[PATCH] D119975: Revert "[PassManager] Run additional LICM before LoopRotate"

Thu Feb 17 00:34:08 PST 2022

nikic added a comment.

> But, apparently, some other improvements to other optimizations passes
> have also solved the problem, because, as the diff shows, reverting D99249 <https://reviews.llvm.org/D99249>
> does not resurface the original problem, since, clearly, the check lines
> for the new-pm and -O2/-O3 are not affected.

Do you have any idea what could have caused this? I find it particularly odd that this is solved for the NewPM but not the OldPM (I don't think we added any pipeline divergences recently...)

================
Comment at: llvm/test/Other/new-pm-defaults.ll:170
 ; CHECK-O-NEXT: Running pass: LoopSimplifyCFGPass
-; CHECK-O-NEXT: Running pass: LICM
+; CHECK-O23SZ-NEXT: Running pass: LICM
 ; CHECK-O-NEXT: Running pass: LoopRotatePass
----------------
A bit confused here: Why does this become O23SZ rather than being dropped entirely?

================
Comment at: llvm/test/Transforms/PhaseOrdering/X86/speculation-vs-tbaa.ll:12

 ; We should retain the TBAA on the load here, not lose it.

----------------
Not strictly necessary, but it would be preferable if this test showed an actual optimization impact (as alluded in https://github.com/llvm/llvm-project/issues/53794#issuecomment-1037460605) rather than just the fact that TBAA is preserved (which is best-effort).

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D119975/new/

https://reviews.llvm.org/D119975