[PATCH] D139275: [SimplifyCFG] `FoldBranchToCommonDest()`: deal with mismatched IV's in PHI's in common successor block

[Roman Lebedev via Phabricator via llvm-commits]

lebedev.ri added inline comments.


================
Comment at: llvm/test/Transforms/SimplifyCFG/X86/SpeculativeExec.ll:15
+; CHECK-NEXT:    [[T3:%.*]] = add i32 [[A]], 1
+; CHECK-NEXT:    [[T4:%.*]] = select i1 [[OR_COND]], i32 [[T3]], i32 [[T4_SEL]]
 ; CHECK-NEXT:    [[T5:%.*]] = sub i32 [[T4]], 1
----------------
nikic wrote:
> nikic wrote:
> > Looks like this test, and many similar tests regress, where we go from two selects to three selects. And we never seem to recover from this in either the middle end or the back end: https://alive2.llvm.org/ce/z/GjCXkB
> Filed https://github.com/llvm/llvm-project/issues/59393. We should at least add an InstCombine fold, but ideally we'd avoid regressing this in SimplifyCFG.
I'll deal with the InstCombine side of this.
I agree that this is a regression, that obviously affects further speculation,
but i'm not sure how much we can avoid it during speculation.
I can take a look, but not in this patch,
so i hope this patch isn't going to be blocked on that.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D139275/new/

https://reviews.llvm.org/D139275