[llvm] [WIP][AMDGPU] Improve the handling of `inreg` arguments (PR #133614)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> It doesn't seem like this optimization can be used at a `tail call` call site, thus we will need to check if a function is only called in a regular call site. This is not something that can be done at function lowering stage.

I don't see why not? The argument list is the same either way 

https://github.com/llvm/llvm-project/pull/133614