[llvm] AMDGPU: Reduce readfirstlane for single demanded vector element (PR #128647)

[Jay Foad via llvm-commits]

jayfoad wrote:

> > Looks OK, but this handling applies _way_ more generally than just readfirstlane (or just amdgcbn lane intrinsics). It applies to any intrinsic that operates elementwise on vectors. Isn't there a generic helper for that somewhere?
> 
> Yes, this should be expanded to cover all of those cases (we should do the same for demanded bits). I was going to leave this as the first sample one, and open an issue to expand to the other cases

Should be able to do it in generic code for any `isTriviallyScalarizable` intrinsic. Then AMDGPU would just need to implement `isTargetIntrinsicTriviallyScalarizable`.

https://github.com/llvm/llvm-project/pull/128647