Chapter 4. The Java Type System

In this chapter, we move beyond basic object-oriented programming with classes and into the additional concepts required to work effectively with Java’s type system.

Note

A statically typed language is one in which variables have definite types, and where it is a compile-time error to assign a value of an incompatible type to a variable. Languages that only check type compatibility at runtime are called dynamically typed.

Java is a fairly classic example of a statically typed language. JavaScript is an example of a dynamically typed language that allows any variable to store any type of value.

The Java type system involves not only classes and primitive types, but also other kinds of reference type that are related to the basic concept of a class, but which differ in some way, and are usually treated in a special way by javac or the JVM.

We have already met arrays and classes, two of Java’s most widely used kinds of reference type. This chapter starts by discussing another very important kind of reference type—interfaces. We then move on to discuss Java’s generics, which have a major role to play in Java’s type system. With these topics under our belts, we can discuss the differences between compile-time and runtime types in Java.

To complete the full picture of Java’s reference types, we look at specialized kinds of classes and interfaces—known as enums and annotations. We conclude the chapter by looking at lambda expressions and nested types, and then reviewing how enhanced type inference has allowed Java’s non-denotable types to become usable by programmers.

Let’s get started by taking a look at interfaces—probably the most important of Java’s reference types after classes, and a key building block for the rest of Java’s type system.

Interfaces

In Chapter 3, we met the idea of inheritance. We also saw that a Java class can only inherit from a single class. This is quite a big restriction on the kinds of object-oriented programs that we want to build. The designers of Java knew this, but they also wanted to ensure that Java’s approach to object-oriented programming was less complex and error-prone than, for example, that of C++.

The solution that they chose was to introduce the concept of an interface to Java. Like a class, an interface defines a new reference type. As its name implies, an interface is intended to represent only an API—so it provides a description of a type, and the methods (and signatures) that classes that implement that API must provide.

In general, a Java interface does not provide any implementation code for the methods that it describes. These methods are considered mandatory—any class that wishes to implement the interface must provide an implementation of these methods.

However, an interface may wish to mark that some API methods are optional, and that implementing classes do not need to implement them if they choose not to. This is done with the default keyword—and the interface must provide an implementation of these optional methods, which will be used by any implementating class that elects not to implement them.

It is not possible to directly instantiate an interface and create a member of the interface type. Instead, a class must implement the interface to provide the necessary method bodies.

Any instances of the implementing class are compatible with both the type defined by the class and the type defined by the interface. This means that the instances may be substituted at any point in the code that requires an instance of either the class type or the interface type. This extends the Liskov principle as seen in “Reference Type Conversions”.

Another way of saying this is that two objects that do not share the same class or superclass may still both be compatible with the same interface type if both objects are instances of classes that implement the interface.

Defining an Interface

An interface definition is much like a class definition in which all the (nondefault) methods are abstract and the keyword class has been replaced with interface. For example, this code shows the definition of an interface named Centered (a Shape class, such as those defined in Chapter 3, might implement this interface if it wants to allow the coordinates of its center to be set and queried):

interface Centered {
  void setCenter(double x, double y);
  double getCenterX();
  double getCenterY();
}

A number of restrictions apply to the members of an interface:

Implementing an Interface

Just as a class uses extends to specify its superclass, it can use implements to name one or more interfaces it supports. The implements keyword can appear in a class declaration following the extends clause. It should be followed by a comma-separated list of interfaces that the class implements.

When a class declares an interface in its implements clause, it is saying that it provides an implementation (i.e., a body) for each mandatory method of that interface. If a class implements an interface but does not provide an implementation for every mandatory interface method, it inherits those unimplemented abstract methods from the interface and must itself be declared abstract. If a class implements more than one interface, it must implement every mandatory method of each interface it implements (or be declared abstract).

The following code shows how we can define a CenteredRectangle class that extends the Rectangle class from Chapter 3 and implements our Centered interface:

public class CenteredRectangle extends Rectangle implements Centered {
  // New instance fields
  private double cx, cy;

  // A constructor
  public CenteredRectangle(double cx, double cy, double w, double h) {
    super(w, h);
    this.cx = cx;
    this.cy = cy;
  }

  // We inherit all the methods of Rectangle but must
  // provide implementations of all the Centered methods.
  public void setCenter(double x, double y) { cx = x; cy = y; }
  public double getCenterX() { return cx; }
  public double getCenterY() { return cy; }
}

Suppose we implement CenteredCircle and CenteredSquare just as we have implemented this CenteredRectangle class. Each class extends Shape, so instances of the classes can be treated as instances of the Shape class, as we saw earlier. Because each class implements the Centered interface, instances can also be treated as instances of that type. The following code demonstrates how objects can be members of both a class type and an interface type:

Shape[] shapes = new Shape[3];      // Create an array to hold shapes

// Create some centered shapes, and store them in the Shape[]
// No cast necessary: these are all compatible assignments
shapes[0] = new CenteredCircle(1.0, 1.0, 1.0);
shapes[1] = new CenteredSquare(2.5, 2, 3);
shapes[2] = new CenteredRectangle(2.3, 4.5, 3, 4);

// Compute average area of the shapes and
// average distance from the origin
double totalArea = 0;
double totalDistance = 0;
for(int i = 0; i < shapes.length; i++) {
  totalArea += shapes[i].area();   // Compute the area of the shapes

  // Be careful, in general, the use of instanceof to determine the
  // runtime type of an object is quite often an indication of a
  // problem with the design
  if (shapes[i] instanceof Centered) { // The shape is a Centered shape
    // Note the required cast from Shape to Centered (no cast would
    // be required to go from CenteredSquare to Centered, however).
    Centered c = (Centered) shapes[i];

    double cx = c.getCenterX();    // Get coordinates of the center
    double cy = c.getCenterY();    // Compute distance from origin
    totalDistance += Math.sqrt(cx*cx + cy*cy);
  }
}
System.out.println("Average area: " + totalArea/shapes.length);
System.out.println("Average distance: " + totalDistance/shapes.length);
Note

Interfaces are data types in Java, just like classes. When a class implements an interface, instances of that class can be assigned to variables of the interface type.

Don’t interpret this example to imply that you must assign a CenteredRectangle object to a Centered variable before you can invoke the setCenter() method or to a Shape variable before invoking the area() method. Instead, because the CenteredRectangle class defines setCenter() and inherits area() from its Rectangle superclass, you can always invoke these methods.

As we could see by examining the bytecode (e.g., by using the javap tool we will meet in Chapter 13), the JVM calls the setCenter() method slightly differently depending on whether the local variable holding the shape is of the type CenteredRectangle or Centered, but this is not a distinction that matters most of the time when you’re writing Java code.

Default Methods

From Java 8 onward, it is possible to declare methods in interfaces that include an implementation. In this section, we’ll discuss these methods, which should be understood as optional methods in the API the interfaces represent—they’re usually called default methods. Let’s start by looking at the reasons why we need the default mechanism in the first place.

Backward compatibility

The Java platform has always been very concerned with backward compatibility. This means that code that was written (or even compiled) for an earlier version of the platform must continue to work with later releases of the platform. This principle allows development groups to have a high degree of confidence that an upgrade of their JDK or JRE will not break currently working applications.

Backward compatibility is a great strength of the Java platform, but in order to achieve it, some constraints are placed on the platform. One of them is that interfaces may not have new mandatory methods added to them in a new release of the interface.

For example, let’s suppose that we want to update the Positionable interface with the ability to add a bottom-left bounding point as well:

public interface Positionable extends Centered {
  void setUpperRightCorner(double x, double y);
  double getUpperRightX();
  double getUpperRightY();
  void setLowerLeftCorner(double x, double y);
  double getLowerLeftX();
  double getLowerLeftY();
}

With this new definition, if we try to use this new interface with code developed for the old, it just won’t work, as the existing code is missing the mandatory methods setLowerLeftCorner(), getLowerLeftX(), and getLowerLeftY().

Note

You can see this effect quite easily in your own code. Compile a class file that depends on an interface. Then add a new mandatory method to the interface, and try to run the program with the new version of the interface, together with your old class file. You should see the program crash with a NoClassDefError.

This limitation was a concern for the designers of Java 8—as one of their goals was to be able to upgrade the core Java Collections libraries, and introduce methods that made use of lambda expressions.

To solve this problem, a new mechanism was needed, essentially to allow interfaces to evolve by allowing new methods to be added without breaking backward compatibility.

Implementation of default methods

Adding new methods to an interface without breaking backward compatibility requires providing some implementation for the older implementations of the interface so that they can continue to work. This mechanism is a default method, and it was first added to the platform in JDK 8.

