Tests for this will be added once the AMDGPU backend enables this
option.
Details
Details
Diff Detail
Diff Detail
- Repository
- rL LLVM
Event Timeline
lib/Transforms/Scalar/StructurizeCFG.cpp | ||
---|---|---|
264 ↗ | (On Diff #46050) | This should not be a required analysis. Even with the false option, it will force running the analysis. I think if the target pass config explicitly adds the analysis pass before StructurizeCFG it should work |
Comment Actions
Do we need to worry about updating the DivergenceAnalysis as the control flow is changed in the non-uniform regions? Sometimes it introduces new xor instructions which won't be in DA's value map.
Comment Actions
I don't think this would be a problem, because we don't use the DivergenceAnalysis any more, once we have started to structurize a region.
lib/Transforms/Scalar/StructurizeCFG.cpp | ||
---|---|---|
264 ↗ | (On Diff #47285) | Explicitly adding DivergenceAnalysis in TargetPassConfig doesn't seem to work, so I've made DivergenceAnalysis required only when SkipUniformRegions is true. |