MindView Inc.
[ Viewing Hints ] [ Revision History ] [ Book Home Page ] [ Free Newsletter ]
[ Seminars ] [ Seminars on CD ROM ] [ Consulting ]

Thinking in C++, 2nd edition, Volume 2
Revision 4.0

by Bruce Eckel & Chuck Allison
©2001 MindView, Inc.

[ Previous Chapter ] [ Short TOC ] [ Table of Contents ] [ Index ] [ Next Chapter ]

10: Multiple inheritance

The basic concept of multiple inheritance (MI) sounds simple enough.

[[[Notes:

  1. Demo of use of MI, using Greenhouse example and different company’s greenhouse controller equipment.
  2. Introduce concept of interfaces; toys and “tuckable” interface
  3. Class Sattelite : public Task, public Displayed {}; highlight(Displayed*); Suspend(Task*)
  4. Slider: Islider, BBslider {} (GUI from a Vendor, MI Decouples)
  5. Barton & Nackman MI Examples
  6. Concrete classes (nonvirtual) vs. interface classes (pure virtual)
  7. class X { int f(x); };
    class Y { int f(y); };
    class Z : X, Y {
    using X::f;
    int f(Z);
    };

]]]

You create a new type by inheriting from more than one base class. The syntax is exactly what you’d expect, and as long as the inheritance diagrams are simple, MI is simple as well.

However, MI can introduce a number of ambiguities and strange situations, which are covered in this chapter. But first, it helps to get a perspective on the subject.

Perspective

Before C++, the most successful object-oriented language was Smalltalk. Smalltalk was created from the ground up as an OO language. It is often referred to as pure, whereas C++, because it was built on top of C, is called hybrid. One of the design decisions made with Smalltalk was that all classes would be derived in a single hierarchy, rooted in a single base class (called Object – this is the model for the object-based hierarchy). You cannot create a new class in Smalltalk without inheriting it from an existing class, which is why it takes a certain amount of time to become productive in Smalltalk – you must learn the class library before you can start making new classes. So the Smalltalk class hierarchy is always a single monolithic tree.

Classes in Smalltalk usually have a number of things in common, and always have some things in common (the characteristics and behaviors of Object), so you almost never run into a situation where you need to inherit from more than one base class. However, with C++ you can create as many hierarchy trees as you want. Therefore, for logical completeness the language must be able to combine more than one class at a time – thus the need for multiple inheritance.

However, this was not a crystal-clear case of a feature that no one could live without, and there was (and still is) a lot of disagreement about whether MI is really essential in C++. MI was added in AT&T cfront release 2.0 and was the first significant change to the language. Since then, a number of other features have been added (notably templates) that change the way we think about programming and place MI in a much less important role. You can think of MI as a “minor” language feature that shouldn’t be involved in your daily design decisions.

One of the most pressing issues that drove MI involved containers. Suppose you want to create a container that everyone can easily use. One approach is to use void* as the type inside the container, as with PStash and Stack. The Smalltalk approach, however, is to make a container that holds Objects. (Remember that Object is the base type of the entire Smalltalk hierarchy.) Because everything in Smalltalk is ultimately derived from Object, any container that holds Objects can hold anything, so this approach works nicely.

Now consider the situation in C++. Suppose vendor A creates an object-based hierarchy that includes a useful set of containers including one you want to use called Holder. Now you come across vendor B’s class hierarchy that contains some other class that is important to you, a BitImage class, for example, which holds graphic images. The only way to make a Holder of BitImages is to inherit a new class from both Object, so it can be held in the Holder, and BitImage:


This was seen as an important reason for MI, and a number of class libraries were built on this model. However, as you saw in Chapter XX, the addition of templates has changed the way containers are created, so this situation isn’t a driving issue for MI.

The other reason you may need MI is logical, related to design. Unlike the above situation, where you don’t have control of the base classes, in this one you do, and you intentionally use MI to make the design more flexible or useful. (At least, you may believe this to be the case.) An example of this is in the original iostream library design:


Both istream and ostream are useful classes by themselves, but they can also be inherited into a class that combines both their characteristics and behaviors.

Regardless of what motivates you to use MI, a number of problems arise in the process, and you need to understand them to use it.

Duplicate subobjects

When you inherit from a base class, you get a copy of all the data members of that base class in your derived class. This copy is referred to as a subobject. If you multiply inherit from class d1 and class d2 into class mi, class mi contains one subobject of d1 and one of d2. So your mi object looks like this:


Now consider what happens if d1 and d2 both inherit from the same base class, called Base:


In the above diagram, both d1 and d2 contain a subobject of Base, so mi contains two subobjects of Base. Because of the path produced in the diagram, this is sometimes called a “diamond” in the inheritance hierarchy. Without diamonds, multiple inheritance is quite straightforward, but as soon as a diamond appears, trouble starts because you have duplicate subobjects in your new class. This takes up extra space, which may or may not be a problem depending on your design. But it also introduces an ambiguity.

