https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124391
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> --- 512.749999999f32 is 512.75 (and with excess precision 512.749999999f16 is also evaluated to the precision of IEEE single) and then rounded for storing into the float16_t.
