[llvm] [AMDGPU] Prefer vector i8s in PHI Nodes (PR #91016)

[Matt Arsenault via llvm-commits]

https://github.com/arsenm commented:

I'm not sure why we need a dedicated phi function. It should essentially be the type legalization cost for any value. Phis aren't a real operation, they are just split to legal registers 

https://github.com/llvm/llvm-project/pull/91016