[llvm] AMDGPU: Allow operand folding between loop body and its preheader (PR #137022)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> I thought this was in the category of stuff we can't really do because we've conflated exec mask mutation with the LLVM CFG. When is this optimisation sound?

Yes, we don't have a proper representation and just have to hack around it. In this case, since the readfirstlanes are explicit in the original IR, it would be better served by #128687

https://github.com/llvm/llvm-project/pull/137022