(In reply to Nicolai H�hnle from comment #16)
> 
> One annoying aspect of this language is that one can reasonably read it as
> non-uniformity only being relevant for non-helper fragments. If a pixel quad
> is partial covered by the original primitive, and discard is used in a way
> that keeps the covered pixels but discard the helper ones, should
> derivatives be defined or not?

That's a good question... My interpretation would be that derivatives should be
undefined in this case if only because otherwise things get even more
complex...