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

[Florian Hahn via llvm-commits]

================
@@ -2185,6 +2273,172 @@ void VPlanTransforms::handleUncountableEarlyExit(
   LatchExitingBranch->eraseFromParent();
 }
 
+/// This function tries to match following pattern to create
+/// VPExtendedReductionRecipe and clamp the \p Range if it is beneficial and
+/// valid. The created VPExtendedReductionRecipe will lower to concrete recipes
+/// before execution.
+///   reduce(ext(...)).
+static VPExtendedReductionRecipe *
+tryToMatchAndCreateExtendedReduction(VPReductionRecipe *Red, VPCostContext &Ctx,
+                                     VFRange &Range) {
+  using namespace VPlanPatternMatch;
+
+  Type *RedTy = Ctx.Types.inferScalarType(Red);
+  VPValue *VecOp = Red->getVecOp();
+  const RecurrenceDescriptor &RdxDesc = Red->getRecurrenceDescriptor();
+
+  // 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, RdxDesc.getFastMathFlags(),
----------------
fhahn wrote:

Should use fast math flags from recipe?

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