[PATCH] D143593: [InstCombine] Don't fold freeze poison when it's used in shufflevector

[Manuel Brito via Phabricator via llvm-commits]

ManuelJBrito added a comment.

Thanks for the feedback !!



================
Comment at: llvm/test/Transforms/InstCombine/shufflevector_freezepoison.ll:3
+
+define <4 x double> @shuffle_op1_freeze_poison(<2 x double> %a) {
+; CHECK-LABEL: @shuffle_op1_freeze_poison(
----------------
spatel wrote:
> This isn't what I was expecting. If we're bailing out for operand 1, then shouldn't we be symmetric and do the same for operand 0 (just in case that logically equivalent pattern was somehow created in the front-end)?
I matched only on op 1, because that's the pattern that is recognized by the backend.
As of right now the symmetric case isn't lower in the same manner : https://godbolt.org/z/K5EWfdbq8, but yes i do think bailing out on both operands makes the most sense.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D143593/new/

https://reviews.llvm.org/D143593