Ambiguous upcasting

What happens, in the above diagram, if you want to cast a pointer to an mi to a pointer to a Base? There are two subobjects of type Base, so which address does the cast produce? Here’s the diagram in code:

//: C10:MultipleInheritance1.cpp
// MI & ambiguity
//{L} ../TestSuite/Test
#include "../purge.h"
#include <iostream>
#include <vector>
using namespace std;

class MBase {
public:
  virtual char* vf() const = 0;
  virtual ~MBase() {}
};

class D1 : public MBase {
public:
  char* vf() const { return "D1"; }
};

class D2 : public MBase {
public:
  char* vf() const { return "D2"; }
};

// Causes error: ambiguous override of vf():
//! class MI : public D1, public D2 {};

int main() {
  vector<MBase*> b;
  b.push_back(new D1);
  b.push_back(new D2);
  // Cannot upcast: which subobject?:
//!  b.push_back(new mi);
  for(int i = 0; i < b.size(); i++)
    cout << b[i]->vf() << endl;
  purge(b);
} ///:~

Two problems occur here. First, you cannot even create the class mi because doing so would cause a clash between the two definitions of vf( ) in D1 and D2.

Second, in the array definition for b[ ] this code attempts to create a new mi and upcast the address to a MBase*. The compiler won’t accept this because it has no way of knowing whether you want to use D1’s subobject MBase or D2’s subobject MBase for the resulting address.

virtual base classes

To solve the first problem, you must explicitly disambiguate the function vf( ) by writing a redefinition in the class mi.

The solution to the second problem is a language extension: The meaning of the virtual keyword is overloaded. If you inherit a base class as virtual, only one subobject of that class will ever appear as a base class. Virtual base classes are implemented by the compiler with pointer magic in a way suggesting the implementation of ordinary virtual functions.

Because only one subobject of a virtual base class will ever appear during multiple inheritance, there is no ambiguity during upcasting. Here’s an example:

//: C10:MultipleInheritance2.cpp
// Virtual base classes
//{L} ../TestSuite/Test
#include "../purge.h"
#include <iostream>
#include <vector>
using namespace std;

class MBase {
public:
  virtual char* vf() const = 0;
  virtual ~MBase() {}
};

class D1 : virtual public MBase {
public:
  char* vf() const { return "D1"; }
};

class D2 : virtual public MBase {
public:
  char* vf() const { return "D2"; }
};

// MUST explicitly disambiguate vf():
class MI : public D1, public D2 {
public:
  char* vf() const { return D1::vf();}
};

int main() {
  vector<MBase*> b;
  b.push_back(new D1);
  b.push_back(new D2);
  b.push_back(new MI); // OK
  for(int i = 0; i < b.size(); i++)
    cout << b[i]->vf() << endl;
  purge(b);
} ///:~

The compiler now accepts the upcast, but notice that you must still explicitly disambiguate the function vf( ) in MI; otherwise the compiler wouldn’t know which version to use.

The "most derived" class and virtual base initialization

The use of virtual base classes isn’t quite as simple as that. The above example uses the (compiler-synthesized) default constructor. If the virtual base has a constructor, things become a bit strange. To understand this, you need a new term: most-derived class.

The most-derived class is the one you’re currently in, and is particularly important when you’re thinking about constructors. In the previous example, MBase is the most-derived class inside the MBase constructor. Inside the D1 constructor, D1 is the most-derived class, and inside the MI constructor, MI is the most-derived class.

When you are using a virtual base class, the most-derived constructor is responsible for initializing that virtual base class. That means any class, no matter how far away it is from the virtual base, is responsible for initializing it. Here’s an example:

