Chapter 14: Polymorphism

Don't hesitate to send in feedback: send an e-mail if you like the C++ Annotations; if you think that important material was omitted; if you find errors or typos in the text or the code examples; or if you just feel like e-mailing. Send your e-mail to Frank B. Brokken.

Please state the document version you're referring to, as found in the title (in this document: 9.9.1) and please state chapter and paragraph name or number you're referring to.

All received mail is processed conscientiously, and received suggestions for improvements are usually processed by the time a new version of the Annotations is released. Except for the incidental case I will normally not acknowledge the receipt of suggestions for improvements. Please don't interpret this as me not appreciating your efforts.

Using inheritance classes may be derived from other classes, called base classes. In the previous chapter we saw that base class pointers may be used to point to derived class objects. We also saw that when a base class pointer points to an object of a derived class it is the type of the pointer rather than the type of the object it points to what determines which member functions are visible. So when a Vehicle *vp, points to a Car object Car's speed or brandName members can't be used.

In the previous chapter two fundamental ways classes may be related to each other were discussed: a class may be implemented-in-terms-of another class and it can be stated that a derived class is-a base class. The former relationship is usually implemented using composition, the latter is usually implemented using a special form of inheritance, called polymorphism, the topic of this chapter.

An is-a relationship between classes allows us to apply the Liskov Substitution Principle (LSP) according to which a derived class object may be passed to and used by code expecting a pointer or reference to a base class object. In the C++ Annotations so far the LSP has been applied many times. Every time an ostringstream, ofstream or fstream was passed to functions expecting an ostream we've been applying this principle. In this chapter we'll discover how to design our own classes accordingly.

LSP is implemented using a technique called polymorphism: although a base class pointer is used it performs actions defined in the (derived) class of the object it actually points to. So, a Vehicle *vp might behave like a Car * when pointing to a Car (In one of the StarTrek movies, Capt. Kirk was in trouble, as usual. He met an extremely beautiful lady who, however, later on changed into a hideous troll. Kirk was quite surprised, but the lady told him: ``Didn't you know I am a polymorph?'').

Polymorphism is implemented using a feature called late binding. It's called that way because the decision which function to call (a base class function or a function of a derived class) cannot be made at compile-time, but is postponed until the program is actually executed: only then it is determined which member function will actually be called.

In C++ late binding is not the default way functions are called. By default static binding (or early binding) is used. With static binding the functions that are called are determined by the compiler, merely using the class types of objects, object pointers or object refences.

Late binding is an inherently different (and slightly slower) process as it is decided at run-time, rather than at compile-time what function is going to be called. As C++ supports both late- and early-binding C++ programmers are offered an option as to what kind of binding to use. Choices can be optimized to the situations at hand. Many other languages offering object oriented facilities (e.g., Java) only or by default offer late binding. C++ programmers should be keenly aware of this. Expecting early binding and getting late binding may easily produce nasty bugs.

Let's look at a simple example to start appreciating the differences between late and early binding. The example merely illustrates. Explanations of why things are as shown are shortly provided.

Consider the following little program:

    #include <iostream>
    using namespace std;

    class Base
    {
        protected:
            void hello()
            {
                cout << "base hello\n";
            }
        public:
            void process()
            {
                hello();
            }
    };
    class Derived: public Base
    {
        protected:
            void hello()
            {
                cout << "derived hello\n";
            }
    };
    int main()
    {
        Derived derived;

        derived.process();
    }
The important characteristic of the above program is the Base::process function, calling hello. As process is the only member that is defined in the public interface it is the only member that can be called by code not belonging to the two classes. The class Derived, derived from Base clearly inherits Base's interface and so process is also available in Derived. So the Derived object in main is able to call process, but not hello.

So far, so good. Nothing new, all this was covered in the previous chapter. One may wonder why Derived was defined at all. It was presumably defined to create an implementation of hello that's appropriate for Derived but differing from Base::hello's implementation. Derived's author's reasoning was as follows: Base's implementation of hello is not appropriate; a Derived class object can remedy that by providing an appropriate implementation. Furthermore our author reasoned:

``since the type of an object determines the interface that is used, process must call Derived::hello as hello is called via process from a Derived class object''.

Unfortunately our author's reasoning is flawed, due to static binding. When Base::process was compiled static binding caused the compiler to bind the hello call to Base::hello().

The author intended to create a Derived class that is-a Base class. That only partially succeeded: Base's interface was inherited, but after that Derived has relinquished all control over what happens. Once we're in process we're only able to see Base's member implementations. Polymorphism offers a way out, allowing us to redefine (in a derived class) members of a base class allowing these redefined members to be used from the base class's interface.

This is the essence of LSP: public inheritance should not be used to reuse the base class members (in derived classes) but to be reused (by the base class, polymorphically using derived class members reimplementing base class members).

Take a second to appreciate the implications of the above little program. The hello and process members aren't too impressive, but the implications of the example are. The process member could implement directory travel, hello could define the action to perform when encountering a file. Base::hello might simply show the name of a file, but Derived::hello might delete the file; might only list its name if its younger than a certain age; might list its name if it contains a certain text; etc., etc.. Up to now Derived would have to implement process's actions itself; Up to now code expecting a Base class reference or pointer could only perform Base's actions. Polymorphism allows us to reimplement members of base classes and to use those reimplemented members in code expecting base class references or pointers. Using polymorphism existing code may be reused by derived classes reimplementing the appropriate members of their base classes. It's about time to uncover how this magic can be realized.

Polymorphism, which is not the default in C++, solves the problem and allows the author of the classes to reach its goal. For the curious reader: prefix void hello() in the Base class with the keyword virtual and recompile. Running the modified program produces the intended and expected derived hello. Why this happens is explained next.

14.1: Virtual functions

By default the behavior of a member function called via a pointer or reference is determined by the implementation of that function in the pointer's or reference's class. E.g., a Vehicle * activates Vehicle's member functions, even when pointing to an object of a derived class. This is known as as early or static binding: the function to call is determined at compile-time. In C++ late or dynamic binding is realized using virtual member functions.

A member function becomes a virtual member function when its declaration starts with the keyword virtual. It is stressed once again that in C++, different from several other object oriented languages, this is not the default situation. By default static binding is used.

Once a function is declared virtual in a base class, it remains virtual in all derived classes. The keyword virtual should not be mentioned with members declared virtual in the base class. In derived classes those members should be provided with the override indicator.

In the vehicle classification system (see section 13.1), let's concentrate on the members mass and setMass. These members define the user interface of the class Vehicle. What we would like to accomplish is that this user interface can be used for Vehicle and for any class inheriting from Vehicle, since objects of those classes are themselves also Vehicles.

If we can define the user interface of our base class (e.g., Vehicle) such that it remains usable irrespective of the classes we derive from Vehicle our software achieves an enormous reusability: we design or software around Vehicle's user interface, and our software will also properly function for derived classes. Using plain inheritance doesn't accomplish this. If we define

    std::ostream &operator<<(std::ostream &out, Vehicle const &vehicle)
    {
        return out << "Vehicle's mass is " << vehicle.mass() << " kg.";
    }
and Vehicle's member mass returns 0, but Car's member mass returns 1000, then twice a mass of 0 is reported when the following program is executed:
    int main()
    {
        Vehicle vehicle;
        Car vw(1000);

        cout << vehicle << '\n' << vw << endl;
    }

We've defined an overloaded insertion operator, but since it only knows about Vehicle's user interface, `cout << vw' will use vw's Vehicle's user interface as well, thus displaying a mass of 0.

