[PATCH] D150388: [CodeGen]Allow targets to use target specific COPY instructions for live range splitting

Sun Jul 16 12:01:28 PDT 2023

cdevadas added a comment.

In D150388#4504380 <https://reviews.llvm.org/D150388#4504380>, @vitalybuka wrote:

> In D150388#4502779 <https://reviews.llvm.org/D150388#4502779>, @yassingh wrote:
>
>> In D150388#4501339 <https://reviews.llvm.org/D150388#4501339>, @asmok-g wrote:
>>
>>> Headsup: This change might be causing a miscompile. We're having some breakages in google because of this change that goes away when we add `__attribute__((optnone)) ` to https://github.com/llvm/llvm-project/blob/main/clang/include/clang/Tooling/Transformer/SourceCode.h#L111. I'm working on confirming it.
>>
>> As Matt suggested, "The most likely source of any issue is the broader recognition of copy-like operations". It's hard to say anything without looking at your internal implementation of isCopyIstrImpl(). For eg MIPS implementation of the same function required some tweaking before this patch was pushed as it wasn't accurately describing COPY like instructions. D151181 <https://reviews.llvm.org/D151181>
>
> It's still regression. Unless someone has a patch ready to land, we need to revert this.

This is part of a chain of patches that fixes an AMDGPU backend issue. D124196 <https://reviews.llvm.org/D124196> being the final one in the chain, the revert won't be that easy.
I'm trying to see if we can avoid the revert.  Can you attach a reproducible IR code?

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D150388/new/

https://reviews.llvm.org/D150388