The basic behavior of default methods is:

  • An implementing class may (but is not required to) implement the default method.

  • If an implementing class implements the default method, then the implementation in the class is used.

  • If no other implementation can be found, then the default implementation is used.

An example default method is the sort() method. It’s been added to the interface java.util.List in JDK 8, and is defined as:

// The <E> syntax is Java's way of writing a generic type-see
// the next section for full details. If you aren't familiar with
// generics, just ignore that syntax for now.
interface List<E> {
  // Other members omitted

  public default void sort(Comparator<? super E> c) {
    Collections.<E>sort(this, c);
  }
}

Thus, from Java 8 upward, any object that implements List has an instance method sort() that can be used to sort the list using a suitable Comparator. As the return type is void, we might expect that this is an in-place sort, and this is indeed the case.

One consequence of default methods is that when implementing multiple interfaces, it’s possible that two or more interfaces may contain a default method with a completely identical name and signature.

For example:

interface Vocal {
  default void call() {
    System.out.println("Hello!");
  }
}

interface Caller {
  default void call() {
    Switchboard.placeCall(this);
  }
}

public class Person implements Vocal, Caller {
  // ... which default is used?
}

These two interfaces have very different default semantics for call() and could cause a potential implementation clash—a colliding default method. In versions of Java prior to 8, this could not occur, as the language only permitted single inheritance of implementation. The introduction of default methods means that Java now permits a limited form of multiple inheritance (but only of method implementations). Java still does not permit (and has no plans to add) multiple inheritance of object state.

Default methods have a simple set of rules to help resolve any potential ambiguities:

  • If a class implements multiple interfaces in such a way as to cause a potential clash of default method implementations, the implementing class must override the clashing method and provide a definition of what is to be done.

  • Syntax is provided to allow the implementing class to simply call one of the interface default methods if that is what is required:

public class Person implements Vocal, Caller {

    public void call() {
        // Can do our own thing
        // or delegate to either interface
        // e.g.,
        // Vocal.super.call();
        // or
        // Caller.super.call();
    }
}

As a side effect of the design of default methods, there is a slight, unavoidable usage issue that may arise in the case of evolving interfaces with colliding methods. Consider the case where a (version 7) class implements two interfaces A and B with versions a.0 and b.0, respectively. As defaults are not available in Java 7, this class will work correctly. However, if at a later time either or both interfaces adopt a default implementation of a colliding method, then compile time breakage can occur.

For example, if version a.1 introduces a default method in A, then the implementing class will pick up the implementation when run with the new version of the dependency. If version b.1 now introduces the same method, it causes a collision:

  • If B introduces the method as a mandatory (i.e., abstract) method, then the implementing class continues to work—both at compile time and at runtime.

  • If B introduces the method as a default method, then this is not safe and the implementing class will fail both at compile and at runtime.

This minor issue is very much a corner case and in practice is a very small price to pay in order to have usable default methods in the language.

When working with default methods, we should be aware that there is a slightly restricted set of operations we can perform from within a default method:

  • Call another method present in the interface’s public API (whether mandatory or optional); some implementation for such methods is guaranteed to be available.

  • Call a private method on the interface (Java 9 and up).

  • Call a static method, whether on the interface or defined elsewhere.

  • Use the this reference (e.g., as an argument to method calls).

The biggest takeaway from these restrictions is that even with default methods, Java interfaces still lack meaningful state; we cannot alter or store state within the interface.

Default methods have had a profound impact on the way that Java practitioners approach object-oriented programming. When combined with the rise of lambda expressions, they have upended many previous conventions of Java coding; we will discuss this in detail in the next chapter.

Marker Interfaces

Occasionally it is useful to define an interface that is entirely empty. A class can implement this interface simply by naming it in its implements clause without having to implement any methods. In this case, any instances of the class become valid instances of the interface as well and can be cast to the type. Java code can check whether an object is an instance of the interface using the instanceof operator, so this technique is a useful way to provide additional information about an object. It can be thought of as providing additional, auxiliary type information about a class.

Tip

Marker interfaces are much less widely used than they once were. Java’s annotations (which we shall meet presently) have largely replaced them due to their much greater flexibility at conveying extended type information.

The interface java.util.RandomAccess is an example of a marker interface: java.util.List implementations use this interface to advertise that they provide fast random access to the elements of the list. For example, ArrayList implements RandomAccess, while LinkedList does not. Algorithms that care about the performance of random-access operations can test for RandomAccess like this:

// Before sorting the elements of a long arbitrary list, we may want
// to make sure that the list allows fast random access.  If not,
// it may be quicker to make a random-access copy of the list before
// sorting it. Note that this is not necessary when using
// java.util.Collections.sort().
List l = ...;  // Some arbitrary list we're given
if (l.size() > 2 && !(l instanceof RandomAccess)) {
    l = new ArrayList(l);
}
sortListInPlace(l);

As we will see later, Java’s type system is very tightly coupled to the names that types have—an approach called nominal typing. A marker interface is a great example of this—it has nothing at all except a name.

Java Generics

One of the great strengths of the Java platform is the standard library that it ships. It provides a great deal of useful functionality—and in particular robust implementations of common data structures. These implementations are relatively simple to develop with and are well documented. The libraries are known as the Java Collections, and we will spend a big chunk of Chapter 8 discussing them. For a far more complete treatment, see the book Java Generics and Collections by Maurice Naftalin and Philip Wadler (O’Reilly).

Although they were still very useful, the earliest versions of the collections had a fairly major limitation, however. This limitation was that the data structure (sometimes called the container) essentially obscured the type of the data being stored in it.

Note

Data hiding and encapsulation is a great principle of object-oriented programming, but in this case, the opaque nature of the container caused a lot of problems for the developer.

Let’s kick off the section by demonstrating the problem, and showing how the introduction of generic types solved it and made life much easier for Java developers.

Introduction to Generics

If we want to build a collection of Shape instances, we can use a List to hold them, like this:

List shapes = new ArrayList();   // Create a List to hold shapes

// Create some centered shapes, and store them in the list
shapes.add(new CenteredCircle(1.0, 1.0, 1.0));
// This is legal Java-but is a very bad design choice
shapes.add(new CenteredSquare(2.5, 2, 3));

// List::get() returns Object, so to get back a
// CenteredCircle we must cast
CenteredCircle c = (CentredCircle)shapes.get(0);

// Next line causes a runtime failure
CenteredCircle c = (CentredCircle)shapes.get(1);

A problem with this code stems from the requirement to perform a cast to get the shape objects back out in a usable form—the List doesn’t know what type of objects it contains. Not only that, but it’s actually possible to put different types of objects into the same container—and everything will work fine until an illegal cast is used, and the program crashes.

What we really want is a form of List that understands what type it contains. Then, javac could detect when an illegal argument was passed to the methods of List and cause a compilation error, rather than deferring the issue to runtime.

Java provides a simple syntax to cater for homogeneous collections—to indicate that a type is a container that holds instances of another reference type, we enclose the payload type that the container holds within angle brackets:

// Create a List-of-CenteredCircle
List<CenteredCircle> shapes = new ArrayList<CenteredCircle>();

// Create some centered shapes, and store them in the list
shapes.add(new CenteredCircle(1.0, 1.0, 1.0));

// Next line will cause a compilation error
shapes.add(new CenteredSquare(2.5, 2, 3));

// List<CenteredCircle>::get() returns a CenteredCircle, no cast needed
CenteredCircle c = shapes.get(0);

This syntax ensures that a large class of unsafe code is caught by the compiler, before it gets anywhere near runtime. This is, of course, the whole point of static type systems—to use compile-time knowledge to help eliminate runtime problems wherever possible.

The resulting types, which combine an enclosing container type and a payload type, are usually called generic types—and they are declared like this:

interface Box<T> {
  void box(T t);
  T unbox();
}

This indicates that the Box interface is a general construct, which can hold any type of payload. It isn’t really a complete interface by itself—it’s more like a general description of a whole family of interfaces, one for each type that can be used in place of T.

Generic Types and Type Parameters

We’ve seen how to use a generic type, to provide enhanced program safety, by using compile-time knowledge to prevent simple type errors. In this section, let’s dig deeper into the properties of generic types.

The syntax <T> has a special name—it’s called a type parameter, and another name for a generic type is a parameterized type. This should convey the sense that the container type (e.g., List) is parameterized by another type (the payload type). When we write a type like Map<String, Integer>, we are assigning concrete values to the type parameters.

