[llvm] [AMDGPU] Add register allocation handoff intrinsic (PR #219842)

[Bangtian Liu via llvm-commits]

bangtianliu wrote:

Closing this draft after investigating the motivating kernel more deeply. The repro here https://github.com/llvm/llvm-project/issues/219377#issuecomment-5510810499 shows that no IR-level handoff or hard VGPR constraint is required;
the difference is localized to pre-RA scheduling/register-pressure decisions. Therefore, the proposed intrinsic and general NoMerge machinery are not justified by this case.

Thanks @qcolombet, @arsenm and @lialan. 


https://github.com/llvm/llvm-project/pull/219842