[llvm] [AMDGPU] Don't cluster lds instrs together (PR #180908)

[via llvm-commits]

lijinpei-amd wrote:

> Did you mean to close this?

Yes. For 2 reasons:

1. there could be 2 cycles penalty between valu and ds instrs on CDNA.
2. I have 3 commits that I am working on, just found I can get similar performance without this one.
see triton internal issue for details: https://github.com/ROCm/triton-internal/issues/1641

https://github.com/llvm/llvm-project/pull/180908