[PATCH] D128800: [AMDGPU] Fix liveness for loops in si-optimize-exec-masking-pre-ra

Wed Jun 29 20:26:10 PDT 2022

critson marked an inline comment as done.
critson added a comment.

In D128800#3618857 <https://reviews.llvm.org/D128800#3618857>, @arsenm wrote:

> I think D128315 <https://reviews.llvm.org/D128315> D128110 <https://reviews.llvm.org/D128110> would also avoid this. I assumed cndmask in a different block was not useful, but presumably you found this in the wild?

Thank you for the quick review!

This was found in a Vulkan CTS test on GFX8 and GFX9.
The result was register reallocation within a loop that caused incorrect results (as s_andn2 source was overwritten).

D128315 <https://reviews.llvm.org/D128315> would fix this issue, but perhaps this change removes the need for D128315 <https://reviews.llvm.org/D128315>?
I am not sure D128110 <https://reviews.llvm.org/D128110> would fix this bug, but there is some overlap between the two changes.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D128800/new/

https://reviews.llvm.org/D128800