[llvm] [VectorCombine] Fold binop(shuffle(V1, Mask), shuffle(V2, Mask)) -> shuffle(binop(V1, V2), Mask) (PR #211530)

[Yingwei Zheng via llvm-commits]

================
@@ -2443,6 +2444,70 @@ bool VectorCombine::foldConcatOfBoolMasks(Instruction &I) {
   return true;
 }
 
+// Fold binop(shuffle(V0, Mask), shuffle(V1, Mask))
+// --> shuffle(binop(V0, V1), Mask)
+bool VectorCombine::foldBinopOfShuffles(Instruction &I) {
+  // It is not safe to transform things like div, urem, etc. because we may
+  // create undefined behavior when executing those on unknown vector elements.
+  if (!isSafeToSpeculativelyExecute(&I))
----------------
dtcxzyw wrote:

Looks like it doesn't guard against FP ops with strictfp, right? This transformation may introduce new floating-point exceptions.


https://github.com/llvm/llvm-project/pull/211530