[PATCH] D136675: [MachineCSE] Allow PRE of instructions that read physical registers

[Vignesh Venkatasubramanian via Phabricator via llvm-commits]

vigneshv added a comment.

In D136675#4217132 <https://reviews.llvm.org/D136675#4217132>, @john.brawn wrote:

> In D136675#4213750 <https://reviews.llvm.org/D136675#4213750>, @vigneshv wrote:
>
>> Please let me know if you need any more details. :)
>
> I quickly tried cross-building that example for a bare-metal environment (because it's something I have the setup to do easily) and the output produced by the executable was identical in current trunk and with this patch reverted. So:
>
> - I'm guessing this failure is happening in some kind of android environment?
> - A description of what's going wrong that's more precise than "doing some of the computations incorrectly" would be helpful
> - A smaller reproducer that doesn't depend on the host environment (e.g. preprocessed with -E) would also be helpful
>
> Also: I'm on holiday tomorrow and next week, so won't be able to look at this until after that.

Thanks for the reply @john.brawn. Turns out it was a bug in libyuv's assembly code and how it was being optimized with -flto.

My bisect may just have been accidentally triggering the edge case in this commit. Sorry about that.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D136675/new/

https://reviews.llvm.org/D136675