arch-independent glibc printf segfault for "special" long double values

[Jim Meyering <jim at meyering dot net>]

This started with Bruno Haible's bug report:

    > printf crashes on some 'long double' values
    > http://sourceware.org/bugzilla/show_bug.cgi?id=4586

I objected to the closure of that BZ:

    http://thread.gmane.org/gmane.comp.lib.glibc.alpha/12394

because I don't want to have to use bulky *printf replacements
forever, just to avoid the potential of this glibc code to make
my applications vulnerable to a user-data-provoked segfault.

Note that I've changed the subject to emphasize that this is
architecture-independent.  Well, at least it is not IA64-specific.
To trigger it you need a "long double" type longer than 8 bytes.

For example, it can cause GNU od to segfault.

On x86_64, you need to use -tf16 to get the long double interpretation.
On an 686-based system, it's a 12-byte quantity, so you need "-tf12".
Both of these systems are using libc-2.5:

On at least an AMD/x86_64:

    $ perl -e 'print pack("LLLL", 0, 0, 32772, 0)' |od -tf16
    [Exit 139 (SEGV)]

On an i686-based system:

    $ perl -e 'print pack("LLLL", 0, 0, 32772, 0)' |od -tf12
    [Exit 139 (SEGV)]

Thanks again to Jakub Jelinek for changing it.
Now, the above should print those values as FP zeros.