When we define a type that has parameters, we need to do so in a way that does not make assumptions about the type parameters. So the List type is declared in a generic way as List<E>, and the type parameter E is used all the way through to stand as a placeholder for the actual type that the programmer will use for the payload when she makes use of the List data structure.

Tip

Type parameters always stand in for reference types. It is not possible to use a primitive type as a value for a type parameter.

The type parameter can be used in the signatures and bodies of methods as though it is a real type, for example:

interface List<E> extends Collection<E> {
  boolean add(E e);
  E get(int index);
  // other methods omitted
}

Note how the type parameter E can be used as a parameter for both return types and method arguments. We don’t assume that the payload type has any specific properties, and only make the basic assumption of consistency—that the type we put in is the same type that we will later get back out.

This enhancement has effectively introduced a new kind of type to Java’s type system—by combining the container type with the value of the type parameter we are making new types.

Type Erasure

In “Default Methods”, we discussed the Java platform’s strong preference for backward compatibility. The addition of generics in Java 5 was another example of where backward compatibility was an issue for a new language feature.

The central question was how to make a type system that allowed older, nongeneric collection classes to be used alongside with newer, generic collections. The design decision was to achieve this by the use of casts:

List someThings = getSomeThings();
// Unsafe cast, but we know that the
// contents of someThings are really strings
List<String> myStrings = (List<String>)someThings;

This means that List and List<String> are compatible as types, at least at some level. Java achieves this compatibility by type erasure. This means that generic type parameters are only visible at compile time—they are stripped out by javac and are not reflected in the bytecode.1

The mechanism of type erasure gives rise to a difference in the type system seen by javac and that seen by the JVM—we will discuss this fully in “Generic Methods”.

Type erasure also prohibits some other definitions, which would otherwise seem legal. In this code, we want to count the orders as represented in two slightly different data structures:

// Won't compile
interface OrderCounter {
  // Name maps to list of order numbers
  int totalOrders(Map<String, List<String>> orders);

  // Name maps to total orders made so far
  int totalOrders(Map<String, Integer> orders);
}

This seems like perfectly legal Java code, but it will not compile. The issue is that although the two methods seem like normal overloads, after type erasure, the signature of both methods becomes:

  int totalOrders(Map);

All that is left after type erasure is the raw type of the container—in this case, Map. The runtime would be unable to distinguish between the methods by signature, and so the language specification makes this syntax illegal.

Bounded Type Parameters

Consider a simple generic box:

public class Box<T> {
    protected T value;

    public void box(T t) {
        value = t;
    }

    public T unbox() {
        T t = value;
        value = null;
        return t;
    }
}

This is a useful abstraction, but suppose we want to have a restricted form of box that only holds numbers. Java allows us to achieve this by using a bound on the type parameter. This is the ability to restrict the types that can be used as the value of a type parameter, for example:

public class NumberBox<T extends Number> extends Box<T> {
    public int intValue() {
        return value.intValue();
    }
}

The type bound T extends Number ensures that T can only be substituted with a type that is compatible with the type Number. As a result of this, the compiler knows that value will definitely have a method intValue() available on it.

Note

Notice that because the value field has protected access, it can be accessed directly in the subclass.

If we attempt to instantiate NumberBox with an invalid value for the type parameter, then the result will be a compilation error, as we can see:

NumberBox<Integer> ni = new NumberBox<>();
// Won't compile
NumberBox<Object> no = new NumberBox<>();

You must take care with raw types when working with type bounds, as the type bound can be evaded, but in doing so, the code is left vulnerable to a runtime exception:

// Compiles
NumberBox n = new NumberBox();
// This is very dangerous
n.box(new Object());
// Runtime error
System.out.println(n.intValue());

The call to intValue() fails with a java.lang.ClassCastException—as javac has inserted an unconditional cast of value to Number before calling the method.

In general, type bounds can be used to write better generic code and libraries. With practice, some fairly complex constructions can be built, for example:

public class ComparingBox<T extends Comparable<T>> extends Box<T>
                            implements Comparable<ComparingBox<T>> {
    @Override
    public int compareTo(ComparingBox<T> o) {
        if (value == null)
            return o.value == null ? 0 : -1;
        return value.compareTo(o.value);
    }
}

The definition might seem daunting, but the ComparingBox is really just a Box that contains a Comparable value. The type also extends the comparison operation to the ComparingBox type itself, by just comparing the contents of the two boxes.

Introducing Covariance

The design of Java’s generics contains the solution to an old problem. In the earliest versions of Java, before the collections libraries were even introduced, the language had been forced to confront a deep-seated type system design issue.

Put simply, the question is this:

Should an array of strings be compatible with a variable of type array-of-object?

In other words, should this code be legal?

String[] words = {"Hello World!"};
Object[] objects = words;

Without this, then even simple methods like Arrays::sort would have been very difficult to write in a useful way, as this would not work as expected:

Arrays.sort(Object[] a);

The method declaration would only work for the type Object[] and not for any other array type. As a result of these complications, the very first version of the Java Language Standard determined that:

If a value of type C can be assigned to a variable of type P then a value of type C[] can be assigned to a variable of type P[].

That is, arrays’ assignment syntax varies with the base type that they hold, or arrays are covariant.

This design decision is rather unfortunate, as it leads to immediate negative consequences:

String[] words = {"Hello", "World!"};
Object[] objects = words;

// Oh, dear, runtime error
objects[0] = new Integer(42);

The assignment to objects[0] attempts to store an Integer into a piece of storage that is expecting to hold a String. This obviously will not work, and will throw an ArrayStoreException.

Warning

The usefulness of covariant arrays led to them being seen as a necessary evil in the very early days of the platform, despite the hole in the static type system that the feature exposes.

However, more recent research on modern open source codebases indicates that array covariance is extremely rarely used and is a language misfeature.2 You should avoided it when writing new code.

When considering the behavior of generics in the Java platform, a very similar question can be asked: “Is List<String> a subtype of List<Object>?” That is, can we write this:

// Is this legal?
List<Object> objects = new ArrayList<String>();

At first glance, this seems entirely reasonable—String is a subclass of Object, so we know that any String element in our collection is also a valid Object.

However, consider the following code (which is just the array covariance code translated to use List):

// Is this legal?
List<Object> objects = new ArrayList<String>();

// What do we do about this?
objects.add(new Object());

As the type of objects was declared to be List<Object>, then it should be legal to add an Object instance to it. However, as the actual instance holds strings, then trying to add an Object would not be compatible, and so this would fail at runtime.

This would have changed nothing from the case of arrays, and so the resolution is to realize that although this is legal:

Object o = new String("X");

that does not mean that the corresponding statement for generic container types is also true, and as a result:

// Won't compile
List<Object> objects = new ArrayList<String>();

Another way of saying this is that List<String> is not a subtype of List<Object> or that generic types are invariant, not covariant. We will have more to say about this when we discuss bounded wildcards.

Wildcards

A parameterized type, such as ArrayList<T>, is not instantiable; we cannot create instances of them. This is because <T> is just a type parameter—merely a placeholder for a genuine type. It is only when we provide a concrete value for the type parameter (e.g., ArrayList<String>) that the type becomes fully formed and we can create objects of that type.

This poses a problem if the type that we want to work with is unknown at compile time. Fortunately, the Java type system is able to accommodate this concept. It does so by having an explicit concept of the unknown type—which is represented as <?>. This is the simplest example of Java’s wildcard types.

We can write expressions that involve the unknown type:

ArrayList<?> mysteryList = unknownList();
Object o = mysteryList.get(0);

This is perfectly valid Java—ArrayList<?> is a complete type that a variable can have, unlike ArrayList<T>. We don’t know anything about mysteryList’s payload type, but that may not be a problem for our code.

For example, when we get an item out of mysteryList, it has a completely unknown type. However, we can be sure that the object is assignable to Object—because all valid values of a generic type parameter are reference types and all reference values can be assigned to a variable of type Object.

On the other hand, when we’re working with the unknown type, there are some limitations on its use in user code. For example, this code will not compile:

// Won't compile
mysteryList.add(new Object());

The reason for this is simple—we don’t know what the payload type of mysteryList is! For example, if mysteryList was really a instance of ArrayList<String>, then we wouldn’t expect to be able to put an Object into it.

The only value that we know we can always insert into a container is null—as we know that null is a possible value for any reference type. This isn’t that useful, and for this reason, the Java language spec also rules out instantiating a container object with the unknown type as payload, for example:

// Won't compile
List<?> unknowns = new ArrayList<?>();

The unknown type may seem to be of limited utility, but one very important use for it is as a starting point for resolving the covariance question. We can use the unknown type if we want to have a subtyping relationship for containers, like this:

// Perfectly legal
List<?> objects = new ArrayList<String>();

This means that List<String> is a subtype of List<?>—although when we use an assignment like the preceding one, we have lost some type information. For example, the return type of get() is now effectively Object.

