MindView Inc.
[ Viewing Hints ] [ Exercise Solutions ] [ Volume 2 ] [ Free Newsletter ]
[ Seminars ] [ Seminars on CD ROM ] [ Consulting ]

Thinking in C++, 2nd ed. Volume 1

©2000 by Bruce Eckel

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

1: Introduction to Objects

The genesis of the computer revolution was in a machine. The genesis of our programming languages thus tends to look like that machine.

But computers are not so much machines as they are mind amplification tools (“bicycles for the mind,” as Steve Jobs is fond of saying) and a different kind of expressive medium. As a result, the tools are beginning to look less like machines and more like parts of our minds, and also like other expressive mediums such as writing, painting, sculpture, animation, and filmmaking. Object-oriented programming is part of this movement toward using the computer as an expressive medium.

This chapter will introduce you to the basic concepts of object-oriented programming (OOP), including an overview of OOP development methods. This chapter, and this book, assume that you have had experience in a procedural programming language, although not necessarily C. If you think you need more preparation in programming and the syntax of C before tackling this book, you should work through the “Thinking in C: Foundations for C++ and Java” training CD ROM, bound in with this book and also available at www.BruceEckel.com.

This chapter is background and supplementary material. Many people do not feel comfortable wading into object-oriented programming without understanding the big picture first. Thus, there are many concepts that are introduced here to give you a solid overview of OOP. However, many other people don’t get the big picture concepts until they’ve seen some of the mechanics first; these people may become bogged down and lost without some code to get their hands on. If you’re part of this latter group and are eager to get to the specifics of the language, feel free to jump past this chapter – skipping it at this point will not prevent you from writing programs or learning the language. However, you will want to come back here eventually to fill in your knowledge so you can understand why objects are important and how to design with them.

The progress of abstraction

All programming languages provide abstractions. It can be argued that the complexity of the problems you’re able to solve is directly related to the kind and quality of abstraction. By “kind” I mean, “What is it that you are abstracting?” Assembly language is a small abstraction of the underlying machine. Many so-called “imperative” languages that followed (such as Fortran, BASIC, and C) were abstractions of assembly language. These languages are big improvements over assembly language, but their primary abstraction still requires you to think in terms of the structure of the computer rather than the structure of the problem you are trying to solve. The programmer must establish the association between the machine model (in the “solution space,” which is the place where you’re modeling that problem, such as a computer) and the model of the problem that is actually being solved (in the “problem space,” which is the place where the problem exists). The effort required to perform this mapping, and the fact that it is extrinsic to the programming language, produces programs that are difficult to write and expensive to maintain, and as a side effect created the entire “programming methods” industry.

The alternative to modeling the machine is to model the problem you’re trying to solve. Early languages such as LISP and APL chose particular views of the world (“All problems are ultimately lists” or “All problems are algorithmic”). PROLOG casts all problems into chains of decisions. Languages have been created for constraint-based programming and for programming exclusively by manipulating graphical symbols. (The latter proved to be too restrictive.) Each of these approaches is a good solution to the particular class of problem they’re designed to solve, but when you step outside of that domain they become awkward.

The object-oriented approach goes a step farther by providing tools for the programmer to represent elements in the problem space. This representation is general enough that the programmer is not constrained to any particular type of problem. We refer to the elements in the problem space and their representations in the solution space as “objects.” (Of course, you will also need other objects that don’t have problem-space analogs.) The idea is that the program is allowed to adapt itself to the lingo of the problem by adding new types of objects, so when you read the code describing the solution, you’re reading words that also express the problem. This is a more flexible and powerful language abstraction than what we’ve had before. Thus, OOP allows you to describe the problem in terms of the problem, rather than in terms of the computer where the solution will run. There’s still a connection back to the computer, though. Each object looks quite a bit like a little computer; it has a state, and it has operations that you can ask it to perform. However, this doesn’t seem like such a bad analogy to objects in the real world; they all have characteristics and behaviors.

Some language designers have decided that object-oriented programming by itself is not adequate to easily solve all programming problems, and advocate the combination of various approaches into multiparadigm programming languages.[4]

Alan Kay summarized five basic characteristics of Smalltalk, the first successful object-oriented language and one of the languages upon which C++ is based. These characteristics represent a pure approach to object-oriented programming:

  1. Everything is an object. Think of an object as a fancy variable; it stores data, but you can “make requests” to that object, asking it to perform operations on itself. In theory, you can take any conceptual component in the problem you’re trying to solve (dogs, buildings, services, etc.) and represent it as an object in your program.
  2. A program is a bunch of objects telling each other what to do by sending messages. To make a request of an object, you “send a message” to that object. More concretely, you can think of a message as a request to call a function that belongs to a particular object.
  3. Each object has its own memory made up of other objects. Put another way, you create a new kind of object by making a package containing existing objects. Thus, you can build complexity in a program while hiding it behind the simplicity of objects.
  4. Every object has a type. Using the parlance, each object is an instance of a class, in which “class” is synonymous with “type.” The most important distinguishing characteristic of a class is “What messages can you send to it?”
  5. All objects of a particular type can receive the same messages. This is actually a loaded statement, as you will see later. Because an object of type “circle” is also an object of type “shape,” a circle is guaranteed to accept shape messages. This means you can write code that talks to shapes and automatically handles anything that fits the description of a shape. This substitutability is one of the most powerful concepts in OOP.

An object has an interface

Aristotle was probably the first to begin a careful study of the concept of type; he spoke of “the class of fishes and the class of birds.” The idea that all objects, while being unique, are also part of a class of objects that have characteristics and behaviors in common was used directly in the first object-oriented language, Simula-67, with its fundamental keyword class that introduces a new type into a program.

Simula, as its name implies, was created for developing simulations such as the classic “bank teller problem[5].” In this, you have a bunch of tellers, customers, accounts, transactions, and units of money – a lot of “objects.” Objects that are identical except for their state during a program’s execution are grouped together into “classes of objects” and that’s where the keyword class came from. Creating abstract data types (classes) is a fundamental concept in object-oriented programming. Abstract data types work almost exactly like built-in types: You can create variables of a type (called objects or instances in object-oriented parlance) and manipulate those variables (called sending messages or requests; you send a message and the object figures out what to do with it). The members (elements) of each class share some commonality: every account has a balance, every teller can accept a deposit, etc. At the same time, each member has its own state, each account has a different balance, each teller has a name. Thus, the tellers, customers, accounts, transactions, etc., can each be represented with a unique entity in the computer program. This entity is the object, and each object belongs to a particular class that defines its characteristics and behaviors.

So, although what we really do in object-oriented programming is create new data types, virtually all object-oriented programming languages use the “class” keyword. When you see the word “type” think “class” and vice versa[6].

Since a class describes a set of objects that have identical characteristics (data elements) and behaviors (functionality), a class is really a data type because a floating point number, for example, also has a set of characteristics and behaviors. The difference is that a programmer defines a class to fit a problem rather than being forced to use an existing data type that was designed to represent a unit of storage in a machine. You extend the programming language by adding new data types specific to your needs. The programming system welcomes the new classes and gives them all the care and type-checking that it gives to built-in types.

The object-oriented approach is not limited to building simulations. Whether or not you agree that any program is a simulation of the system you’re designing, the use of OOP techniques can easily reduce a large set of problems to a simple solution.

Once a class is established, you can make as many objects of that class as you like, and then manipulate those objects as if they are the elements that exist in the problem you are trying to solve. Indeed, one of the challenges of object-oriented programming is to create a one-to-one mapping between the elements in the problem space and objects in the solution space.

But how do you get an object to do useful work for you? There must be a way to make a request of the object so that it will do something, such as complete a transaction, draw something on the screen or turn on a switch. And each object can satisfy only certain requests. The requests you can make of an object are defined by its interface, and the type is what determines the interface. A simple example might be a representation of a light bulb:


Light lt;
lt.on(); 

The interface establishes what requests you can make for a particular object. However, there must be code somewhere to satisfy that request. This, along with the hidden data, comprises the implementation. From a procedural programming standpoint, it’s not that complicated. A type has a function associated with each possible request, and when you make a particular request to an object, that function is called. This process is usually summarized by saying that you “send a message” (make a request) to an object, and the object figures out what to do with that message (it executes code).

Here, the name of the type/class is Light, the name of this particular Light object is lt, and the requests that you can make of a Light object are to turn it on, turn it off, make it brighter or make it dimmer. You create a Light object by declaring a name (lt) for that object. To send a message to the object, you state the name of the object and connect it to the message request with a period (dot). From the standpoint of the user of a pre-defined class, that’s pretty much all there is to programming with objects.

The diagram shown above follows the format of the Unified Modeling Language (UML). Each class is represented by a box, with the type name in the top portion of the box, any data members that you care to describe in the middle portion of the box, and the member functions (the functions that belong to this object, which receive any messages you send to that object) in the bottom portion of the box. Often, only the name of the class and the public member functions are shown in UML design diagrams, and so the middle portion is not shown. If you’re interested only in the class name, then the bottom portion doesn’t need to be shown, either.

