[llvm] [LV] Bundle partial reductions inside VPExpressionRecipe (PR #147302)

[Sam Tebbs via llvm-commits]

================
@@ -3514,16 +3528,41 @@ tryToMatchAndCreateMulAccumulateReduction(VPReductionRecipe *Red,
 
   // Clamp the range if using multiply-accumulate-reduction is profitable.
   auto IsMulAccValidAndClampRange =
-      [&](bool isZExt, VPWidenRecipe *Mul, VPWidenCastRecipe *Ext0,
-          VPWidenCastRecipe *Ext1, VPWidenCastRecipe *OuterExt) -> bool {
+      [&](VPWidenRecipe *Mul, VPWidenCastRecipe *Ext0, VPWidenCastRecipe *Ext1,
+          VPWidenCastRecipe *OuterExt) -> bool {
     return LoopVectorizationPlanner::getDecisionAndClampRange(
         [&](ElementCount VF) {
           TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
           Type *SrcTy =
               Ext0 ? Ctx.Types.inferScalarType(Ext0->getOperand(0)) : RedTy;
-          auto *SrcVecTy = cast<VectorType>(toVectorTy(SrcTy, VF));
-          InstructionCost MulAccCost = Ctx.TTI.getMulAccReductionCost(
-              isZExt, Opcode, RedTy, SrcVecTy, CostKind);
+          InstructionCost MulAccCost;
+
+          if (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);
----------------
SamTebbs33 wrote:

Thanks, I'll re-add the assert as part of #162503 .

https://github.com/llvm/llvm-project/pull/147302