https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119614
--- Comment #27 from Martin Jambor <jamborm at gcc dot gnu.org> --- Created attachment 61116 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=61116&action=edit WIP and only mildly tested patch This is my current WIP patch, so far only mildly tested, but which fixes the test from comment #22. A logical continuation of this might be to ditch ipa_return_value_sum and store the return value VR always to clone_info since even with this patch it becomes a misnomer because it will contain optimization information about nodes which are not a clone. But since the information needs to be available to the callers of the function they describe too, I cannot think of a more appropriate place (apart from creating another variant of clone_info which does not seem to be justified either). The fact that the VR would then go from clone_info to ipa_node_ref and then back to clone_info would actually be reasonable, since another logical extension is to have IPA-CP propagate VRs through return pass-through jump functions. In any event, comments welcome.
