> > This patch fixes the latest 58678 testcase by avoiding speculative > > devirtualization to the destructor of an abstract class, which can > > never succeed: you can't create an object of an abstract class, so > > the pointer must point to an object of a derived class, and the > > derived class necessarily has its own destructor. Other virtual > > member functions of an abstract class are OK for devirtualization: > > the destructor is the only virtual function that is always > > overridden in every class.] > > > > We could also detect an abstract class by searching through the > > vtable for __cxa_pure_virtual, but I figured it was easy enough for > > the front end to set a flag on the BINFO. > > > > Tested x86_64-pc-linux-gnu. OK for trunk? > > > commit b64f52066f3f4cdc9d5a30e2d48aaf6dd5efd3d4 > > Author: Jason Merrill <ja...@redhat.com> > > Date: Wed Mar 5 11:35:07 2014 -0500 > > > > PR c++/58678 > > gcc/ > > * tree.h (BINFO_ABSTRACT_P): New. > > * ipa-devirt.c (abstract_class_dtor_p): New. > > (likely_target_p): Check it. > > I think the abstract classes probably shuld never be considered in the type > inheritance graph walk as possible instances. That is we probably want to test > them in record_targets_from_bases, possible_polymorphic_call_targets and > record_target_from_binfo and simply never call maybe_record_node when the type > considered is abstract without concluding that list is incomplete as we do > when > method can not be referred. If the type has derivations that do not override > something, we will record the methods when walking derrived type, so i do not > think we need to make difference in between destructor and other type. > > We also want to sanity check that after get_class_context the outer_type > is not abstract or maybe_derived_type is set. > > gcc/cp/ > > * search.c (get_pure_virtuals): Set BINFO_ABSTRACT_P. > > * tree.c (copy_binfo): Copy it. > > You need to also save and restre the flag in LTO streaming. > > I can prepare patch tomorrow if you preffer, thanks for looking into this!
BTW it would conserve little bit of memory & streaming if we also stripped DECL_INITIAL of those vtables. They should never be needed. (We still need DECL_VTABLE to be able to work out ODR equivalence, since the types must be in type inheritance tree) Honza