The hidden implementation

It is helpful to break up the playing field into class creators (those who create new data types) and client programmers[7] (the class consumers who use the data types in their applications). The goal of the client programmer is to collect a toolbox full of classes to use for rapid application development. The goal of the class creator is to build a class that exposes only what’s necessary to the client programmer and keeps everything else hidden. Why? Because if it’s hidden, the client programmer can’t use it, which means that the class creator can change the hidden portion at will without worrying about the impact to anyone else. The hidden portion usually represents the tender insides of an object that could easily be corrupted by a careless or uninformed client programmer, so hiding the implementation reduces program bugs. The concept of implementation hiding cannot be overemphasized.

In any relationship it’s important to have boundaries that are respected by all parties involved. When you create a library, you establish a relationship with the client programmer, who is also a programmer, but one who is putting together an application by using your library, possibly to build a bigger library.

If all the members of a class are available to everyone, then the client programmer can do anything with that class and there’s no way to enforce rules. Even though you might really prefer that the client programmer not directly manipulate some of the members of your class, without access control there’s no way to prevent it. Everything’s naked to the world.

So the first reason for access control is to keep client programmers’ hands off portions they shouldn’t touch – parts that are necessary for the internal machinations of the data type but not part of the interface that users need in order to solve their particular problems. This is actually a service to users because they can easily see what’s important to them and what they can ignore.

The second reason for access control is to allow the library designer to change the internal workings of the class without worrying about how it will affect the client programmer. For example, you might implement a particular class in a simple fashion to ease development, and then later discover that you need to rewrite it in order to make it run faster. If the interface and implementation are clearly separated and protected, you can accomplish this easily and require only a relink by the user.

C++ uses three explicit keywords to set the boundaries in a class: public, private, and protected. Their use and meaning are quite straightforward. These access specifiers determine who can use the definitions that follow. public means the following definitions are available to everyone. The private keyword, on the other hand, means that no one can access those definitions except you, the creator of the type, inside member functions of that type. private is a brick wall between you and the client programmer. If someone tries to access a private member, they’ll get a compile-time error. protected acts just like private, with the exception that an inheriting class has access to protected members, but not private members. Inheritance will be introduced shortly.

Reusing the implementation

Once a class has been created and tested, it should (ideally) represent a useful unit of code. It turns out that this reusability is not nearly so easy to achieve as many would hope; it takes experience and insight to produce a good design. But once you have such a design, it begs to be reused. Code reuse is one of the greatest advantages that object-oriented programming languages provide.

The simplest way to reuse a class is to just use an object of that class directly, but you can also place an object of that class inside a new class. We call this “creating a member object.” Your new class can be made up of any number and type of other objects, in any combination that you need to achieve the functionality desired in your new class. Because you are composing a new class from existing classes, this concept is called composition (or more generally, aggregation). Composition is often referred to as a “has-a” relationship, as in “a car has an engine.”


(The above UML diagram indicates composition with the filled diamond, which states there is one car. I will typically use a simpler form: just a line, without the diamond, to indicate an association.[8])

Composition comes with a great deal of flexibility. The member objects of your new class are usually private, making them inaccessible to the client programmers who are using the class. This allows you to change those members without disturbing existing client code. You can also change the member objects at runtime, to dynamically change the behavior of your program. Inheritance, which is described next, does not have this flexibility since the compiler must place compile-time restrictions on classes created with inheritance.

Because inheritance is so important in object-oriented programming it is often highly emphasized, and the new programmer can get the idea that inheritance should be used everywhere. This can result in awkward and overly-complicated designs. Instead, you should first look to composition when creating new classes, since it is simpler and more flexible. If you take this approach, your designs will stay cleaner. Once you’ve had some experience, it will be reasonably obvious when you need inheritance.

Inheritance:
reusing the interface

By itself, the idea of an object is a convenient tool. It allows you to package data and functionality together by concept, so you can represent an appropriate problem-space idea rather than being forced to use the idioms of the underlying machine. These concepts are expressed as fundamental units in the programming language by using the class keyword.

It seems a pity, however, to go to all the trouble to create a class and then be forced to create a brand new one that might have similar functionality. It’s nicer if we can take the existing class, clone it, and then make additions and modifications to the clone. This is effectively what you get with inheritance, with the exception that if the original class (called the base or super or parent class) is changed, the modified “clone” (called the derived or inherited or sub or child class) also reflects those changes.


(The arrow in the above UML diagram points from the derived class to the base class. As you will see, there can be more than one derived class.)

A type does more than describe the constraints on a set of objects; it also has a relationship with other types. Two types can have characteristics and behaviors in common, but one type may contain more characteristics than another and may also handle more messages (or handle them differently). Inheritance expresses this similarity between types using the concept of base types and derived types. A base type contains all of the characteristics and behaviors that are shared among the types derived from it. You create a base type to represent the core of your ideas about some objects in your system. From the base type, you derive other types to express the different ways that this core can be realized.

For example, a trash-recycling machine sorts pieces of trash. The base type is “trash,” and each piece of trash has a weight, a value, and so on, and can be shredded, melted, or decomposed. From this, more specific types of trash are derived that may have additional characteristics (a bottle has a color) or behaviors (an aluminum can may be crushed, a steel can is magnetic). In addition, some behaviors may be different (the value of paper depends on its type and condition). Using inheritance, you can build a type hierarchy that expresses the problem you’re trying to solve in terms of its types.

A second example is the classic “shape” example, perhaps used in a computer-aided design system or game simulation. The base type is “shape,” and each shape has a size, a color, a position, and so on. Each shape can be drawn, erased, moved, colored, etc. From this, specific types of shapes are derived (inherited): circle, square, triangle, and so on, each of which may have additional characteristics and behaviors. Certain shapes can be flipped, for example. Some behaviors may be different, such as when you want to calculate the area of a shape. The type hierarchy embodies both the similarities and differences between the shapes.


Casting the solution in the same terms as the problem is tremendously beneficial because you don’t need a lot of intermediate models to get from a description of the problem to a description of the solution. With objects, the type hierarchy is the primary model, so you go directly from the description of the system in the real world to the description of the system in code. Indeed, one of the difficulties people have with object-oriented design is that it’s too simple to get from the beginning to the end. A mind trained to look for complex solutions is often stumped by this simplicity at first.

When you inherit from an existing type, you create a new type. This new type contains not only all the members of the existing type (although the private ones are hidden away and inaccessible), but more importantly it duplicates the interface of the base class. That is, all the messages you can send to objects of the base class you can also send to objects of the derived class. Since we know the type of a class by the messages we can send to it, this means that the derived class is the same type as the base class. In the previous example, “a circle is a shape.” This type equivalence via inheritance is one of the fundamental gateways in understanding the meaning of object-oriented programming.

Since both the base class and derived class have the same interface, there must be some implementation to go along with that interface. That is, there must be some code to execute when an object receives a particular message. If you simply inherit a class and don’t do anything else, the methods from the base-class interface come right along into the derived class. That means objects of the derived class have not only the same type, they also have the same behavior, which isn’t particularly interesting.

You have two ways to differentiate your new derived class from the original base class. The first is quite straightforward: You simply add brand new functions to the derived class. These new functions are not part of the base class interface. This means that the base class simply didn’t do as much as you wanted it to, so you added more functions. This simple and primitive use for inheritance is, at times, the perfect solution to your problem. However, you should look closely for the possibility that your base class might also need these additional functions. This process of discovery and iteration of your design happens regularly in object-oriented programming.


Although inheritance may sometimes imply that you are going to add new functions to the interface, that’s not necessarily true. The second and more important way to differentiate your new class is to change the behavior of an existing base-class function. This is referred to as overriding that function.


To override a function, you simply create a new definition for the function in the derived class. You’re saying, “I’m using the same interface function here, but I want it to do something different for my new type.”

Is-a vs. is-like-a relationships

There’s a certain debate that can occur about inheritance: Should inheritance override only base-class functions (and not add new member functions that aren’t in the base class)? This would mean that the derived type is exactly the same type as the base class since it has exactly the same interface. As a result, you can exactly substitute an object of the derived class for an object of the base class. This can be thought of as pure substitution, and it’s often referred to as the substitution principle. In a sense, this is the ideal way to treat inheritance. We often refer to the relationship between the base class and derived classes in this case as an is-a relationship, because you can say “a circle is a shape.” A test for inheritance is to determine whether you can state the is-a relationship about the classes and have it make sense.

There are times when you must add new interface elements to a derived type, thus extending the interface and creating a new type. The new type can still be substituted for the base type, but the substitution isn’t perfect because your new functions are not accessible from the base type. This can be described as an is-like-a relationship; the new type has the interface of the old type but it also contains other functions, so you can’t really say it’s exactly the same. For example, consider an air conditioner. Suppose your house is wired with all the controls for cooling; that is, it has an interface that allows you to control cooling. Imagine that the air conditioner breaks down and you replace it with a heat pump, which can both heat and cool. The heat pump is-like-an air conditioner, but it can do more. Because the control system of your house is designed only to control cooling, it is restricted to communication with the cooling part of the new object. The interface of the new object has been extended, and the existing system doesn’t know about anything except the original interface.


Of course, once you see this design it becomes clear that the base class “cooling system” is not general enough, and should be renamed to “temperature control system” so that it can also include heating – at which point the substitution principle will work. However, the diagram above is an example of what can happen in design and in the real world.

When you see the substitution principle it’s easy to feel like this approach (pure substitution) is the only way to do things, and in fact it is nice if your design works out that way. But you’ll find that there are times when it’s equally clear that you must add new functions to the interface of a derived class. With inspection both cases should be reasonably obvious.

