| Issue |
139250
|
| Summary |
AMDGPU fmed3 of fpext optimization should look through source modifiers
|
| Labels |
backend:AMDGPU,
missed-optimization
|
| Assignees |
|
| Reporter |
arsenm
|
The combine to fold `fmed3.f32(fpext x, fpext y, fpext z) ` [here](https://github.com/llvm/llvm-project/blob/b291cfcad4815568dc1eaca58185d25dceed3f1c/llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp#L919) should account for intermediate source modifiers.
e.g. it should handle
`fmed3.f32(fneg (fpext x)), fabs (fpext y), fneg (fabs (fpext z))`
Additionally this should preserve the fast math flags on the new fpext
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
