[llvm] [AMDGPU] Split the true16 fcopysign pattern by uniformity (PR #226898)

[Guo Chen via llvm-commits]

https://github.com/broxigarchen commented:

As Jay mentioned the uniform 16bit value could still be in vgpr16. i.e. an i16 load. Could you check if there are still bad copy/reg_sequence generated from isel?

FYI I am also working on another version of fix in this stack pr https://github.com/llvm/llvm-project/pull/220275 that trying to remove all bad copy/reg_sequence regarding this reg size issue

https://github.com/llvm/llvm-project/pull/226898