C++

Revised 10 September 1999


[990910 IBM -- Brian] Added more examples, split out the two kinds of adjustments in Table 1a, and added a summary of the component counts for the two approaches.


Table 1a: Example Code and Call Semantics 
Declarations  Call  Callee  Call-site 
Adjustment 
Thunk/Entry-point 
Adjustment
struct A {
  virtual void f ();
  virtual void g ();
  virtual void h ();
  int ia;
};

A *pa;
pa->f() A::f()  none  none
pa->g() A::g()  none  none
pa->h() A::h() none none
struct B: public virtual A {
  void f ();
  void h ();
  int ib;
};

B *pb;
A *pa_in_b = pb;
pb->f() B::f()  none  none
pb->A::f() A::f()  B => A none
pb->g() A::g()  B => A  none
pb->h() B::h() none none
pa_in_b->f() B::f()  none  A => B
pa_in_b->g() A::g()  none  none
pa_in_b->h() B::h()  none  A => B
pa_in_b->A::f() A::f()  none  none
struct C: public virtual A {
  void g ();
  void h ();
  int ic;
};

C *pc;
A *pa_in_c = pc;
pc->f() A::f()  C => A  none
pc->g() C::g()  none  none
pc->A::g() A::g()  C => A  none
pc->h() C::h()  none  none
pa_in_c->f() A::f()  none  none
pa_in_c->g() C::g()  none A => C
pa_in_c->h() C::h()  none A => C
pa_in_c->A::g() A::g()  none  none
struct D: public B, public C {
  int id;
  void h();
};

D *pd;


A *pa_in_d = pd;
B *pb_in_d = pd;
C *pc_in_d = pd;


A *pa_in_b_in_d = pb_in_d;
A *pa_in_c_in_d = pc_in_d;
pd->f() B::f()  none [D => B]  none
pd->g() C::g()  D => C  none
pd->h() D::h()  none  none
pa_in_d->f() B::f()  none  A => B
pa_in_d->g() C::g()  none A => C
pa_in_d->h() D::h()  none A => D
pb_in_d->f() B::f()  none  none
pb_in_d->g() C::g()  B => A A => C
pb_in_d->h() D::h()  none  B => D
pc_in_d->f() B::f()  C => A A => B
pc_in_d->g() C::g()  none  none
pc_in_d->h() D::h()  none  C => D
pa_in_b_in_d->f() same as for pa_in_d 
pa_in_b_in_d->g()
pa_in_b_in_d->h()
pa_in_c_in_d->f()
pa_in_c_in_d->g()
pa_in_c_in_d->h()
p...d->A::f() A::f()  ... => A  none
p...d->A::g() A::g()  ... => A  none
p...d->A::h() A::g()  ... => A  none
struct X {
  int ix;
  virtual void x();
};
struct E : X, D {
  int ie;
  void f();
  void h();
};
pe->f() E::f() none none
pe->g() C::g() E => C none
pe->h() E::h() none  none
pe->x() X::x() none [E=>X] none
pa_in_e->f()
E::f() none A => E
pa_in_e->g() C::g() none A => C
pa_in_e->h() E::h() none A => E
pb_in_e->f() E::f() none B => E
pb_in_e->g() C::g() B => A A => C
pb_in_e->h() E::h() none  B => E
pc_in_e->f() E::f() C => A A => E
pc_in_e->g() C::g() none none
pc_in_e->h() E::h() none C => E
pd_in_e->f() E::f() none [D=>B] B => E
pd_in_e->g() C::g() D => C none
pd_in_e->h() E::h() none D => E

