https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126914
--- Comment #6 from Jeffrey A. Law <law at gcc dot gnu.org> --- Unfortunately, it appears that the x86 cost model is broken and if we go with the approach in c#2 we get several failures where the costs claim one sequence is better, but it looks significantly worse in practice. I'm going to have to rethink...
