[clang] [llvm] [AMDGPU][clang][CodeGen][opt] Add late-resolved feature identifying predicates (PR #134016)

[Jon Chesterfield via cfe-commits]

JonChesterfield wrote:

I strongly oppose the magic globals approach despite rocm being enthusiastic users of it. I tracked down some of the authors of that to ask why there are IR files containing a single constant being linked in using llvm-link in the middle of clang and the best answer I got was that it seemed easier than creating the constant directly, which is nonsense.

I want this feature. If we let it in with the magic globals that rocm loves, they'll stay forever, and people will fight us when we try to replace them with intrinsics because it's different to rocm, so in practice that hack lives forever and ever. The commit gate is the highest leverage point we have to not adopt that mistake.

https://github.com/llvm/llvm-project/pull/134016