[llvm] [AMDGPU] SIWholeQuadMode: avoid execz effects in exact regions (PR #101157)
Carl Ritson via llvm-commits
llvm-commits at lists.llvm.org
Wed Jul 31 01:57:02 PDT 2024
perlfu wrote:
- Rebased on to pre-commit tests
- Tighten the condition for shortening exact regions - makes change more targeted and thus has essentially no impact on any existing shaders unimpacted by this issue
In the main we might possibly have some more missed EXECZ regions, but they can only be from WQM to exact transitions.
Strict WWM enables extra lanes for short sections, so will never be EXECZ.
Strict WQM can only be EXECZ if it was entered into from an existing EXECZ region, so in principle this might be possible.
This would have to be full shader WQM -> exact -> strict WQM -- which is low probability.
Again I think this is probably best addressed after control flow lowering is reworked (i.e. wave transform).
https://github.com/llvm/llvm-project/pull/101157
More information about the llvm-commits
mailing list