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

Tue Jul 25 16:55:38 PDT 2023

alexfh added a comment.

In D150388#4529498 <https://reviews.llvm.org/D150388#4529498>, @arsenm wrote:

> In D150388#4529445 <https://reviews.llvm.org/D150388#4529445>, @arsenm wrote:
>
>> In D150388#4529268 <https://reviews.llvm.org/D150388#4529268>, @vitalybuka wrote:
>>
>>> So if we have not fix, just discussion, should we revert? Even if it's preexisted unknown bug, I don't think it's appropriate to keep regression in existing functionality, for sake of fixing previously broken or missing functionally. 
>>> Is it possible to workaround/revert some change with "if (x86) do old way" ?
>>
>> Try https://reviews.llvm.org/D156164 as a workaround. The high-bits-must-be-zero invariant disappears somewhere in the coalescer, this just hacks the resulting broken mov such that it hides the problem as it was before
>
> Much better, probably real fix incoming

There is still neither a fix nor even a workaround in the tree. I think, the right solution in such cases is reverting to the last known good state and investigating asynchronously. I'm not saying this would be the best approach right now, given there is a workaround in review. But it would have been the right choice almost two weeks ago, when the issue was reported.

Can we have at least the workaround submitted?

Repository:
  rG LLVM Github Monorepo

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

https://reviews.llvm.org/D150388