[llvm] AMDGPU/GFX12: Do not wait unnecessarily before barriers (PR #154970)

[Nicolai Hähnle via llvm-commits]

nhaehnle wrote:

> We could get the same effect by saying GFX12 has FeatureBackOffBarrier, right? But perhaps your way is more honest.

Yeah, that was pretty much my thinking. The split barriers in GFX12 eliminate the need to "back off" from the barrier.

> I have a slight concern that this could trigger the same kind of problems we have seen on GFX10 and GFX11, which we worked around by removing FeatureBackOffBarrier from them. But if not then this seems fine to me.

Yes, there is a risk with this change that there is code out there which doesn't have the correct fences ("memory barriers") around these (control) barriers.

We've had people asking about this though because it is a performance problem for correctly written code.

https://github.com/llvm/llvm-project/pull/154970