[llvm] [AMDGPU] Fix code sequence for barrier start in GFX10+ CU Mode (PR #160501)

[Nicolai Hähnle via llvm-commits]

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

I don't think we should ever unconditionally insert waits purely based on a barrier. This pessimizes LDS-only barriers too much.

Quite frankly, I'd rather admit that trying to reduce CU-mode release fences to only VM_VSRC waits was a mistake, and go back to using the same waits there as for WGP-mode.

https://github.com/llvm/llvm-project/pull/160501