Reusablility is enhanced if we add a redefinable interface to the base class's interface. A redefinable interface allows derived classes to fill in their own implementation, without affecting the user interface. At the same time the user interface will behave according to the derived class's wishes, and not just to the base class's default implementation.

Members of the reusable interface should be declared in the class's private sections: conceptually they merely belong to their own classes (cf. section 14.7). In the base class these members should be declared virtual. These members can be redefined (overridden) by derived classes, and should there be provided with override indicators.

We keep our user interface (mass), and add the redefinable member vmass to Vehicle's interface:

    class Vehicle
    {
        public:
            int mass() const;
            int si_mass() const;    // see below

        private:
            virtual vmass() const;
    };
Separating the user interface from the redefinable interface is a sensible thing to do. It allows us to fine-tune the user interface (only one point of maintenance), while at the same time allowing us to standardize the expected behavior of the members of the redefinable interface. E.g., in many countries the International system of units is used, using the kilogram as the unit for mass. Some countries use other units (like the lbs: 1 kg being approx. 2.2046 lbs). By separating the user interface from the redefinable interface we can use one standard for the redefinable interface, and keep the flexibility of transforming the information ad-lib in the user interface.

Just to maintain a clean separation of user- and redefinable interface we might consider adding another accessor to Vehicle, providing the si_mass, simply implemented like this:

    int Vehicle::si_mass() const
    {
        return vmass();
    }

If Vehicle supports a member d_massFactor then its mass member can be implemented like this:

    int Vehicle::mass()
    {
        return d_massFactor * si_mass();
    }
Vehicle itself could define vmass so that it returns a token value. E.g.,
    int Vehicle::vmass()
    {
        return 0;
    }

Now let's have a look at the class Car. It is derived from Vehicle, and it inherits Vehicle's user interface. It also has a data member int d_mass, and it implements its own reusable interface:

    class Car: public Vehicle
    {
        ...
        private:
            int vmass() override;
    }
If Car constructors require us to specify the car's mass (stored in d_mass), then Car simply implements its vmass member like this:
    int Car::vmass() const
    {
        return d_mass;
    }

The class Truck, inheriting from Car needs two mass values: the tractor's mass and the trailer's mass. The tractor's mass is passed to its Car base class, the trailor's mass is passed to its Vehicle d_trailor data member. Truck, too, overrides vmass, this time returning the sum of its tractor and trailor masses:

    int Truck::vmass() const
    {
        return Car::si_mass() + d_trailer.si_mass();
    }

Once a class member has been declared virtual it becomes a virtual member in all derived classes, whether or not these members are provided with the override indicator. But override should be used, as it allows to compiler to catch typos when writing down the derived class interface.

A member function may be declared virtual anywhere in a class hierarchy, but this probably defeats the underlying polymorphic class design, as the original base class is no longer capable of completely covering the redefinable interfaces of derived classes. If, e.g, mass is declared virtual in Car, but not in Vehicle, then the specific characteristics of virtual member functions would only be available for Car objects and for objects of classes derived from Car. For a Vehicle pointer or reference static binding would remain to be used.

The effect of late binding (polymorphism) is illustrated below:

    void showInfo(Vehicle &vehicle)
    {
        cout << "Info: " << vehicle << '\n';
    }

    int main()
    {
        Car car(1200);            // car with mass 1200
        Truck truck(6000, 115,      // truck with cabin mass 6000, 
              "Scania", 15000);     // speed 115, make Scania, 
                                    // trailer mass 15000

        showInfo(car);             // see (1) below
        showInfo(truck);            // see (2) below

        Vehicle *vp = &truck;
        cout << vp->speed() << '\n';// see (3) below
    }
Now that mass is defined virtual, late binding is used: The example illustrates that when a pointer to a class is used only the members of that class can be called. A member's virtual characteristic only influences the type of binding (early vs. late), not the set of member functions that is visible to the pointer.

Through virtual members derived classes may redefine the behavior performed by functions called from base class members or from pointers or references to base class objects. This redefinition of base class members by derived classes is called overriding members.

14.2: Virtual destructors

When an object ceases to exist the object's destructor is called. Now consider the following code fragment (cf. section 13.1):
    Vehicle *vp = new Land(1000, 120);

    delete vp;          // object destroyed
Here delete is applied to a base class pointer. As the base class defines the available interface delete vp calls ~Vehicle and ~Land remains out of sight. Assuming that Land allocates memory a memory leak results. Freeing memory is not the only action destructors can perform. In general they may perform any action that's necessary when an object ceases to exist. But here none of the actions defined by ~Land are performed. Bad news....

In C++ this problem is solved by virtual destructors. A destructor can be declared virtual. When a base class destructor is declared virtual then the destructor of the actual class pointed to by a base class pointer bp is going to be called when delete bp is executed. Thus, late binding is realized for destructors even though the destructors of derived classes have unique names. Example:

    class Vehicle
    {
        public:
            virtual ~Vehicle();     // all derived class destructors are
                                    // now virtual as well.
    };
By declaring a virtual destructor, the above delete operation (delete vp) correctly calls Land's destructor, rather than Vehicle's destructor.

Once a destructor is called it performs as usual, whether or not it is a virtual destructor. So, ~Land first executes its own statements and then calls ~Vehicle. Thus, the above delete vp statement uses late binding to call ~Vehicle and from this point on the object destruction proceeds as usual.

Destructors should always be defined virtual in classes designed as a base class from which other classes are going to be derived. Often those destructors themselves have no tasks to perform. In these cases the virtual destructor is given an empty body. For example, the definition of Vehicle::~Vehicle() may be as simple as:

    Vehicle::~Vehicle()
    {}
Resist the temptation to define virtual destructors (even empty destructors) inline as this complicates class maintenance. Section 14.11 discusses the reason behind this rule of thumb.

14.3: Pure virtual functions

The base class Vehicle is provided with its own concrete implementations of its virtual members (mass and setMass). However, virtual member functions do not necessarily have to be implemented in base classes.

When the implementations of virtual members are omitted from base classes the class imposes requirements upon derived classes. The derived classes are required to provide the `missing implementations'.

This approach, in some languages (like C#, Delphi and Java) known as an interface, defines a protocol. Derived classes must obey the protocol by implementing the as yet not implemented members. If a class contains at least one member whose implementation is missing no objects of that class can be defined.

Such incompletely defined classes are always base classes. They enforce a protocol by merely declaring names, return values and arguments of some of their members. These classes are call abstract classes or abstract base classes. Derived classes become non-abtract classes by implementing the as yet not implemented members.

Abstract base classes are the foundation of many design patterns (cf. Gamma et al. (1995)) , allowing the programmer to create highly reusable software. Some of these design patterns are covered by the C++ Annotations (e.g, the Template Method in section 23.2), but for a thorough discussion of design patterns the reader is referred to Gamma et al.'s book.

Members that are merely declared in base classes are called pure virtual functions. A virtual member becomes a pure virtual member by postfixing = 0 to its declaration (i.e., by replacing the semicolon ending its declaration by `= 0;'). Example:

    #include <iosfwd>
    class Base
    {
        public:
            virtual ~Base();
            virtual std::ostream &insertInto(std::ostream &out) const = 0;
    };
    inline std::ostream &operator<<(std::ostream &out, Base const &base)
    {
        return base.insertInto(out);
    }
All classes derived from Base must implement the insertInto member function, or their objects cannot be constructed. This is neat: all objects of class types derived from Base can now always be inserted into ostream objects.

