https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91635
--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Certainly both register_operand predicate as well as the usual register constraints don't rule out subregs even after reload. can_create_pseudo_p () is false after reload of course. In theory, what you could do after reload if the operand is a paradoxical subreg would be just use SUBREG_REG in that case instead of the subreg itself.