> Hi, > > I have made a mistake when I decided to redirect member-pointer calls > to builtin_unreachable because I wrongly thought there are two calls, > one for the VMT lookup and other for non-virtual members and thought > only the second got redirected. However, even though there are two > different lookups, there is only one call and thus we must refrain > from killing it when the value coming from a caller is not an address > of a function. I have pondered about overloading the polymorphic flag > for marking member pointer calls but eventually decided to introduce a > new one, as there are plenty bits still available in indirect info and > the added complexity of overloading the flag quickly turned ugly. > > Bootstrapped and tested on x86_64-linux. OK for trunk? > > Thanks and sorry for the confusion, > > Martin > > > 2013-06-21 Martin Jambor <mjam...@suse.cz> > > PR middle-end/57670 > * cgraph.h (cgraph_indirect_call_info): New flag member_ptr. > * ipa-prop.c (ipa_print_node_jump_functions): Mark member pointer > calls in the dump. > (ipa_note_param_call): Initialize member_ptr flag. > (ipa_analyze_indirect_call_uses): Set member_ptr flag. > (ipa_make_edge_direct_to_target): Bail out if member_ptr is set. > (ipa_write_indirect_edge_info): Stream member_ptr flag. > (ipa_read_indirect_edge_info): Likewise. > > testsuite/ > * g++.dg/ipa/pr57670.C (H): New test.
OK, thanks! Honza
