https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102785
--- Comment #2 from Jeffrey A. Law <law at gcc dot gnu.org> --- Yea, it could well be a representational problem in the RTL. I didn't try to debug it at all beyond reduction and noting that cse1 was where the two compilers diverged in behavior. I don't personally care about bfin, my biggest concern with this stuff is to make sure there's not a generic problem. If it's a bfin issue and fixing it is a major lift, then we can probably disable the patterns/builtins or just let it be.
