> Ah, so there's an assumption that NaNs are handled earlier and never reach
> this place? That's probably a safe assumption. I haven't thought about that,
> it simply seemed suspicious that the code mixes direct comparisons and
> float8_mi() calls.
The comparison functions handle NaNs. The arithmetic functions handle
returning error on underflow, overflow and division by zero. I
assumed we want to return error on those in any case, but we don't
want to handle NaNs at every place.
> Not sure, I'll leave that up to you. I don't mind doing it in a separate
> patch (I'd probably prefer that over mixing it into unrelated patch).
It is attached separately.