[llvm] [SLP] Match poison as instruction with the same opcode (PR #115946)

Alexey Bataev via llvm-commits llvm-commits at lists.llvm.org
Mon Jan 13 09:49:36 PST 2025


alexey-bataev wrote:

> This commit (...somehow...) broke Triton/PyTorch on CUDA (see [triton-lang/triton#5591](https://github.com/triton-lang/triton/pull/5591) for a description).
> 
> I have attached the failing and passing LLIR:
> 
> [failing_triton_poi_fused_33.llir.txt](https://github.com/user-attachments/files/18391665/failing_triton_poi_fused_33.llir.txt)
> 
> [passing_triton_poi_fused_33.llir.txt](https://github.com/user-attachments/files/18391661/passing_triton_poi_fused_33.llir.txt)
> 
> @alexey-bataev any guidance would be appreciated (I am not familiar with the SLP vectorizer....).



> This commit (...somehow...) broke Triton/PyTorch on CUDA (see [triton-lang/triton#5591](https://github.com/triton-lang/triton/pull/5591) for a description).
> 
> I have attached the failing and passing LLIR:
> 
> [failing_triton_poi_fused_33.llir.txt](https://github.com/user-attachments/files/18391665/failing_triton_poi_fused_33.llir.txt)
> 
> [passing_triton_poi_fused_33.llir.txt](https://github.com/user-attachments/files/18391661/passing_triton_poi_fused_33.llir.txt)
> 
> @alexey-bataev any guidance would be appreciated (I am not familiar with the SLP vectorizer....).

Try to check the latest head. I don't quite understand what is the problem. What shall I do to reproduce the issue?

https://github.com/llvm/llvm-project/pull/115946


More information about the llvm-commits mailing list