Update: Replaced the destructor example with a straight up method call example.
Hi,
If I have the following code:
class a
{
public:
virtual void func0(); // a has a VTable now
void func1();
};
class b : public a
{
public:
void func0() { a::func0(); }
void func2();
};
Is there a VTable in B? B has no virtual functions but calls a::func0() from b::func0()
Does func1 reside in a VTable? It's not virtual.
Does func2 reside in a VTable?
Will the answers to the above be different if there wasn't a a::func0() call in b::func0()?
Thanks
If you declare virtual functions you should also declare your destructor virtual ;-).
B has a virtual table, because it has a virtual function, namely func0(). If you declare a function (including a destructor) virtual in a base class, all its derived classes will have the function with same signature virtual as well. And it will cause them to have a vtable. Moreover, B would have the vtable even if you didn't declare func0 in it explicitly.
Non-virtual functions are not referenced through vtables.
See 2.
No. Classes' vtables are constructed based on class declarations. The bodies of class' functions (let alone other functions) are not taken into account. Therefore, B has a vtable, because its function func0() is virtual.
There also is a tricky detail, although it's not the gist of your question. You declared your function B::func0() as inline. In gcc compiler, if a virtual function is declared inline, it retains its slot in virtual table, the slot pointing to a special function emitted for that inline one (that counts as taking its address, which makes the inline emitted). That means, whether the funciton is inline doesn't influence amount of slots in vtable and its necessity for a class.
Yes, because its base class has one; also its destructor is virtual (even though you didn't declare it virtual) because the destructor of the base class is virtual.
No
No.
No. Actually I don't think the current code is legal: the compiler will invoke the A destructor after it invokes the B destructor even if you don't explicitly call ~A from ~B; so I don't think you should invoke ~A from ~B even if the compiler allows you to.
Referring to the updated example:
Yes, b has a vtable. Note that b::func0() is virtual (overrides a::func0()), even though you didn't explicitly label it as virtual. Weird C++ "loophole", I guess.
No. Non-virtual functions do not reside in the vtable.
See 2.
No. You've overridden a:func0(); it doesn't matter if you call a::func0() or not.
A few additional notes (compiler dependent, but these are pretty common generalizations):
Every instance of b will have a pointer to a vtable, because you're derived from a class that has virtual functions.
This would be the case even if you didn't define b::func0().
In this situation, the compiler might have instances of b point to a's static vtable, or it might create a static vtable for b and fill it with pointers to pointers to members of a.
But it's still required, so that you can properly access an instance of b via a pointer to a.
If base class function is virtual then if you override that function in derived class it is implicitly virtual even if you don't specify explicitly. If class has a virtual function then it has a v table.
Only virtual functions present in vtable, function1 will not reside in vtable
function2 will not reside in vtable same reason above
Creation of vtable does not depend on whether you call that function from base class or from somewhere else. Function call does not decide creation of vtable.
Related
Is there a good reason why a virtual function is virtual by default in a derived class and it is not even possible to remove virtuality completely in the derived function?
The reason why I want this behaviour is the same reason I do not want every function to be virtual by default. The compiler might generate vtables which costs performance.
There should be a vtable for the base class (I want polymorphism there), but no vtable for the derived class (I do not want polymorphism there, why should I want that, only because it derives of a class with polymorphism?).
The problem I want to solve is to better understand virtual functions. I just do not get this design decision and wonder if there is any reason for this.
A vtable for a particular class has pointers to the virtual functions for that particular class. A vtable for a base class does not point to the virtual functions overridden by a derived class. There could be many derived classes, and the base class knows about precisely zero of them, so there's no way to have pointers to those functions. Only a derived class's vtable may have pointers to the derived class's functions.
The point of a virtual function is that, if you convert an object of the derived class to a base class pointer/reference, then calling any virtual function declared in the base class must call the most derived class's version of that function for that object. The only way to do that is if the vtable for that object has a pointer to the most-derived-class's version of that function. And that can only happen if the vtable for that object is the derived class's vtable.
