(In reply to Marek Olšák from comment #8)
> If KILL_IF masks out some but not all invocations in a 2x2 quad, the
> subsequent DDX and DDY opcodes can result in undefined values, resulting in
> garbage on the output.

I think Nicolai is making some improvements for the EXEC mask handling for
shader images. Maybe that'll help for this as well.

Note that AFAIK using things like derivatives in non-uniform control flow isn't
supported by GLSL. What's the original GLSL shader?