[llvm] [AMDGPU] Insert readfirstlane in the function returns in sgpr. (PR #135326)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> > I reverted it.
> 
> Thanks, but why does it fail after merging? I can see all the checks passed before the merge.

The target constants values will shuffle around depending on what targets are built (and on any future intrinsic change). You need to make the check fuzzier to avoid depending on the exact value 

https://github.com/llvm/llvm-project/pull/135326