[PATCH] D93065: [InstCombine] Disable optimizations of select instructions that causes propagation of poison values

Tue Jan 19 18:23:07 PST 2021

aqjune added a comment.

In D93065#2508119 <https://reviews.llvm.org/D93065#2508119>, @nikic wrote:

> Current diff for flipping the flag: https://gist.github.com/nikic/f65b36adb70c93f5da9bfe3422fd8904 There's still a number of cases that can be handled.

I think I can handle or_andn_cmp_1_logical  and its family, llvm.umul.with.overflow.i64 related transformations, and PR41069_logical.

> I also noticed that SimplifyCFG is creating and/or when merging branch conditions in https://github.com/llvm/llvm-project/blob/ecf696641e6ce4b22e8c8ea3c7476b9c1f0f200b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L2959. I think it would make sense to switch that code over to use the select form first. That will make the transform correct in SimplifyCFG, but still let it be folded by InstCombine. As other passes have already been adjusted to recognize logical and/or, this should be low impact.

+1, this very makes sense to me; InstCombine happens immediately after SimplifyCFG in many places.
I believe non-InstCombine passes coming after SimplifyCFG are already familiar with unoptimized select instructions as well because SimplifyCFG is already generating such ops (see test6f <https://web.ist.utl.pt/nuno.lopes/alive2/index.php?hash=a480a669a361f91d&test=Transforms%2FSimplifyCFG%2Fbasictest.ll>)
I'll make a patch for this.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D93065/new/

https://reviews.llvm.org/D93065