| Issue |
134569
|
| Summary |
[DirectX] The `dx.dot2/3/4` and `dx.dot2add` intrinsics leave vectors around post-scalarization
|
| Labels |
backend:DirectX
|
| Assignees |
bogner
|
| Reporter |
bogner
|
The `dx.dot2`, `dot3`, `dot4`, and `dot2add` intrinsics are defined as taking vector arguments and have special handling to scalarize those arguments in `DXILOpLowering`. This causes issues, as the vectors don't get entirely cleaned up before we generate DXIL.
>From https://github.com/llvm-beanz/offload-test-suite/pull/61:
```
# .---command stdout------------
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%17 = insertelement <3 x float> undef, float %16, i32 0' in block '#0' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%18 = insertelement <3 x float> %17, float %15, i32 1' in block '#0' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%19 = insertelement <3 x float> %18, float %14, i32 2' in block '#0' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%20 = extractelement <3 x float> %19, i32 0' in block '#0' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%21 = extractelement <3 x float> %19, i32 1' in block '#0' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%22 = extractelement <3 x float> %19, i32 2' in block '#0' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%23 = extractelement <3 x float> %19, i32 0' in block '#0' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%24 = extractelement <3 x float> %19, i32 1' in block '#0' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%25 = extractelement <3 x float> %19, i32 2' in block '#0' of function 'main'.
# | Validation failed.
```
It would be simpler to make these intrinsics match their DXIL ops more closely and simply take scalar arguments. This will allow the scalarizer to do its job earlier in the pipeline.
Alternatively, we could improve the vectorArgExpansion logic in DXILOpLowering to clean up any remaining vector artifacts, but this seems like it would be redundant with the logic in the scalarizer.
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
