[llvm] [AMDGPU] Add option to bias SGPR allocation to reduce read hazards (PR #129869)

[Carl Ritson via llvm-commits]

perlfu wrote:

> Why does this need to be a dynamic property built on top of hints? We can statically change the allocation order in tablegen

Only the reversed allocation order could be generated as the others are based on the actual intersection of VALU and SALU register usage.
We could generate that order and switch all GFX12.
It has the side effect that all waves require maximum SGPR allocation, although this is not an issue for RDNA* where these are always available.


https://github.com/llvm/llvm-project/pull/129869