//: C10:MultipleInheritance3.cpp
// Virtual base initialization.
// Virtual base classes must always be
// Initialized by the "most-derived" class.
//{L} ../TestSuite/Test
#include "../purge.h"
#include <iostream>
#include <vector>
using namespace std;

class MBase {
public:
  MBase(int) {}
  virtual char* vf() const = 0;
  virtual ~MBase() {}
};

class D1 : virtual public MBase {
public:
  D1() : MBase(1) {}
  char* vf() const { return "D1"; }
};

class D2 : virtual public MBase {
public:
  D2() : MBase(2) {}
  char* vf() const { return "D2"; }
};

class MI : public D1, public D2 {
public:
  MI() : MBase(3) {}
  char* vf() const {
    return D1::vf(); // MUST disambiguate
  }
};

class X : public MI {
public:
  // You must ALWAYS init the virtual base:
  X() : MBase(4) {}
};

int main() {
  vector<MBase*> b;
  b.push_back(new D1);
  b.push_back(new D2);
  b.push_back(new MI); // OK
  b.push_back(new X);
  for(int i = 0; i < b.size(); i++)
    cout << b[i]->vf() << endl;
  purge(b);
} ///:~

As you would expect, both D1 and D2 must initialize MBase in their constructor. But so must MI and X, even though they are more than one layer away! That’s because each one in turn becomes the most-derived class. The compiler can’t know whether to use D1’s initialization of MBase or to use D2’s version. Thus you are always forced to do it in the most-derived class. Note that only the single selected virtual base constructor is called.

"Tying off" virtual bases with a default constructor

Forcing the most-derived class to initialize a virtual base that may be buried deep in the class hierarchy can seem like a tedious and confusing task to put upon the user of your class. It’s better to make this invisible, which is done by creating a default constructor for the virtual base class, like this:

//: C10:MultipleInheritance4.cpp
// "Tying off" virtual bases so you don't have
// to worry about them in derived classes.
//{L} ../TestSuite/Test
#include "../purge.h"
#include <iostream>
#include <vector>
using namespace std;

class MBase {
public:
 // Default constructor removes responsibility:
  MBase(int = 0) {}
  virtual char* vf() const = 0;
  virtual ~MBase() {}
};

class D1 : virtual public MBase {
public:
  D1() : MBase(1) {}
  char* vf() const { return "D1"; }
};

class D2 : virtual public MBase {
public:
  D2() : MBase(2) {}
  char* vf() const { return "D2"; }
};

class MI : public D1, public D2 {
public:
  MI() {} // Calls default constructor for MBase
  char* vf() const {
    return D1::vf(); // MUST disambiguate
  }
};

class X : public MI {
public:
  X() {} // Calls default constructor for MBase
};

int main() {
  vector<MBase*> b;
  b.push_back(new D1);
  b.push_back(new D2);
  b.push_back(new MI); // OK
  b.push_back(new X);
  for(int i = 0; i < b.size(); i++)
    cout << b[i]->vf() << endl;
  purge(b);
} ///:~

If you can always arrange for a virtual base class to have a default constructor, you’ll make things much easier for anyone who inherits from that class.

Overhead

The term “pointer magic” has been used to describe the way virtual inheritance is implemented. You can see the physical overhead of virtual inheritance with the following program:

//: C10:Overhead.cpp
// Virtual base class overhead
//{L} ../TestSuite/Test
#include <fstream>
using namespace std;
ofstream out("overhead.out");

class MBase {
public:
  virtual void f() const {};
  virtual ~MBase() {}
};

class NonVirtualInheritance
  : public MBase {};

class VirtualInheritance
  : virtual public MBase {};

class VirtualInheritance2
  : virtual public MBase {};

class MI
  : public VirtualInheritance,
    public VirtualInheritance2 {};

#define WRITE(ARG) \
out << #ARG << " = " << ARG << endl;

int main() {
  MBase b;
  WRITE(sizeof(b));
  NonVirtualInheritance nonv_inheritance;
  WRITE(sizeof(nonv_inheritance));
  VirtualInheritance v_inheritance;
  WRITE(sizeof(v_inheritance));
  MI mi;
  WRITE(sizeof(mi));
} ///:~

Each of these classes only contains a single byte, and the “core size” is that byte. Because all these classes contain virtual functions, you expect the object size to be bigger than the core size by a pointer (at least – your compiler may also pad extra bytes into an object for alignment). The results are a bit surprising (these are from one particular compiler; yours may do it differently):

sizeof(b) = 2
sizeof(nonv_inheritance) = 2
sizeof(v_inheritance) = 6
sizeof(MI) = 12

Both b and nonv_inheritance contain the extra pointer, as expected. But when virtual inheritance is added, it would appear that the VPTR plus two extra pointers are added! By the time the multiple inheritance is performed, the object appears to contain five extra pointers (however, one of these is probably a second VPTR for the second multiply inherited subobject).

The curious can certainly probe into your particular implementation and look at the assembly language for member selection to determine exactly what these extra bytes are for, and the cost of member selection with multiple inheritance[24]. The rest of you have probably seen enough to guess that quite a bit more goes on with virtual multiple inheritance, so it should be used sparingly (or avoided) when efficiency is an issue.

