__builtin_isnanl() and invalid x87 80-bit floating point numbers

[Vincent Lefevre]

On 2017-05-15 22:35:33 +0800, Liu Hao wrote:
> Yeah but `printf()` from glibc doesn't think it is a NaN... From
> your point of view, is this a glibc bug?

I think so. If I understand correctly, the output of floating-point
data in glibc is handled in stdio-common/printf_fp.c, which uses:

extern mp_size_t __mpn_extract_long_double (mp_ptr res_ptr, mp_size_t size,
                                            int *expt, int *is_neg,
                                            long double value);

The code is one of (not sure):
  sysdeps/i386/ldbl2mpn.c
  sysdeps/ieee754/ldbl-96/ldbl2mpn.c

In sysdeps/i386/ldbl2mpn.c only, there is a special case for
"pseudo zero" (no comments about the choice of the behavior,
though), which was added here:

commit 835abc5c0dfd1ba8aabeb52d46793b13702c708b
Author: Ulrich Drepper <drepper@redhat.com>
Date:   2007-06-08 02:50:59 +0000

    [BZ #4586]
    
    2007-06-06  Jakub Jelinek  <jakub@redhat.com>
            BZ #4586
            * sysdeps/i386/ldbl2mpn.c (__mpn_extract_long_double): Treat
            pseudo-zeros as zero.
            * sysdeps/x86_64/ldbl2mpn.c: New file.
            * sysdeps/ia64/ldbl2mpn.c: New file.

This corresponds to Bruno Haible's bug report:

  https://sourceware.org/bugzilla/show_bug.cgi?id=4586

(though this was for IA64). He said:

  Additionally, the printf results for pseudo-NaN and pseudo-Inf
  should better be "nan", because these numbers behave like NaNs
  in comparisons, as you can see from the program's output.

Not sure about the reason of the current choice.

-- 
Vincent Lefèvre <vincent@vinc17.net> - Web: <https://www.vinc17.net/>
100% accessible validated (X)HTML - Blog: <https://www.vinc17.net/blog/>
Work: CR INRIA - computer arithmetic / AriC project (LIP, ENS-Lyon)