That's how vtable's work.
If the derived class doesn't have a vtable of its own, which points to its own overridden member functions rather than the base class functions, then virtual dispatch can't work.
I do not want polymorphism there, why should I want that, only because it derives of a class with polymorphism?
But you do want polymorphism. You asked for it. If you inherit from a polymorphic base class and override one of its virtual functions, you are declaring the intent to use polymorphism. If you didn't want polymorphism, you wouldn't inherit from a polymorphic base class.
I think you are confused about what a vtable does. A vtable for a class C deriving from B basically maps functions specifiers to function pointers for class C. So if you do
B* ptr = new C;
ptr->foo();
we have to look up at runtime, which function to call. In the end, we need a function pointer to C::foo since B cannot possibly know where that lies, since it does not know about C at all. Hence, here we need C's vtable to make the call. Now you could argue that we do not need Bs vtable here. But ptr could also to an Object of type B, so there you need the B vtable.
Arguably, one could do something like this (of course, pseudo code):
virtual_call_to_foo(B* ptr) {
if(ptr->is_actually_base_class) {
ptr->B::foo(); // non-virtual call
} else {
ptr->vtable["foo"](); // indirect call
}
}
But that wouldn't be better than just also using a vtable for B since you already have the indirection there.
Further, note that a call through a pointer to C won't be virtual if C::foo is final:
B* ptr = new C;
ptr->foo(); // virtual call
static_cast<C*>(ptr)->foo(); // non-virtual call, the compiler can inline this
When a method is declared, the declaring class decides if it is virtual, according to the nature of what the method does.
When you override such a method, your overriding method is a specialized version of doing the exact same thing. Your overriding method cannot decide that it is doing something that cannot be specialized, the base class has already made that decision. The nature of the method is decided by the class that originally declared it, the base class.
Vtables are an implementation detail that should not affect any of this.
I was trying to familiarize myself with the OOP concepts but could not quite understand the concept of virtual.
One can create a virtual destructor but not a virtual constructor. Why?
How are virtual destructors handled internally? I mean the link Virtual Destructors illustrates the concept but my question is how the vptr of both the vtables (Derived and Base) are called? (In case of virtual member functions when such a scenario occurs generally the function that vptr of Derived class points to is only called)
Are there any other scenarios where one may need to use a virtual destructor?
Can anyone please help me understand the above concepts with links/examples?
First, a little about the difference between virtual functions and non-virtual functions:
Every non-virtual function-call that you have in your code can be resolved during compilation or linkage.
By resolved, we mean that the address of the function can be computed by the compiler or the linker.
So in the object code created, the function-call can be replaced with an op-code for jumping to the address of that function in memory.
With virtual functions, you have the ability to invoke functions which can be resolved only during runtime.
Instead of explaining it, let's run through a simple scenario:
class Animal
{
virtual void Eat(int amount) = 0;
};
class Lion : public Animal
{
virtual void Eat(int amount) { ... }
};
class Tiger : public Animal
{
virtual void Eat(int amount) { ... }
};
class Tigon : public Animal
{
virtual void Eat(int amount) { ... }
};
class Liger : public Animal
{
virtual void Eat(int amount) { ... }
};
void Safari(Animal* animals[], int numOfAnimals, int amount)
{
for (int i=0; i<numOfAnimals; i++)
animals[i]->Eat(amount);
// A different function may execute at each iteration
}
As you can probably understand, the Safari function allows you to be flexible and feed different animals.
But since the exact type of each animal is not known until runtime, so is the exact Eat function to be called.
The constructor of a class cannot be virtual because:
Calling a virtual function of an object is performed through the V-Table of the object's class.
Every object holds a pointer to the V-Table of its class, but this pointer is initialized only at runtime, when the object is created.
In other words, this pointer is initialized only when the constructor is called, and therefore the constructor itself cannot be virtual.
Besides that, there is no sense for the constructor to be virtual in the first place.