Table 1b: Example Data Layout 
Declarations  Size  Offset Member 
struct A {
  virtual void f ();
  virtual void g ();
  virtual void h ();
  int ia;
};
16  0 A::vptr
ia
struct B: public virtual A {
  void f ();
  void h ();
  int ib;
};
32  0 B::vptr
ib
16  A::vptr
24  ia
struct C: public virtual A {
  void g ();
  void h ();
  int ic;
};
32  0 C::vptr
ic
16  A::vptr
24  ia
struct D: public B, public C {
  void h ();
  int id;
};
48  0 D/B::vptr
ib
16  C::vptr
24  ic
28  id
32  A::vptr
40  ia
struct X {
  int ix;
  virtual void x();
};
struct E : X, D {
  void f ();
  void h (); 
  int ie; 
};
64 0 X/E::vptr
8 ix
16 D/B::vptr
24 ib
32 C::vptr
40 ic
48 id
56 A::vptr
64 ia

Table 1c: Example Vtable Layout 
Declarations  Vtable (HP) 1,2,3 Vtable (Cygnus/IBM)
struct A {
  virtual void f ();
  virtual void g ();
  virtual void h ();
  int ia;
};
A::offset_to_top (0)
A::rtti
-- A vtable address --
A::f() []
A::g() []
A::h() []
A::offset_to_top (0)
A::rtti
-- A vtable address --
A::f() []
A::g() []
A::h() []
struct B: public virtual A {
  void f ();
  void h ();
  int ib;
};
B::offset_to_A (16)
B::offset_to_top (0)
B::rtti
-- B vtable address --
B::f() []
B::h() []

A::offset_to_top (-16)
A::rtti
-- A-in-B vtable address --
B::f() [[-72] B::offset_to_A : thunk]
A::g() []
B::h() [[-72] B::offset_to_A : thunk]
B::offset_to_A (16)
B::offset_to_top (0)
B::rtti
-- B vtable address --
B::f() []
B::h() []

A::offset_for_h (-16)
A::offset_for_g (0)
A::offset_for_f (-16)
A::offset_to_top (-16)
A::rtti
-- A-in-B vtable address --
B::f() [[-24]offset_for_f]
A::g() []
B::h() [[-40]offset_for_h]
struct C: public virtual A {
  void g ();
  void h ();
  int ic;
};
C::offset_to_A (16)
C::offset_to_top (0)
C::rtti
-- C vtable address --
C::g() []
C::h() []

A::offset_to_top (-16)
A::rtti
-- A-in-C vtable address --
A::f() []
C::g() [[-72] C::offset_to_A : thunk]
C::h() [[-72] C::offset_to_A : thunk]
total size 15*8 = 120 bytes
C::offset_to_A (16)
C::offset_to_top (0)
C::rtti
C vtable address --
C::g() []
C::h() []

A::offset_for_h (-16)
A::offset_for_g (-16)
A::offset_for_f (0)
A::offset_to_top (-16)
A::rtti
A-in-C vtable address --
A::f() []
C::g() [[-32] offset_for_g]
C::h() [[-40] offset_for_h]
total size 18*8 = 144 bytes
struct D: public B, public C {
  void h ();
  int id;
};
D::offset_to_C (16)
D::offset_to_A (32)
D::offset_to_top (0)
D::rtti
-- D, B-in-D vtable address --
B::f() []
D::h() []

C::offset_to_A (16)
C::offset_to_top (-16)
C::rtti
-- C-in-D vtable address --
C::g() []
D::h() [[-88] D::offset_to_C]

A::offset_to_top (-32)
A::rtti
-- A-in-D vtable address --
B::f() [[-128] D::offset_to_A : thunk]
C::g() [[-72] C::offset_to_A : thunk]
D::h() [[-128] D::offset_to_A : thunk]

total size 23*8 = 184 bytes
D::offset_to_A (32)
D::offset_to_top (0)
D::rtti
-- D, B-in-D vtable address --
B::f() []
D::h() []

C::offset_to_A (16)
C::offset_to_top (-16)
C::rtti
-- C-in-D vtable address --
C::g() []
D::h() [-16]