Could the virtual destructor of a base class ever be a pure virtual function? The answer to this question is no. First of all, there is no need to enforce the availability of destructors in derived classes as destructors are provided by default (unless a destructor is declared with the = delete attribute using the new C++11 standard). Second, if it is a pure virtual member its implementation does not exist. However, derived class destructors eventually call their base class destructors. How could they call base class destructors if their implementations are lacking? More about this in the next section.

Often, but not necessarily, pure virtual member functions are const member functions. This allows the construction of constant derived class objects. In other situations this might not be necessary (or realistic), and non-constant member functions might be required. The general rule for const member functions also applies to pure virtual functions: if the member function alters the object's data members, it cannot be a const member function.

Abstract base classes frequently don't have data members. However, once a base class declares a pure virtual member it must be declared identically in derived classes. If the implementation of a pure virtual function in a derived class alters the derived class object's data, then that function cannot be declared as a const member. Therefore, the author of an abstract base class should carefully consider whether a pure virtual member function should be a const member function or not.

14.3.1: Implementing pure virtual functions

Pure virtual member functions may be implemented. To implement a pure virtual member function, provide it with its normal = 0; specification, but implement it as well. Since the = 0; ends in a semicolon, the pure virtual member is always at most a declaration in its class, but an implementation may either be provided outside from its interface (maybe using inline).

Pure virtual member functions may be called from derived class objects or from its class or derived class members by specifying the base class and scope resolution operator together with the member to be called. Example:

#include <iostream>

class Base
{
    public:
        virtual ~Base();
        virtual void pureimp() = 0;
};
Base::~Base()
{}
void Base::pureimp()
{
    std::cout << "Base::pureimp() called\n";
}
class Derived: public Base
{
    public:
        virtual void pureimp();
};
inline void Derived::pureimp()
{
    Base::pureimp();
    std::cout << "Derived::pureimp() called\n";
}
int main()
{
    Derived derived;

    derived.pureimp();
    derived.Base::pureimp();

    Derived *dp = &derived;

    dp->pureimp();
    dp->Base::pureimp();
}
// Output:
//      Base::pureimp() called
//      Derived::pureimp() called
//      Base::pureimp() called
//      Base::pureimp() called
//      Derived::pureimp() called
//      Base::pureimp() called

Implementing a pure virtual member has limited use. One could argue that the pure virtual member function's implementation may be used to perform tasks that can already be performed at the base class level. However, there is no guarantee that the base class virtual member function is actually going to be called. Therefore base class specific tasks could as well be offered by a separate member, without blurring the distinction between a member doing some work and a pure virtual member enforcing a protocol.

14.4: Explicit virtual overrides

Consider the following situations: C++11 allows the use of two special identifiers, final and override to realize the above. These identifiers are special in the sense that they only require their special meanings in specific contexts. Outside of this context they are just plain identifiers, allowing the programmer to define a variable like bool final.

The identifier final can be applied to class declarations to indicate that the class cannot be used as a base class. E.g.:

    class Base1 final               // cannot be a base class
    {};
    class Derived1: public Base1    // ERR: Base1 is final
    {};

    class Base2                     // OK as base class
    {};
    class Derived2 final: public Base2  // OK, but Derived2 can't be
    {};                                 //     used as a base class
    class Derived: public Derived2      // ERR: Derived2 is final
    {};
The identifier final can also be added to virtual member declarations. This indicates that those virtual members cannot be overridden by derived classes. The restricted polymorphic character of a class, mentioned above, can thus be realized as follows:
    class Base
    {
        virtual int v_process();    // define polymorphic behavior
        virtual int v_call();
        virtual int v_display();
    };
    class Derived: public Base      // Derived restricts polymorphism
    {                               // to v_call and v_display
        virtual int v_process() final;
    };
    class Derived2: public Derived
    {
        // int v_process();            No go: Derived:v_process is final
        virtual int v_display();    // OK to override
    };
To allow the compiler to detect typos, differences in parameter types, or differences in member function modifiers (e.g., const vs. non-const) the identifier override can (should) be appended to derived class members overriding base class members. E.g.,
    class Base
    {
        virtual int v_process();
        virtual int v_call() const;
        virtual int v_display(std::ostream &out);
    };
    class Derived: public Base
    {
        virtual int v_proces() override;    // ERR: v_proces != v_process
        virtual int v_call() override;      // ERR: not const
                                            // ERR: parameter types differ
        virtual int v_display(std::istream &out) override;
    };

14.5: Virtual functions and multiple inheritance

In chapter 6 we encountered the class fstream, one class offering features of ifstream and ofstream. In chapter 13 we learned that a class may be derived from multiple base classes. Such a derived class inherits the properties of all its base classes. Polymorphism can also be used in combination with multiple inheritance.

Consider what would happen if more than one `path' leads from the derived class up to its (base) classes. This is illustrated in the next (fictitious) example where a class Derived is doubly derived from Base:

    class Base
    {
        int d_field;
        public:
            void setfield(int val);
            int field() const;
    };
    inline void Base::setfield(int val)
    {
        d_field = val;
    }
    inline int Base::field() const
    {
        return d_field;
    }

    class Derived: public Base, public Base
    {
    };
Due to the double derivation, Base's functionality now occurs twice in Derived. This results in ambiguity: when the function setfield() is called for a Derived class object, which function will that be as there are two of them? The scope resolution operator won't come to the rescue and so the C++ compiler cannot compile the above example and (correctly) identifies an error.

The above code clearly duplicates its base class in the derivation, which can of course easily be avoided by not doubly deriving from Base (or by using composition (!)). But duplication of a base class can also occur through nested inheritance, where an object is derived from, e.g., a Car and from an Air (cf. section 13.1). Such a class would be needed to represent, e.g., a flying car (such as the one in James Bond vs. the Man with the Golden Gun...). An AirCar would ultimately contain two Vehicles, and hence two mass fields, two setMass() functions and two mass() functions. Is this what we want?

14.5.1: Ambiguity in multiple inheritance

Let's investigate closer why an AirCar introduces ambiguity, when derived from Car and Air. The duplication of Vehicle data is further illustrated in Figure 14.

Figure 14 is shown here.
Figure 14: Duplication of a base class in multiple derivation.


The internal organization of an AirCar is shown in Figure 15

Figure 15 is shown here.
Figure 15: Internal organization of an AirCar object.


The C++ compiler detects the ambiguity in an AirCar object, and will therefore not compile statements like:

    AirCar jBond;
    cout << jBond.mass() << '\n';
Which member function mass to call cannot be determined by the compiler but the programmer has two possibilities to resolve the ambiguity for the compiler: The second possibility is preferred as it does not require the compiler to flag an error; nor does it require the programmer using the class AirCar to take special precautions.

However, there exists a more elegant solution, discussed in the next section.

14.5.2: Virtual base classes

As illustrated in Figure 15, an AirCar represents two Vehicles. This not only results in an ambiguity about which function to use to access the mass data, but it also defines two mass fields in an AirCar. This is slightly redundant, since we can assume that an AirCar has but one mass.

It is, however, possible to define an AirCar as a class consisting of but one Vehicle and yet using multiple derivation. This is realized by defining the base classes that are multiply mentioned in a derived class's inheritance tree as a virtual base class.

