[PATCH] D98218: [LSR] fix a issue that LoopStrengthReduction drop debug location unnecessarily

[Jeremy Morse via Phabricator via llvm-commits]

jmorse added a comment.

Happily we can consult https://llvm.org/docs/HowToUpdateDebugInfo.html -- the example that you've given looks like's a good test for the "positive" case, where we should keep the source location attached to the compare instruction. The comparison doesn't get moved between blocks, so we match "When to preserve an instruction location".

Ideally we should have a test for the negative case, i.e. a scenario where the source location _should_ be dropped. That would happen if we moved the comparison between blocks, across a conditional branch. That's what I was originally concerned about earlier... however after some study I'm less certain now that this can ever occur with loop strength reduction. I put some assertions into LoopStrengthReduce.cpp to look for those scenarios happening and did a stage2 build of clang, and it never detected any. Plus, LoopStrengthReduction only operates on loops that have been transformed to only have one backedge. Thus, I reckon we don't need a negative test here.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D98218/new/

https://reviews.llvm.org/D98218