[llvm] [VPlan] Impl VPlan-based pattern match for ExtendedRed and MulAccRed (PR #113903)

Sander de Smalen via llvm-commits llvm-commits at lists.llvm.org
Wed Apr 23 07:14:38 PDT 2025


================
@@ -2548,6 +2630,172 @@ void VPlanTransforms::handleUncountableEarlyExit(
   LatchExitingBranch->eraseFromParent();
 }
 
+/// This function tries convert extended in-loop reductions to
+/// VPExtendedReductionRecipe and clamp the \p Range if it is beneficial and
+/// valid. The created recipe must be lowered to concrete
+/// recipes before execution.
+static VPExtendedReductionRecipe *
+tryToMatchAndCreateExtendedReduction(VPReductionRecipe *Red, VPCostContext &Ctx,
+                                     VFRange &Range) {
+  using namespace VPlanPatternMatch;
+
+  Type *RedTy = Ctx.Types.inferScalarType(Red);
+  VPValue *VecOp = Red->getVecOp();
+
+  // Test if using extended-reduction is profitable and clamp the range.
+  auto IsExtendedRedValidAndClampRange = [&](unsigned Opcode, bool isZExt,
+                                             Type *SrcTy) -> bool {
+    return LoopVectorizationPlanner::getDecisionAndClampRange(
+        [&](ElementCount VF) {
+          auto *SrcVecTy = cast<VectorType>(toVectorTy(SrcTy, VF));
+          TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
+          InstructionCost ExtRedCost = Ctx.TTI.getExtendedReductionCost(
+              Opcode, isZExt, RedTy, SrcVecTy, Red->getFastMathFlags(),
+              CostKind);
+          InstructionCost ExtCost =
+              cast<VPWidenCastRecipe>(VecOp)->computeCost(VF, Ctx);
+          InstructionCost RedCost = Red->computeCost(VF, Ctx);
+          return ExtRedCost.isValid() && ExtRedCost < ExtCost + RedCost;
+        },
+        Range);
+  };
+
+  VPValue *A;
+  // Match reduce(ext)).
+  if (match(VecOp, m_ZExtOrSExt(m_VPValue(A))) &&
+      IsExtendedRedValidAndClampRange(
+          RecurrenceDescriptor::getOpcode(Red->getRecurrenceKind()),
+          cast<VPWidenCastRecipe>(VecOp)->getOpcode() ==
+              Instruction::CastOps::ZExt,
+          Ctx.Types.inferScalarType(A)))
+    return new VPExtendedReductionRecipe(Red, cast<VPWidenCastRecipe>(VecOp));
+
+  return nullptr;
+}
+
+/// This function tries convert extended in-loop reductions to
+/// VPMulAccumulateReductionRecipe and clamp the \p Range if it is beneficial
+/// and valid. The created VPExtendedReductionRecipe must be lower to concrete
+/// recipes before execution. Patterns of MulAccumulateReduction:
+///   reduce.add(mul(...)),
+///   reduce.add(mul(ext(A), ext(B))),
+///   reduce.add(ext(mul(ext(A), ext(B)))).
+static VPMulAccumulateReductionRecipe *
+tryToMatchAndCreateMulAccumulateReduction(VPReductionRecipe *Red,
+                                          VPCostContext &Ctx, VFRange &Range) {
+  using namespace VPlanPatternMatch;
+
+  Type *RedTy = Ctx.Types.inferScalarType(Red);
+
+  // Test if using mulutiply-accumulate-reduction is beneficial and clamp the
+  // range.
----------------
sdesmalen-arm wrote:

nit:
```suggestion
  // Clamp the range if using mulutiply-accumulate-reduction is beneficial.
```

https://github.com/llvm/llvm-project/pull/113903


More information about the llvm-commits mailing list