[llvm] [JumpThreading] Don't thread into blocks with convergent operations (PR #213870)

[Antonio Frighetto via llvm-commits]

https://github.com/antoniofrighetto requested changes to this pull request.

The pass should never run on NVPTX, where hasBranchDivergence() is defaulted to true:

https://github.com/llvm/llvm-project/blob/705aa2a864d1573a17e62cfbcffbafdab7745a6d/llvm/lib/Transforms/Scalar/JumpThreading.cpp#L240-L245

Are we correctly targeting NVPTX (`-mtriple=nvptx64-nvidia-cuda` to opt) in the Triton compilation pipeline (thus ensuring TTI is set up correctly)?

https://github.com/llvm/llvm-project/pull/213870