(In reply to Marek Olšák from comment #13)
> (In reply to Tom Stellard from comment #12)
> > 
> > I think the best solution here would be to teach the backend how to do a
> > scalar select based on value of vccz.
> 
> You are missing the point. The condition code (VCC) isn't the same across
> all threads. The problem is that the conditional assignment is transformed
> into a form that makes descriptor load addresses non-uniform (dependent on a
> VGPR). There is nothing the AMDGPU backend can do about it. It's not a
> problem in the backend.

Well the backend could iterate over all the variants to handle that, but I have
strong doubts that this would result in optimal performance.

So I think the best solution for now would be to mark the texture intrinsic in
a way which disallows such optimizations.