Interchangeable objects
with polymorphism

When dealing with type hierarchies, you often want to treat an object not as the specific type that it is but instead as its base type. This allows you to write code that doesn’t depend on specific types. In the shape example, functions manipulate generic shapes without respect to whether they’re circles, squares, triangles, and so on. All shapes can be drawn, erased, and moved, so these functions simply send a message to a shape object; they don’t worry about how the object copes with the message.

Such code is unaffected by the addition of new types, and adding new types is the most common way to extend an object-oriented program to handle new situations. For example, you can derive a new subtype of shape called pentagon without modifying the functions that deal only with generic shapes. This ability to extend a program easily by deriving new subtypes is important because it greatly improves designs while reducing the cost of software maintenance.

There’s a problem, however, with attempting to treat derived-type objects as their generic base types (circles as shapes, bicycles as vehicles, cormorants as birds, etc.). If a function is going to tell a generic shape to draw itself, or a generic vehicle to steer, or a generic bird to move, the compiler cannot know at compile-time precisely what piece of code will be executed. That’s the whole point – when the message is sent, the programmer doesn’t want to know what piece of code will be executed; the draw function can be applied equally to a circle, a square, or a triangle, and the object will execute the proper code depending on its specific type. If you don’t have to know what piece of code will be executed, then when you add a new subtype, the code it executes can be different without requiring changes to the function call. Therefore, the compiler cannot know precisely what piece of code is executed, so what does it do? For example, in the following diagram the BirdController object just works with generic Bird objects, and does not know what exact type they are. This is convenient from BirdController’s perspective, because it doesn’t have to write special code to determine the exact type of Bird it’s working with, or that Bird’s behavior. So how does it happen that, when move( ) is called while ignoring the specific type of Bird, the right behavior will occur (a Goose runs, flies, or swims, and a Penguin runs or swims)?


The answer is the primary twist in object-oriented programming: The compiler cannot make a function call in the traditional sense. The function call generated by a non-OOP compiler causes what is called early binding, a term you may not have heard before because you’ve never thought about it any other way. It means the compiler generates a call to a specific function name, and the linker resolves this call to the absolute address of the code to be executed. In OOP, the program cannot determine the address of the code until runtime, so some other scheme is necessary when a message is sent to a generic object.

To solve the problem, object-oriented languages use the concept of late binding. When you send a message to an object, the code being called isn’t determined until runtime. The compiler does ensure that the function exists and performs type checking on the arguments and return value (a language in which this isn’t true is called weakly typed), but it doesn’t know the exact code to execute.

To perform late binding, the C++ compiler inserts a special bit of code in lieu of the absolute call. This code calculates the address of the function body, using information stored in the object (this process is covered in great detail in Chapter 15). Thus, each object can behave differently according to the contents of that special bit of code. When you send a message to an object, the object actually does figure out what to do with that message.

You state that you want a function to have the flexibility of late-binding properties using the keyword virtual. You don’t need to understand the mechanics of virtual to use it, but without it you can’t do object-oriented programming in C++. In C++, you must remember to add the virtual keyword because, by default, member functions are not dynamically bound. Virtual functions allow you to express the differences in behavior of classes in the same family. Those differences are what cause polymorphic behavior.

Consider the shape example. The family of classes (all based on the same uniform interface) was diagrammed earlier in the chapter. To demonstrate polymorphism, we want to write a single piece of code that ignores the specific details of type and talks only to the base class. That code is decoupled from type-specific information, and thus is simpler to write and easier to understand. And, if a new type – a Hexagon, for example – is added through inheritance, the code you write will work just as well for the new type of Shape as it did on the existing types. Thus, the program is extensible.

If you write a function in C++ (as you will soon learn how to do):

void doStuff(Shape& s) {
  s.erase();
  // ...
  s.draw();
} 

This function speaks to any Shape, so it is independent of the specific type of object that it’s drawing and erasing (the ‘&’ means “Take the address of the object that’s passed to doStuff( ),” but it’s not important that you understand the details of that right now). If in some other part of the program we use the doStuff( ) function:

Circle c;
Triangle t;
Line l;
doStuff(c);
doStuff(t);
doStuff(l); 

The calls to doStuff( ) automatically work right, regardless of the exact type of the object.

This is actually a pretty amazing trick. Consider the line:

doStuff(c);

What’s happening here is that a Circle is being passed into a function that’s expecting a Shape. Since a Circle is a Shape it can be treated as one by doStuff( ). That is, any message that doStuff( ) can send to a Shape, a Circle can accept. So it is a completely safe and logical thing to do.

We call this process of treating a derived type as though it were its base type upcasting. The name cast is used in the sense of casting into a mold and the up comes from the way the inheritance diagram is typically arranged, with the base type at the top and the derived classes fanning out downward. Thus, casting to a base type is moving up the inheritance diagram: “upcasting.”


An object-oriented program contains some upcasting somewhere, because that’s how you decouple yourself from knowing about the exact type you’re working with. Look at the code in doStuff( ):

  s.erase();
  // ...
  s.draw();

Notice that it doesn’t say “If you’re a Circle, do this, if you’re a Square, do that, etc.” If you write that kind of code, which checks for all the possible types that a Shape can actually be, it’s messy and you need to change it every time you add a new kind of Shape. Here, you just say “You’re a shape, I know you can erase( ) and draw( ) yourself, do it, and take care of the details correctly.”

What’s impressive about the code in doStuff( ) is that, somehow, the right thing happens. Calling draw( ) for Circle causes different code to be executed than when calling draw( ) for a Square or a Line, but when the draw( ) message is sent to an anonymous Shape, the correct behavior occurs based on the actual type of the Shape. This is amazing because, as mentioned earlier, when the C++ compiler is compiling the code for doStuff( ), it cannot know exactly what types it is dealing with. So ordinarily, you’d expect it to end up calling the version of erase( ) and draw( ) for Shape, and not for the specific Circle, Square, or Line. And yet the right thing happens because of polymorphism. The compiler and runtime system handle the details; all you need to know is that it happens and more importantly how to design with it. If a member function is virtual, then when you send a message to an object, the object will do the right thing, even when upcasting is involved.

Creating and destroying objects

Technically, the domain of OOP is abstract data typing, inheritance, and polymorphism, but other issues can be at least as important. This section gives an overview of these issues.

Especially important is the way objects are created and destroyed. Where is the data for an object and how is the lifetime of that object controlled? Different programming languages use different philosophies here. C++ takes the approach that control of efficiency is the most important issue, so it gives the programmer a choice. For maximum runtime speed, the storage and lifetime can be determined while the program is being written, by placing the objects on the stack or in static storage. The stack is an area in memory that is used directly by the microprocessor to store data during program execution. Variables on the stack are sometimes called automatic or scoped variables. The static storage area is simply a fixed patch of memory that is allocated before the program begins to run. Using the stack or static storage area places a priority on the speed of storage allocation and release, which can be valuable in some situations. However, you sacrifice flexibility because you must know the exact quantity, lifetime, and type of objects while you’re writing the program. If you are trying to solve a more general problem, such as computer-aided design, warehouse management, or air-traffic control, this is too restrictive.

The second approach is to create objects dynamically in a pool of memory called the heap. In this approach you don’t know until runtime how many objects you need, what their lifetime is, or what their exact type is. Those decisions are made at the spur of the moment while the program is running. If you need a new object, you simply make it on the heap when you need it, using the new keyword. When you’re finished with the storage, you must release it using the delete keyword.

Because the storage is managed dynamically at runtime, the amount of time required to allocate storage on the heap is significantly longer than the time to create storage on the stack. (Creating storage on the stack is often a single microprocessor instruction to move the stack pointer down, and another to move it back up.) The dynamic approach makes the generally logical assumption that objects tend to be complicated, so the extra overhead of finding storage and releasing that storage will not have an important impact on the creation of an object. In addition, the greater flexibility is essential to solve general programming problems.

There’s another issue, however, and that’s the lifetime of an object. If you create an object on the stack or in static storage, the compiler determines how long the object lasts and can automatically destroy it. However, if you create it on the heap, the compiler has no knowledge of its lifetime. In C++, the programmer must determine programmatically when to destroy the object, and then perform the destruction using the delete keyword. As an alternative, the environment can provide a feature called a garbage collector that automatically discovers when an object is no longer in use and destroys it. Of course, writing programs using a garbage collector is much more convenient, but it requires that all applications must be able to tolerate the existence of the garbage collector and the overhead for garbage collection. This does not meet the design requirements of the C++ language and so it was not included, although third-party garbage collectors exist for C++.

Exception handling:
dealing with errors

Ever since the beginning of programming languages, error handling has been one of the most difficult issues. Because it’s so hard to design a good error-handling scheme, many languages simply ignore the issue, passing the problem on to library designers who come up with halfway measures that can work in many situations but can easily be circumvented, generally by just ignoring them. A major problem with most error-handling schemes is that they rely on programmer vigilance in following an agreed-upon convention that is not enforced by the language. If programmers are not vigilant, which often occurs when they are in a hurry, these schemes can easily be forgotten.