Upcasting

When you embed subobjects of a class inside a new class, whether you do it by creating member objects or through inheritance, each subobject is placed within the new object by the compiler. Of course, each subobject has its own this pointer, and as long as you’re dealing with member objects, everything is quite straightforward. But as soon as multiple inheritance is introduced, a funny thing occurs: An object can have more than one this pointer because the object represents more than one type during upcasting. The following example demonstrates this point:

//: C10:Mithis.cpp
// MI and the "this" pointer
//{L} ../TestSuite/Test
#include <fstream>
using namespace std;
ofstream out("mithis.out");

class Base1 {
  char c[0x10];
public:
  void printthis1() {
    out << "Base1 this = " << this << endl;
  }
};

class Base2 {
  char c[0x10];
public:
  void printthis2() {
    out << "Base2 this = " << this << endl;
  }
};

class Member1 {
  char c[0x10];
public:
  void printthism1() {
    out << "Member1 this = " << this << endl;
  }
};

class Member2 {
  char c[0x10];
public:
  void printthism2() {
    out << "Member2 this = " << this << endl;
  }
};

class MI : public Base1, public Base2 {
  Member1 m1;
  Member2 m2;
public:
  void printthis() {
    out << "MI this = " << this << endl;
    printthis1();
    printthis2();
    m1.printthism1();
    m2.printthism2();
  }
};

int main() {
  MI mi;
  out << "sizeof(mi) = "
    << hex << sizeof(mi) << " hex" << endl;
  mi.printthis();
  // A second demonstration:
  Base1* b1 = &mi; // Upcast
  Base2* b2 = &mi; // Upcast
  out << "Base 1 pointer = " << b1 << endl;
  out << "Base 2 pointer = " << b2 << endl;
} ///:~

The arrays of bytes inside each class are created with hexadecimal sizes, so the output addresses (which are printed in hex) are easy to read. Each class has a function that prints its this pointer, and these classes are assembled with both multiple inheritance and composition into the class MI, which prints its own address and the addresses of all the other subobjects. This function is called in main( ). You can clearly see that you get two different this pointers for the same object. The address of the MI object is taken and upcast to the two different types. Here’s the output:[25]

sizeof(mi) = 40 hex
mi this = 0x223e
Base1 this = 0x223e
Base2 this = 0x224e
Member1 this = 0x225e
Member2 this = 0x226e
Base 1 pointer = 0x223e
Base 2 pointer = 0x224e

Although object layouts vary from compiler to compiler and are not specified in Standard C++, this one is fairly typical. The starting address of the object corresponds to the address of the first class in the base-class list. Then the second inherited class is placed, followed by the member objects in order of declaration.

When the upcast to the Base1 and Base2 pointers occur, you can see that, even though they’re ostensibly pointing to the same object, they must actually have different this pointers, so the proper starting address can be passed to the member functions of each subobject. The only way things can work correctly is if this implicit upcasting takes place when you call a member function for a multiply inherited subobject.

Persistence

Normally this isn’t a problem, because you want to call member functions that are concerned with that subobject of the multiply inherited object. However, if your member function needs to know the true starting address of the object, multiple inheritance causes problems. Ironically, this happens in one of the situations where multiple inheritance seems to be useful: persistence.

The lifetime of a local object is the scope in which it is defined. The lifetime of a global object is the lifetime of the program. A persistent object lives between invocations of a program: You can normally think of it as existing on disk instead of in memory. One definition of an object-oriented database is “a collection of persistent objects.”

To implement persistence, you must move a persistent object from disk into memory in order to call functions for it, and later store it to disk before the program expires. Four issues arise when storing an object on disk:

  1. The object must be converted from its representation in memory to a series of bytes on disk.
  2. Because the values of any pointers in memory won’t have meaning the next time the program is invoked, these pointers must be converted to something meaningful.
  3. What the pointers point to must also be stored and retrieved.
  4. When restoring an object from disk, the virtual pointers in the object must be respected.

Because the object must be converted back and forth between a layout in memory and a serial representation on disk, the process is called serialization (to write an object to disk) and deserialization (to restore an object from disk). Although it would be very convenient, these processes require too much overhead to support directly in the language. Class libraries will often build in support for serialization and deserialization by adding special member functions and placing requirements on new classes. (Usually some sort of serialize( ) function must be written for each new class.) Also, persistence is generally not automatic; you must usually explicitly write and read the objects.