Note

List<?> is not a subtype of any List<T>, for any value of T.

The unknown type sometimes confuses developers—provoking questions like, “Why wouldn’t you just use Object instead of the unknown type?” However, as we’ve seen, the need to have subtyping relationships between generic types essentially requires us to have a notion of the unknown type.

Bounded wildcards

In fact, Java’s wildcard types extend beyond just the unknown type, with the concept of bounded wildcards.

They are used to describe the inheritance hierarchy of a mostly unknown type—effectively making statements like, for example, “I don’t know anything about this type, except that it must implement List.”

This would be written as ? extends List in the type parameter. This provides a useful lifeline to the programmer—instead of being restricted to the totally unknown type, she knows that at least the capabilities of the type bound are available.

Warning

The extends keyword is always used, regardless of whether the constraining type is a class or interface type.

This is an example of a concept called type variance, which is the general theory of how inheritance between container types relates to the inheritance of their payload types.

Type covariance

This means that the container types have the same relationship to each other as the payload types do. This is expressed using the extends keyword.

Type contravariance

This means that the container types have the inverse relationship to each other as the payload types. This is expressed using the super keyword.

These ideas tend to appear when discussing container types. For example, if Cat extends Pet, then List<Cat> is a subtype of List<? extends Pet>, and so:

List<Cat> cats = new ArrayList<Cat>();
List<? extends Pet> pets = cats;

However, this differs from the array case, because type safety is maintained in the following way:

pets.add(new Cat()); // won't compile
pets.add(new Pet()); // won't compile
cats.add(new Cat());

The compiler cannot prove that the storage pointed at by pets is capable of storing a Cat and so it rejects the call to add(). However, as cats definitely points at a list of Cat objects, then it must be acceptable to add a new one to the list.

As a result, it is very commonplace to see these types of generic constructions with types that act as producers or consumers of payload types.

For example, when the List is acting as a producer of Pet objects, then the appropriate keyword is extends.

Pet p = pets.get(0);

Note that for the producer case, the payload type appears as the return type of the producer method.

For a container type that is acting purely as a consumer of instances of a type, we would use the super keyword, and we would expect to see the payload type as the type of a method argument.

As discussed in Chapter 8, we see both covariance and contravariance throughout the Java Collections. They largely exist to ensure that the generics just “do the right thing” and behave in a manner that should not surprise the developer.

Generic Methods

A generic method is a method that is able to take instances of any reference type.

For example, this method emulates the behavior of the , (comma) operator from the C language, which is usually used to combine expressions with side effects together:

// Note that this class is not generic
public class Utils
  public static <T> T comma(T a, T b) {
    return a;
  }
}

Even though a type parameter is used in the definition of the method, the class it is defined in need not be generic—instead, the syntax is used to indicate that the method can be used freely, and that the return type is the same as the argument.

Let’s look at another example, from the Java Collections library. In the ArrayList class we can find a method to create a new array object from an arraylist instance:

@SuppressWarnings("unchecked")
public <T> T[] toArray(T[] a) {
    if (a.length < size)
        // Make a new array of a's runtime type, but my contents:
        return (T[]) Arrays.copyOf(elementData, size, a.getClass());
    System.arraycopy(elementData, 0, a, 0, size);
    if (a.length > size)
        a[size] = null;
    return a;
}

This method uses the low-level arraycopy() method to do the actual work.

Note

If we look at the class definition for ArrayList we can see that it is a generic class—but the type parameter is <E>, not <T>, and the type parameter <E> does not appear at all in the definition of toArray().

The toArray() method provides one half of a bridge API between the collections and Java’s original arrays. The other half of the API—moving from arrays to collections—involves a few additional subtleties, as we will discuss in Chapter 8.

Enums and Annotations

Java has specialized forms of classes and interfaces that are used to fulfill specific roles in the type system. They are known as enumerated types and annotation types, or normally just called enums and annotations.

Enums

Enums are a variation of classes that have limited functionality and that have only a small number of possible values that the type permits.

For example, suppose we want to define a type to represent the primary colors of red, green, and blue, and we want these to be the only possible values of the type. We can do this by making use of the enum keyword:

public enum PrimaryColor {
  // The ; is not required at the end of the list of instances
  RED, GREEN, BLUE
}

Instances of the type PrimaryColor can then be referenced as though they were static fields: PrimaryColor.RED, PrimaryColor.GREEN, and PrimaryColor.BLUE.

As enums are specialized classes, enums can have member fields and methods. If they do have a body (consisting of fields or methods), then the semicolon at the end of the list of instances is required, and the list of enum constants must precede the methods and fields.

For example, suppose that we want to have an enum that encompasses the first few regular polygons (shapes with all sides and all angles equal), and we want them to have some behavior (in the form of methods). We could achieve this by using an enum that takes a value as a parameter, like this:

public enum RegularPolygon {
  // The ; is mandatory for enums that have parameters
  TRIANGLE(3), SQUARE(4), PENTAGON(5), HEXAGON(6);

  private Shape shape;

  public Shape getShape() {
    return shape;
  }

  private RegularPolygon(int sides) {
    switch (sides) {
      case 3:
        // We assume that we have some general constructors
        // for shapes that take the side length and
        // angles in degrees as parameters
        shape = new Triangle(1,1,1,60,60,60);
        break;
      case 4:
        shape = new Rectangle(1,1);
        break;
      case 5:
        shape = new Pentagon(1,1,1,1,1,108,108,108,108,108);
        break;
      case 6:
        shape = new Hexagon(1,1,1,1,1,1,120,120,120,120,120,120);
        break;
    }
  }
}

These parameters (only one of them in this example) are passed to the constructor to create the individual enum instances. As the enum instances are created by the Java runtime, and can’t be instantiated from outside, the constructor is declared as private.

Enums have some special properties:

  • All (implicitly) extend java.lang.Enum

  • May not be generic

  • May implement interfaces

  • Cannot be extended

  • May only have abstract methods if all enum values provide an implementation body

  • May not be directly instantiated by new

Annotations

Annotations are a specialized kind of interface that, as the name suggests, annotate some part of a Java program.

For example, consider the @Override annotation. You may have seen it on some methods in some of the earlier examples, and may have asked the following question: what does it do?

The short, and perhaps surprising, answer is that it does nothing at all.

The less short (and flippant) answer is that, like all annotations, it has no direct effect, but instead acts as additional information about the method that it annotates; in this case, it denotes that a method overrides a superclass method.

This acts as a useful hint to compilers and integrated development environments (IDEs)—if a developer has misspelled the name of a method that she intended to be an override of a superclass method, then the presence of the @Override annotation on the misspelled method (which does not override anything) alerts the compiler to the fact that something is not right.

Annotations, as originally conceived, were not supposed to alter program semantics; instead, they were to provide optional metadata. In its strictest sense, this means that they should not affect program execution and instead should only provide information for compilers and other pre-execution phases.

In practice, modern Java applications make heavy use of annotations, and this now includes many use cases that essentially render the annotated classes useless without additional runtime support.

For example, classes bearing annotations such as @Inject, @Test, or @Autowired cannot realistically be used outside of a suitable container. As a result, it is difficult to argue that such annotations do not violate the “no semantic meaning” rule.

The platform defines a small number of basic annotations in java.lang. The original set were @Override, @Deprecated, and @SuppressWarnings—which were used to indicate that a method was overriden, deprecated, or that it generated some compiler warnings that should be suppressed.

These were augmented by @SafeVarargs in Java 7 (which provides extended warning suppression for varargs methods) and @FunctionalInterface in Java 8.

This last annotation indicates an interface can be used as a target for a lambda expression—it is a useful marker annotation although not mandatory, as we will see.

Annotations have some special properties, compared to regular interfaces:

  • All (implicitly) extend java.lang.annotation.Annotation

  • May not be generic

  • May not extend any other interface

  • May only define zero-arg methods

  • May not define methods that throw exceptions

  • Have restrictions on the return types of methods

  • Can have a default return value for methods

In practice, annotations do not typically have a great deal of functionality and instead are a fairly simple language concept.

Defining Custom Annotations

Defining custom annotation types for use in your own code is not that hard. The @interface keyword allows the developer to define a new annotation type, in much the same way that class or interface is used.

The meta-annotations are defined in java.lang.annotation and allow the developer to specify policy for where the new annotation type is to be used, and how it will be treated by the compiler and runtime.