For the class AirCar this implies a small change when deriving an AirCar from Land and Air classes:

    class Land: virtual public Vehicle
    {
        // etc
    };
    class Car: public Land
    {
        // etc
    };
    class Air: virtual public Vehicle
    {
        // etc
    };
    class AirCar: public Car, public Air
    {
    };
Virtual derivation ensures that a Vehicle is only added once to a derived class. This means that the route along which a Vehicle is added to an AirCar is no longer depending on its direct base classes; we can only state that an AirCar is a Vehicle. The internal organization of an AirCar after virtual derivation is shown in Figure 16.

Figure 16 is shown here.
Figure 16: Internal organization of an AirCar object when the base classes are virtual.


When a class Third inherits from a base class Second which in turn inherits from a base class First then the First class constructor called by the Second class constructor is also used when this Second constructor is used when constructing a Third object. Example:

    class First
    {
        public:
            First(int x);
    };
    class Second: public First
    {
        public:
            Second(int x)
            :
                First(x)
            {}
    };
    class Third: public Second
    {
        public:
            Third(int x)
            :
                Second(x)           // calls First(x)
            {}
    };
The above no longer holds true when Second uses virtual derivation. When Second uses virtual derivation its base class constructor is ignored when Second's constructor is called from Third. Instead Second by default calls First's default constructor. This is illustrated by the next example:
    class First
    {
        public:
            First()
            {
                cout << "First()\n";
            }
            First(int x);
    };
    class Second: public virtual First      // note: virtual
    {
        public:
            Second(int x)
            :
                First(x)
            {}
    };
    class Third: public Second
    {
        public:
            Third(int x)
            :
                Second(x)
            {}
    };
    int main()
    {
        Third third(3);     // displays `First()'
    }

When constructing Third First's default constructor is used by default. Third's constructor, however, may overrule this default behavior by explicitly specifying the constructor to use. Since the First object must be available before Second can be constructed it must be specified first. To call First(int) when constructing Third(int) the latter constructor can be defined as follows:

    class Third: public Second
    {
        public:
            Third(int x)
            :
                First(x),           // now First(int) is called.
                Second(x)
            {}
    };
This behavior may seem puzzling when simple linear inheritance is used but it makes sense when multiple inheritance is used with base classes using virtual inheritance. Consider AirCar: when Air and Car both virtually inherit from Vehicle will Air and Car both initialize the common Vehicle object? If so, which one is going to be called first? What if Air and Car use different Vehicle constructors? All these questions can be avoided by passing the responsibility for the initialization of a common base class to the class eventually using the common base class object. In the above example Third. Hence Third is provided an opportunity to specify the constructor to use when initializing First.

Multiple inheritance may also be used to inherit from classes that do not all use virtual inheritance. Assume we have two classes, Derived1 and Derived2, both (possibly virtually) derived from Base.

We now address the question which constructors will be called when calling a constructor of the class Final: public Derived1, public Derived2.

To distinguish the involved constructors Base1 indicates the Base class constructor called as base class initializer for Derived1 (and analogously: Base2 called from Derived2). A plain Base indicates Base's default constructor.

Derived1 and Derived2 indicate the base class initializers used when constructing a Final object.

Now we're ready to distinguish the various cases when constructing an object of the class Final: public Derived1, public Derived2:

Virtual derivation is, in contrast to virtual functions, a pure compile-time issue. Virtual inheritance merely defines how the compiler defines a class's data organization and construction process.

14.5.3: When virtual derivation is not appropriate

Virtual inheritance can be used to merge multiply occurring base classes. However, situations may be encountered where multiple occurrences of base classes is appropriate. Consider the definition of a Truck (cf. section 13.5):
    class Truck: public Car
    {
        int d_trailer_mass;

        public:
            Truck();
            Truck(int engine_mass, int sp, char const *nm,
                   int trailer_mass);

            void setMass(int engine_mass, int trailer_mass);
            int mass() const;
    };
    Truck::Truck(int engine_mass, int sp, char const *nm,
                  int trailer_mass)
    :
        Car(engine_mass, sp, nm)
    {
        d_trailer_mass = trailer_mass;
    }
    int Truck::mass() const
    {
        return                  // sum of:
            Car::mass() +    //   engine part plus
            trailer_mass;         //   the trailer
    }
This definition shows how a Truck object is constructed to contain two mass fields: one via its derivation from Car and one via its own int d_trailer_mass data member. Such a definition is of course valid, but it could also be rewritten. We could derive a Truck from a Car and from a Vehicle, thereby explicitly requesting the double presence of a Vehicle; one for the mass of the engine and cabin, and one for the mass of the trailer. A slight complication is that a class organization like
    class Truck: public Car, public Vehicle
is not accepted by the C++ compiler. As a Vehicle is already part of a Car, it is therefore not needed once again. This organzation may, however be forced using a small trick. By creating an additional class inheriting from Vehicle and deriving Truck from that additional class rather than directly from Vehicle the problem is solved. Simply derive a class TrailerVeh from Vehicle, and then Truck from Car and TrailerVeh:
    class TrailerVeh: public Vehicle
    {
        public:
            TrailerVeh(int mass)
            :
                Vehicle(mass)
            {}
    };
    class Truck: public Car, public TrailerVeh
    {
        public:
            Truck();
            Truck(int engine_mass, int sp, char const *nm, int trailer_mass);
            void setMass(int engine_mass, int trailer_mass);
            int mass() const;
    };
    inline Truck::Truck(int engine_mass, int sp, char const *nm,
                        int trailer_mass)
    :
        Car(engine_mass, sp, nm),
        TrailerVeh(trailer_mass)
    {}
    inline int Truck::mass() const
    {
        return                      // sum of:
            Car::mass() +        //   engine part plus
            TrailerVeh::mass();   //   the trailer
    }

14.6: Run-time type identification

C++ offers two ways to (run-time) retrieve the type of objects and expressions. The possibilities of C++'s run-time type identification are limited compared to languages like Java. Usually static type checking and static type identification is used in C++. Static type checking is possibly safer and certainly more efficient than run-time type identification and should therefore be preferred over run-time type identification. But situations exist where run-time type identification is appropriate. C++ offers run-time type identification through the dynamic cast and typeid operators. These operators can be used with objects of classes having at least one virtual member function.

14.6.1: The dynamic_cast operator

The dynamic_cast<> operator is used to convert a base class pointer or reference to, respectively, a derived class pointer or reference. This is also called down-casting as direction of the cast is down the inheritance tree.

A dynamic cast's actions are determined run-time; it can only be used if the base class declares at least one virtual member function. For the dynamic cast to succeed, the destination class's Vtable must be equal to the Vtable to which the dynamic cast's argument refers to, lest the cast fails and returns 0 (if a dynamic cast of a pointer was requested) or throws a std::bad_cast exception (if a dynamic cast of a reference was requested).

In the following example a pointer to the class Derived is obtained from the Base class pointer bp:

    class Base
    {
        public:
            virtual ~Base();
    };
    class Derived: public Base
    {
        public:
            char const *toString();
    };
    inline char const *Derived::toString()
    {
        return "Derived object";
    }
    int main()
    {
        Base *bp;
        Derived *dp,
        Derived d;

        bp = &d;

        dp = dynamic_cast<Derived *>(bp);

        if (dp)
            cout << dp->toString() << '\n';
        else
            cout << "dynamic cast conversion failed\n";
    }
In the condition of the above if statement the success of the dynamic cast is verified. This verification is performed at run-time, as the actual class of the objects to which the pointer points is only known by then.

