[Mlir-commits] [mlir] [mlir][amdgpu] Introduce `assume_subgroup_uniform` op (PR #152740)

[Matt Arsenault]

https://github.com/arsenm commented:

This is something that should be handled in LLVM IR, not MLIR. This is also path dependent, so should look more like an assume call than something propagating a value. If you're just going to insert readfirstlane you can just directly insert readfirstlane (i.e. I don't see what a wrapper around it does for you) 

https://github.com/llvm/llvm-project/pull/152740