help requested with long double issue on 32-bit FreeBSD

[Zimmermann Paul]

       Hi,

[this mail was sent on the MPFR list, but Steve Kargl suggested to post it
also here]

we need help to investigate an error we get with the development version of
MPFR on i686-freebsd. The error occurs on the hydra.nixos.org continuous
integration platform. Unfortunately we don't have direct access to the
corresponding computer, thus isolating the issue is not easy.

The error we get can be seen at http://hydra.nixos.org/build/5666171/log/raw.
Starting from the long double value in 3, and subtracting e=0.25, we get the
long double value in 4, which is obviously wrong in the last bits:

3: d=2.632048527199790448306230411512629530e-01
4:  d=1.320485271997904469509776959057489876e-02
                        ^^^^^^^^^^^^^^^^^^^^

We believe this is a compiler bug (GCC 4.2.1 was used in this test).

If you have access to a 32-bit computer under FreeBSD, please could you try
to compile the development version and run the "tset_ld" test, after doing
export GMP_CHECK_RANDOMIZE=1376689137?

If you can reproduce the error, please tell Vincent Lefèvre and myself
(Vincent.Lefevre at ens-lyon.fr and Paul.Zimmermann at inria.fr); we'll help
you to isolate the problem (e.g., trying different optimization levels).

If you can't reproduce (on a 32-bit FreeBSD machine), please tell us too;
this might indicate the bug was fixed in a later gcc version.

Thank you in advance,
Paul Zimmermann