If a base class pointer is provided, the dynamic cast operator returns 0 on failure and a pointer to the requested derived class on success.

Assume a vector<Base *> is used. Such a vector's pointers may point to objects of various classes, all derived from Base. A dynamic cast returns a pointer to the specified class if the base class pointer indeed points to an object of the specified class and returns 0 otherwise.

We could determine the actual class of an object a pointer points to by performing a series of checks to find the derived class to which a base class pointer points. Example:

    class Base
    {
        public:
            virtual ~Base();
    };
    class Derived1: public Base;
    class Derived2: public Base;

    int main()
    {
        vector<Base *> vb(initializeBase());

        Base *bp = vb.front();

        if (dynamic_cast<Derived1 *>(bp))
            cout << "bp points to a Derived1 class object\n";
        else if (dynamic_cast<Derived2 *>(bp))
            cout << "bp points to a Derived2 class object\n";
    }

Alternatively, a reference to a base class object may be available. In this case the dynamic_cast operator throws an exception if the down casting fails. Example:

    #include <iostream>
    #include <typeinfo>

    class Base
    {
        public:
            virtual ~Base();
            virtual char const *toString();
    };
    inline char const *Base::toString()
    {
        return "Base::toString() called";
    }
    class Derived1: public Base
    {};
    class Derived2: public Base
    {};

    Base::~Base()
    {}
    void process(Base &b)
    {
        try
        {
            std::cout << dynamic_cast<Derived1 &>(b).toString() << '\n';
        }
        catch (std::bad_cast)
        {}
        try
        {
            std::cout << dynamic_cast<Derived2 &>(b).toString() << '\n';
        }
        catch (std::bad_cast)
        {
            std::cout << "Bad cast to Derived2\n";
        }
    }
    int main()
    {
        Derived1 d;
        process(d);
    }
    /*
        Generated output:

        Base::toString() called
        Bad cast to Derived2
    */
In this example the value std::bad_cast is used. A std::bad_cast exception is thrown if the dynamic cast of a reference to a derived class object fails.

Note the form of the catch clause: bad_cast is the name of a type. Section 17.4.1 describes how such a type can be defined.

The dynamic cast operator is a useful tool when an existing base class cannot or should not be modified (e.g., when the sources are not available), and a derived class may be modified instead. Code receiving a base class pointer or reference may then perform a dynamic cast to the derived class to access the derived class's functionality.

You may wonder in what way the behavior of the dynamic_cast differs from that of the static_cast.

When the static_cast is used, we tell the compiler that it must convert a pointer or reference to its expression type to a pointer or reference of its destination type. This holds true whether the base class declares virtual members or not. Consequently, all the static_cast's actions can be determined by the compiler, and the following compiles fine:

    class Base
    {
        // maybe or not virtual members
    };
    class Derived1: public Base
    {};
    class Derived2: public Base
    {};

    int main()
    {
        Derived1 derived1;
        Base *bp = &derived1;

        Derived1 &d1ref = static_cast<Derived1 &>(*bp);
        Derived2 &d2ref = static_cast<Derived2 &>(*bp);
    }
Pay attention to the second static_cast: here the Base class object is cast to a Derived2 class reference. The compiler has no problems with this, as Base and Derived2 are related by inheritance.

Semantically, however, it makes no sense as bp in fact points to a Derived1 class object. This is detected by a dynamic_cast. A dynamic_cast, like the static_cast, converts related pointer or reference types, but the dynamic_cast provides a run-time safeguard. The dynamic cast fails when the requested type doesn't match the actual type of the object we're pointing at. In addition, the dynamic_cast's use is much more restricted than the static_cast's use, as the dynamic_cast can only be used for downcasting to derived classes having virtual members.

In the end a dynamic cast is a cast, and casts should be avoided whenever possible. When the need for dynamic casting arises ask yourself whether the base class has correctly been designed. In situations where code expects a base class reference or pointer the base class interface should be all that is required and using a dynamic cast should not be necessary. Maybe the base class's virtual interface can be modified so as to prevent the use of dynamic casts. Start frowning when encountering code using dynamic casts. When using dynamic casts in your own code always properly document why the dynamic cast was appropriately used and was not avoided (for an example where a dynamic cast is used on purpose, see section 23.8.3).

14.6.2: The `typeid' operator

As with the dynamic_cast operator, typeid is usually applied to references to base class objects that refer to derived class objects. Typeid should only be used with base classes offering virtual members. Before using typeid the <typeinfo> header file must have been included.

The typeid operator returns an object of type type_info. Different compilers may offer different implementations of the class type_info, but at the very least typeid must offer the following interface:

    class type_info
    {
        public:
            virtual ~type_info();
            int operator==(type_info const &other) const;
            int operator!=(type_info const &other) const;
            bool before(type_info const &rhs) const
            char const *name() const;
        private:
            type_info(type_info const &other);
            type_info &operator=(type_info const &other);
    };
Note that this class has a private copy constructor and a private overloaded assignment operator. This prevents code from constructing type_info objects and prevents code from assigning type_info objects to each other. Instead, type_info objects are constructed and returned by the typeid operator.

If the typeid operator is passed a base class reference it is able to return the actual name of the type the reference refers to. Example:

    class Base;
    class Derived: public Base;

    Derived d;
    Base    &br = d;

    cout << typeid(br).name() << '\n';
In this example the typeid operator is given a base class reference. It prints the text ``Derived'', being the class name of the class br actually refers to. If Base does not contain virtual functions, the text ``Base'' is printed.

The typeid operator can be used to determine the name of the actual type of expressions, not just of class type objects. For example:

    cout << typeid(12).name() << '\n';     // prints:  int
    cout << typeid(12.23).name() << '\n';  // prints:  double
Note, however, that the above example is suggestive at most. It may print int and double, but this is not necessarily the case. If portability is required, make sure no tests against these static, built-in text-strings are required. Check out what your compiler produces in case of doubt.

In situations where the typeid operator is applied to determine the type of a derived class, a base class reference should be used as the argument of the typeid operator. Consider the following example:

    class Base;     // contains at least one virtual function
    class Derived: public Base;

    Base *bp = new Derived;     // base class pointer to derived object

    if (typeid(bp) == typeid(Derived *))    // 1: false
        ...
    if (typeid(bp) == typeid(Base *))       // 2: true
        ...
    if (typeid(bp) == typeid(Derived))      // 3: false
        ...
    if (typeid(bp) == typeid(Base))         // 4: false
        ...
    if (typeid(*bp) == typeid(Derived))     // 5: true
        ...
    if (typeid(*bp) == typeid(Base))        // 6: false
        ...

    Base &br = *bp;

    if (typeid(br) == typeid(Derived))      // 7: true
        ...
    if (typeid(br) == typeid(Base))         // 8: false
        ...
Here, (1) returns false as a Base * is not a Derived *. (2) returns true, as the two pointer types are the same, (3) and (4) return false as pointers to objects are not the objects themselves.

On the other hand, if *bp is used in the above expressions, then (1) and (2) return false as an object (or reference to an object) is not a pointer to an object, whereas (5) now returns true: *bp actually refers to a Derived class object, and typeid(*bp) returns typeid(Derived). A similar result is obtained if a base class reference is used: 7 returning true and 8 returning false.