MI-based persistence

Consider sidestepping the pointer issues for now and creating a class that installs persistence into simple objects using multiple inheritance. By inheriting the persistence class along with your new class, you automatically create classes that can be read from and written to disk. Although this sounds great, the use of multiple inheritance introduces a pitfall, as seen in the following example.

//: C10:Persist1.cpp
// Simple persistence with MI
//{L} ../TestSuite/Test
//{-g++3} dumps core
#include "../require.h"
#include <iostream>
#include <fstream>
using namespace std;

class Persistent {
  int objSize; // Size of stored object
public:
  Persistent(int sz) : objSize(sz) {}
  void write(ostream& out) const {
    out.write((char*)this, objSize);
  }
  void read(istream& in) {
    in.read((char*)this, objSize);
  }
};

class Data {
  float f[3];
public:
  Data(float f0 = 0.0, float f1 = 0.0,
    float f2 = 0.0) {
    f[0] = f0;
    f[1] = f1;
    f[2] = f2;
  }
  void print(const char* msg = "") const {
    if(*msg) cout << msg << "   ";
    for(int i = 0; i < 3; i++)
      cout << "f[" << i << "] = "
           << f[i] << endl;
  }
};

class WData1 : public Persistent, public Data {
public:
  WData1(float f0 = 0.0, float f1 = 0.0,
    float f2 = 0.0) : Data(f0, f1, f2),
    Persistent(sizeof(WData1)) {}
};

class WData2 : public Data, public Persistent {
public:
  WData2(float f0 = 0.0, float f1 = 0.0,
    float f2 = 0.0) : Data(f0, f1, f2),
    Persistent(sizeof(WData2)) {}
};

int main() {
  {
    ofstream f1("f1.dat"), f2("f2.dat");
    assure(f1, "f1.dat"); assure(f2, "f2.dat");
    WData1 d1(1.1, 2.2, 3.3);
    WData2 d2(4.4, 5.5, 6.6);
    d1.print("d1 before storage");
    d2.print("d2 before storage");
    d1.write(f1);
    d2.write(f2);
  } // Closes files
  ifstream f1("f1.dat"), f2("f2.dat");
  assure(f1, "f1.dat"); assure(f2, "f2.dat");
  WData1 d1;
  WData2 d2;
  d1.read(f1);
  d2.read(f2);
  d1.print("d1 after storage");
  d2.print("d2 after storage");
} ///:~

In this very simple version, the Persistent::read( ) and Persistent::write( ) functions take the this pointer and call iostream read( ) and write( ) functions. (Note that any type of iostream can be used). A more sophisticated Persistent class would call a virtual write( ) function for each subobject.

With the language features covered so far in the book, the number of bytes in the object cannot be known by the Persistent class so it is inserted as a constructor argument. (In Chapter XX, run-time type identification shows how you can find the exact type of an object given only a base pointer; once you have the exact type you can find out the correct size with the sizeof operator.)

The Data class contains no pointers or VPTR, so there is no danger in simply writing it to disk and reading it back again. And it works fine in class WData1 when, in main( ), it’s written to file F1.DAT and later read back again. However, when Persistent is second in the inheritance list of WData2, the this pointer for Persistent is offset to the end of the object, so it reads and writes past the end of the object. This not only produces garbage when reading the object from the file, it’s dangerous because it walks over any storage that occurs after the object.

This problem occurs in multiple inheritance any time a class must produce the this pointer for the actual object from a subobject’s this pointer. Of course, if you know your compiler always lays out objects in order of declaration in the inheritance list, you can ensure that you always put the critical class at the beginning of the list (assuming there’s only one critical class). However, such a class may exist in the inheritance hierarchy of another class and you may unwittingly put it in the wrong place during multiple inheritance. Fortunately, using run-time type identification (the subject of Chapter XX) will produce the proper pointer to the actual object, even if multiple inheritance is used.

Improved persistence

A more practical approach to persistence, and one you will see employed more often, is to create virtual functions in the base class for reading and writing and then require the creator of any new class that must be streamed to redefine these functions. The argument to the function is the stream object to write to or read from.[26] Then the creator of the class, who knows best how the new parts should be read or written, is responsible for making the correct function calls. This doesn’t have the “magical” quality of the previous example, and it requires more coding and knowledge on the part of the user, but it works and doesn’t break when pointers are present:

//: C10:Persist2.cpp
// Improved MI persistence
//{L} ../TestSuite/Test
#include "../require.h"
#include <iostream>
#include <fstream>
#include <cstring>
using namespace std;

class Persistent {
public:
  virtual void write(ostream& out) const = 0;
  virtual void read(istream& in) = 0;
  virtual ~Persistent() {}
};

class Data {
protected:
  float f[3];
public:
  Data(float f0 = 0.0, float f1 = 0.0,
    float f2 = 0.0) {
    f[0] = f0;
    f[1] = f1;
    f[2] = f2;
  }
  void print(const char* msg = "") const {
    if(*msg) cout << msg << endl;
    for(int i = 0; i < 3; i++)
      cout << "f[" << i << "] = "
           << f[i] << endl;
  }
};

class WData1 : public Persistent, public Data {
public:
  WData1(float f0 = 0.0, float f1 = 0.0,
    float f2 = 0.0) : Data(f0, f1, f2) {}
  void write(ostream& out) const {
    out << f[0] << " " 
      << f[1] << " " << f[2] << " ";
  }
  void read(istream& in) {
    in >> f[0] >> f[1] >> f[2];
  }
};

class WData2 : public Data, public Persistent {
public:
  WData2(float f0 = 0.0, float f1 = 0.0,
    float f2 = 0.0) : Data(f0, f1, f2) {}
  void write(ostream& out) const {
    out << f[0] << " " 
      << f[1] << " " << f[2] << " ";
  }
  void read(istream& in) {
    in >> f[0] >> f[1] >> f[2];
  }
};

class Conglomerate : public Data,
public Persistent {
  char* name; // Contains a pointer
  WData1 d1;
  WData2 d2;
public:
  Conglomerate(const char* nm = "",
    float f0 = 0.0, float f1 = 0.0,
    float f2 = 0.0, float f3 = 0.0,
    float f4 = 0.0, float f5 = 0.0,
    float f6 = 0.0, float f7 = 0.0,
    float f8= 0.0) : Data(f0, f1, f2),
    d1(f3, f4, f5), d2(f6, f7, f8) {
    name = new char[strlen(nm) + 1];
    strcpy(name, nm);
  }
  void write(ostream& out) const {
    int i = strlen(name) + 1;
    out << i << " "; // Store size of string
    out << name << endl;
    d1.write(out);
    d2.write(out);
    out << f[0] << " " << f[1] << " " << f[2];
  }
  // Must read in same order as write:
  void read(istream& in) {
    delete []name; // Remove old storage
    int i;
    in >> i >> ws; // Get int, strip whitespace
    name = new char[i];
    in.getline(name, i);
    d1.read(in);
    d2.read(in);
    in >> f[0] >> f[1] >> f[2];
  }
  void print() const {
    Data::print(name);
    d1.print();
    d2.print();
  }
};

int main() {
  {
    ofstream data("data.dat");
    assure(data, "data.dat");
    Conglomerate C("This is Conglomerate C",
      1.1, 2.2, 3.3, 4.4, 5.5,
      6.6, 7.7, 8.8, 9.9);
    cout << "C before storage" << endl;
    C.print();
    C.write(data);
  } // Closes file
  ifstream data("data.dat");
  assure(data, "data.dat");
  Conglomerate C;
  C.read(data);
  cout << "after storage: " << endl;
  C.print();
} ///:~

The pure virtual functions in Persistent must be redefined in the derived classes to perform the proper reading and writing. If you already knew that Data would be persistent, you could inherit directly from Persistent and redefine the functions there, thus eliminating the need for multiple inheritance. This example is based on the idea that you don’t own the code for Data, that it was created elsewhere and may be part of another class hierarchy so you don’t have control over its inheritance. However, for this scheme to work correctly you must have access to the underlying implementation so it can be stored; thus the use of protected.

The classes WData1 and WData2 use familiar iostream inserters and extractors to store and retrieve the protected data in Data to and from the iostream object. In write( ), you can see that spaces are added after each floating point number is written; these are necessary to allow parsing of the data on input.

The class Conglomerate not only inherits from Data, it also has member objects of type WData1 and WData2, as well as a pointer to a character string. In addition, all the classes that inherit from Persistent also contain a VPTR, so this example shows the kind of problem you’ll actually encounter when using persistence.

When you create write( ) and read( ) function pairs, the read( ) must exactly mirror what happens during the write( ), so read( ) pulls the bits off the disk the same way they were placed there by write( ). Here, the first problem that’s tackled is the char*, which points to a string of any length. The size of the string is calculated and stored on disk as an int (followed by a space to enable parsing) to allow the read( ) function to allocate the correct amount of storage.

When you have subobjects that have read( ) and write( ) member functions, all you need to do is call those functions in the new read( ) and write( ) functions. This is followed by direct storage of the members in the base class.

