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

[Jay Foad via llvm-commits]

jayfoad wrote:

> > This sounds like a bad idea. How do you know it's safe to clobber lane 0 of some VGPR, when lane 0 might not be active? Would you introduce a whole wave spill of the VGPR?
> 
> Yes, this needs to use WWM values

OK then it is not at all clear that this is a desirable change to the ABI. It seems like it could make the code perform worse in a lot of cases.

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