Hi, On Thu, Aug 22, 2013 at 05:33:50PM +0200, Jan Hubicka wrote: > Hi, > this problem was noticed by my verifier that binfo walks are not across type > hiearchy. ipa_intraprocedural_devirtualization is one remaining place where we > take class of object from OBJ_TYPE_REF_OBJECT instead of > obj_type_ref_class_type. > > Unforutnately I noticed that this problem is propagated quite further across > ipa-prop > design. We assume that types of pointers taken from gimple call arguments > are types of pointers to classes pass down to the callee. This is not true > after propagation. > > I did not fix the all places, only places needed to get parameter > for ipa_set_jf_known_type and detect_type_change_ssa right. Also I > modified ipa_set_jf_known_type to not record non-polymorphic > type. It is a waste of memory and LTO streaming bandwidth. > > Bootstrapped/regtesed x86_64-linux. Martin, please can you review the change? > > * ipa-prop.c (ipa_set_jf_known_type): Check that component type > is a record type with BINFO. > (detect_type_change_ssa): Add comp_type argument. > (compute_complex_assign_jump_func): Add param_type argument; pass > it down to detect_type_change_ssa. > (compute_known_type_jump_func): Add expected_type parameter. > Do not bother tracking a non-polymorphic type. > (ipa_get_callee_param_type): New function. > (ipa_compute_jump_functions_for_edge): Pass down calle parm types. > (ipa_analyze_virtual_call_uses): Use class typee as argument > of detect_type_change_1. > (ipa_intraprocedural_devirtualization): Pass down class type.
Hopefully I'll get rid of the component_types in the jump functions and most of this won't be necesary. Meanwhile, this is OK. Thanks, Martin