There are two primary meta-annotations that are both essentially required when creating a new annotation type—@Target and @Retention. These both take values that are represented as enums.

The @Target meta-annotation indicates where the new custom annotation can be legally placed within Java source code. The enum ElementType has the possible values TYPE, FIELD, METHOD, PARAMETER, CONSTRUCTOR, LOCAL_VARIABLE, ANNOTATION_TYPE, PACKAGE, TYPE_PARAMETER, and TYPE_USE, and annotations can indicate that they intend to be used at one or more of these locations.

The other meta-annotation is @Retention, which indicates how javac and the Java runtime should process the custom annotation type. It can have one of three values, which are represented by the enum RetentionPolicy:

SOURCE

Annotations with this retention policy are discarded by javac during compilation.

CLASS

This means that the annotation will be present in the class file, but will not necessarily be accessible at runtime by the JVM. This is rarely used, but is sometimes seen in tools that do offline analysis of JVM bytecode.

RUNTIME

This indicates that the annotation will be available for user code to access at runtime (by using reflection).

Let’s take a look at an example, a simple annotation called @Nickname, which allows the developer to define a nickname for a method, which can then be used to find the method reflectively at runtime:

@Target(ElementType.METHOD)
@Retention(RetentionPolicy.RUNTIME)
public @interface Nickname {
    String[] value() default {};
}

This is all that’s required to define the annotation—a syntax element where the annotation can appear, a retention policy, and the name of the element. As we need to be able to state the nickname we’re assigning to the method, we also need to define a method on the annotation. Despite this, defining new custom annotations is a remarkably compact undertaking.

In addition to the two primary meta-annotations, there are also the @Inherited and @Documented meta-annotations. These are much less frequently encountered in practice, and details on them can be found in the platform documentation.

Lambda Expressions

One of the most eagerly anticated features of Java 8 was the introduction of lambda expressions (frequently referred to just as lambdas).

This was a major upgrade to the Java platform and was driven by five goals, in roughly descending order of priority:

  • More expressive programming

  • Better libraries

  • Concise code

  • Improved programming safety

  • Potentially increased data parallelism

Lambdas have three key aspects that help define the essential nature of the feature:

  • They allow small bits of code to be written inline as literals in a program.

  • They relax the strict naming rules of Java code by using type inference.

  • They are intended to facilitate a more functional style of programming Java.

As we saw in Chapter 2, the syntax for a lambda expression is to take a list of parameters (the types of which are typically inferred), and to attach that to a method body, like this:

(p, q) -> { /* method body */ }

This can provide a very compact way to represent what is effectively a single method. It is also a major departure from earlier versions of Java—until now, we have always had to have a class declaration and then a complete method declaration, all of which adds to the verboseness of the code.

In fact, before the arrival of lambdas, the only way to approximate this coding style was to use anonymous classes, which we will discuss later in this chapter. However, since Java 8, lambdas have proved to be very popular with Java programmers and now have mostly taken over the role of anonymous classes wherever they are able to do so.

Lambda expressions represent the creation of an object of a specific type. The type of the instance that is created is known as the target type of the lambda.

Only certain types are eligible to be the target of a lamba.

Target types are also called functional interfaces and they must:

  • Be interfaces

  • Have only one nondefault method (but may have other methods that are default)

Some developers also like to use the single abstract method (or SAM) type to refer to the interface type that the lambda is converted into. This draws attention to the fact that to be usable by the lambda expression mechanism, an interface must have only a single nondefault method.

Note

A lambda expression has almost all of the component parts of a method, with the obvious exception that a lambda doesn’t have a name. In fact, many developers like to think of lambdas as “anonymous methods.”

As a result, this means that the single line of code:

Runnable r  = () -> System.out.println("Hello");

actually represents the creation of an object, which is assigned to a variable r, of type Runnable.

Lambda Expression Conversion

When javac encounters a lambda expression, it interprets it as the body of a method with a specific signature—but which method?

To resolve this question, javac looks at the surrounding code. To be legal Java code, the lambda expression must satisfy the following properties:

  • The lambda must appear where an instance of an interface type is expected.

  • The expected interface type should have exactly one mandatory method.

  • The expected interface method should have a signature that exactly matches that of the lambda expression.

If this is the case, then an instance is created of a type that implements the expected interface, and uses the lambda body as the implementation for the mandatory method.

This slightly complex conversion approach comes from the desire to keep Java’s type system as purely nominative (based on names). The lambda expression is said to be converted to an instance of the correct interface type.

From this discussion, we can see that although Java 8 has added lambda expressions, they have been specifically designed to fit into Java’s existing type system—which has a very strong emphasis on nominal types (rather than the other possible sorts of types that exist in some other programming languages).

Let’s consider an example of lambda conversion—the list() method of the java.io.File class. This method lists the files in a directory. Before it returns the list, though, it passes the name of each file to a FilenameFilter object that the programmer must supply. This FilenameFilter object accepts or rejects each file, and is a SAM type defined in the java.io package:

@FunctionalInterface
public interface FilenameFilter {
    boolean accept(File dir, String name);
}

The type FilenameFilter carries the @FunctionalInterface to indicate that it is a suitable type to be used as the target type for a lambda. However, this annotation is not required and any type that meets the requirements (by being an interface and a SAM type) can be used as a target type.

This is because the JDK and the existing corpus of Java code already had a huge number of SAM types available before Java 8 was released. To require potential target types to carry the annotation would have prevented lambdas from being retrofitted to existing code for no real benefit.

Tip

In code that you write, you should always try to indicate when your types are usable as target types, which you can do by adding the @FunctionalInterface to them. This aids readability and can help some automated tools as well.

Here’s how we can define a FilenameFilter class to list only those files whose names end with .java, using a lambda:

File dir = new File("/src");      // The directory to list

String[] filelist = dir.list((d, fName) -> fName.endsWith(".java"));

For each file in the list, the block of code in the lambda expression is evaluated. If the method returns true (which happens if the filename ends in .java), then the file is included in the output—which ends up in the array filelist.

This pattern, where a block of code is used to test if an element of a container matches a condition, and to only return the elements that pass the condition, is called a filter idiom—and is one of the standard techniques of functional programming, which we will discuss in more depth presently.

Method References

Recall that we can think of lambda expressions as objects representing methods that don’t have names. Now, consider this lambda expression:

// In real code this would probably be
// shorter because of type inference
(MyObject myObj) -> myObj.toString()

This will be autoconverted to an implementation of a @FunctionalInterface type that has a single nondefault method that takes a single MyObject and returns a String—specifically, the string obtained by calling toString() on the instance of MyObject. However, this seems like excessive boilerplate, and so Java 8 provides a syntax for making this easier to read and write:

MyObject::toString

This is a shorthand, known as a method reference, that uses an existing method as a lambda expression. The method reference syntax is completely equivalent to the previous form expressed as a lambda. It can be thought of as using an existing method, but ignoring the name of the method, so it can be used as a lambda and then autoconverted in the usual way. Java defines four types of method reference, which are equivalent to four slightly different lambda expression forms (see Table 4-1).

Table 4-1. Method references
Name Method reference Equivalent lambda

Unbound

Trade::getPrice

trade -> trade.getPrice()

Bound

System.out::println

s -> System.out.println(s)

Static

System::getProperty

key -> System.getProperty(key)

Constructor

Trade::new

price -> new Trade(price)

The form we originally introduced can be seen to be an unbound method reference. When we use an unbound method reference, it is equivalent to a lambda that is expecting an instance of the type that contains the method reference—in Table 4-1 that is a Trade object.

It is called an unbound method reference because the receiver object needs to be supplied (as the first argument to the lambda) when the method reference is used. That is, we are going to call getPrice() on some Trade object, but the supplier of the method reference has not defined which one—that is left up to the user of the reference.

By contrast, a bound method reference always includes the receiver as part of the instantiation of the method reference. In Table 4-1, the receiver is System.out—so when the reference is used, the println() method will always be called on System.out, and all the parameters of the lambda will be used as method parameters to println().

We will discuss use cases for method references versus lambda expressions in more detail in the next chapter.

Functional Programming

Java is fundamentally an object-oriented lanaguage. However, with the arrival of lambda expressions, it becomes much easier to write code that is closer to the functional approach.

Note

There’s no single definition of exactly what constitutes a functional language—but there is at least a consensus that it should at minimum contain the ability to represent a function as a value that can be put into a variable.

Java has always (since version 1.1) been able to represent functions via inner classes, but the syntax was complex and lacking in clarity. Lambda expressions greatly simplify that syntax, and so it is only natural that more developers will be seeking to use aspects of functional programming in their Java code, now that it is considerably easier to do so.

