https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109008
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Well, we already do that half ulp fuzzing in frange_arithmetic. The thing is that for forward [1., 1.] - [1., -1.] it is really correctly [0., 0.], the subtraction is exact in that case. It is just the reverse operation which perhaps need some flag that would be propagated to frange_arithmetic and would make it fuzz even more.