Re: [Maxima-discuss] SIGN1 strangeness, was: Problem with bfloat

[David B. <dbm...@gm...>]

On 3/09/2014 8:21 AM, Robert Dodier wrote:
> On 2014-09-01, David Billinghurst <dbm...@gm...> wrote:
>
>>      I:87^611$
>>      M:91^211$
>>      sign(sqrt(I)-M);
> Thanks for the example. Here's a proposed fix. This causes $FLOAT to
> always fail (even for Lisps which otherwise allow float infinity)
> and therefore SIGN1 takes a different route and sign returns 'pos.
>
> The patch assumes that (float inf) minus (float inf) != 0.
> I'm pretty sure that's required by IEEE 754 (in fact I think the
> result is supposed to be NaN) but I can't quote chapter & verse
> at the moment.
>
> best
>
> Robert Dodier
>
> PS.
> $ git diff src/comm.lisp
> diff --git a/src/comm.lisp b/src/comm.lisp
> index 42bdc12..1d6d178 100644
> --- a/src/comm.lisp
> +++ b/src/comm.lisp
> @@ -1178,7 +1178,7 @@
>              (car result))))))
>              
>   (defmfun $float (e)
> -  (cond ((numberp e) (float e))
> +  (cond ((numberp e) (let ((e1 (float e))) (if (= (- e1 e1) 0) e1 (signal 'floating-point-overflow))))
>          ((and (symbolp e) (mget e '$numer)))
>          ((or (atom e) (member 'array (cdar e) :test #'eq)) e)
>          ((eq (caar e) 'rat) (fpcofrat e))
This works for me with gcl-2.6.10 on Windows XP.  Thanks.  All 
regression tests pass.