https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83565
--- Comment #32 from Eric Botcazou <ebotcazou at gcc dot gnu.org> --- > However I'm not sure whether exposing the distinction is really the way to > go so I'm going to evaluate the pessimization that would be introduced on > SPARC 64-bit by disregarding WORD_REGISTER_OPERATIONS for shift operations. Pessimization is real under the form of redundant sign/zero-extensions, so I'm going to test a stopgap fix instead on IA-64 and SPARC64.