Why parent classes' vptrs are stored in derived objects

[KishorAditya]

Hi All,
Consider the following scenario:
class Top { };
class Left: virtual public Top { };
class Right: virtual public Top { };
class Bottom: public Left, public Right {};

Many books propose that object of Bottom contains three vptrs, one for
Left and Bottom, one for Right and one for Top.
Now my question is why the compiler is storing superclasses' vptrs in
derived class object. After all Bottom also has a vptr pointing to its
vtable from which it will call all virtual functions.
Thanks in advance.

 

[udontknowhoami]

Hi All,
Consider the following scenario:
class Top { };
class Left: virtual public Top { };
class Right: virtual public Top { };
class Bottom: public Left, public Right {};

Many books propose that object of Bottom contains three vptrs, one for
Left and Bottom, one for Right and one for Top.
Now my question is why the compiler is storing superclasses' vptrs in
derived class object. After all Bottom also has a vptr pointing to its
vtable from which it will call all virtual functions.
Thanks in advance.

Well as far as you use derived object alone it works fine. But what
happens when you dynamic cast Bottom pointer to either of Left, Right
or Top like:
Bottom * b = new Bottom;
Left * something = <dynamic_cast*>b;
In this case you type casting you should use something as if it is a
original Left object; for that you need to access Left class' vtbl.
Thats why you need to store base classes vptrs in the derived objects.
Hope its clear.

 

[Ron House]

Well as far as you use derived object alone it works fine. But what
happens when you dynamic cast Bottom pointer to either of Left, Right
or Top like:
Bottom * b = new Bottom;
Left * something = <dynamic_cast*>b;
In this case you type casting you should use something as if it is a
original Left object; for that you need to access Left class' vtbl.
Thats why you need to store base classes vptrs in the derived objects.
Hope its clear.

No that's not clear. When calling a method from a Left*, if it points to
a Bottom, then Bottom's virtual functions should be used, not Left's.

 

[udontknowhoami]

Yes, you are right Ron. Bit misunderstood the question.
May be this explanation is clear:

class D : public A, public B, public C {.........}

If we dont store the base classes' vptrs in derived objects, that
means you have to club all the virtual functions of all the base
classes (in case of multiple inheritance) into one vtbl that has a
single vptr that will be stored in your object. Now, no problem in
accessing virtual functions of first subobject (like class A in the
above code snippet) whose virtual function pointers are stored in
starting location of vtbl. But what happens in accessing other
subobjects (like classes B and C in the above code snippet) virtual
functions. Their offsets will get disturbed as you have clubbed all
the virtual functions into one vtbl. After all, compiler will access
virtual functions with their offsets only. Now you dont know what will
be get called.

Thats why in case of single or pure multilevel inheritance you dont hv
to store the base classes' vptrs in derived objects.

Now the question: Can not compiler determine these changes in offsets
and adjust virtual functions pointers accordingly? It may, it depends
on the compiler. But most of the compilers adopt this strategy and may
some compilers do this optimization.

 

[Ron House]

Yes, you are right Ron. Bit misunderstood the question.
May be this explanation is clear:

class D : public A, public B, public C {.........}

If we dont store the base classes' vptrs in derived objects, that
means you have to club all the virtual functions of all the base
classes (in case of multiple inheritance) into one vtbl that has a
single vptr that will be stored in your object. Now, no problem in
accessing virtual functions of first subobject (like class A in the
above code snippet) whose virtual function pointers are stored in
starting location of vtbl. But what happens in accessing other
subobjects (like classes B and C in the above code snippet) virtual
functions. Their offsets will get disturbed as you have clubbed all
the virtual functions into one vtbl. After all, compiler will access
virtual functions with their offsets only. Now you dont know what will
be get called.

Thats why in case of single or pure multilevel inheritance you dont hv
to store the base classes' vptrs in derived objects.

Now the question: Can not compiler determine these changes in offsets
and adjust virtual functions pointers accordingly? It may, it depends
on the compiler. But most of the compilers adopt this strategy and may
some compilers do this optimization.

Yes, with multiple inheritance, obviously A, B, and C cannot all share
the same starting address as D, so yes, there must be an offset on the
address of the _data_ of the object, when calling superclass functions
(of all but 1); and there must be distinct vtable addresses for all but
1 also. But like you, I cannot think why this cannot be calculated. My
only suggestion is that perhaps it is a performance issue. But even
here, surely all such objects will have the same relationships, so the
extra addresses could be stored in the vtable, I would have thought.

We must be overlooking something (unless the book authors are)!