Exception handling wires error handling directly into the programming language and sometimes even the operating system. An exception is an object that is “thrown” from the site of the error and can be “caught” by an appropriate exception handler designed to handle that particular type of error. It’s as if exception handling is a different, parallel path of execution that can be taken when things go wrong. And because it uses a separate execution path, it doesn’t need to interfere with your normally-executing code. This makes that code simpler to write since you aren’t constantly forced to check for errors. In addition, a thrown exception is unlike an error value that’s returned from a function or a flag that’s set by a function in order to indicate an error condition – these can be ignored. An exception cannot be ignored so it’s guaranteed to be dealt with at some point. Finally, exceptions provide a way to recover reliably from a bad situation. Instead of just exiting the program, you are often able to set things right and restore the execution of a program, which produces much more robust systems.

It’s worth noting that exception handling isn’t an object-oriented feature, although in object-oriented languages the exception is normally represented with an object. Exception handling existed before object-oriented languages.

Exception handling is only lightly introduced and used in this Volume; Volume 2 (available from www.BruceEckel.com) has thorough coverage of exception handling.

Analysis and design

The object-oriented paradigm is a new and different way of thinking about programming and many folks have trouble at first knowing how to approach an OOP project. Once you know that everything is supposed to be an object, and as you learn to think more in an object-oriented style, you can begin to create “good” designs that take advantage of all the benefits that OOP has to offer.

A method (often called a methodology) is a set of processes and heuristics used to break down the complexity of a programming problem. Many OOP methods have been formulated since the dawn of object-oriented programming. This section will give you a feel for what you’re trying to accomplish when using a method.

Especially in OOP, methodology is a field of many experiments, so it is important to understand what problem the method is trying to solve before you consider adopting one. This is particularly true with C++, in which the programming language is intended to reduce the complexity (compared to C) involved in expressing a program. This may in fact alleviate the need for ever-more-complex methodologies. Instead, simpler ones may suffice in C++ for a much larger class of problems than you could handle using simple methodologies with procedural languages.

It’s also important to realize that the term “methodology” is often too grand and promises too much. Whatever you do now when you design and write a program is a method. It may be your own method, and you may not be conscious of doing it, but it is a process you go through as you create. If it is an effective process, it may need only a small tune-up to work with C++. If you are not satisfied with your productivity and the way your programs turn out, you may want to consider adopting a formal method, or choosing pieces from among the many formal methods.

While you’re going through the development process, the most important issue is this: Don’t get lost. It’s easy to do. Most of the analysis and design methods are intended to solve the largest of problems. Remember that most projects don’t fit into that category, so you can usually have successful analysis and design with a relatively small subset of what a method recommends[9]. But some sort of process, no matter how limited, will generally get you on your way in a much better fashion than simply beginning to code.

It’s also easy to get stuck, to fall into “analysis paralysis,” where you feel like you can’t move forward because you haven’t nailed down every little detail at the current stage. Remember, no matter how much analysis you do, there are some things about a system that won’t reveal themselves until design time, and more things that won’t reveal themselves until you’re coding, or not even until a program is up and running. Because of this, it’s crucial to move fairly quickly through analysis and design, and to implement a test of the proposed system.

This point is worth emphasizing. Because of the history we’ve had with procedural languages, it is commendable that a team will want to proceed carefully and understand every minute detail before moving to design and implementation. Certainly, when creating a DBMS, it pays to understand a customer’s needs thoroughly. But a DBMS is in a class of problems that is very well-posed and well-understood; in many such programs, the database structure is the problem to be tackled. The class of programming problem discussed in this chapter is of the “wild-card” (my term) variety, in which the solution isn’t simply re-forming a well-known solution, but instead involves one or more “wild-card factors” – elements for which there is no well-understood previous solution, and for which research is necessary[10]. Attempting to thoroughly analyze a wild-card problem before moving into design and implementation results in analysis paralysis because you don’t have enough information to solve this kind of problem during the analysis phase. Solving such a problem requires iteration through the whole cycle, and that requires risk-taking behavior (which makes sense, because you’re trying to do something new and the potential rewards are higher). It may seem like the risk is compounded by “rushing” into a preliminary implementation, but it can instead reduce the risk in a wild-card project because you’re finding out early whether a particular approach to the problem is viable. Product development is risk management.

It’s often proposed that you “build one to throw away.” With OOP, you may still throw part of it away, but because code is encapsulated into classes, during the first iteration you will inevitably produce some useful class designs and develop some worthwhile ideas about the system design that do not need to be thrown away. Thus, the first rapid pass at a problem not only produces critical information for the next analysis, design, and implementation iteration, it also creates a code foundation for that iteration.

That said, if you’re looking at a methodology that contains tremendous detail and suggests many steps and documents, it’s still difficult to know when to stop. Keep in mind what you’re trying to discover:

  1. What are the objects? (How do you partition your project into its component parts?)
  2. What are their interfaces? (What messages do you need to be able to send to each object?)

If you come up with nothing more than the objects and their interfaces, then you can write a program. For various reasons you might need more descriptions and documents than this, but you can’t get away with any less.

The process can be undertaken in five phases, and a phase 0 that is just the initial commitment to using some kind of structure.

Phase 0: Make a plan

You must first decide what steps you’re going to have in your process. It sounds simple (in fact, all of this sounds simple) and yet people often don’t make this decision before they start coding. If your plan is “let’s jump in and start coding,” fine. (Sometimes that’s appropriate when you have a well-understood problem.) At least agree that this is the plan.

You might also decide at this phase that some additional process structure is necessary, but not the whole nine yards. Understandably enough, some programmers like to work in “vacation mode” in which no structure is imposed on the process of developing their work; “It will be done when it’s done.” This can be appealing for awhile, but I’ve found that having a few milestones along the way helps to focus and galvanize your efforts around those milestones instead of being stuck with the single goal of “finish the project.” In addition, it divides the project into more bite-sized pieces and makes it seem less threatening (plus the milestones offer more opportunities for celebration).

When I began to study story structure (so that I will someday write a novel) I was initially resistant to the idea of structure, feeling that when I wrote I simply let it flow onto the page. But I later realized that when I write about computers the structure is clear enough so that I don’t think much about it. But I still structure my work, albeit only semi-consciously in my head. So even if you think that your plan is to just start coding, you still somehow go through the subsequent phases while asking and answering certain questions.

The mission statement

Any system you build, no matter how complicated, has a fundamental purpose, the business that it’s in, the basic need that it satisfies. If you can look past the user interface, the hardware- or system-specific details, the coding algorithms and the efficiency problems, you will eventually find the core of its being, simple and straightforward. Like the so-called high concept from a Hollywood movie, you can describe it in one or two sentences. This pure description is the starting point.

The high concept is quite important because it sets the tone for your project; it’s a mission statement. You won’t necessarily get it right the first time (you may be in a later phase of the project before it becomes completely clear), but keep trying until it feels right. For example, in an air-traffic control system you may start out with a high concept focused on the system that you’re building: “The tower program keeps track of the aircraft.” But consider what happens when you shrink the system to a very small airfield; perhaps there’s only a human controller or none at all. A more useful model won’t concern the solution you’re creating as much as it describes the problem: “Aircraft arrive, unload, service and reload, and depart.”

Phase 1: What are we making?

In the previous generation of program design (called procedural design), this is called “creating the requirements analysis and system specification.” These, of course, were places to get lost; intimidatingly-named documents that could become big projects in their own right. Their intention was good, however. The requirements analysis says “Make a list of the guidelines we will use to know when the job is done and the customer is satisfied.” The system specification says “Here’s a description of what the program will do (not how) to satisfy the requirements.” The requirements analysis is really a contract between you and the customer (even if the customer works within your company or is some other object or system). The system specification is a top-level exploration into the problem and in some sense a discovery of whether it can be done and how long it will take. Since both of these will require consensus among people (and because they will usually change over time), I think it’s best to keep them as bare as possible – ideally, to lists and basic diagrams – to save time. You might have other constraints that require you to expand them into bigger documents, but by keeping the initial document small and concise, it can be created in a few sessions of group brainstorming with a leader who dynamically creates the description. This not only solicits input from everyone, it also fosters initial buy-in and agreement by everyone on the team. Perhaps most importantly, it can kick off a project with a lot of enthusiasm.

It’s necessary to stay focused on the heart of what you’re trying to accomplish in this phase: determine what the system is supposed to do. The most valuable tool for this is a collection of what are called “use cases.” Use cases identify key features in the system that will reveal some of the fundamental classes you’ll be using. These are essentially descriptive answers to questions like[11]:

If you are designing an auto-teller, for example, the use case for a particular aspect of the functionality of the system is able to describe what the auto-teller does in every possible situation. Each of these “situations” is referred to as a scenario, and a use case can be considered a collection of scenarios. You can think of a scenario as a question that starts with: “What does the system do if...?” For example, “What does the auto-teller do if a customer has just deposited a check within 24 hours and there’s not enough in the account without the check to provide the desired withdrawal?”

Use case diagrams are intentionally simple to prevent you from getting bogged down in system implementation details prematurely:


Each stick person represents an “actor,” which is typically a human or some other kind of free agent. (These can even be other computer systems, as is the case with “ATM.”) The box represents the boundary of your system. The ellipses represent the use cases, which are descriptions of valuable work that can be performed with the system. The lines between the actors and the use cases represent the interactions.

It doesn’t matter how the system is actually implemented, as long as it looks like this to the user.