People have gone to great lengths to automate persistence, for example, by creating modified preprocessors to support a “persistent” keyword to be applied when defining a class. One can imagine a more elegant approach than the one shown here for implementing persistence, but it has the advantage that it works under all implementations of C++, doesn’t require special language extensions, and is relatively bulletproof.

Avoiding MI

The need for multiple inheritance in Persist2.cpp is contrived, based on the concept that you don’t have control of some of the code in the project. Upon examination of the example, you can see that MI can be easily avoided by using member objects of type Data, and putting the virtual read( )and write( ) members inside Data or WData1 and WData2 rather than in a separate class. There are many situations like this one where multiple inheritance may be avoided; the language feature is included for unusual, special-case situations that would otherwise be difficult or impossible to handle. But when the question of whether to use multiple inheritance comes up, you should ask two questions:

  1. Do I need to show the public interfaces of both these classes, or could one class be embedded with some of its interface produced with member functions in the new class?
  2. Do I need to upcast to both of the base classes? (This applies when you have more than two base classes, of course.)

If you can’t answer “no” to both questions, you can avoid using MI and should probably do so.

One situation to watch for is when one class only needs to be upcast as a function argument. In that case, the class can be embedded and an automatic type conversion operator provided in your new class to produce a reference to the embedded object. Any time you use an object of your new class as an argument to a function that expects the embedded object, the type conversion operator is used. However, type conversion can’t be used for normal member selection; that requires inheritance.

Mixin types

Rodents & pets(play)

interfaces in general

Repairing an interface

One of the best arguments for multiple inheritance involves code that’s out of your control. Suppose you’ve acquired a library that consists of a header file and compiled member functions, but no source code for member functions. This library is a class hierarchy with virtual functions, and it contains some global functions that take pointers to the base class of the library; that is, it uses the library objects polymorphically. Now suppose you build an application around this library, and write your own code that uses the base class polymorphically.

Later in the development of the project or sometime during its maintenance, you discover that the base-class interface provided by the vendor is incomplete: A function may be nonvirtual and you need it to be virtual, or a virtual function is completely missing in the interface, but essential to the solution of your problem. If you had the source code, you could go back and put it in. But you don’t, and you have a lot of existing code that depends on the original interface. Here, multiple inheritance is the perfect solution.

For example, here’s the header file for a library you acquire:

//: C10:Vendor.h
// Vendor-supplied class header
// You only get this & the compiled Vendor.obj
#ifndef VENDOR_H
#define VENDOR_H

class Vendor {
public:
  virtual void v() const;
  void f() const;
  ~Vendor();
};

class Vendor1 : public Vendor {
public:
  void v() const;
  void f() const;
  ~Vendor1();
};

void A(const Vendor&);
void B(const Vendor&);
// Etc.
#endif // VENDOR_H ///:~

Assume the library is much bigger, with more derived classes and a larger interface. Notice that it also includes the functions A( ) and B( ), which take a base pointer and treat it polymorphically. Here’s the implementation file for the library:

//: C10:Vendor.cpp {O}
// Implementation of VENDOR.H
// This is compiled and unavailable to you
#include "Vendor.h"
#include <fstream>
using namespace std;

extern ofstream out; // For trace info

void Vendor::v() const {
  out << "Vendor::v()\n";
}

void Vendor::f() const {
  out << "Vendor::f()\n";
}

Vendor::~Vendor() {
  out << "~Vendor()\n";
}

void Vendor1::v() const {
  out << "Vendor1::v()\n";
}

void Vendor1::f() const {
  out << "Vendor1::f()\n";
}

Vendor1::~Vendor1() {
  out << "~Vendor1()\n";
}

void A(const Vendor& V) {
  // ...
  V.v();
  V.f();
  //..
}

void B(const Vendor& V) {
  // ...
  V.v();
  V.f();
  //..
} ///:~

In your project, this source code is unavailable to you. Instead, you get a compiled file as Vendor.obj or Vendor.lib (or the equivalent for your system).

The problem occurs in the use of this library. First, the destructor isn’t virtual. This is actually a design error on the part of the library creator. In addition, f( ) was not made virtual; assume the library creator decided it wouldn’t need to be. And you discover that the interface to the base class is missing a function essential to the solution of your problem. Also suppose you’ve already written a fair amount of code using the existing interface (not to mention the functions A( ) and B( ), which are out of your control), and you don’t want to change it.