The first taste of functional programming that Java developers are likely to encounter are three basic idioms that are remarkably useful:

map()

The map idiom is used with lists and list-like containers. The idea is that a function is passed in that is applied to each element in the collection, and a new collection is created—consisting of the results of applying the function to each element in turn. This means that a map idiom converts a collection of one type to a collection of potentially a different type.

filter()

We have already met an example of the filter idiom, when we discussed how to replace an anonymous implementation of FilenameFilter with a lambda. The filter idiom is used for producing a new subset of a collection, based on some selection criteria. Note that in functional programming, it is normal to produce a new collection, rather than modifying an existing one in-place.

reduce()

The reduce idiom has several different guises. It is an aggregation operation, which can be called fold, accumulate, or aggregate as well as reduce. The basic idea is to take an initial value, and an aggregation (or reduction) function, and apply the reduction function to each element in turn, building up a final result for the whole collection by making a series of intermediate results—similar to a “running total”—as the reduce operation traverses the collection.

Java has full support for these key functional idioms (and several others). The implementation is explained in some depth in Chapter 8, where we discuss Java’s data structures and collections, and in particular the stream abstraction, which makes all of this possible.

Let’s conclude this introduction with some words of caution. It’s worth noting that Java is best regarded as having support for “slightly functional programming.” It is not an especially functional language, nor does it try to be. Some particular aspects of Java that militate against any claims to being a functional language include the following:

  • Java has no structural types, which means no “true” function types. Every lambda is automatically converted to the appropriate nominal target type.

  • Type erasure causes problems for functional programming—type safety can be lost for higher-order functions.

  • Java is inherently mutable (as we’ll discuss in Chapter 6)—mutability is often regarded as highly undesirable for functional languages.

  • The Java collections are imperative, not functional. Collections must be converted to streams to use functional style.

Despite this, easy access to the basics of functional programing—and especially idioms such as map, filter, and reduce—is a huge step forward for the Java community. These idioms are so useful that a large majority of Java developers will never need or miss the more advanced capabilities provided by languages with a more thoroughbred functional pedigree.

In truth, many of these techniques were possible using nested types, via patterns like callbacks and handlers, but the syntax was always quite cumbersome, especially given that you had to explicitly define a completely new type even when you only needed to express a single line of code in the callback.

Lexical Scoping and Local Variables

A local variable is defined within a block of code that defines its scope, and outside of that scope, a local variable cannot be accessed and ceases to exist. Only code within the curly braces that define the boundaries of a block can use local variables defined in that block. This type of scoping is known as lexical scoping, and just defines a section of source code within which a variable can be used.

It is common for programmers to think of such a scope as temporal instead—that is, to think of a local variable as existing from the time the JVM begins executing the block until the time control exits the block. This is usually a reasonable way to think about local variables and their scope. However, lambda expressions (and anonymous and local classes, which we will meet later) have the ability to bend or break this intuition somewhat.

This can cause effects that some developers initially find surprising. This is because lambdas can use local variables, and so they can contain copies of values from lexical scopes that no longer exist. This can been seen in the following code:

public interface IntHolder {
    public int getValue();
}

public class Weird {
    public static void main(String[] args) {
        IntHolder[] holders = new IntHolder[10];
        for (int i = 0; i < 10; i++) {
            final int fi = i;

            holders[i] = () -> {
                return fi;
            };
        }
  // The lambda is now out of scope, but we have 10 valid instances
  // of the class the lambda has been converted to in our array.
  // The local variable fi is not in our scope here, but is still
  // in scope for the getValue() method of each of those 10 objects.
  // So call getValue() for each object and print it out.
  // This prints the digits 0 to 9.
        for (int i = 0; i < 10; i++) {
            System.out.println(holders[i].getValue());
        }
    }
}

Each instance of a lambda has an automatically created private copy of each of the final local variables it uses, so, in effect, it has its own private copy of the scope that existed when it was created. This is sometimes referred to as a captured variable.

Lambdas that capture variables like this are referred to as closures, and the variables are said to have been closed over.

Warning

Other programming languages may have a slightly different definition of a closure. In fact, some theorists would dispute that Java’s mechanism counts as a closure because, technically, it is the contents of the variable (a value) and not the variable itself that is captured.

In practice, the preceding closure example is more verbose than it needs to be in two separate ways:

  • The lambda has an explicit scope {} and return statement.

  • The variable fi is explicitly declared final.

The compiler javac helps with both of these.

Lambdas that only return the value of a single expression need not include a scope or return; instead, the body of the lambda is just the expression without the need for curly braces. In our example we have explicitly included the braces and return statement to spell out that the lambda is defining its own scope.

In early versions of Java there were two hard requirements when closing over a variable:

  • The captures must not be modified after they have been captured (e.g., after the lambda)

  • The captured variables must be declared final

However, in recent Java versions, javac can analyze the code and detect whether the programmer attempts to modify the captured variable after the scope of the lambda. If not, then the final qualifier on the captured variable can be omitted (such a variable is said to be effectively final). If the final qualifier is omitted, then it is a compile-time error to attempt to modify a captured variable after the lambda’s scope.

The reason for this is that Java implements closures by copying the bit pattern of the contents of the variable into the scope created by the closure. Further changes to the contents of the closed-over variable would not be reflected in the copy contained in closure scope, so the design decision was made to make such changes illegal, and a compile-time error.

These assists from javac mean that we can rewrite the inner loop of the preceding example to the very compact form:

for (int i = 0; i < 10; i++) {
    int fi = i;
    holders[i] = () -> fi;
}

Closures are very useful in some styles of programming, and different programming languages define and implement closures in different ways. Java implements closures as lambda expressions, but local classes and anonymous classes can also capture state—and in fact this is how Java implemented closures before lambdas were available.

Nested Types

The classes, interfaces, and enum types we have seen so far in this book have all been defined as top-level types. This means that they are direct members of packages, defined independently of other types. However, type definitions can also be nested within other type definitions. These nested types, commonly known as “inner classes,” are a powerful feature of the Java language.

In general, nested types are used for two separate purposes, both related to encapsulation. First, a type may be nested because it needs especially intimate access to the internals of another type. By being a nested type, it has access in the same way that member variables and methods do. This means that nested types have privileged access and can be thought of as “slightly bending the rules of encapsulation.”

Another way of thinking about this use case of nested types is that they are types that are somehow tied together with another type. This means that they don’t really have a completely independent existence as an entity, and only live in coexist.

Alternatively, a type may be only required for a very specific reason, and in a very small section of code. This means that it should be tightly localized, as it is really part of the implementation detail.

In older versions of Java, the only way to do this was with a nested type, such as an anonymous implementation of an interface. In practice, with the advent of Java 8, this use case has substantially been taken over by lambda expressions and the use of anonymous types as closely localized types has dramatically declined, although it still persists for some cases.

Types can be nested within another type in four different ways:

Static member types

A static member type is any type defined as a static member of another type. Nested interfaces, enums, and annotations are always static (even if you don’t use the keyword).

Nonstatic member classes

A “nonstatic member type” is simply a member type that is not declared static. Only classes can be nonstatic member types.

Local classes

A local class is a class that is defined and only visible within a block of Java code. Interfaces, enums, and annotations may not be defined locally.

Anonymous classes

An anonymous class is a kind of local class that has no meaningful name in the Java language. Interfaces, enums, and annotations cannot be defined anonymously.

The term “nested types,” while a correct and precise usage, is not widely used by developers. Instead, most Java programmers use the much vaguer term “inner class.” Depending on the situation, this can refer to a nonstatic member class, local class, or anonymous class, but not a static member type, with no real way to distinguish between them.

Fortunately, although the terminology for describing nested types is not always clear, the syntax for working with them is, and it is usually apparent from context which kind of nested type is being discussed.

Let’s move on to describe each of the four kinds of nested types in greater detail. Each section describes the features of the nested type, the restrictions on its use, and any special Java syntax used with the type.

Static Member Types

A static member type is much like a regular top-level type. For convenience, however, it is nested within the namespace of another type. Static member types have the following basic properties:

  • A static member type is like the other static members of a class: static fields and static methods.

  • A static member type is not associated with any instance of the containing class (i.e., there is no this object).

  • A static member type can access (only) the static members of the class that contains it.

  • A static member type has access to all the static members (including any other static member types) of its containing type.

  • Nested interfaces, enums, and annotations are implicitly static, whether or not the static keyword appears.

  • Any type nested within an interface or annotation is also implicitly static.

