[llvm-branch-commits] [llvm] [LV] Choose best reduction for VPlan (PR #166138)

[Sushant Gokhale via llvm-branch-commits]

================
@@ -5679,6 +5679,18 @@ InstructionCost AArch64TTIImpl::getPartialReductionCost(
   if (CostKind != TTI::TCK_RecipThroughput)
     return Invalid;
 
+  unsigned Ratio =
+      AccumType->getScalarSizeInBits() / InputTypeA->getScalarSizeInBits();
+
+  // A ratio of 1 would mean it's similar to a regular add, e.g.
+  // v4i64 partial.reduce(v4i64 %acc, v4i64 %vec)
+  // <=> add v4i64 %acc, %vec
+  if (Ratio == 1) {
----------------
sushgokh wrote:

rather than checking for ratio=1,  you should either replace
1. https://github.com/llvm/llvm-project/blob/main/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp#L323 with getArithmeticInstructionCost (I believe this line is simply checking for the simple add reduction)

OR 

2. https://github.com/llvm/llvm-project/blob/main/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp#L3552 with getAirthmeticInstructionCost

https://github.com/llvm/llvm-project/pull/166138