[llvm] Extend vector.reduce.add and splat transform to scalable vectors (PR #161101)

[Gábor Spaits via llvm-commits]

================
@@ -3785,13 +3785,19 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
 
       // vector.reduce.add.vNiM(splat(%x)) -> mul(%x, N)
       if (Value *Splat = getSplatValue(Arg)) {
-        ElementCount VecToReduceCount =
-            cast<VectorType>(Arg->getType())->getElementCount();
+        VectorType *VecToReduceTy = cast<VectorType>(Arg->getType());
+        ElementCount VecToReduceCount = VecToReduceTy->getElementCount();
+        Value *RHS;
         if (VecToReduceCount.isFixed()) {
           unsigned VectorSize = VecToReduceCount.getFixedValue();
-          return BinaryOperator::CreateMul(
-              Splat, ConstantInt::get(Splat->getType(), VectorSize));
+          RHS = ConstantInt::get(Splat->getType(), VectorSize);
----------------
spaits wrote:

I see. Inetersting that it did not change the exepcted output for the other cases (I mean the disregarding of the result coming from this branch). Probably later CreateElementCount returns wit a constant and that is propagated (or folded? I am not 100% sure that is the right word to use here) to the mul.

https://github.com/llvm/llvm-project/pull/161101