| Issue |
97554
|
| Summary |
AMDGPUAtomicOptimizer: re-enable uniform path for fadd/fsub with result
|
| Labels |
backend:AMDGPU
|
| Assignees |
|
| Reporter |
jayfoad
|
See #96479: the uniform path for fadd/fsub was disabled because it can return the wrong value in the first active lane of the result in the presence of NaNs and signed zeros.
We should find a way to fix and re-enable it, e.g. by overwriting the first active lane of `%y * +0.0` with -0.0. And maybe this can be optimized more if certain values are known not to be zero or nan.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