The type_info::before(type_info const &rhs) member is used to determine the collating order of classes. This is useful when comparing two types for equality. The function returns a nonzero value if *this precedes rhs in the hierarchy or collating order of the used types. When a derived class is compared to its base class the comparison returns 0, otherwise a non-zero value. E.g.:

    cout << typeid(ifstream).before(typeid(istream)) << '\n' << // not 0
            typeid(istream).before(typeid(ifstream)) << '\n';   // 0
With built-in types the implementor may implement that non-0 is returned when a `wider' type is compared to a `smaller' type and 0 otherwise:
    cout << typeid(double).before(typeid(int)) << '\n' <<   // not 0
            typeid(int).before(typeid(double)) << '\n';     // 0
When two equal types are compared, 0 is returned:
    cout << typeid(ifstream).before(typeid(ifstream)) << '\n';   // 0
When a 0-pointer is passed to the operator typeid a bad_typeid exception is thrown.

14.7: Inheritance: when to use to achieve what?

Inheritance should not be applied automatically and thoughtlessly. Often composition can be used instead, improving on a class's design by reducing coupling. When inheritance is used public inheritance should not automatically be used but the type of inheritance that is selected should match the programmer's intent.

We've seen that polymorphic classes on the one hand offer interface members defining the functionality that can be requested of base classes and on the other hand offer virtual members that can be overridden. One of the signs of good class design is that member functions are designed according to the principle of `one function, one task'. In the current context: a class member should either be a member of the class's public or protected interface or it should be available as a virtual member for reimplementation by derived classes. Often this boils down to virtual members that are defined in the base class's private section. Those functions shouldn't be called by code using the base class, but they exist to be overridden by derived classes using polymorphism to redefine the base class's behavior.

The underlying principle was mentioned before in the introductional paragraph of this chapter: according to the Liskov Substitution Principle (LSP) an is-a relationship between classes (indicating that a derived class object is a base class object) implies that a derived class object may be used in code expecting a base class object.

In this case inheritance is used not to let the derived class use the facilities already implemented by the base class but to reuse the base class polymorphically by reimplementing the base class's virtual members in the derived class.

In this section we'll discuss the reasons for using inheritance. Why should inheritance (not) be used? If it is used what do we try to accomplish by it?

Inheritance often competes with composition. Consider the following two alternative class designs:

    class Derived: // derived from Base
    { ... };

    class Composed
    {
        Base d_base;
        ...
    };
Why and when prefer Derived over Composed and vice versa? What kind of inheritance should be used when designing the class Derived?

Private inheritance should also be used when a derived class is-a certain type of base class, but in order to initialize that base class an object of another class type must be available. Example: a new istream class-type (say: a stream IRandStream from which random numbers can be extracted) is derived from std::istream. Although an istream can be constructed empty (receiving its streambuf later using its rdbuf member), it is clearly preferable to initialize the istream base class right away.

Assuming that a Randbuffer: public std::streambuf has been created for generating random numbers then IRandStream can be derived from Randbuffer and std::istream. That way the istream base class can be initialized using the Randbuffer base class.

As a RandStream is definitely not a Randbuffer public inheritance is not appropriate. In this case IRandStream is-implemented-in-terms-of a Randbuffer and so private inheritance should be used.

IRandStream's class interface should therefore start like this:

    class IRandStream: private Randbuffer, public std::istream
    {
        public:
            IRandStream(int lowest, int highest)    // defines the range
            :
                Randbuffer(lowest, highest),
                std::istream(this)                  // passes &Randbuffer
            {}
        ...
    };

Public inheritance should be reserved for classes for which the LSP holds true. In those cases the derived classes can always be used instead of the base class from which they derive by code merely using base class references, pointers or members (I.e., conceptually the derived class is-a base class). This most often applies to classes derived from base classes offering virtual members. To separate the user interface from the redefinable interface the base class's public interface should not contain virtual members (except for the virtual destructor) and the virtual members should all be in the base class's private section. Such virtual members can still be overridden by derived classes (this should not come as a surprise, considering how polymorphism is implemented) and this design offers the base class full control over the context in which the redefined members are used. Often the public interface merely calls a virtual member, but those members can always be redefined to perform additional duties.

The prototypical form of a base class therefore looks like this:

    class Base
    {
        public:
            virtual ~Base()
            void process();             // calls virtual members (e.g.,
                                        // v_process)
        private:
            virtual void v_process();   // overridden by derived classes
    };
Alternatively a base class may offer a non-virtual destructor, which should then be protected. It shouldn't be public to prevent deleting objects through their base class pointers (in which case virtual destructors should be used). It should be protected to allow derived class destructors to call their base class destructors. Such base classes should, for the same reasons, have non-public constructors and overloaded assignment operators.

14.8: The `streambuf' class

The class std::streambuf receives the character sequences processed by streams and defines the interface between stream objects and devices (like a file on disk). A streambuf object is usually not directly constructed, but usually it is used as base class of some derived class implementing the communication with some concrete device.

The primary reason for existence of the class streambuf is to decouple the stream classes from the devices they operate upon. The rationale here is to add an extra layer between the classes allowing us to communicate with devices and the devices themselves. This implements a chain of command which is seen regularly in software design.

The chain of command is considered a generic pattern when designing reusable software, encountered also in, e.g., the TCP/IP stack.

A streambuf can be considered yet another example of the chain of command pattern. Here the program talks to stream objects, which in turn forward their requests to streambuf objects, which in turn communicate with the devices. Thus, as we will see shortly, we are able to do in user-software what had to be done via (expensive) system calls before.

The class streambuf has no public constructor, but does make available several public member functions. In addition to these public member functions, several member functions are only available to classes derived from streambuf. In section 14.8.2 a predefined specialization of the class streambuf is introduced. All public members of streambuf discussed here are also available in filebuf.

The next section shows the streambuf members that may be overridden when deriving classes from streambuf. Chapter 23 offers concrete examples of classes derived from streambuf.

The class streambuf is used by streams performing input operations and by streams performing output operations and their member functions can be ordered likewise. The type std::streamsize used below may, for all practical purposes, be considered equal to the type size_t.

When inserting information into ostream objects the information is eventually passed on to the ostream's streambuf. The streambuf may decide to throw an exception. However, this exception does not leave the ostream using the streambuf. Rather, the exception is caught by the ostream, which sets its ios::bad_bit. Exception raised by manipulators inserted into ostream objects are not caught by the ostream objects.

Public members for input operations

Public members for output operations

Public members for miscellaneous operations

The next three members are normally only used by classes derived from streambuf.

14.8.1: Protected `streambuf' members

The protected members of the class streambuf are important for understanding and using streambuf objects. Although there are both protected data members and protected member functions defined in the class streambuf the protected data members are not mentioned here as using them would violate the principle of data hiding. As streambuf's set of member functions is quite extensive, it is hardly ever necessary to use its data members directly. The following subsections do not even list all protected member functions but only those are covered that are useful for constructing specializations.

Streambuf objects control a buffer, used for input and/or output, for which begin-, actual- and end-pointers have been defined, as depicted in figure 17.

Figure 17 is shown here.
Figure 17: Input- and output buffer pointers of the class `streambuf'


Streambuf offers one protected constructor:

14.8.1.1: Protected members for input operations

Several protected member functions are available for input operations. The member functions marked virtual may or course be redefined in derived classes:

14.8.1.2: Protected members for output operations

The following protected members are available for output operations. Again, some members may be overridden by derived classes:

14.8.1.3: Protected members for buffer manipulation

Several protected members are related to buffer management and positioning:

14.8.1.4: Deriving classes from `streambuf'

When classes are derived from streambuf at least underflow should be overridden by classes intending to read information from devices, and overflow should be overridden by classes intending to write information to devices. Several examples of classes derived from streambuf are provided in chapter 23.

