[clang] [llvm] [AMDGPU] Select a kernel argument preload range (PR #211059)

[Yaxun Liu via llvm-commits]

yxsamliu wrote:

@arsenm @Pierre-vh Gentle ping. CI is green now and all inline comments are addressed.

The current version uses an offset plus SGPR-count limit, so it selects a contiguous range of kernel arguments to preload. This roughly matches Matt’s suggestion to choose which arguments are preloaded, while still fitting what the runtime descriptor and backend can express today.

Does this design look acceptable, or is there anything else you would like changed before review?

https://github.com/llvm/llvm-project/pull/211059