (ExtendedOp.CastRecipes[1])
-        RegularCost += ExtendedOp.CastRecipes[1]->computeCost(VF, CostCtx);
+      for (VPWidenCastRecipe *Extend : ExtendedOp.CastRecipes)
+        if (Extend)
+          RegularCost += Extend->computeCost(VF, CostCtx);
     }
     return PartialCost.isValid() && PartialCost <= RegularCost;
   };



More information about the llvm-branch-commits mailing list