[llvm] [AMDGPU] Skip the iterative atomic scan on native LDS atomics (PR #227222)

[Yaxun Liu via llvm-commits]

yxsamliu wrote:

The code change looks reasonable to me. Could you share timings for `threadfence-hip` and `aop-hip` comparing main with this patch? The numbers above compare main with the whole atomic optimizer disabled.

Could you also share a small synthetic benchmark comparing main with this patch at a few different active lane counts, from one lane to a full wave? That would help confirm the answer to Perlfu’s question.

https://github.com/llvm/llvm-project/pull/227222