Re: isnanl failure on MirBSD

[Eric Blake]

On 09/13/2010 02:08 PM, Eric Blake wrote:
> Bruno, what else do you need to help debug this?  It's crashing on an
> i686 installation of the most recent MirBSD (version 10, 20100815),
> using mgcc 3.4.6 (propolice; gpc; MirOS 0AAB). I don't yet know whether
> to suspect a compiler error, a buggy <float.h>, or faulty configure
> results; but this failure in turn means that vasprintf-posix fails.
>
> $ grep '^#.#EXPBIT' config.log
> #define FLT_EXPBIT0_WORD 0
> #define FLT_EXPBIT0_BIT 23
> #define LDBL_EXPBIT0_WORD 2
> #define LDBL_EXPBIT0_BIT 0
> $ gdb gltests/test-isnanl-nolibm
> ...
> 99 exponent = (m.word[EXPBIT0_WORD] >> EXPBIT0_BIT) & EXP_MASK;
> (gdb) p m
> $1 = {value = 3.1410000000000000142108547152020037, word = {790273982,
> 3372623069, 16384}}
...
> (gdb) p sizeof (long double)
> 12

Also, EXP_MASK is 2047 (0x7ff).

Some interesting comparisons with Linux:
(gdb) p m
$1 = {value = 3.1409999999999999998993860383933452, word = {790273982,
    3372623069, 16384, 0}}
(gdb) p sizeof(long double)
$2 = 16

(Obviously some rounding differences between the two builds of gdb in 
how to display the long double, but the value is identical).

EXP_MASK is 32767 (0x7fff).

Therefore, the bug lies in MirBSD's <float.h>.  Sure enough, on Linux:
$ printf '#include <float.h>\nLDBL_MAX_EXP LDBL_MIN_EXP\n' | gcc -E - \
  | tail -n1
16384 (-16381)

But on MirBSD:

$ printf '#include <float.h>\nLDBL_MAX_EXP LDBL_MIN_EXP\n' | mgcc -E - \
  | tail -n1
1024 (-1021)

--
Eric Blake   address@hidden    +1-801-349-2682
Libvirt virtualization library http://libvirt.org