Upstream LLVM SLP vectorizer change requires the correct triple

[pclove1]

Describe the issue

We recently encountered a performance regression in one of our triton kernels and identified the culprit as the upstream LLVM SLP vectorization pass change (llvm/llvm-project@88e7b8b / llvm/llvm-project#125725).

So, I filed llvm/llvm-project#127244 but got a response from the original author (@alexey-bataev) saying that the correct triple is required to get it vectorized properly now.

Digging triton code suggests that triton doesn't pass the target machine to SLP vectorization on purpose.

Any chance/plan to address TODO item in the linked code? (i.e. pass NVPTX target)

Environment details

GPU: Nvidia A100