[Gcl-devel] Re: Building GCL 2.6.1 on FreeBSD 4.8

[Magnus Henoch]

Camm Maguire <address@hidden> writes:

> Greetings!  I have two questions about this:
>
> 1) The (finite(a) && (a) != 0.0) expression is already the default in
>    gclincl.h.in, no?

Yes, it is.

> 2)  This doesnt' seem to be the right default in any case.  Consider
>     this program:

[...]

>   double d=1.0;
>
>   for (;d;d*=0.5)
>     printf("%e %-64.64s %u %u\n",d,bit_pattern(d),isnormal(d) ? 1 : 
> 0,finite(d));

[...]

> 8.900295e-308 
> 0000000000000000000000000000000000000000000000000000110000000000 1 1
> 4.450148e-308 
> 0000000000000000000000000000000000000000000000000000010000000000 1 1
> 2.225074e-308 
> 0000000000000000000000000000000000000000000000000000100000000000 1 1
> 1.112537e-308 
> 0000000000000000000000000000000000000000000000000001000000000000 0 1
> 5.562685e-309 
> 0000000000000000000000000000000000000000000000000010000000000000 0 1
> 2.781342e-309 
> 0000000000000000000000000000000000000000000000000100000000000000 0 1
> =============================================================================

Subnormal numbers... I didn't know such a thing existed, but FreeBSD
(5.1 and up) and glibc both know them (gathered from source, not
experiments).

> Does FreeBSD/OpenBSD have an alternate isnormal that reproduces the
> GNU behavior?

As far as I can see, no.

I believe a small replacement function would blend in nicely into
num_co.c.  The comments in there already describe the bit patterns of
doubles and floats for a couple of architectures; the new isnormal
function would test for finite(a) and non-zero exponent part of a.

I could do this some time next week.

Regards,
Magnus