The idea behind virtual functions is that you can call them without knowing the exact type of the object with which they are called.
When you create an object (i.e., when you implicitly call a constructor), you know exactly what type of object you are creating, so you have no need for this mechanism.
The destructor of a base-class has to be virtual because:
When you statically allocate an object whose class inherits from the base-class, then at the end of the function (if the object is local) or the program (if the object is global), the destructor of the class is automatically invoked, and in turn, invokes the destructor of the base-class.
In this case, there is no meaning to the fact that the destructor is virtual.
On the other hand, when you dynamically allocate (new) an object whose class inherits from the base-class, then you need to dynamically deallocate (delete) it at some later point in the execution of the program.
The delete operator takes a pointer to the object, where the pointer's type may be the base-class itself.
In such case, if the destructor is virtual, then the delete operator invokes the destructor of the class, which in turn invokes the destructor of the base-class.
But if the destructor is not virtual, then the delete operator invokes the destructor of the base-class, and the destructor of the actual class is never invoked.
Consider the following example:
class A
{
A() {...}
~A() {...}
};
class B: public A
{
B() {...}
~B() {...}
};
void func()
{
A* b = new B(); // must invoke the destructor of class 'B' at some later point
...
delete b; // the destructor of class 'B' is never invoked
}
One can create a virtual destructor but not a virtual constructor. Why?
Virtual functions are dispatched according to the type of the object they're called on. When a constructor is called, there is no object - it's the constructor's job to create one. Without an object, there's no possibility of virtual dispatch, so the constructor can't be virtual.
How are virtual destructors handled internally?
The internal details of virtual dispatch are implementation-defined; the language doesn't specify the implementation, only the behaviour. Typically, the destructor is called via a vtable just like any virtual function.
how the vptr of both the vtables (Derived and Base) are called?
Only the most-derived destructor will be called virtually. All destructors, virtual or not, will implicitly call the destructors of all member and direct base-class subobjects. (The situation is slightly more complicated in the presence of virtual inheritance; but that's beyond the scope of this question).
Are there any other scenarios where one may need to use a virtual destructor?
You need one in order to support polymorphic deletion; that is, to be able to delete an object of derived type via a pointer to a base type. Without a virtual destructor for the base type, that's not allowed, and will give undefined behaviour.
because a Virtual function is invoked at runtime phase, however constructors are invoked at initialization phase, object is not constructed. So it's meaningless to have a virtual constructor.
a. the reason why only the base class desctructor is invoked in your link, the destructor is not marked as virtual, so the desctructor address is linked to Base class destructor at compile/link time, and obviously the type of the pointer is Base instead of Derived at compile time.
b. for why both of Base and Derived constructors are invoked after adding virtual to Base desctructor. It's same behavior like below:
Derived d; // when d exits the lifecycle, both Derived and Base's desctructor will be invoked.
Suppose when you have at least one virtual function, you should have a virtual desctructor.
One can create a virtual destructor but not a virtual constructor.
Why?
I'll try and explain this in layman's terms.
A class in c++ only exists after it's constructor completes. Each base class exists prior to initialisation of derived class and its members (vtable links included). Hence, having a virtual constructor does not make sense (since to construct, you need to know the type). Furthermore (in c++), calling virtual functions from a constructor does not work (as the derived class's vtable part has not been set up). If one thinks about it carefully, allowing virtual functions to be called from a contructor opens up a can of worms (such as what if virtual functions of derived classes are called prior to member initialization).
As far as destructors are concerned, at the point of destruction, the vtable is "intact", and we (c++ runtime) are fully aware of the type (so to speak). The destructor of the most derived part of the type is found (if virtual, through vtable), and therefore that destructor, and naturally that of all bases can be called.
How are virtual destructors handled internally? I mean the link
Virtual Destructors illustrates the concept but my question is how the
vptr of both the vtables (Derived and Base) are called?
