[llvm] [AMDGPU] Illegal VGPR to SGPR copy (PR #144008)

[Matt Arsenault via llvm-commits]

https://github.com/arsenm requested changes to this pull request.

This is just hacking around the issue at the failure point. You cannot meaningful do anything with a physical register copy. This needs to be a semantic, context dependent choice at the point the copy originated. e.g. if this was an SGPR call argument, we either needed to insert the readfirstlane or do a waterfall loop 

https://github.com/llvm/llvm-project/pull/144008