(In reply to Roland Scheidegger from comment #7)
> I guess though the question is why other mesa drivers render it correctly
> and how they avoid the NaNs if they also use ieee conformant behavior (if
> they actually render it correctly...).

If you can get me a trace, I can have a look to see what happens with nouveau.

There's additional implications from having NaN's in RT outputs, i.e. what
happens in blending functions. I believe that nouveau configures the GPU for
the "zero wins" style of multiplication in the blending unit. And separately,
same thing for color clamps (I believe NVIDIA hw maps NaN -> 0 unconditionally
for that case).