A use case does not need to be terribly complex, even if the underlying system is complex. It is only intended to show the system as it appears to the user. For example:


The use cases produce the requirements specifications by determining all the interactions that the user may have with the system. You try to discover a full set of use cases for your system, and once you’ve done that you have the core of what the system is supposed to do. The nice thing about focusing on use cases is that they always bring you back to the essentials and keep you from drifting off into issues that aren’t critical for getting the job done. That is, if you have a full set of use cases you can describe your system and move onto the next phase. You probably won’t get it all figured out perfectly on the first try, but that’s OK. Everything will reveal itself in time, and if you demand a perfect system specification at this point you’ll get stuck.

If you get stuck, you can kick-start this phase by using a rough approximation tool: describe the system in a few paragraphs and then look for nouns and verbs. The nouns can suggest actors, context of the use case (e.g. “lobby”), or artifacts manipulated in the use case. Verbs can suggest interactions between actors and use cases, and specify steps within the use case. You’ll also discover that nouns and verbs produce objects and messages during the design phase (and note that use cases describe interactions between subsystems, so the “noun and verb” technique can be used only as a brainstorming tool as it does not generate use cases) [12].

The boundary between a use case and an actor can point out the existence of a user interface, but it does not define such a user interface. For a process of defining and creating user interfaces, see Software for Use by Larry Constantine and Lucy Lockwood, (Addison Wesley Longman, 1999) or go to www.ForUse.com.

Although it’s a black art, at this point some kind of basic scheduling is important. You now have an overview of what you’re building so you’ll probably be able to get some idea of how long it will take. A lot of factors come into play here. If you estimate a long schedule then the company might decide not to build it (and thus use their resources on something more reasonable – that’s a good thing). Or a manager might have already decided how long the project should take and will try to influence your estimate. But it’s best to have an honest schedule from the beginning and deal with the tough decisions early. There have been a lot of attempts to come up with accurate scheduling techniques (like techniques to predict the stock market), but probably the best approach is to rely on your experience and intuition. Get a gut feeling for how long it will really take, then double that and add 10 percent. Your gut feeling is probably correct; you can get something working in that time. The “doubling” will turn that into something decent, and the 10 percent will deal with the final polishing and details[13]. However you want to explain it, and regardless of the moans and manipulations that happen when you reveal such a schedule, it just seems to work out that way.

Phase 2: How will we build it?

In this phase you must come up with a design that describes what the classes look like and how they will interact. An excellent technique in determining classes and interactions is the Class-Responsibility-Collaboration (CRC) card. Part of the value of this tool is that it’s so low-tech: you start out with a set of blank 3” by 5” cards, and you write on them. Each card represents a single class, and on the card you write:

  1. The name of the class. It’s important that this name capture the essence of what the class does, so that it makes sense at a glance.
  2. The “responsibilities” of the class: what it should do. This can typically be summarized by just stating the names of the member functions (since those names should be descriptive in a good design), but it does not preclude other notes. If you need to seed the process, look at the problem from a lazy programmer’s standpoint: What objects would you like to magically appear to solve your problem?
  3. The “collaborations” of the class: what other classes does it interact with? “Interact” is an intentionally broad term; it could mean aggregation or simply that some other object exists that will perform services for an object of the class. Collaborations should also consider the audience for this class. For example, if you create a class Firecracker, who is going to observe it, a Chemist or a Spectator? The former will want to know what chemicals go into the construction, and the latter will respond to the colors and shapes released when it explodes.

You may feel like the cards should be bigger because of all the information you’d like to get on them, but they are intentionally small, not only to keep your classes small but also to keep you from getting into too much detail too early. If you can’t fit all you need to know about a class on a small card, the class is too complex (either you’re getting too detailed, or you should create more than one class). The ideal class should be understood at a glance. The idea of CRC cards is to assist you in coming up with a first cut of the design so that you can get the big picture and then refine your design.

One of the great benefits of CRC cards is in communication. It’s best done real-time, in a group, without computers. Each person takes responsibility for several classes (which at first have no names or other information). You run a live simulation by solving one scenario at a time, deciding which messages are sent to the various objects to satisfy each scenario. As you go through this process, you discover the classes that you need along with their responsibilities and collaborations, and you fill out the cards as you do this. When you’ve moved through all the use cases, you should have a fairly complete first cut of your design.

Before I began using CRC cards, the most successful consulting experiences I had when coming up with an initial design involved standing in front of a team, who hadn’t built an OOP project before, and drawing objects on a whiteboard. We talked about how the objects should communicate with each other, and erased some of them and replaced them with other objects. Effectively, I was managing all the “CRC cards” on the whiteboard. The team (who knew what the project was supposed to do) actually created the design; they “owned” the design rather than having it given to them. All I was doing was guiding the process by asking the right questions, trying out the assumptions, and taking the feedback from the team to modify those assumptions. The true beauty of the process was that the team learned how to do object-oriented design not by reviewing abstract examples, but by working on the one design that was most interesting to them at that moment: theirs.

Once you’ve come up with a set of CRC cards, you may want to create a more formal description of your design using UML[14]. You don’t need to use UML, but it can be helpful, especially if you want to put up a diagram on the wall for everyone to ponder, which is a good idea. An alternative to UML is a textual description of the objects and their interfaces, or, depending on your programming language, the code itself[15].

UML also provides an additional diagramming notation for describing the dynamic model of your system. This is helpful in situations in which the state transitions of a system or subsystem are dominant enough that they need their own diagrams (such as in a control system). You may also need to describe the data structures, for systems or subsystems in which data is a dominant factor (such as a database).

You’ll know you’re done with phase 2 when you have described the objects and their interfaces. Well, most of them – there are usually a few that slip through the cracks and don’t make themselves known until phase 3. But that’s OK. All you are concerned with is that you eventually discover all of your objects. It’s nice to discover them early in the process but OOP provides enough structure so that it’s not so bad if you discover them later. In fact, the design of an object tends to happen in five stages, throughout the process of program development.

Five stages of object design

The design life of an object is not limited to the time when you’re writing the program. Instead, the design of an object appears over a sequence of stages. It’s helpful to have this perspective because you stop expecting perfection right away; instead, you realize that the understanding of what an object does and what it should look like happens over time. This view also applies to the design of various types of programs; the pattern for a particular type of program emerges through struggling again and again with that problem (Design Patterns are covered in Volume 2). Objects, too, have their patterns that emerge through understanding, use, and reuse.

1. Object discovery. This stage occurs during the initial analysis of a program. Objects may be discovered by looking for external factors and boundaries, duplication of elements in the system, and the smallest conceptual units. Some objects are obvious if you already have a set of class libraries. Commonality between classes suggesting base classes and inheritance may appear right away, or later in the design process.

2. Object assembly. As you’re building an object you’ll discover the need for new members that didn’t appear during discovery. The internal needs of the object may require other classes to support it.

3. System construction. Once again, more requirements for an object may appear at this later stage. As you learn, you evolve your objects. The need for communication and interconnection with other objects in the system may change the needs of your classes or require new classes. For example, you may discover the need for facilitator or helper classes, such as a linked list, that contain little or no state information and simply help other classes function.

4. System extension. As you add new features to a system you may discover that your previous design doesn’t support easy system extension. With this new information, you can restructure parts of the system, possibly adding new classes or class hierarchies.

5. Object reuse. This is the real stress test for a class. If someone tries to reuse it in an entirely new situation, they’ll probably discover some shortcomings. As you change a class to adapt to more new programs, the general principles of the class will become clearer, until you have a truly reusable type. However, don’t expect most objects from a system design to be reusable – it is perfectly acceptable for the bulk of your objects to be system-specific. Reusable types tend to be less common, and they must solve more general problems in order to be reusable.

Guidelines for object development

These stages suggest some guidelines when thinking about developing your classes:

  1. Let a specific problem generate a class, then let the class grow and mature during the solution of other problems.
  2. Remember, discovering the classes you need (and their interfaces) is the majority of the system design. If you already had those classes, this would be an easy project.
  3. Don’t force yourself to know everything at the beginning; learn as you go. This will happen anyway.
  4. Start programming; get something working so you can prove or disprove your design. Don’t fear that you’ll end up with procedural-style spaghetti code – classes partition the problem and help control anarchy and entropy. Bad classes do not break good classes.
  5. Always keep it simple. Little clean objects with obvious utility are better than big complicated interfaces. When decision points come up, use an Occam’s Razor approach: Consider the choices and select the one that is simplest, because simple classes are almost always best. Start small and simple, and you can expand the class interface when you understand it better, but as time goes on, it’s difficult to remove elements from a class.

Phase 3: Build the core

This is the initial conversion from the rough design into a compiling and executing body of code that can be tested, and especially that will prove or disprove your architecture. This is not a one-pass process, but rather the beginning of a series of steps that will iteratively build the system, as you’ll see in phase 4.

Your goal is to find the core of your system architecture that needs to be implemented in order to generate a running system, no matter how incomplete that system is in this initial pass. You’re creating a framework that you can build upon with further iterations. You’re also performing the first of many system integrations and tests, and giving the stakeholders feedback about what their system will look like and how it is progressing. Ideally, you are also exposing some of the critical risks. You’ll probably also discover changes and improvements that can be made to your original architecture – things you would not have learned without implementing the system.

Part of building the system is the reality check that you get from testing against your requirements analysis and system specification (in whatever form they exist). Make sure that your tests verify the requirements and use cases. When the core of the system is stable, you’re ready to move on and add more functionality.

