[Mlir-commits] [mlir] [mlir][VectorToLLVM] emit `inbounds|nuw` GEP flags when lowering `vector.load/store` (PR #202118)

[Fabian Mora]

fabianmcg wrote:

> So, as far as I'm concerned, we should add `inbounds` and `nneg` flags to `vector.load`, and not emit inbounds or nusw by default...

AFAIK, the LLVM lang ref is sort of vague about this, but reading out of bounds is UB -the Lang Ref should get some clarifications though.  Therefore, AFAIK the only way that AMDGPU buffer address space is well-defined, is if it's considered that everything is in bounds at the LLVM level, then the HW using the metadata can return 0 if outside of the buffer bounds. Otherwise, the result should be poison and not 0.

https://github.com/llvm/llvm-project/pull/202118