Jim Meyering pushed for a quick resolution:
> I'm hoping to make a coreutils test release RSN,
> and one of the few things missing is the macro to help
> work around the freebsd 6.1 long double printf bug.
Your test proggie results revealed that the bug is that although
'long double' generally has 53 bits of precision (in compile-time
expressions, in <float.h>, and in some run-time computations), it sometimes
has 64 bits of precision (in some other run-time computations). To me,
that counts as "excess precision".
I planned to conditionally disable the check in vasnprintf.c; but I don't
have much time right now and you are pushing, so I disable it unconditionally.
2007-11-26 Bruno Haible <[EMAIL PROTECTED]>
* lib/vasnprintf.c (decode_long_double): Don't abort if the
'long double' type has excess precision.
Reported by Jim Meyering in
<http://lists.gnu.org/archive/html/bug-gnulib/2007-11/msg00120.html>.
--- lib/vasnprintf.c
+++ lib/vasnprintf.c
@@ -871,8 +871,11 @@ decode_long_double (long double x, int *ep, mpn_t *mp)
abort ();
m.limbs[--i] = (hi << (GMP_LIMB_BITS / 2)) | lo;
}
+#if 0 /* On FreeBSD 6.1/x86, 'long double' numbers sometimes have excess
+ precision. */
if (!(y == 0.0L))
abort ();
+#endif
/* Normalise. */
while (m.nlimbs > 0 && m.limbs[m.nlimbs - 1] == 0)
m.nlimbs--;
