[Mlir-commits] [mlir] [mlir][nvgpu] update commit group and wait async ops (PR #130482)

[Guray Ozen]

grypp wrote:

> They should be reordered yes. 

Typo: they should not be reordered.

To prevent them from being reordered, we need to make them side-effecting. Otherwise, MLIR—or any downstream compiler—is free to reorder them, which is not what we want.

That said, making them side-effecting is a big hammer. Preserving SSA semantics is, in my opinion, a less intrusive approach.

Your PR makes sense if we didn’t have SSA. I can totally understand the reasoning if these ops are lowered directly to PTX, where there are no tokens. However, PTX isn’t SSA anyway, so I’m not sure it’s the best comparison.

If you look at vectors or structs at the NVGPU level, you’ll see the same thing: SSA semantics are preserved.

https://github.com/llvm/llvm-project/pull/130482