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

[Sander de Smalen via llvm-commits]

================
@@ -2189,6 +2277,172 @@ 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 VPExtendedReductionRecipe must be lower 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 beneficial 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;
+  // Matched 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.
+  auto IsMulAccValidAndClampRange =
+      [&](bool isZExt, 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, RedTy, SrcVecTy, CostKind);
+          InstructionCost MulCost = Mul->computeCost(VF, Ctx);
+          InstructionCost RedCost = Red->computeCost(VF, Ctx);
+          InstructionCost ExtCost = 0;
+          if (Ext0)
+            ExtCost += Ext0->computeCost(VF, Ctx);
+          if (Ext1)
+            ExtCost += Ext1->computeCost(VF, Ctx);
+          if (OuterExt)
+            ExtCost += OuterExt->computeCost(VF, Ctx);
+
+          return MulAccCost.isValid() &&
+                 MulAccCost < ExtCost + MulCost + RedCost;
+        },
+        Range);
+  };
+
+  unsigned Opcode = RecurrenceDescriptor::getOpcode(Red->getRecurrenceKind());
+  if (Opcode != Instruction::Add)
+    return nullptr;
+
+  VPValue *VecOp = Red->getVecOp();
+  VPValue *A, *B;
+  // Try to match reduce.add(mul(...))
+  if (match(VecOp, m_Mul(m_VPValue(A), m_VPValue(B)))) {
+    auto *RecipeA =
+        dyn_cast_if_present<VPWidenCastRecipe>(A->getDefiningRecipe());
+    auto *RecipeB =
+        dyn_cast_if_present<VPWidenCastRecipe>(B->getDefiningRecipe());
+    auto *Mul = cast<VPWidenRecipe>(VecOp->getDefiningRecipe());
+
+    // Matched reduce.add(mul(ext, ext))
+    if (RecipeA && RecipeB &&
+        (RecipeA->getOpcode() == RecipeB->getOpcode() || A == B) &&
+        match(RecipeA, m_ZExtOrSExt(m_VPValue())) &&
+        match(RecipeB, m_ZExtOrSExt(m_VPValue())) &&
+        IsMulAccValidAndClampRange(RecipeA->getOpcode() ==
+                                       Instruction::CastOps::ZExt,
+                                   Mul, RecipeA, RecipeB, nullptr))
+      return new VPMulAccumulateReductionRecipe(Red, Mul, RecipeA, RecipeB,
+                                                RecipeA->getResultType());
+    // Matched reduce.add(mul)
+    if (IsMulAccValidAndClampRange(true, Mul, nullptr, nullptr, nullptr))
+      return new VPMulAccumulateReductionRecipe(Red, Mul);
+  }
+  // Matched reduce.add(ext(mul(ext(A), ext(B))))
+  // All extend recipes must have same opcode or A == B
----------------
sdesmalen-arm wrote:

Okay it doesn't sound like that will be too hard to relax in some follow up patch, thanks!

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