https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67815
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Comment on attachment 36487 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=36487 r2 I think for copysign we don't want/need the first_pass_instance guard, and likely also not flag_unsafe_math_optimizations, we want only be able to reassociate, which can_reassociate_p should already guarantee, and then what Joseph talked about for inexact.