https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93934
--- Comment #17 from Richard Biener <rguenth at gcc dot gnu.org> --- It might help to provide a option or tunable to disable the use of fcmov which should reduce the attack surface a bit. I don't see any way to avoid using fld for loading x87 float or double values (well, load into GPR, do softfp convert to long double, spill and fld as long double ...). For the reporter using -fno-if-conversion should also reliably avoid the situation in the testcase.