[clang] [llvm] [AMDGPU] Add per-kernel kernarg preload count (PR #211059)

[Yaxun Liu via cfe-commits]

yxsamliu wrote:

@kerbowa Re: SGPR budget

Thanks. I changed the attribute to `amdgpu_kernarg_preload_sgpr_count(N)`. It now caps the total SGPRs used by the contiguous preload range, including padding and hidden arguments. The backend may use fewer SGPRs when needed.

I left out the offset because the current backend only preloads from offset zero. We can add it later if the backend supports other offsets.

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