Destructors are handled the same as normal virtual functions (that is, there addresses are looked up in a vtable if they are virtual at the expense of one (perhaps 2?) extra level/s of indirection). Furthermore, c++ guarantees that all base destructors shall execute (in opposite order of construction which relies on order of declaration) after completion of a derived destructor.
One can mimick/simulate virtual construction by using patterns such as the prototype pattern (or cloning), or by using factories. In such cases either an instance of the real type exists (to be used polymorphically), or a factory exists (deriving from abstract factory), that creates a type (via virtual function) based on some knowledge provided.
Hope this helps.
I assume we have a Base class A, and it's derived B.
1.: You can delete B via an A pointer, and then the correct method is to call the B destructor too.
However, you just can't say, that a B object should be created while you actually just call the A constructor. There is just not such a case.
You can say:
A* a = new B ();
or
B b;
But both directly call the B's constructor.
2.: Well, i am not entirely sure, but i guess it will iterate through the relevant part of the class hierarchy, and search for the closest call of the function. If a function is not virtual, it stop iterating and call it.
3.: You should always use virtual destructor, if you want to inherit something from that class. If it's a final class, you shouldn't.
I wasted a couple of days trying to discover why my derived virtual destructors were not being called before discovering the answer so hopefully I can save other a lot of grief with this reply.
I started using derived classes three and four levels deep in my project. Virtual functions seemed to work fine but then I discovered I had massive memory leaks because my destructors were not being called. No compiler or runtime error - the destructors just were not being called.
There is a ton of documentation and examples about this on the web but none of it was useful because my syntax was correct.
I decided that if the compiler wasn't going to call my destructors, I needed to create my own virtual destruct method to call. Then I got the compiler error that solved the problem - "class if a Forward Reference". Adding an include for the derived class header files in the base class solved the problem. The compiler needs the class definition to call the destructor!
I suggest when creating a new derived class, include the header file in the base and intermediate classes. Probably also a good idea to add conditional debug code to your destructors to check that they are bing called.
Bob Rice
Consider this code:
class A {
public:
void fun() {}
};
class B : public A {
public:
void fun() {}
};
int main()
{
A *p = new B;
delete p;
}
Classes A and B are not polymorphic, and neither class declares a destructor. If I compile this code with g++ -Wall, the GCC compiler happily compiles the code.
But if I add virtual to void fun() in A, the compiler issues this warning: "deleting object of polymorphic class type ‘A’ which has non-virtual destructor might cause undefined behavior".
I'm quite aware of the dangers of using non-virtual destructors. But the code above makes me wonder about two things:
Why do I need to write an empty virtual destructor in the base class when I'm not using destructors at all?
Why is the empty virtual destructor not required if the base class contains no other virtual functions?
EDIT
It appears that I need to clarify the thing that bothers me:
The above code declares no destructors.
If I declare a virtual function, the compiler complains about the missing virtual destructor. My conclusion: If the class is polymorphic, I need to write a virtual destructor if delete p is to work correctly.
But if I declare no virtual function (as in the initial example above), the compiler does not complain about a missing virtual destructor. My conclusion: If the class is not polymorphic, I do not need to write a virtual desctructor, and delete p will work correctly anyway.
But that last conclusion sounds intuitively wrong to me. Is it wrong? Should the compiler have complained in both cases?
Following up on PaulMcKenzie's and KerrekSB's comments, here is the answer to the two questions in the original post:
The class always has a destructor, even when the programmer doesn't explicitly write one. It is necessary to declare an empty virtual destructor in order to prevent the system from automatically generating a non-virtual one.
In the sample code provided, you do need a virtual destructor, even when there is no other virtual function in the class. The fact that GCC doesn't complain in this case is probably a bug in the compiler (or at least a shortcoming).
The background for this is found in §5.3.5 of the C++11 standard, which says, "if the static type of the object to be deleted is different from its dynamic type, the static type shall be a base class of the dynamic type of the object to be deleted and the static type shall have a virtual destructor or the behavior is undefined." (Italics mine.)
