https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84938
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> --- int &&a = 1 / ~-1; is enough. When we have int &&b = 1 / 0; this doesn't ICE. When creating TRUNC_DIV_EXPR for the latter, we don't set the constant bit on it, because the last argument is integer_zerop. That is not the case with the former testcase, but both arguments are TREE_CONSTANT and so we set TREE_CONSTANT on the TRUNC_DIV_EXPR as well. I'm surprised we never cp_fold this initializer.