A::offset_for_h (-32)
A::offset_for_g (-16)
A::offset_for_f (-32)
A::offset_to_top (-32)
A::rtti
-- A-in-D vtable address --
B::f() [[-24] offset_for_f]
C::g() [[-32] offset_for_g]
D::h() [[-40] offset_for_h]
total size 25*8 = 200 bytes
struct X {
  int ix;
  virtual void x();
};
struct E : X, D {
  int ie;
  void f();
  void h ();
};
E::offset_to_D (16)
not used 
not used
not used
not used
E::offset_to_C (32)
E::offset_to_A (56)
E::offset_to_top (0)
E::rtti
-- E, X-in-E vtable address --
X::x() []
E::f() []
E::h() []

D::offset_to_A (40)
D::offset_to_top (-16)
D::rtti
-- D, B-in-E vtable address --
E::f() [[-144] E::offset_to_D]
E::h() [[-144] E::offset_to_D]

C::offset_to_A (24)
C::offset_to_top (-32)
C::rtti
-- C-in-E vtable address --
C::g() []
E::h() [[-144] E::offset_to_C]

A::offset_to_top (-56)
A::rtti
-- A-in-E vtable address --
E::f() [[-200] E::offset_to_A : thunk]
C::g() [[-72] C::offset_to_A : thunk]
E::h() [[-200] E::offset_to_A : thunk]
total size 37*8 = 296 bytes
E::offset_to_A (56)
E::offset_to_top (0)
E::rtti
-- E, X-in-E vtable address --
X::x() []
E::f() []
E::h() []

D::offset_to_A (40)
D::offset_to_top (-16)
D::rtti
-- D, B-in-E vtable address --
E::f() [-16]
E::h() [-16]

C::offset_to_A (24)
C::offset_to_top (-32)
C::rtti
-- C-in-E vtable address --
C::g() []
E::h() [-32]

A::offset_for_h (-56)
A::offset_for_g (-24)
A::offset_for_f (-56)
A::offset_to_top (-56)
A::rtti
-- A-in-E vtable address --
E::f() [[-24] A::offset_for_f ]
C::g() [[-32] A::offset_for_g ]
E::h() [[-40] A::offset_for_h ]
total size 34*8 = 272 bytes
  1. Numbers in parentheses after offset_to_top entries are actual values.
  2. Class prefixes for functions identify class where defined.
  3. Information in square brackets after function pointer entries indicates entry-point adjustment:
    [] no adjustment required, use primary entry point
    [n] use adjusting entry point that adds "n" to this
    [[n] blurb]  use adjusting entry point that dereferences vptr+n and subtracts (HP) or adds (Cygnus/IBM)
        that value to this. blurb is the name of the accessed  field
    [[n] blub : thunk]  use adjusting 3rd party thunk that dereferences vptr+n and subtracts that value from this

Notes: 1) Each function descriptor in the vtable is 16 bytes but the offset and data pointers are only 8, the earlier versions of this table didn't take that into account
2) In the HP column for struct E, I have omitted the D::offset_to_C field because the overrides in E render it unnecessary.  However, if maintaining navigability inside the nonvirtual parts of the vtable is important then this "cleanup" can only be done for direct nonvirtual bases and not for more deeply nested ones.
3) I have taken Christophe at his word that thunks are used for adjusting vtable entries in virtual bases in the HP proposal. Some of them could be done with entry points though.

When all is said and done we have
 
x/y/z
x = # direct secondary entries
y = # "reach back" secondary entries
z = # 3rd-party thunks
Function HP Cygnus/IBM
A::f 0/0/0 0/0/0
A::g 0/0/0 0/0/0
A::h 0/0/0 0/0/0
B::f 0/0/2 0/1/0
B::h 0/0/1 0/1/0
C::g 0/0/1 0/1/0
C::h 0/0/1 0/1/0
D::h 0/1/1 1/1/0
E::f 0/1/1 1/1/0
E::h 0/1/1 2/1/0