http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55939
--- Comment #13 from Richard Henderson <rth at gcc dot gnu.org> 2013-01-29 17:46:21 UTC --- All hail extended precision, and lack of proper rounding operations (until the 68040 anyway). Aldy's analysis in #c11 suggests that compiling (all of gmp, including the testcase) with -ffloat-store will probably fix this test case. I can't immediately see the bug in gmp, as it looks as if most of the double extraction should be done via integer arithmetic, but perhaps if something went wrong in the configuration one of the fp arithmetic fallback paths is being used accidentially...
