[llvm] [InstCombine] Fold vector.reduce.op(vector.reverse(X)) -> vector.reduce.op(X) (PR #91743)

[Paul Walker via llvm-commits]

================
@@ -1435,6 +1435,51 @@ static Instruction *foldBitOrderCrossLogicOp(Value *V,
   return nullptr;
 }
 
+Instruction *InstCombinerImpl::simplifyReductionOfShuffle(IntrinsicInst *II) {
+  Intrinsic::ID IID = II->getIntrinsicID();
+  bool CanBeReassociated = (IID != Intrinsic::vector_reduce_fadd &&
+                            IID != Intrinsic::vector_reduce_fmul) ||
+                           II->hasAllowReassoc();
+
+  if (!CanBeReassociated)
+    return nullptr;
+
+  const unsigned ArgIdx = (IID == Intrinsic::vector_reduce_fadd ||
+                           IID == Intrinsic::vector_reduce_fmul)
+                              ? 1
+                              : 0;
+  Value *Arg = II->getArgOperand(ArgIdx);
+  Value *V;
+
----------------
paulwalker-arm wrote:

That's not quite what I meant.  I was more thinking `simplifyReductionOperand` would take the reduction's operand rather than the reduction itself.  This way the reduction specific parts remain at the call site.

https://github.com/llvm/llvm-project/pull/91743