http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59176
Martin Jambor <jamborm at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jamborm at gcc dot gnu.org
Assignee|mjambor at suse dot cz |unassigned at gcc dot
gnu.org
--- Comment #9 from Martin Jambor <jamborm at gcc dot gnu.org> ---
(In reply to Jan Hubicka from comment #8)
>
> I believe it is bug somewhere in logic skipping thunks. Martin, this is
> probably yours.
I on the other hand think this is either a verification false positive
or some sort of cgraph node removal bug.
What happens in IPA-CP is this: m_fn3/8 originally calls B::~B()/5 which
is an alias (not a thunk as far as I can see but I do not see how that
matters even if I was wrong) to B::~B()/4. IPA-CP does not gather
constants for aliases or thunks separately (though it is careful not to
propagate first parameters through thunks), so IPA-CP is only interested
in the /4 real thing. Furthermore, it decides to clone B::~B()/4 and
redirect the call from m_fn3/8 to the new clone.
In fact, the modified edge passes verify_edge_corresponds_to_fndecl
just fine several times because of
!clone_of_p (cgraph_function_or_thunk_node (node, NULL), e->callee))
part of the checking correctly detects this case.
But then, after being successfully verified a number of times, the above
fails, because the cgraph_function_or_thunk_node (node, NULL) part stops
returning the /4 node but returns the argument itself which is /5. When
I now dump the node I cannot see it being an alias and that is because
symtab_remove_unreachable_nodes had zapped it.
I think that either symtab_remove_unreachable_nodes should remove nodes
it cripples from the decl->symtab hash (why does it not call
cgraph_remove_node?) or the verifier needs to understand that these
nodes are no good and ignore them.
