https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122058
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P1 |P2
CC| |jakub at gcc dot gnu.org
--- Comment #46 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #45)
> (In reply to Richard Biener from comment #43)
> > (In reply to Jan Hubicka from comment #41)
> > > + PROF_64 = [obj_type_ref] OBJ_TYPE_REF(_12;(struct
> > > ModifierStore)prephitmp_74->2B);
> > > + if (PROF_64 ==
> > > _ZThn8_NK3icu6number4impl15LongNameHandler11getModifierENS1_6SignumENS_14Stan
> > > dardPlural4FormE.localalias)
> > > + goto <bb 8>; [24.00%]
> > > + else
> > > + goto <bb 9>; [76.00%]
> > > +
> > > + <bb 8> [local count: 229067787]:
> > >
> > > ....
> > >
> > > + goto <bb 13>; [100.00%]
> > > +
> > > + <bb 9> [local count: 725381317]:
> > > + PROF_68 = [obj_type_ref] OBJ_TYPE_REF(_12;(struct
> > > ModifierStore)prephitmp_74->2B);
> > >
> > > We definitely miss optimization here; PROF_68 is the same as PROF_64. I
> > > suppose FRE is confused by OBJ_TYPE_REF?
> >
> > FRE doesn't handle OBJ_TYPE_REF in any specific way. It's semantics
> > were never very clear to me and we seem to rely on its presence and thus
> > lowering it like RTL expansion does on its own isn't wanted. OBJ_TYPE_REF
> > is GIMPLE_SINGLE_RHS and VN simply drops OBJ_TYPE_REF to varying.
>
> Btw, for virtual calls we embed the OBJ_TYPE_REF as gimple_call_fn,
> and we also consider virtual tables writable?
>
> <bb 2> :
> # VUSE <.MEM_5(D)>
> _1 = a_6(D)->_vptr.A;
> # VUSE <.MEM_5(D)>
> _2 = *_1;
> # .MEM_7 = VDEF <.MEM_5(D)>
> OBJ_TYPE_REF(_2;(struct A)a_6(D)->0B) (a_6(D));
> # VUSE <.MEM_7>
> _3 = a_6(D)->_vptr.A;
> # VUSE <.MEM_7>
> _4 = *_3;
> # .MEM_8 = VDEF <.MEM_7>
> OBJ_TYPE_REF(_4;(struct A)a_6(D)->0B) (a_6(D));
>
> so we do not CSE a virtual table lookup here. I think we should see to
> improve the GIMPLE representation before improving CSE.
The vtables themselves are read-only (.rodata or .data.rel.ro for PIC), so I
guess
we could change _1 and _2 types above from something like whatever (*)
(whatever) * and whatever (*) (whatever) to their const variants (though maybe
they already are and we just don't bother printing that). But a_6(D)->_vptr.A
can change (though it does change only in constructors or destructors, so I bet
it can't change in between the calls above, if it was a ctor or dtor, we'd see
the change in the IL, ctors are never virtual and while dtors can be, I think
it isn't valid to call a dtor this way and then call it again and random
methods on a class shouldn't destruct the class under it (though, perhaps I'm
wrong).
Anyway, per the above discussions downgrading this to P2 as it is likely latent
for several releases.