[llvm] [AMDGPU] Allocate i1 argument to SGPRs (PR #72461)

[Jun Wang via llvm-commits]

jwanggit86 wrote:

The latest solution is quite different from the previous one. For calling convention, a custom function is created for i1 arg/return, which allocates i1 to either a 32b SGPR for wavesize32 or a 64b SGPR for wavesize64.

Two additional changes are made: (1) in LowerFormalArgument(), previously it's assumed the all registers are from a 32b regclass. Now it's possible that it is from a 64b regclass. (2) In lowerCopiesToI1(), for i1 return value, previously it is assumed it's only from a 32b register. Now it can be from a 64b register.

https://github.com/llvm/llvm-project/pull/72461