https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114522
--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> --- I think I have a fix for aes_xor_combine.c . Just testing it right now. In this case we have a 3->2 reduction and a simplification that has happened due to the const_vector of 0. And the new pattern no longer references const_vector so combine should be able to accept it without exploding the algorithm since the new insn no longer references back to i2 here.