Phase 4: Iterate the use cases

Once the core framework is running, each feature set you add is a small project in itself. You add a feature set during an iteration, a reasonably short period of development.

How big is an iteration? Ideally, each iteration lasts one to three weeks (this can vary based on the implementation language). At the end of that period, you have an integrated, tested system with more functionality than it had before. But what’s particularly interesting is the basis for the iteration: a single use case. Each use case is a package of related functionality that you build into the system all at once, during one iteration. Not only does this give you a better idea of what the scope of a use case should be, but it also gives more validation to the idea of a use case, since the concept isn’t discarded after analysis and design, but instead it is a fundamental unit of development throughout the software-building process.

You stop iterating when you achieve target functionality or an external deadline arrives and the customer can be satisfied with the current version. (Remember, software is a subscription business.) Because the process is iterative, you have many opportunities to ship a product instead of a single endpoint; open-source projects work exclusively in an iterative, high-feedback environment, which is precisely what makes them successful.

An iterative development process is valuable for many reasons. You can reveal and resolve critical risks early, the customers have ample opportunity to change their minds, programmer satisfaction is higher, and the project can be steered with more precision. But an additional important benefit is the feedback to the stakeholders, who can see by the current state of the product exactly where everything lies. This may reduce or eliminate the need for mind-numbing status meetings and increase the confidence and support from the stakeholders.

Phase 5: Evolution

This is the point in the development cycle that has traditionally been called “maintenance,” a catch-all term that can mean everything from “getting it to work the way it was really supposed to in the first place” to “adding features that the customer forgot to mention” to the more traditional “fixing the bugs that show up” and “adding new features as the need arises.” So many misconceptions have been applied to the term “maintenance” that it has taken on a slightly deceiving quality, partly because it suggests that you’ve actually built a pristine program and all you need to do is change parts, oil it, and keep it from rusting. Perhaps there’s a better term to describe what’s going on.

I’ll use the term evolution[16]. That is, “You won’t get it right the first time, so give yourself the latitude to learn and to go back and make changes.” You might need to make a lot of changes as you learn and understand the problem more deeply. The elegance you’ll produce if you evolve until you get it right will pay off, both in the short and the long term. Evolution is where your program goes from good to great, and where those issues that you didn’t really understand in the first pass become clear. It’s also where your classes can evolve from single-project usage to reusable resources.

What it means to “get it right” isn’t just that the program works according to the requirements and the use cases. It also means that the internal structure of the code makes sense to you, and feels like it fits together well, with no awkward syntax, oversized objects, or ungainly exposed bits of code. In addition, you must have some sense that the program structure will survive the changes that it will inevitably go through during its lifetime, and that those changes can be made easily and cleanly. This is no small feat. You must not only understand what you’re building, but also how the program will evolve (what I call the vector of change[17]). Fortunately, object-oriented programming languages are particularly adept at supporting this kind of continuing modification – the boundaries created by the objects are what tend to keep the structure from breaking down. They also allow you to make changes – ones that would seem drastic in a procedural program – without causing earthquakes throughout your code. In fact, support for evolution might be the most important benefit of OOP.

With evolution, you create something that at least approximates what you think you’re building, and then you kick the tires, compare it to your requirements and see where it falls short. Then you can go back and fix it by redesigning and re-implementing the portions of the program that didn’t work right[18]. You might actually need to solve the problem, or an aspect of the problem, several times before you hit on the right solution. (A study of Design Patterns, described in Volume 2, is usually helpful here.)

Evolution also occurs when you build a system, see that it matches your requirements, and then discover it wasn’t actually what you wanted. When you see the system in operation, you find that you really wanted to solve a different problem. If you think this kind of evolution is going to happen, then you owe it to yourself to build your first version as quickly as possible so you can find out if it is indeed what you want.

Perhaps the most important thing to remember is that by default – by definition, really – if you modify a class then its super- and subclasses will still function. You need not fear modification (especially if you have a built-in set of unit tests to verify the correctness of your modifications). Modification won’t necessarily break the program, and any change in the outcome will be limited to subclasses and/or specific collaborators of the class you change.

Plans pay off

Of course you wouldn’t build a house without a lot of carefully-drawn plans. If you build a deck or a dog house, your plans won’t be so elaborate but you’ll probably still start with some kind of sketches to guide you on your way. Software development has gone to extremes. For a long time, people didn’t have much structure in their development, but then big projects began failing. In reaction, we ended up with methodologies that had an intimidating amount of structure and detail, primarily intended for those big projects. These methodologies were too scary to use – it looked like you’d spend all your time writing documents and no time programming. (This was often the case.) I hope that what I’ve shown you here suggests a middle path – a sliding scale. Use an approach that fits your needs (and your personality). No matter how minimal you choose to make it, some kind of plan will make a big improvement in your project as opposed to no plan at all. Remember that, by most estimates, over 50 percent of projects fail (some estimates go up to 70 percent!).

By following a plan – preferably one that is simple and brief – and coming up with design structure before coding, you’ll discover that things fall together far more easily than if you dive in and start hacking, and you’ll also realize a great deal of satisfaction. It’s my experience that coming up with an elegant solution is deeply satisfying at an entirely different level; it feels closer to art than technology. And elegance always pays off; it’s not a frivolous pursuit. Not only does it give you a program that’s easier to build and debug, but it’s also easier to understand and maintain, and that’s where the financial value lies.

Extreme programming

I have studied analysis and design techniques, on and off, since I was in graduate school. The concept of Extreme Programming (XP) is the most radical, and delightful, that I’ve seen. You can find it chronicled in Extreme Programming Explained by Kent Beck (Addison-Wesley 2000) and on the Web at www.xprogramming.com.

XP is both a philosophy about programming work and a set of guidelines to do it. Some of these guidelines are reflected in other recent methodologies, but the two most important and distinct contributions, in my opinion, are “write tests first” and “pair programming.” Although he argues strongly for the whole process, Beck points out that if you adopt only these two practices you’ll greatly improve your productivity and reliability.

Write tests first

Testing has traditionally been relegated to the last part of a project, after you’ve “gotten everything working, but just to be sure.” It’s implicitly had a low priority, and people who specialize in it have not been given a lot of status and have often even been cordoned off in a basement, away from the “real programmers.” Test teams have responded in kind, going so far as to wear black clothing and cackling with glee whenever they broke something (to be honest, I’ve had this feeling myself when breaking C++ compilers).

XP completely revolutionizes the concept of testing by giving it equal (or even greater) priority than the code. In fact, you write the tests before you write the code that’s being tested, and the tests stay with the code forever. The tests must be executed successfully every time you do an integration of the project (which is often, sometimes more than once a day).

Writing tests first has two extremely important effects.

First, it forces a clear definition of the interface of a class. I’ve often suggested that people “imagine the perfect class to solve a particular problem” as a tool when trying to design the system. The XP testing strategy goes further than that – it specifies exactly what the class must look like, to the consumer of that class, and exactly how the class must behave. In no uncertain terms. You can write all the prose, or create all the diagrams you want describing how a class should behave and what it looks like, but nothing is as real as a set of tests. The former is a wish list, but the tests are a contract that is enforced by the compiler and the running program. It’s hard to imagine a more concrete description of a class than the tests.

While creating the tests, you are forced to completely think out the class and will often discover needed functionality that might be missed during the thought experiments of UML diagrams, CRC cards, use cases, etc.

The second important effect of writing the tests first comes from running the tests every time you do a build of your software. This activity gives you the other half of the testing that’s performed by the compiler. If you look at the evolution of programming languages from this perspective, you’ll see that the real improvements in the technology have actually revolved around testing. Assembly language checked only for syntax, but C imposed some semantic restrictions, and these prevented you from making certain types of mistakes. OOP languages impose even more semantic restrictions, which if you think about it are actually forms of testing. “Is this data type being used properly? Is this function being called properly?” are the kinds of tests that are being performed by the compiler or run-time system. We’ve seen the results of having these tests built into the language: people have been able to write more complex systems, and get them to work, with much less time and effort. I’ve puzzled over why this is, but now I realize it’s the tests: you do something wrong, and the safety net of the built-in tests tells you there’s a problem and points you to where it is.

But the built-in testing afforded by the design of the language can only go so far. At some point, you must step in and add the rest of the tests that produce a full suite (in cooperation with the compiler and run-time system) that verifies all of your program. And, just like having a compiler watching over your shoulder, wouldn’t you want these tests helping you right from the beginning? That’s why you write them first, and run them automatically with every build of your system. Your tests become an extension of the safety net provided by the language.

One of the things that I’ve discovered about the use of more and more powerful programming languages is that I am emboldened to try more brazen experiments, because I know that the language will keep me from wasting my time chasing bugs. The XP test scheme does the same thing for your entire project. Because you know your tests will always catch any problems that you introduce (and you regularly add any new tests as you think of them), you can make big changes when you need to without worrying that you’ll throw the whole project into complete disarray. This is incredibly powerful.