You are making an upcasting, in other words: a polymorphic use of the class B. If the class A has not virtual members, the compiler does not generate VTABLE for class A and it is not require a virtual destructor (note that your upcasting has no sense without polymorphism use). While if the class A declares virtual members, a VTABLE is generated by compiler, in this case you should provide a virtual destructor.
If your want polymorphic behaviour you need to define at least one virtual function in order compiler should generate v-table for your class.
Because C++ class contains two special functions (constructor and destructor) which used for every object it a good choice to make destructor virtual.
When you write delete p you really call a destructor for object associated with pointer p. If you not declare a destructor as virtual you've get error prone code.
Before your declare at least one of member function as virtual compiler did not expect that you intend using your class as polymorphic. In C++ phylosophy: "You should not pay for functionallity that you will never use". E.g. in simple case destructor should not be virtual.
I have a class declared as follow:
class TestFoo {
public:
TestFoo();
virtual void virtualFunction();
void nonVirtualFunction();
};
that I try to implement this way
TestFoo::TestFoo(){}
void TestFoo::nonVirtualFunction(){}
which on compilation returns the error:
undefined reference to vtable for TestFoo
I tried :
TestFoo::TestFoo(){}
void TestFoo::nonVirtualFunction(){}
void TestFoo::virtualFunction(){}
which compiles ok which is consistent to the answers to these posts:
Undefined reference to vtable
undefined reference to vtable
What confuses me is that I thought the whole point of declaring a virtual function is that I would not need to define it. In this example, I am not planning to create any instance of TestFoo, but to create instances of (concrete) classes inheriting from TestFoo. But still, I want to define the function nonVirtualFunction for every subclasses of TestFoo.
Something I did not get right ?
Thanks !
the whole point of declaring a virtual function is that I would not
need to define it
Not quite, it says "I may want to replace the implementation of this function by something else in a derived class."
I may have misunderstood your question, but you seem to imply that you don't think you can define pure virtual member function in C++, which you can. You can declare one as follows.
virtual void virtualFunction() = 0;
Normally, a pure virtual function won't be defined, but of course you can. That says "There is no default implementation of this function because it won't always make sense, but I'll provide you with an implementation that you can opt into."
By the way, if a class has any virtual functions, you should also define a virtual destructor, as it is perfectly legal (and often recommended) to have a base class (smart) pointer to a derived class - without a virtual destructor, the object may not be deleted correctly.
... I thought the whole point of declaring a
virtual function is that I would not need to define it ...
For that facility you have a feature called pure virtual methods:
virtual void virtualFunction() = 0; // no linking error now
Note that, a virtual method cannot remain unimplemented. The reason is that for every virtual method declared inside a class body there has to be a vtable entry. Failing to find its body results in linking error.
Purpose of this restriction:
Unless a class is abstract - that is it has at least one virtual function - there is no way you can guarantee to the compiler that you are not going to declare an object of TestFoo. What happens when you do following:
DerivedOfTestFoo obj1;
TestFoo obj2 = obj1, *p = &obj2; // object slicing
p->virtualFunction(); // where is the body?
Other situation; in constructor there is no virtual mechanism:
TestFoo::TestFoo () {
this->virtualFunction(); // where is the body?
}
We can conclude that, compilers follow the rule, "Better to be safe than sorry". :)
Your description matches perfectly with the case of an abstract class. Declare your virtual function as:
virtual void VirtualFunction () = 0;
This means that you are not implementing the function in this class. As a result, the class becomes abstract. That is, no bare objects of this class can be instantiated.
Also, you should provide a virtual destructor.
Update: Some clarifications...
The language allows you to redefine a non-virtual function. Though, the wrong version might be called in some cases:
derived D; // rB is a reference to base class but it
base & rB=D; // points to an object of the derived class
rB.NonVirtualFunction (); // The base-class version is called
For this reason, redefining a non-virtual function is strongly discouraged nowadays. See Scott Meyers' "Effective C++, Third Edition: 55 Specific Ways to Improve Your Programs and Designs", item 36: "Never redefine an inherited non-virtual function."
