Re: [avr-libc-dev] Bug in math library

[Joerg Wunsch]

As Colin O'Flynn wrote:

> Perhaps this is something that we could just use #define's to work
> with? It should be trivial to have a +inf put in place of a NaN,
> then there is the option of compiling with either NaN generated or
> +inf.

Why make it optional?  What does IEEE say about it?

Without reading the IEEE standard, I'd say: 0/0 => NaN, ±number/0 =>
±Inf (that's the way the mathematicians handle it).

Yeah, look here:

http://stevehollasch.com/cgindex/coding/ieeefloat.html

(under Special Operations).

> The only other problem: avr-libc doesn't support -0 I think (going
> from source code comments), which would mean there could be no -inf
> option.

See savannah bug #1929.

I'm not sure whether it's worth the while to support denormalized
operation, but supporting these Special Operations seems useful to me.

Well, SNaNs cannot be really supported, as there's no kind of
signalling available to us.  So it boils down to full QNaN support.

-- 
cheers, J"org               .-.-.   --... ...--   -.. .  DL8DTL

http://www.sax.de/~joerg/                        NIC: JW11-RIPE
Never trust an operating system you don't have sources for. ;-)