To repair the problem, create your own class interface and multiply inherit a new set of derived classes from your interface and from the existing classes:

//: C10:Paste.cpp
// Fixing a mess with MI
//{L} Vendor ../TestSuite/Test
#include "Vendor.h"
#include <fstream>
using namespace std;

ofstream out("paste.out");

class MyBase { // Repair Vendor interface
public:
  virtual void v() const = 0;
  virtual void f() const = 0;
  // New interface function:
  virtual void g() const = 0;
  virtual ~MyBase() { out << "~MyBase()\n"; }
};

class Paste1 : public MyBase, public Vendor1 {
public:
  void v() const {
    out << "Paste1::v()\n";
    Vendor1::v();
  }
  void f() const {
    out << "Paste1::f()\n";
    Vendor1::f();
  }
  void g() const {
    out << "Paste1::g()\n";
  }
  ~Paste1() { out << "~Paste1()\n"; }
};

int main() {
  Paste1& p1p = *new Paste1;
  MyBase& mp = p1p; // Upcast
  out << "calling f()\n";
  mp.f();  // Right behavior
  out << "calling g()\n";
  mp.g(); // New behavior
  out << "calling A(p1p)\n";
  A(p1p); // Same old behavior
  out << "calling B(p1p)\n";
  B(p1p);  // Same old behavior
  out << "delete mp\n";
  // Deleting a reference to a heap object:
  delete &mp; // Right behavior
} ///:~

In MyBase (which does not use MI), both f( ) and the destructor are now virtual, and a new virtual function g( ) has been added to the interface. Now each of the derived classes in the original library must be recreated, mixing in the new interface with MI. The functions Paste1::v( ) and Paste1::f( )need to call only the original base-class versions of their functions. But now, if you upcast to MyBase as in main( )

MyBase* mp = p1p; // Upcast

any function calls made through mp will be polymorphic, including delete. Also, the new interface function g( ) can be called through mp. Here’s the output of the program:

calling f()
Paste1::f()
Vendor1::f()
calling g()
Paste1::g()
calling A(p1p)
Paste1::v()
Vendor1::v()
Vendor::f()
calling B(p1p)
Paste1::v()
Vendor1::v()
Vendor::f()
delete mp
~Paste1()
~Vendor1()
~Vendor()
~MyBase()

The original library functions A( ) and B( ) still work the same (assuming the new v( ) calls its base-class version). The destructor is now virtual and exhibits the correct behavior.

Although this is a messy example, it does occur in practice and it’s a good demonstration of where multiple inheritance is clearly necessary: You must be able to upcast to both base classes.

Summary

The reason MI exists in C++ and not in other OOP languages is that C++ is a hybrid language and couldn’t enforce a single monolithic class hierarchy the way Smalltalk does. Instead, C++ allows many inheritance trees to be formed, so sometimes you may need to combine the interfaces from two or more trees into a new class.

If no “diamonds” appear in your class hierarchy, MI is fairly simple (although identical function signatures in base classes must be resolved). If a diamond appears, then you must deal with the problems of duplicate subobjects by introducing virtual base classes. This not only adds confusion, but the underlying representation becomes more complex and less efficient.

Multiple inheritance has been called the “goto of the 90’s”.[27] This seems appropriate because, like a goto, MI is best avoided in normal programming, but can occasionally be very useful. It’s a “minor” but more advanced feature of C++, designed to solve problems that arise in special situations. If you find yourself using it often, you may want to take a look at your reasoning. A good Occam’s Razor is to ask, “Must I upcast to all of the base classes?” If not, your life will be easier if you embed instances of all the classes you don’t need to upcast to.

Exercises

  1. These exercises will take you step-by-step through the traps of MI. Create a base class X with a single constructor that takes an int argument and a member function f( ), that takes no arguments and returns void. Now inherit X into Y and Z, creating constructors for each of them that takes a single int argument. Now multiply inherit Y and Z into A. Create an object of class A, and call f( ) for that object. Fix the problem with explicit disambiguation.
  2. Starting with the results of exercise 1, create a pointer to an X called px, and assign to it the address of the object of type A you created before. Fix the problem using a virtual base class. Now fix X so you no longer have to call the constructor for X inside A.
  3. Starting with the results of exercise 2, remove the explicit disambiguation for f( ), and see if you can call f( ) through px. Trace it to see which function gets called. Fix the problem so the correct function will be called in a class hierarchy.

[24] See also Jan Gray, “C++ Under the Hood”, a chapter in Black Belt C++ (edited by Bruce Eckel, M&T Press, 1995).

[25] For easy readability the code was generated for a small-model Intel processor.

[26] Sometimes there’s only a single function for streaming, and the argument contains information about whether you’re reading or writing.

[27] A phrase coined by Zack Urlocker.

[ Previous Chapter ] [ Short TOC ] [ Table of Contents ] [ Index ] [ Next Chapter ]
Last Update:09/26/2001