Pair programming

Pair programming goes against the rugged individualism that we’ve been indoctrinated into from the beginning, through school (where we succeed or fail on our own, and working with our neighbors is considered “cheating”) and media, especially Hollywood movies in which the hero is usually fighting against mindless conformity[19]. Programmers, too, are considered paragons of individuality – “cowboy coders” as Larry Constantine likes to say. And yet XP, which is itself battling against conventional thinking, says that code should be written with two people per workstation. And that this should be done in an area with a group of workstations, without the barriers that the facilities design people are so fond of. In fact, Beck says that the first task of converting to XP is to arrive with screwdrivers and Allen wrenches and take apart everything that gets in the way.[20] (This will require a manager who can deflect the ire of the facilities department.)

The value of pair programming is that one person is actually doing the coding while the other is thinking about it. The thinker keeps the big picture in mind, not only the picture of the problem at hand, but the guidelines of XP. If two people are working, it’s less likely that one of them will get away with saying, “I don’t want to write the tests first,” for example. And if the coder gets stuck, they can swap places. If both of them get stuck, their musings may be overheard by someone else in the work area who can contribute. Working in pairs keeps things flowing and on track. Probably more important, it makes programming a lot more social and fun.

I’ve begun using pair programming during the exercise periods in some of my seminars and it seems to significantly improve everyone’s experience.

Why C++ succeeds

Part of the reason C++ has been so successful is that the goal was not just to turn C into an OOP language (although it started that way), but also to solve many other problems facing developers today, especially those who have large investments in C. Traditionally, OOP languages have suffered from the attitude that you should abandon everything you know and start from scratch with a new set of concepts and a new syntax, arguing that it’s better in the long run to lose all the old baggage that comes with procedural languages. This may be true, in the long run. But in the short run, a lot of that baggage was valuable. The most valuable elements may not be the existing code base (which, given adequate tools, could be translated), but instead the existing mind base. If you’re a functioning C programmer and must drop everything you know about C in order to adopt a new language, you immediately become much less productive for many months, until your mind fits around the new paradigm. Whereas if you can leverage off of your existing C knowledge and expand on it, you can continue to be productive with what you already know while moving into the world of object-oriented programming. As everyone has his or her own mental model of programming, this move is messy enough as it is without the added expense of starting with a new language model from square one. So the reason for the success of C++, in a nutshell, is economic: It still costs to move to OOP, but C++ may cost less[21].

The goal of C++ is improved productivity. This productivity comes in many ways, but the language is designed to aid you as much as possible, while hindering you as little as possible with arbitrary rules or any requirement that you use a particular set of features. C++ is designed to be practical; C++ language design decisions were based on providing the maximum benefits to the programmer (at least, from the world view of C).

A better C

You get an instant win even if you continue to write C code because C++ has closed many holes in the C language and provides better type checking and compile-time analysis. You’re forced to declare functions so that the compiler can check their use. The need for the preprocessor has virtually been eliminated for value substitution and macros, which removes a set of difficult-to-find bugs. C++ has a feature called references that allows more convenient handling of addresses for function arguments and return values. The handling of names is improved through a feature called function overloading, which allows you to use the same name for different functions. A feature called namespaces also improves the control of names. There are numerous smaller features that improve the safety of C.

You’re already on the learning curve

The problem with learning a new language is productivity. No company can afford to suddenly lose a productive software engineer because he or she is learning a new language. C++ is an extension to C, not a complete new syntax and programming model. It allows you to continue creating useful code, applying the features gradually as you learn and understand them. This may be one of the most important reasons for the success of C++.

In addition, all of your existing C code is still viable in C++, but because the C++ compiler is pickier, you’ll often find hidden C errors when recompiling the code in C++.

Efficiency

Sometimes it is appropriate to trade execution speed for programmer productivity. A financial model, for example, may be useful for only a short period of time, so it’s more important to create the model rapidly than to execute it rapidly. However, most applications require some degree of efficiency, so C++ always errs on the side of greater efficiency. Because C programmers tend to be very efficiency-conscious, this is also a way to ensure that they won’t be able to argue that the language is too fat and slow. A number of features in C++ are intended to allow you to tune for performance when the generated code isn’t efficient enough.

Not only do you have the same low-level control as in C (and the ability to directly write assembly language within a C++ program), but anecdotal evidence suggests that the program speed for an object-oriented C++ program tends to be within ±10% of a program written in C, and often much closer[22]. The design produced for an OOP program may actually be more efficient than the C counterpart.

Systems are easier
to express and understand

Classes designed to fit the problem tend to express it better. This means that when you write the code, you’re describing your solution in the terms of the problem space (“Put the grommet in the bin”) rather than the terms of the computer, which is the solution space (“Set the bit in the chip that means that the relay will close”). You deal with higher-level concepts and can do much more with a single line of code.

The other benefit of this ease of expression is maintenance, which (if reports can be believed) takes a huge portion of the cost over a program’s lifetime. If a program is easier to understand, then it’s easier to maintain. This can also reduce the cost of creating and maintaining the documentation.

Maximal leverage with libraries

The fastest way to create a program is to use code that’s already written: a library. A major goal in C++ is to make library use easier. This is accomplished by casting libraries into new data types (classes), so that bringing in a library means adding new types to the language. Because the C++ compiler takes care of how the library is used – guaranteeing proper initialization and cleanup, and ensuring that functions are called properly – you can focus on what you want the library to do, not how you have to do it.

Because names can be sequestered to portions of your program via C++ namespaces, you can use as many libraries as you want without the kinds of name clashes you’d run into with C.

Source-code reuse with templates

There is a significant class of types that require source-code modification in order to reuse them effectively. The template feature in C++ performs the source code modification automatically, making it an especially powerful tool for reusing library code. A type that you design using templates will work effortlessly with many other types. Templates are especially nice because they hide the complexity of this kind of code reuse from the client programmer.

Error handling

Error handling in C is a notorious problem, and one that is often ignored – finger-crossing is usually involved. If you’re building a large, complex program, there’s nothing worse than having an error buried somewhere with no clue as to where it came from. C++ exception handling (introduced in this Volume, and fully covered in Volume 2, which is downloadable from www.BruceEckel.com) is a way to guarantee that an error is noticed and that something happens as a result.

Programming in the large

Many traditional languages have built-in limitations to program size and complexity. BASIC, for example, can be great for pulling together quick solutions for certain classes of problems, but if the program gets more than a few pages long or ventures out of the normal problem domain of that language, it’s like trying to swim through an ever-more viscous fluid. C, too, has these limitations. For example, when a program gets beyond perhaps 50,000 lines of code, name collisions start to become a problem – effectively, you run out of function and variable names. Another particularly bad problem is the little holes in the C language – errors buried in a large program can be extremely difficult to find.

There’s no clear line that tells you when your language is failing you, and even if there were, you’d ignore it. You don’t say, “My BASIC program just got too big; I’ll have to rewrite it in C!” Instead, you try to shoehorn a few more lines in to add that one new feature. So the extra costs come creeping up on you.

C++ is designed to aid programming in the large, that is, to erase those creeping-complexity boundaries between a small program and a large one. You certainly don’t need to use OOP, templates, namespaces, and exception handling when you’re writing a hello-world style utility program, but those features are there when you need them. And the compiler is aggressive about ferreting out bug-producing errors for small and large programs alike.


Strategies for transition

If you buy into OOP, your next question is probably, “How can I get my manager/colleagues/department/peers to start using objects?” Think about how you – one independent programmer – would go about learning to use a new language and a new programming paradigm. You’ve done it before. First comes education and examples; then comes a trial project to give you a feel for the basics without doing anything too confusing. Then comes a “real world” project that actually does something useful. Throughout your first projects you continue your education by reading, asking questions of experts, and trading hints with friends. This is the approach many experienced programmers suggest for the switch from C to C++. Switching an entire company will of course introduce certain group dynamics, but it will help at each step to remember how one person would do it.

Guidelines

Here are some guidelines to consider when making the transition to OOP and C++:

1. Training

The first step is some form of education. Remember the company’s investment in plain C code, and try not to throw everything into disarray for six to nine months while everyone puzzles over how multiple inheritance works. Pick a small group for indoctrination, preferably one composed of people who are curious, work well together, and can function as their own support network while they’re learning C++.

An alternative approach that is sometimes suggested is the education of all company levels at once, including overview courses for strategic managers as well as design and programming courses for project builders. This is especially good for smaller companies making fundamental shifts in the way they do things, or at the division level of larger companies. Because the cost is higher, however, some may choose to start with project-level training, do a pilot project (possibly with an outside mentor), and let the project team become the teachers for the rest of the company.

2. Low-risk project

Try a low-risk project first and allow for mistakes. Once you’ve gained some experience, you can either seed other projects from members of this first team or use the team members as an OOP technical support staff. This first project may not work right the first time, so it should not be mission-critical for the company. It should be simple, self-contained, and instructive; this means that it should involve creating classes that will be meaningful to the other programmers in the company when they get their turn to learn C++.

3. Model from success

Seek out examples of good object-oriented design before starting from scratch. There’s a good probability that someone has solved your problem already, and if they haven’t solved it exactly you can probably apply what you’ve learned about abstraction to modify an existing design to fit your needs. This is the general concept of design patterns, covered in Volume 2.

4. Use existing class libraries

The primary economic motivation for switching to OOP is the easy use of existing code in the form of class libraries (in particular, the Standard C++ libraries, which are covered in depth in Volume two of this book). The shortest application development cycle will result when you don’t have to write anything but main( ), creating and using objects from off-the-shelf libraries. However, some new programmers don’t understand this, are unaware of existing class libraries, or, through fascination with the language, desire to write classes that may already exist. Your success with OOP and C++ will be optimized if you make an effort to seek out and reuse other people’s code early in the transition process.

