https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125974
--- Comment #1 from mjr19 at cam dot ac.uk --- Looking further, this might not be a bug. -O3 produces the code I would expect, and -O3 -ffast-math or -Ofast produces the more verbose code. But the more verbose code has much better ILP. It might not actually be any slower, although I am still struggling to believe that it is any faster.
