https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68930
--- Comment #6 from Jan Hubicka <hubicka at ucw dot cz> --- For the Martin's answer. I poked about this a bit. One can add call to node->get_body to the end of cgraph_materialize_clone which after removing the associated assert will call to the ipa-prop transform method. It will also call inliner's transform which will in turn remove the cgraph and break. I suppose I can fix this by making IPA transforms function speicfic (they should be) and making inliner to only drop its transform method on functions non-inline functions and make ipa-cp add transforms only to the clones. It should improve compile time, too. Still we will get missed optimizations, so keeping this info for GVN seems like resonable thing to do.