  • Static member types may be defined within top-level types or nested to any depth within other static member types.

  • A static member type may not be defined within any other kind of nested type.

Let’s look at a quick example of the syntax for static member types. Example 4-1 shows a helper interface defined as a static member of a containing class.

Example 4-1. Defining and using a static member interface
// A class that implements a stack as a linked list
public class LinkedStack {

    // This static member interface defines how objects are linked
    // The static keyword is optional: all nested interfaces are static
    static interface Linkable {
        public Linkable getNext();
        public void setNext(Linkable node);
    }

    // The head of the list is a Linkable object
    Linkable head;

    // Method bodies omitted
    public void push(Linkable node) { ... }

    public Object pop() { ... }
}

// This class implements the static member interface
class LinkableInteger implements LinkedStack.Linkable {
    // Here's the node's data and constructor
    int i;
    public LinkableInteger(int i) { this.i = i; }

    // Here are the data and methods required to implement the interface
    LinkedStack.Linkable next;

    public LinkedStack.Linkable getNext() { return next; }

    public void setNext(LinkedStack.Linkable node) { next = node; }
}

The example also shows how this interface is used both within the class that contains it and by external classes. Note the use of its hierarchical name in the external class.

Features of static member types

A static member type has access to all static members of its containing type, including private members. The reverse is true as well: the methods of the containing type have access to all members of a static member type, including the private members. A static member type even has access to all the members of any other static member types, including the private members of those types. A static member type can use any other static member without qualifying its name with the name of the containing type.

Top-level types can be declared as either public or package-private (if they’re declared without the public keyword). But declaring top-level types as private and protected wouldn’t make a great deal of sense—protected would just mean the same as package-private and a private top-level class would be unable to be accessed by any other type.

Static member types, on the other hand, are members and so can use any access control modifiers that other members of the containing type can. These modifiers have the same meanings for static member types as they do for other members of a type.

For example, in Example 4-1, the Linkable interface is declared public, so it can be implemented by any class that is interested in being stored on a LinkedStack.

In code outside the containing class, a static member type is named by combining the name of the outer type with that of the inner (e.g., LinkedStack.Linkable).

Under most circumstances, this syntax provides a helpful reminder that the inner class is interconnected with its containing type. However, the Java language does permit you to use the import directive to directly import a static member type:

import pkg.LinkedStack.Linkable;  // Import a specific nested type
// Import all nested types of LinkedStack
import pkg.LinkedStack.*;

You can then reference the nested type without including the name of its enclosing type (e.g., just as Linkable).

Note

You can also use the import static directive to import a static member type. See “Packages and the Java Namespace” in Chapter 2 for details on import and import static.

However, importing a nested type obscures the fact that that type is closely associated with its containing type—which is usually important information—and as a result it is not commonly done.

Nonstatic Member Classes

A nonstatic member class is a class that is declared as a member of a containing class or enumerated type without the static keyword:

  • If a static member type is analogous to a class field or class method, a nonstatic member class is analogous to an instance field or instance method.

  • Only classes can be nonstatic member types.

  • An instance of a nonstatic member class is always associated with an instance of the enclosing type.

  • The code of a nonstatic member class has access to all the fields and methods (both static and non-static) of its enclosing type.

  • Several features of Java syntax exist specifically to work with the enclosing instance of a nonstatic member class.

Example 4-2 shows how a member class can be defined and used. This example extends the previous LinkedStack example to allow enumeration of the elements on the stack by defining an iterator() method that returns an implementation of the java.util.Iterator interface. The implementation of this interface is defined as a member class.

Example 4-2. An iterator implemented as a member class
import java.util.Iterator;

public class LinkedStack {

    // Our static member interface
    public interface Linkable {
        public Linkable getNext();
        public void setNext(Linkable node);
    }

    // The head of the list
    private Linkable head;

    // Method bodies omitted here
    public void push(Linkable node) { ... }
    public Linkable pop() { ... }

    // This method returns an Iterator object for this LinkedStack
    public Iterator<Linkable> iterator() { return new LinkedIterator(); }

    // Here is the implementation of the Iterator interface,
    // defined as a nonstatic member class.
    protected class LinkedIterator implements Iterator<Linkable> {
        Linkable current;

        // The constructor uses a private field of the containing class
        public LinkedIterator() { current = head; }

        // The following three methods are defined
        // by the Iterator interface
        public boolean hasNext() {  return current != null; }

        public Linkable next() {
            if (current == null)
              throw new java.util.NoSuchElementException();
            Linkable value = current;
            current = current.getNext();
            return value;
        }

        public void remove() { throw new UnsupportedOperationException(); }
    }
}

Notice how the LinkedIterator class is nested within the LinkedStack class. Because LinkedIterator is a helper class used only within LinkedStack, having it defined so close to where it is used by the containing class makes for a clean design, just as we discussed when we introduced nested types.

Features of member classes

Like instance fields and instance methods, every instance of a nonstatic member class is associated with an instance of the class in which it is defined. This means that the code of a member class has access to all the instance fields and instance methods (as well as the static members) of the containing instance, including any that are declared private.

This crucial feature was already illustrated in Example 4-2. Here is the LinkedStack.LinkedIterator() constructor again:

public LinkedIterator() { current = head; }

This single line of code sets the current field of the inner class to the value of the head field of the containing class. The code works as shown, even though head is declared as a private field in the containing class.

A nonstatic member class, like any member of a class, can be assigned one of the standard access control modifiers. In Example 4-2, the LinkedIterator class is declared protected, so it is inaccessible to code (in a different package) that uses the LinkedStack class but is accessible to any class that subclasses LinkedStack.

Member classes have two important restrictions:

  • A nonstatic member class cannot have the same name as any containing class or package. This is an important rule, one that is not shared by fields and methods.

  • Nonstatic member classes cannot contain any static fields, methods, or types, except for constant fields declared both static and final.

Local Classes

A local class is declared locally within a block of Java code rather than as a member of a class. Only classes may be defined locally: interfaces, enumerated types, and annotation types must be top-level or static member types. Typically, a local class is defined within a method, but it can also be defined within a static initializer or instance initializer of a class.

Just as all blocks of Java code appear within class definitions, all local classes are nested within containing blocks. For this reason, local classes share many of the features of member classes. It is usually more appropriate to think of them as an entirely separate kind of nested type.

Note

See Chapter 5 for details as to when it’s appropriate to choose a local class versus a lambda expression.

The defining characteristic of a local class is that it is local to a block of code. Like a local variable, a local class is valid only within the scope defined by its enclosing block. Example 4-3 illustrates how we can modify the iterator() method of the LinkedStack class so it defines LinkedIterator as a local class instead of a member class.

By doing this, we move the definition of the class even closer to where it is used and hopefully improve the clarity of the code even further. For brevity, Example 4-3 shows only the iterator() method, not the entire LinkedStack class that contains it.

Example 4-3. Defining and using a local class
// This method returns an Iterator object for this LinkedStack
public Iterator<Linkable> iterator() {
    // Here's the definition of LinkedIterator as a local class
    class LinkedIterator implements Iterator<Linkable> {
        Linkable current;

        // The constructor uses a private field of the containing class
        public LinkedIterator() { current = head; }

        // The following three methods are defined
        // by the Iterator interface
        public boolean hasNext() {  return current != null; }

        public Linkable next() {
            if (current == null)
              throw new java.util.NoSuchElementException();
            Linkable value = current;
            current = current.getNext();
            return value;
        }

        public void remove() { throw new UnsupportedOperationException(); }
    }

    // Create and return an instance of the class we just defined
    return new LinkedIterator();
}

Scope of a local class

In discussing nonstatic member classes, we saw that a member class can access any members inherited from superclasses and any members defined by its containing classes.

The same is true for local classes, but local classes can also behave like lambdas and access effectively final local variables and parameters. Example 4-4 illustrates the different kinds of fields and variables that may be accessible to a local class (or a lambda, for that matter):

Example 4-4. Fields and variables available to a local class
class A { protected char a = 'a'; }
class B { protected char b = 'b'; }

public class C extends A {
  private char c = 'c';         // Private fields visible to local class
  public static char d = 'd';
  public void createLocalObject(final char e)
  {
    final char f = 'f';
    int i = 0;                  // i not final; not usable by local class
    class Local extends B
    {
      char g = 'g';
      public void printVars()
      {
        // All of these fields and variables are accessible to this class
        System.out.println(g);  // (this.g) g is a field of this class
        System.out.println(f);  // f is a final local variable
        System.out.println(e);  // e is a final local parameter
        System.out.println(d);  // (C.this.d) d field of containing class
        System.out.println(c);  // (C.this.c) c field of containing class
        System.out.println(b);  // b is inherited by this class
        System.out.println(a);  // a is inherited by the containing class
      }
    }
    Local l = new Local();      // Create an instance of the local class
    l.printVars();              // and call its printVars() method.
  }
}

Local classes have quite a complex scoping structure, therefore. To see why, notice that instances of a local class can have a lifetime that extends past the time that the JVM exits the block where the local class is defined.

