http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55939
--- Comment #14 from Mikael Pettersson <mikpe at it dot uu.se> 2013-01-29 18:03:04 UTC --- The wrong-code occurs in the test wrapper rather than in gmp proper, and the test wrapper does contain a helper function (my_ldexp) which computes and returns 'double'. So the problem might simply be caused by excess precision.