Fstream class type objects use a combined input/output buffer. This is a result from that istream and ostream being virtually derived from ios, which class contains the streambuf. To construct a class supporting both input and output using separate buffers, the streambuf itself may define two buffers. When seekoff is called for reading, a mode parameter can be set to ios::in, otherwise to ios::out. Thus the derived class knows whether it should access the read buffer or the write buffer. Of course, underflow and overflow do not have to inspect the mode flag as they by implication know on which buffer they should operate.

14.8.2: The class `filebuf'

The class filebuf is a specialization of streambuf used by the file stream classes. Before using a filebuf the header file <fstream> must have been included.

In addition to the (public) members that are available through the class streambuf, filebuf offers the following (public) members:

14.9: A polymorphic exception class

Earlier in the C++ Annotations (section 10.3.1) we hinted at the possibility of designing a class Exception whose process member would behave differently, depending on the kind of exception that was thrown. Now that we've introduced polymorphism we can further develop this example.

It probably does not come as a surprise that our class Exception should be a polymorphic base class from which special exception handling classes can be derived. In section 10.3.1 a member severity was used offering functionality that may be replaced by members of the Exception base class.

The base class Exception may be designed as follows:

    #ifndef INCLUDED_EXCEPTION_H_
    #define INCLUDED_EXCEPTION_H_
    #include <iostream>
    #include <string>

    class Exception
    {
        std::string d_reason;

        public:
            Exception(std::string const &reason);
            virtual ~Exception();

            std::ostream &insertInto(std::ostream &out) const;
            void handle() const;

        private:
            virtual void action() const;
    };

    inline void Exception::action() const
    {
        throw;
    }
    inline Exception::Exception(std::string const &reason)
    :
        d_reason(reason)
    {}
    inline void Exception::handle() const
    {
        action();
    }
    inline std::ostream &Exception::insertInto(std::ostream &out) const
    {
        return out << d_reason;
    }

    inline std::ostream &operator<<(std::ostream &out, Exception const &e)
    {
        return e.insertInto(out);
    }

    #endif
Objects of this class may be inserted into ostreams but the core element of this class is the virtual member function action, by default rethrowing an exception.

A derived class Warning simply prefixes the thrown warning text by the text Warning:, but a derived class Fatal overrides Exception::action by calling std::terminate, forcefully terminating the program.

Here are the classes Warning and Fatal

    #ifndef WARNINGEXCEPTION_H_
    #define WARNINGEXCEPTION_H_

    #include "exception.h"

    class Warning: public Exception
    {
        public:
            Warning(std::string const &reason)
            :
                Exception("Warning: " + reason)
            {}
    };
    #endif
    #ifndef FATAL_H_
    #define FATAL_H_

    #include "exception.h"

    class Fatal: public Exception
    {
        public:
            Fatal(std::string  const &reason);
        private:
            virtual void action() const;
    };

    inline Fatal::Fatal(std::string  const &reason)
    :
        Exception(reason)
    {}

    inline void Fatal::action() const
    {
        std::cout << "Fatal::action() terminates" << '\n';
        std::terminate();
    }

    #endif

When the example program is started without arguments it throws a Fatal exception, otherwise it throws a Warning exception. Of course, additional exception types could also easily be defined. To make the example compilable the Exception destructor is defined above main. The default destructor cannot be used, as it is a virtual destructor. In practice the destructor should be defined in its own little source file:

    #include "warning.h"
    #include "fatal.h"

    Exception::~Exception()
    {}

    using namespace std;

    int main(int argc, char **argv)
    try
    {
        try
        {
            if (argc == 1)
                throw Fatal("Missing Argument") ;
            else
                throw Warning("the argument is ignored");
        }
        catch (Exception const &e)
        {
            cout << e << '\n';
            e.handle();
        }
    }
    catch(...)
    {
        cout << "caught rethrown exception\n";
    }

14.10: How polymorphism is implemented

This section briefly describes how polymorphism is implemented in C++. It is not necessary to understand how polymorphism is implemented if you just want to use polymorphism. However, we think it's nice to know how polymorphism is possible. Also, knowing how polymorphism is implemented clarifies why there is a (small) penalty to using polymorphism in terms of memory usage and efficiency.

The fundamental idea behind polymorphism is that the compiler does not know which function to call at compile-time. The appropriate function is selected at run-time. That means that the address of the function must be available somewhere, to be looked up prior to the actual call. This `somewhere' place must be accessible to the object in question. So when a Vehicle *vp points to a Truck object, then vp->mass() calls Truck's member function. the address of this function is obtained through the actual object to which vp points.

Polymorphism is commonly implemented as follows: an object containing virtual member functions also contains, usually as its first data member a hidden data member, pointing to an array containing the addresses of the class's virtual member functions. The hidden data member is usually called the vpointer, the array of virtual member function addresses the vtable.

The class's vtable is shared by all objects of that class. The overhead of polymorphism in terms of memory consumption is therefore:

Consequently, a statement like vp->mass first inspects the hidden data member of the object pointed to by vp. In the case of the vehicle classification system, this data member points to a table containing two addresses: one pointer to the function mass and one pointer to the function setMass (three pointers if the class also defines (as it should) a virtual destructor). The actually called function is determined from this table.

The internal organization of the objects having virtual functions is illustrated in figures Figure 18 and Figure 19 (originals provided by Guillaume Caumon).

Figure 18 is shown here.
Figure 18: Internal organization objects when virtual functions are defined.


Figure 19 is shown here.
Figure 19: Complementary figure, provided by Guillaume Caumon


As shown by figures Figure 18 and Figure 19, objects potentially using virtual member functions must have one (hidden) data member to address a table of function pointers. The objects of the classes Vehicle and Car both address the same table. The class Truck, however, overrides mass. Consequently, Truck needs its own vtable.

A small complication arises when a class is derived from multiple base classes, each defining virtual functions. Consider the following example:

    class Base1
    {
        public:
            virtual ~Base1();
            void fun1();        // calls vOne and vTwo
        private:
            virtual void vOne();
            virtual void vTwo();
    };
    class Base2
    {
        public:
            virtual ~Base2();
            void fun2();        // calls vThree
        private:
            virtual void vThree();
    };
    class Derived: public Base1, public Base2
    {
        public:
            virtual ~Derived();
        private:
            virtual ~vOne();
            virtual ~vThree();
    };
In the example Derived is multiply derived from Base1 and Base2, each supporting virtual functions. Because of this, Derived also has virtual functions, and so Derived has a vtable allowing a base class pointer or reference to access the proper virtual member.

When Derived::fun1 is called (or a Base1 pointer pointing to fun1 calls fun1) then fun1 calls Derived::vOne and Base1::vTwo. Likewise, when Derived::fun2 is called Derived::vThree is called.

The complication occurs with Derived's vtable. When fun1 is called its class type determines the vtable to use and hence which virtual member to call. So when vOne is called from fun1, it is presumably the second entry in Derived's vtable, as it must match the second entry in Base1's vtable. However, when fun2 calls vThree it apparently is also the second entry in Derived's vtable as it must match the second entry in Base2's vtable.