Note

In other words, if you create an instance of a local class, that instance does not automatically go away when the JVM finishes executing the block that defines the class. So, even though the definition of the class was local, instances of that class can escape out of the place they were defined.

Local classe, therefore, behave like lambdas in many regards, although the use case of local classes is more general than that of lambdas. However, in practice the extra generality is rarely required, and lambdas are preferred wherever possible.

Anonymous Classes

An anonymous class is a local class without a name. It is defined and instantiated in a single expression using the new operator. While a local class definition is a statement in a block of Java code, an anonymous class definition is an expression, which means that it can be included as part of a larger expression, such as a method call.

Consider Example 4-5, which shows the LinkedIterator class implemented as an anonymous class within the iterator() method of the LinkedStack class. Compare it with Example 4-4, which shows the same class implemented as a local class.

Example 4-5. An enumeration implemented with an anonymous class
public Iterator<Linkable> iterator() {
    // The anonymous class is defined as part of the return statement
    return new Iterator<Linkable>() {
        Linkable current;
        // Replace constructor with an instance initializer
        { current = head; }

        // The following three methods are defined
        // by the Iterator interface
        public boolean hasNext() {  return current != null; }
        public Linkable next() {
            if (current == null)
              throw new java.util.NoSuchElementException();
            Linkable value = current;
            current = current.getNext();
            return value;
        }
        public void remove() { throw new UnsupportedOperationException(); }
    };  // Note the required semicolon. It terminates the return statement
}

As you can see, the syntax for defining an anonymous class and creating an instance of that class uses the new keyword, followed by the name of a type and a class body definition in curly braces. If the name following the new keyword is the name of a class, the anonymous class is a subclass of the named class. If the name following new specifies an interface, as in the two previous examples, the anonymous class implements that interface and extends Object.

Note

The syntax for anonymous classes does not include any way to specify an extends clause, an implements clause, or a name for the class.

Because an anonymous class has no name, it is not possible to define a constructor for it within the class body. This is one of the basic restrictions on anonymous classes. Any arguments you specify between the parentheses following the superclass name in an anonymous class definition are implicitly passed to the superclass constructor. Anonymous classes are commonly used to subclass simple classes that do not take any constructor arguments, so the parentheses in the anonymous class definition syntax are often empty.

Because an anonymous class is just a type of local class, anonymous classes and local classes share the same restrictions. An anonymous class cannot define any static fields, methods, or classes, except for static final constants. Interfaces, enumerated types, and annotation types cannot be defined anonymously. Also, like local classes, anonymous classes cannot be public, private, protected, or static.

The syntax for defining an anonymous class combines definition with instantiation, similar to a lambda expression. Using an anonymous class instead of a local class is not appropriate if you need to create more than a single instance of the class each time the containing block is executed.

Because an anonymous class has no name, it is not possible to define a constructor for an anonymous class. If your class requires a constructor, you must use a local class instead.

Non-Denotable Types and var

One of the only new language features to arrive in Java 10 is Local Variable Type Inference, otherwise known as var. This is an enhancement to Java’s type inference capabilities that may prove to be more significant than it first appears. In the simplest case, it allows code such as:

var ls = new ArrayList<String>();

which moves the inference from the type of values to the type of variables.

The implementation in Java 10 achieves this by making var a reserved type name rather than a keyword. This means that code can still use var as a variable, method, or package name without being affected by the new syntax. However, code that has previously used var as the name of a type will have to be recompiled.

This simple case is designed to reduce verbosity and to make programmers coming to Java from other languages (especially Scala, .NET, and JavaScript) feel more comfortable. However, it does carry the risk that overuse will potentially obscure the intent of the code being written, so it should be used sparingly.

As well as the simple cases, var actually permits programming constructs that were not possible before. To see the differences, let’s consider that javac has always permitted a very limited form of type inference:

public class Test {
    public static void main(String[] args) {
        (new Object() {
            public void bar() {
                System.out.println("bar!");
            }
        }).bar();
    }
}

The code will compile and run, printing out bar!. This slightly counterintuitive result occurs because javac preserves enough type information about the anonymous class (i.e., that it has a bar() method) for just long enough that the compiler can conclude that the call to bar() is valid.

In fact, this edge case has been known in the Java community since at least 2009, long before the arrival of even Java 7.

The problem with this form of type inference is that it has no real practical applications—the type of “Object-with-a-bar-method” exists within the compiler, but the type is impossible to express as the type of a variable—it is not a denotable type. This means that before Java 10 the existence of this type is restricted to a single expression and cannot be used in a larger scope.

With the arrival of Java 10, however, the type of variables does not always need to be made explicit. Instead, we can use var to allow us to preserve the static type information by avoiding denoting the type.

This means we can now modify our example and write:

var o = new Object() {
    public void bar() {
        System.out.println("bar!");
    }
};

o.bar();

This has allowed us to preserve the true type of o beyond a single expression. The type of o cannot be denoted, and so it cannot appear as the type of either a method parameter or return type. This means that the type is still limited to only a single method, but it is still useful to express some constructions that would be awkward or impossible otherwise.

This use of var as a “magic type” allows the programmer to preserve type information for each distinct usage of var, in a way that is somewhat reminiscent of bounded wildcards from Java’s generics.

More advanced usages of var with non-denotable types are possible. While the feature is not able to satisfy every criticism of Java’s type system, it does represent a definite (if cautious) step forward.

Summary

By examining Java’s type system, we have been able to build up a clear picture of the worldview that the Java platform has about data types. Java’s type system can be characterized as:

Static

All Java variables have types that are known at compile time.

Nominal

The name of a Java type is of paramount importance. Java does not permit structural types and has only limited support for non-denotable types.

Object/imperative

Java code is object-oriented, and all code must live inside methods, which must live inside classes. However, Java’s primitive types prevent full adoption of the “everything is an object” worldview.

Slightly functional

Java provides support for some of the more common functional idioms, but more as a convenience to programmers than anything else.

Type-inferred

Java is optimized for readability (even by novice progammers) and prefers to be explicit, but uses type inference to reduce boilerplate where it does not impact the legibility of the code.

Strongly backward compatible

Java is primarily a business-focused language, and backward compatibility and protection of existing codebases is a very high priority.

Type erased

Java permits parameterized types, but this information is not available at runtime.

Java’s type system has evolved (albeit slowly and cautiously) over the years—and is now on a par with the type systems of other mainstream programming languages. Lambda expressions, along with default methods, represent the greatest transformation since the advent of Java 5, and the introduction of generics, annotations, and related innovations.

Default methods represent a major shift in Java’s approach to object-oriented programming—perhaps the biggest since the language’s inception. From Java 8 onward, interfaces can contain implementation code. This fundamentally changes Java’s nature—previously a single-inherited language, Java is now multiply inherited (but only for behavior—there is still no multiple inheritance of state).

Despite all of these innovations, Java’s type system is not (and is not intended to be) equipped with the power of the type systems of languages such as Scala or Haskell. Instead, Java’s type system is strongly biased in favor of simplicity, readability, and a simple learning curve for newcomers.

Java has also benefited enormously from the approaches to types developed in other languages over the last 10 years. Scala’s example of a statically typed language that nevertheless achieves much of the feel of a dynamically typed language by the use of type inference has been a good source of ideas for features to add to Java, even though the languages have quite different design philosophies.

One remaining question is whether the modest support for functional idioms that lambda expressions provide in Java is sufficient for the majority of Java programmers.

Note

The long-term direction of Java’s type system is being explored in research projects such as Valhalla, where concepts such as data classes, pattern matching, and sealed classes are being explored.

It remains to be seen whether the majority of ordinary Java programmers require the added power—and attendant complexity—that comes from an advanced (and much less nominal) type system such as Scala’s, or whether the “slightly functional programming” introduced in Java 8 (e.g., map, filter, reduce, and their peers) will suffice for most developers’ needs.

1 Some small traces of generics remain, which can be seen at runtime via reflection.

2 Raoul-Gabriel Urma and Janina Voigt, “Using the OpenJDK to Investigate Covariance in Java,” Java Magazine (May/June 2012): 44–47.