https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63476
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ah,
#5 0x0000000000c263c2 in ipa_polymorphic_call_context::get_dynamic_type (
this=0x7fffffffd790, instance=<ssa_name 0x7ffff687b678>,
otr_object=<ssa_name 0x7ffff687b678>,
otr_type=<record_type 0x7ffff6867b28 D>, call=<gimple_call 0x7ffff6895498>)
at /space/rguenther/src/svn/trunk/gcc/ipa-polymorphic-call.c:1606
(gdb) l
1601 tci.multiple_types_encountered = false;
1602 tci.speculative = false;
1603 tci.seen_unanalyzed_store = false;
1604
1605 walk_aliased_vdefs (&ao, gimple_vuse (stmt),
check_stmt_for_type_change,
1606 &tci, NULL, &function_entry_reached);
this walks on gimple_vuse!
(gdb) p debug_gimple_stmt (stmt)
# VUSE <.MEM>
pretmp_5 = this_10(D)->D.2797._vptr.A;
VUSEs are not kept renamed by load PRE insertion phase so this will not work.
The function is called on the call
# .MEM_45 = VDEF <.MEM_6(D)>
OBJ_TYPE_REF(pretmp_1;(struct D)this_10(D)->0) (this_10(D), p1_11(D), p2_14(D),
0);
Of course it would be nice to keep virtual SSA form up-to-date during
PRE insertion, but as of now that was not necessary (and we can't simply
rename VOPs as that will break the SSA value-numbering lattice which includes
virtuals). Keeping virtual SSA form up-to-date requires computing the
life virtual operand on edges (which means BB exit).
I'm going to try that.