Of course this cannot be realized by a single vtable. Therefore, when multiple inheritance is used (each base class defining virtual members) another approach is followed to determine which virtual function to call. In this situation (cf. figure Figure 20) the class Derived receives two vtables, one for each of its base classes and each Derived class object harbors two hidden vpointers, each one pointing to its corresponding vtable.

Figure 20 is shown here.
Figure 20: Vtables and vpointers with multiple base classes


Since base class pointers, base class references, or base class interface members unambiguously refer to one of the base classes the compiler can determine which vpointer to use.

The following therefore holds true for classes multiply derived from base classes offering virtual member functions:

14.11: Undefined reference to vtable ...

Occasionaly, the linker generates an error like the following:
    In function `Derived::Derived()':
        : undefined reference to `vtable for Derived'
This error is generated when a virtual function's implementation is missing in a derived class, but the function is mentioned in the derived class's interface.

Such a situation is easily encountered:

Here is an example producing the error:
    class Base
    {
        virtual void member();
    };
    inline void Base::member()
    {}
    class Derived: public Base
    {
        virtual void member();      // only declared
    };
    int main()
    {
        Derived d;  // Will compile, since all members were declared.
                    // Linking will fail, since we don't have the
                    // implementation of Derived::member()
    }
It's of course easy to correct the error: implement the derived class's missing virtual member function.

Virtual functions should never be implemented inline. Since the vtable contains the addresses of the class's virtual functions, these functions must have addresses and so they must have been compiled as real (out-of-line) functions. By defining virtual functions inline you run the risk that the compiler simply overlooks those functions as they may very well never be explicitly called (but only polymorphically, from a base class pointer or reference). As a result their addresses may never enter their class's vtables (and even the vtable itself might remain undefined), causing linkage problems or resulting in programs showing unexpected behavior. All these kinds of problems are simply avoided: never define virtual members inline (see also section 7.8.2.1).

14.12: Virtual constructors

In section 14.2 we learned that C++ supports virtual destructors. Like many other object oriented languages (e.g., Java), however, the notion of a virtual constructor is not supported. Not having virtual constructors becomes a liability when only base class references or pointers are available, and a copy of a derived class object is required. Gamma et al. (1995) discuss the Prototype design pattern to deal with this situation.

According to the Prototype Design Pattern each derived class is given the responsibility of implementing a member function returning a pointer to a copy of the object for which the member is called. The usual name for this function is clone. Separating the user interface from the reimplementation interface clone is made part of the interface and newCopy is defined in the reimplementation interface. A base class supporting `cloning' defines a virtual destructor, clone, returning newCopy's return value and the virtual copy constructor, a pure virtual function, having the prototype virtual Base *newCopy() const = 0. As newCopy is a pure virtual function all derived classes must now implement their own `virtual constructor'.

This setup suffices in most situations where we have a pointer or reference to a base class, but it fails when used with abstract containers. We can't create a vector<Base>, with Base featuring the pure virtual copy member in its interface, as Base is called to initialize new elements of such a vector. This is impossible as newCopy is a pure virtual function, so a Base object can't be constructed.

The intuitive solution, providing newCopy with a default implementation, defining it as an ordinary virtual function, fails too as the container calls Base(Base const &other), which would have to call newCopy to copy other. At this point it is unclear what to do with that copy, as the new Base object already exists, and contains no Base pointer or reference data member to assign newCopy's return value to.

Alternatively (and preferred) the original Base class (defined as an abstract base class) is kept as-is and a wrapper class Clonable is used to manage the Base class pointers returned by newCopy. In chapter 17 ways to merge Base and Clonable into one class are discussed, but for now we'll define Base and Clonable as separate classes.

The class Clonable is a very standard class. It contains a pointer member so it needs a copy constructor, destructor, and overloaded assignment operator. It's given at least one non-standard member: Base &base() const, returning a reference to the derived object to which Clonable's Base * data member refers. It is also provided with an additional constructor to initialize its Base * data member.

Any non-abstract class derived from Base must implement Base *newCopy(), returning a pointer to a newly created (allocated) copy of the object for which newCopy is called.

Once we have defined a derived class (e.g., Derived1), we can put our Clonable and Base facilities to good use. In the next example we see main defining a vector<Clonable>. An anonymous Derived1 object is then inserted into the vector using the following steps:

In this sequence, only the Clonable object containing the Derived1 * is used. No additional copies need to be made (or destroyed).

Next, the base member is used in combination with typeid to show the actual type of the Base & object: a Derived1 object.

Main then contains the interesting definition vector<Clonable> v2(bv). Here a copy of bv is created. This copy construction observes the actual types of the Base references, making sure that the appropriate types appear in the vector's copy.

At the end of the program, we have created two Derived1 objects, which are correctly deleted by the vector's destructors. Here is the full program, illustrating the `virtual constructor' concept ( Jesse van den Kieboom created an alternative implementation of a class Clonable, implemented as a class template. His implementation is found here.):

    #include <iostream>
    #include <vector>
    #include <algorithm>
    #include <typeinfo>

// Base and its inline member:
    class Base
    {
        public:
            virtual ~Base();
            Base *clone() const;
        private:
            virtual Base *newCopy() const = 0;
    };
    inline Base *Base::clone() const
    {
        return newCopy();
    }

// Clonable and its inline members:
    class Clonable
    {
        Base *d_bp;

        public:
            Clonable();
            explicit Clonable(Base *base);
            ~Clonable();
            Clonable(Clonable const &other);
            Clonable(Clonable &&tmp);
            Clonable &operator=(Clonable const &other);
            Clonable &operator=(Clonable &&tmp);

            Base &base() const;
    };
    inline Clonable::Clonable()
    :
        d_bp(0)
    {}
    inline Clonable::Clonable(Base *bp)
    :
        d_bp(bp)
    {}
    inline Clonable::Clonable(Clonable const &other)
    :
        d_bp(other.d_bp->clone())
    {}
    inline Clonable::Clonable(Clonable &&tmp)
    :
        d_bp(tmp.d_bp)
    {
        tmp.d_bp = 0;
    }
    inline Clonable::~Clonable()
    {
        delete d_bp;
    }
    inline Base &Clonable::base() const
    {
        return *d_bp;
    }

// Derived and its inline member:
    class Derived1: public Base
    {
        public:
            ~Derived1();
        private:
            virtual Base *newCopy() const;
    };
    inline Base *Derived1::newCopy() const
    {
        return new Derived1(*this);
    }

// Members not implemented inline:
    Base::~Base()
    {}
    Clonable &Clonable::operator=(Clonable const &other)
    {
        Clonable tmp(other);
        std::swap(d_bp, tmp.d_bp);
        return *this;
    }
    Clonable &Clonable::operator=(Clonable &&tmp)
    {
        std::swap(d_bp, tmp.d_bp);
        return *this;
    }
    Derived1::~Derived1()
    {
        std::cout << "~Derived1() called\n";
    }

// The main function:
    using namespace std;

    int main()
    {
        vector<Clonable> bv;

        bv.push_back(Clonable(new Derived1()));
        cout << "bv[0].name: " << typeid(bv[0].base()).name() << '\n';

        vector<Clonable> v2(bv);
        cout << "v2[0].name: " << typeid(v2[0].base()).name() << '\n';
    }
    /*
        Output:
            bv[0].name: 8Derived1
            v2[0].name: 8Derived1
            ~Derived1() called
            ~Derived1() called
    */