https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122552
--- Comment #4 from Jeffrey A. Law <law at gcc dot gnu.org> --- Ah, that clarifies things. Thanks Pan. So ISTM that we want the initial pattern to match the .vv form, then an alternate pattern to match the vx form when combine decides the costs are such that the .vx form is cheaper? Is that how the other cases are currently handled? The other approach one could try would be to define a predicate that accepts both forms, then use that in the named insn. That would result in the initial code generation using whatever seemed the most appropriate from the vectorizer's viewpoint. A predicate can actually match more than just a terminal node. We use this a lot for matching variable shift counts for scalar ops. Not sure if it'd help here, but I mention it just in case it might help.
