[PATCH] D117075: [AMDGPU][WQM] Set the lanes that should be deactivated with VCC.

[Carl Ritson via Phabricator via llvm-commits]

critson added a comment.

Thanks for finding and fixing this bug.
In principle this change looks fine, but could you add a description of what was changed/fixed.
Perhaps something along the lines of:
WQM pass incorrectly uses VCC for vector comparisons in Wave32 mode. Make register selection wave size dependent.



================
Comment at: llvm/test/CodeGen/AMDGPU/WQM_set_killed_lanes.ll:1
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mcpu=gfx1030 < %s 2>&1 | FileCheck --check-prefix=CHECK %s
----------------
This test can probably be further simplified, and the filename should be all lowercase to match existing naming convention.
Also I think this is covered by the existing test code, but never run on GFX10 or with Wave32.
This has made sense so far as we usually compile pixel shaders Wave64, which I guess is why we did not see this bug sooner.
Perhaps we should simply be extending the existing wqm.ll to test Wave32?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D117075/new/

https://reviews.llvm.org/D117075