[llvm] [AMDGPU] Filter candidates of LiveRegOptimizer for profitable cases (PR #124624)

[via llvm-commits]

choikwa wrote:

> > is beneficial for operations that natively work on illegal vector types to prevent flip-flopping between SGPR and VGPR.
> 
> This isn't how SGPR vs VGPR works. It's one direction only, you can't just copy to SGPR from VGPR. What example are you talking about?

Apologies. corrected to 'unpacked and packed'

https://github.com/llvm/llvm-project/pull/124624