See also item 7: "Declare destructors virtual in polymorphic base classes". An example:
base * pB = new derived;
delete pB; // If base's destructor is not virtual,
// ~derived() will not be called.
In case you wonder why isn't everything virtual by default, the reason is that calling a virtual function is slightly slower than calling a non-virtual one. Oh, and objects of classes with virtual functions occupy a few more bytes each.
If you want make this virtual function as pure virtual function,do not want to define it then, virtual void virtualFunction() = 0;
I will type an example :
class A
{
public:
virtual ~A(){}
};
class B: public A
{
public:
~B()
{
}
};
int main(void)
{
A * a = new B;
delete a;
return 0;
}
Now in Above Example , destructors will be called recursively bottom to up .
My Question is how Compiler do this MAGIC .
There are two different pieces of magic in your question. The first one is how does the compiler call the final overrider for the destructor and the second one is how does it then call all the other destructors in order.
Disclaimer: The standard does not mandate any particular way of performing this operations, it only mandates what the behavior of the operations at a higher level are. These are implementation details that are common to various implementations, but not mandated by the standard.
How does the compiler dispatch to the final overrider?
The first answer is the simple one, the same dynamic dispatch mechanism that is used for other virtual functions is used for destructors. To refresh it, each object stores a pointer (vptr) to each of its vtables (in the event of multiple inheritance there can be more than one), when the compiler sees a call to any virtual function, it follows the vptr of the static type of the pointer to find the vtable and then uses the pointer in that table to forward the call. In most cases the call can be directly dispatched, in others (multiple inheritance) it calls some intermediate code (thunk) that fixes the this pointer to refer to the type of the final overrider for that function.
How does the compiler then call the base destructors?
The process of destructing an object takes more operations than those you write inside the body of the destructor. When the compiler generates the code for the destructor, it adds extra code both before and after the user defined code.
Before the first line of a user defined destructor is called, the compiler injects code that will make the type of the object be that of the destructor being called. That is, right before ~derived is entered, the compiler adds code that will modify the vptr to refer to the vtable of derived, so that effectively, the runtime type of the object becomes derived (*).
After the last line of your user defined code, the compiler injects calls to the member destructors as well as base destructor(s). This is performed disabling dynamic dispatch, which means that it will no longer come all the way down to the just executed destructor. It is the equivalent of adding this->~mybase(); for each base of the object (in reverse order of declaration of the bases) at the end of the destructor.
With virtual inheritance, things get a bit more complex, but overall they follow this pattern.
EDIT (forgot the (*)):
(*) The standard mandates in §12/3:
When a virtual function is called directly or indirectly from a constructor (including from the mem-initializer for a data member) or from a destructor, and the object to which the call applies is the object under construction or destruction, the function called is the one defined in the constructor or destructor’s own class or in one of its bases, but not a function overriding it in a class derived from the con- structor or destructor’s class, or overriding it in one of the other base classes of the most derived object.
That requirement implies that the runtime type of the object is that of the class being constructed/destructed at this time, even if the original object that is being constructed/destructed is of a derived type. A simple test to verify this implementation can be:
struct base {
virtual ~base() { f(); }
virtual void f() { std::cout << "base"; }
};
struct derived : base {
void f() { std::cout << "derived"; }
};
int main() {
base * p = new derived;
delete p;
}
A virtual destructor is treated in the same way as any other virtual function. I note that you've correctly maked the base class's destructor as virtual. As such, it is no way different than any other virtual function, as far as dynamic dispatch is concerned. The most derived class destructor gets called through dynamic dispatch but it also automatically results in calls to Base class destructors of the class1.
Most compiler implements this feature using vtable and vptr, though the language specification does not mandate it. There can be a compiler which does this differently, without using vtable and vptr.
