https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122958
--- Comment #2 from Jonathan Wakely <redi at gcc dot gnu.org> --- Presumably the backport for PR c++/50852 to 4.7.3 caused the accepts-invalid regression there (but not an ICE). That's academic now anyway, as we only need to care about the ICE on current branches.
