[llvm] [AMDGPU] Add dynamic threshold for DPP atomic optimizer on integer LDS atomics (PR #186762)

[YaFan Tao via llvm-commits]

================

----------------
taoyafan wrote:

Thanks for the feedback. I agree on removing the cl::opt, and I have done that.

I kept a function attribute because this threshold is not stable enough to hardcode: the best value depends on other optimizations (for example loop unrolling and memory clause formation), and can vary by function. We also need a way to fully disable the DPP path for integer LDS atomics in specific cases; setting the threshold to a value >= wavefront size (for example 64 on wave64) provides that control.

Per your comment, I have also added documentation for this attribute in AMDGPUUsage.rst.

https://github.com/llvm/llvm-project/pull/186762