Anyway, as it true for most compilers, it is worth knowing what vtable is. So vtable is a table contains pointers of all virtual functions the class defines, and the compiler adds vptr to the class as hidden pointer which points to the correct vtable, so the compiler uses correct index, calculated at compile-time, to the vtable so as to dispatch the correct virtual function at runtime.
1. The italicized text is taken from #Als's comment. Thanks to him. It makes things more clear.
A suitable implementation of (virtual) destructors the compiler might use would be (in pseudocode)
class Base {
...
virtual void __destruct(bool should_delete);
...
};
void Base::__destruct(bool should_delete)
{
this->__vptr = &Base::vtable; // Base is now the most derived subobject
... your destructor code ...
members::__destruct(false); // if any, in the reverse order of declaration
base_classes::__destruct(false); // if any
if(should_delete)
operator delete(this); // this would call operator delete defined here, or inherited
}
This function gets defined even if you didn't define a destructor. Your code would just be empty in that case.
Now all derived classes would override (automatically) this virtual function:
class Der : public Base {
...
virtual void __destruct(bool should_delete);
...
};
void Der::__destruct(bool should_delete)
{
this->__vptr = &Der::vtable;
... your destructor code ...
members::__destruct(false);
Base::__destruct(false);
if(should_delete)
operator delete(this);
}
A call delete x, where x is of pointer to class type, would be translated as
x->__destruct(true);
and any other destructor call (implicit due to variable going out of scope, explicit x.~T()) would be
x.__destruct(false);
This results in
the most derived destructor always being called (for virtual destructors)
operator delete from the most derived object being called
all members' and base classes' destructors being called.
HTH. This should be understandable if you understand virtual functions.
As usual with virtual functions there will be some implementation mechanism (like a vtable pointer) that will let the compiler find which destructor to run first depending on the type of the object. Once the most derived class destructor is run it will in turn run the base class destructor and so on.
It's up to the compiler how to implement it and typically it's done with the same mechanism as other virtual methods. In other words there's nothing special about destructors that requires a virtual method dispatch mechanism that is distinct from that used by normal methods.
A virtual destructor has an entry in the virtual table just as other virtual functions do. When the destructor is invoked -either manually or automatically from a call to delete- the most derived version is invoked. A destructor also automatically calls the destructor for its base classes, so that in combination with the virtual dispatch is what causes the magic.
Unlike other virtual functions, when you override a virtual destructor, your object's virtual destructor is called in addition to any inherited virtual destructors.
Technically this can be achieved by whatever means the compiler chooses, but almost all compilers achieve this via static memory called a vtable, which permits polymorphism on functions and destructors. For each class in your source code, a static constant vtable is generated for it at compile time. When an object of type T is constructed at runtime, the object's memory is initialized with a hidden vtable pointer which points to the T's vtable in ROM. Inside the vtable is a list of member function pointers and a list of destructor function pointers. When a variable of any type that has a vtable goes out of scope or is deleted with delete or delete[], all of the destructor pointers in vtable the object points to are all invoked. (Some compilers choose to only store the most derived destructor pointer in the table, and then include a hidden invocation of the superclass's destructor in the body of every virtual destructor if one exists. This results in equivalent behavior.)
Additional magic is needed for virtual and nonvirtual multiple inheritance. Assume I am deleting a pointer p, where p is of the type of a base-class. We need to invoke the destructor of the sub-classes with this=p. But using multiple inheritance, p and the start of the derived object may not be the same! There is a fixed offset which must be applied. There is one such offset stored in the vtable for each class that is inherited, as well as a set of inherited offsets.
When you have a pointer to an object, it points to a block of memory that has both the data for that object and a 'vtable pointer.' In microsoft compilers, the vtable pointer is the first piece of data in the object. In Borland compilers, it is the last. Either way, it points to a vtable that represents a list of function vectors corresponding to the virtual methods that can be invoked for that object/class. The virtual destructor is just another vector in that list of function pointer vectors.