5. Don’t rewrite existing code in C++

Although compiling your C code with a C++ compiler usually produces (sometimes tremendous) benefits by finding problems in the old code, it is not usually the best use of your time to take existing, functional code and rewrite it in C++. (If you must turn it into objects, you can “wrap” the C code in C++ classes.) There are incremental benefits, especially if the code is slated for reuse. But chances are you aren’t going to see the dramatic increases in productivity that you hope for in your first few projects unless that project is a new one. C++ and OOP shine best when taking a project from concept to reality.


Management obstacles

If you’re a manager, your job is to acquire resources for your team, to overcome barriers to your team’s success, and in general to try to provide the most productive and enjoyable environment so your team is most likely to perform those miracles that are always being asked of you. Moving to C++ falls in all three of these categories, and it would be wonderful if it didn’t cost you anything as well. Although moving to C++ may be cheaper – depending on your constraints[23] – than the OOP alternatives for a team of C programmers (and probably for programmers in other procedural languages), it isn’t free, and there are obstacles you should be aware of before trying to sell the move to C++ within your company and embarking on the move itself.


Startup costs

The cost of moving to C++ is more than just the acquisition of C++ compilers (the GNU C++ compiler, one of the very best, is free). Your medium- and long-term costs will be minimized if you invest in training (and possibly mentoring for your first project) and also if you identify and purchase class libraries that solve your problem rather than trying to build those libraries yourself. These are hard-money costs that must be factored into a realistic proposal. In addition, there are the hidden costs in loss of productivity while learning a new language and possibly a new programming environment. Training and mentoring can certainly minimize these, but team members must overcome their own struggles to understand the new technology. During this process they will make more mistakes (this is a feature, because acknowledged mistakes are the fastest path to learning) and be less productive. Even then, with some types of programming problems, the right classes, and the right development environment, it’s possible to be more productive while you’re learning C++ (even considering that you’re making more mistakes and writing fewer lines of code per day) than if you’d stayed with C.


Performance issues

A common question is, “Doesn’t OOP automatically make my programs a lot bigger and slower?” The answer is, “It depends.” Most traditional OOP languages were designed with experimentation and rapid prototyping in mind rather than lean-and-mean operation. Thus, they virtually guaranteed a significant increase in size and decrease in speed. C++, however, is designed with production programming in mind. When your focus is on rapid prototyping, you can throw together components as fast as possible while ignoring efficiency issues. If you’re using any third party libraries, these are usually already optimized by their vendors; in any case it’s not an issue while you’re in rapid-development mode. When you have a system that you like, if it’s small and fast enough, then you’re done. If not, you begin tuning with a profiling tool, looking first for speedups that can be done with simple applications of built-in C++ features. If that doesn’t help, you look for modifications that can be made in the underlying implementation so no code that uses a particular class needs to be changed. Only if nothing else solves the problem do you need to change the design. The fact that performance is so critical in that portion of the design is an indicator that it must be part of the primary design criteria. You have the benefit of finding this out early using rapid development.

As mentioned earlier, the number that is most often given for the difference in size and speed between C and C++ is ±10%, and often much closer to par. You might even get a significant improvement in size and speed when using C++ rather than C because the design you make for C++ could be quite different from the one you’d make for C.

The evidence for size and speed comparisons between C and C++ tends to be anecdotal and is likely to remain so. Regardless of the number of people who suggest that a company try the same project using C and C++, no company is likely to waste money that way unless it’s very big and interested in such research projects. Even then, it seems like the money could be better spent. Almost universally, programmers who have moved from C (or some other procedural language) to C++ (or some other OOP language) have had the personal experience of a great acceleration in their programming productivity, and that’s the most compelling argument you can find.

Common design errors

When starting your team into OOP and C++, programmers will typically go through a series of common design errors. This often happens because of too little feedback from experts during the design and implementation of early projects, because no experts have been developed within the company and there may be resistance to retaining consultants. It’s easy to feel that you understand OOP too early in the cycle and go off on a bad tangent. Something that’s obvious to someone experienced with the language may be a subject of great internal debate for a novice. Much of this trauma can be skipped by using an experienced outside expert for training and mentoring.

On the other hand, the fact that it is easy to make these design errors points to C++’s main drawback: its backward compatibility with C (of course, that’s also its main strength). To accomplish the feat of being able to compile C code, the language had to make some compromises, which have resulted in a number of “dark corners.” These are a reality, and comprise much of the learning curve for the language. In this book and the subsequent volume (and in other books; see Appendix C), I try to reveal most of the pitfalls you are likely to encounter when working with C++. You should always be aware that there are some holes in the safety net.

Summary

This chapter attempts to give you a feel for the broad issues of object-oriented programming and C++, including why OOP is different, and why C++ in particular is different, concepts of OOP methodologies, and finally the kinds of issues you will encounter when moving your own company to OOP and C++.

OOP and C++ may not be for everyone. It’s important to evaluate your own needs and decide whether C++ will optimally satisfy those needs, or if you might be better off with another programming system (including the one you’re currently using). If you know that your needs will be very specialized for the foreseeable future and if you have specific constraints that may not be satisfied by C++, then you owe it to yourself to investigate the alternatives[24]. Even if you eventually choose C++ as your language, you’ll at least understand what the options were and have a clear vision of why you took that direction.

You know what a procedural program looks like: data definitions and function calls. To find the meaning of such a program you have to work a little, looking through the function calls and low-level concepts to create a model in your mind. This is the reason we need intermediate representations when designing procedural programs – by themselves, these programs tend to be confusing because the terms of expression are oriented more toward the computer than to the problem you’re solving.

Because C++ adds many new concepts to the C language, your natural assumption may be that the main( ) in a C++ program will be far more complicated than for the equivalent C program. Here, you’ll be pleasantly surprised: A well-written C++ program is generally far simpler and much easier to understand than the equivalent C program. What you’ll see are the definitions of the objects that represent concepts in your problem space (rather than the issues of the computer representation) and messages sent to those objects to represent the activities in that space. One of the delights of object-oriented programming is that, with a well-designed program, it’s easy to understand the code by reading it. Usually there’s a lot less code, as well, because many of your problems will be solved by reusing existing library code.


2: Making & Using Objects


[4] See Multiparadigm Programming in Leda by Timothy Budd (Addison-Wesley 1995).

[5] You can find an interesting implementation of this problem in Volume 2 of this book, available at www.BruceEckel.com.

[6] Some people make a distinction, stating that type determines the interface while class is a particular implementation of that interface.

[7] I’m indebted to my friend Scott Meyers for this term.

[8] This is usually enough detail for most diagrams, and you don’t need to get specific about whether you’re using aggregation or composition.

[9] An excellent example of this is UML Distilled, by Martin Fowler (Addison-Wesley 2000), which reduces the sometimes-overwhelming UML process to a manageable subset.

[10] My rule of thumb for estimating such projects: If there’s more than one wild card, don’t even try to plan how long it’s going to take or how much it will cost until you’ve created a working prototype. There are too many degrees of freedom.

[11] Thanks for help from James H Jarrett.

[12] More information on use cases can be found in Applying Use Cases by Schneider & Winters (Addison-Wesley 1998) and Use Case Driven Object Modeling with UML by Rosenberg (Addison-Wesley 1999).

[13] My personal take on this has changed lately. Doubling and adding 10 percent will give you a reasonably accurate estimate (assuming there are not too many wild-card factors), but you still have to work quite diligently to finish in that time. If you want time to really make it elegant and to enjoy yourself in the process, the correct multiplier is more like three or four times, I believe.

[14] For starters, I recommend the aforementioned UML Distilled.

[15] Python (www.Python.org) is often used as “executable pseudocode.”

[16] At least one aspect of evolution is covered in Martin Fowler’s book Refactoring: improving the design of existing code (Addison-Wesley 1999). Be forewarned that this book uses Java examples exclusively.

[17] This term is explored in the Design Patterns chapter in Volume 2.

[18] This is something like “rapid prototyping,” where you were supposed to build a quick-and-dirty version so that you could learn about the system, and then throw away your prototype and build it right. The trouble with rapid prototyping is that people didn’t throw away the prototype, but instead built upon it. Combined with the lack of structure in procedural programming, this often produced messy systems that were expensive to maintain.

[19] Although this may be a more American perspective, the stories of Hollywood reach everywhere.

[20] Including (especially) the PA system. I once worked in a company that insisted on broadcasting every phone call that arrived for every executive, and it constantly interrupted our productivity (but the managers couldn’t begin to conceive of stifling such an important service as the PA). Finally, when no one was looking I started snipping speaker wires.

[21] I say “may” because, due to the complexity of C++, it might actually be cheaper to move to Java. But the decision of which language to choose has many factors, and in this book I’ll assume that you’ve chosen C++.

[22] However, look at Dan Saks’ columns in the C/C++ User’s Journal for some important investigations into C++ library performance.

[23] Because of its productivity improvements, the Java language should also be considered here.

[24] In particular, I recommend looking at Java (http://java.sun.com) and Python (http://www.Python.org).

[ Previous Chapter ] [ Table of Contents ] [ Index ] [ Next Chapter ]
Last Update:09/27/2001