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

[Yaxun Liu via cfe-commits]

yxsamliu wrote:

@arsenm Re: preload offset and SGPR budget

I initially thought offset was unsupported because the existing LLVM path always preloaded from offset zero, including the compatibility prologue. After checking the runtime code, I found that the descriptor offset is already supported. Only the compiler compatibility prologue needed a small fix, which is now included.

I also ran E2E tests and measurements on gfx950. They confirmed that the offset works and that both the offset and SGPR cap can affect performance. I changed the attribute to select an argument range so users can focus the limited preload SGPR budget on performance-sensitive arguments and avoid preloading unused or less useful ones.

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