[llvm] [AMDGPU] Prioritize allocation of low 256 VGPR classes (PR #167978)

[Stanislav Mekhanoshin via llvm-commits]

rampitec wrote:

Even this is not a magic bullet. I still can seduce RA to spill in the pathological very big case. Still correct, but will spill on lo256 operands. I guess this is the nature of having limited VGPR classes and inability to encode all operands to use all possible registers.

https://github.com/llvm/llvm-project/pull/167978