[llvm] [AMDGPU] Merge 32-bit VGPR copies into 64-bit copies (PR #193270)

[Jay Foad via llvm-commits]

jayfoad wrote:

> I think this is the wrong solution for the vast majority of cases. We should be emitting the wider operations up front, not reassembling them later

I'm ambivalent about this. I agree that it would be better to try to create wider operations up-front, and then see if this new pass is still required. On the other hand this pass does seem to be giving a nice tangible benefit, so it is hard to argue against it.

https://github.com/llvm/llvm-project/pull/193270