[llvm] [AMDGPU] Shrink to SOPK with 32-bit signed literals (PR #70263)

[Stanislav Mekhanoshin via llvm-commits]

rampitec wrote:

> I'm a bit surprised it doesn't affect any codegen tests.
> 
> Anyway makes sense to me since the src operand of S_MOVK_I32 is signed but the src operand of S_MOV_B32 is not (B32 is neither signed nor unsigned, but I guess we generally treat it like unsigned).

I was surprised too, but this is all because we do not select 64-bit moves. All of that is a step to select these.

https://github.com/llvm/llvm-project/pull/70263