How do inner classes affect the organization of the Java Virtual Machine?
There are no changes to the class file format as processed by the Java Virtual Machine, or to the standard class libraries. The new features are implemented by the compiler. The organization of the resulting bytecodes is specified with enough precision that all 1.1-conforming compilers will produce binary compatible class files.
A single file of Java source code can compile to many class files. Although this is not a new phenomenon, the power of the inner class notation means that the programmer can end up creating a larger number of class files with relatively less code. In addition, adapter classes tend to be very simple, with few methods. This means that a Java program which uses many inner classes will compile to many small class files. Packaging technologies for such classes process them reasonably efficiently. For example, the class file for the example class FixedStack.Enumeration occupies about three quarters of a kilobyte, of which about 40% is directly required to implement its code. This ratio is likely to improve over time as file formats are tuned. The memory usage patterns in the virtual machine are comparable.
Class name transformations
Names of nested classes are transformed as necessary by the compiler to avoid conflicts with identical names in other scopes. Names are encoded to the virtual machine by taking their source form, qualified with dots, and changing each dot `.' after a class name into a dollar sign `$'. (Mechanical translators are allowed to use dollar signs in Java.)
When a class name is private or local to a block, it is globally inaccessible. A compiler may opt to code such an inaccessible name by using an accessible enclosing class name as a prefix, followed by a `$' separator and a locally unique decimal number. Anonymous classes must be encoded this way.
So, an inner class pkg.Foo.Bar gets a run-time name of pkg.Foo$Bar, or perhaps something like pkg.Foo$23, if Bar is a private member or local class. Implementations must conform to the format of names, even globally inaccessible ones, so that debuggers and similar tools can recognize them.
Any class file which defines or uses a transformed name also contains an attribute (as supported by the 1.0 file format) recording the transformation. These attributes are ignorable by the virtual machine and by 1.0 compilers. The format of this attribute is described in the section on binary compatibility.
Names of generated variables and methods
As we have seen previously, if an inner class uses a variable from an enclosing scope, the name expression will be transformed, into a reference either to a field of an enclosing instance, or to a field of the current instance which provides the value of a final local variable. A reference to an enclosing instance, in turn, is transformed into a reference to a field in a more accessible current instance. These techniques require that the compiler synthesize hidden fields in inner classes.
There is one more category of compiler-generated members. A private member m of a class C may be used by another class D, if one class encloses the other, or if they are enclosed by a common class. Since the virtual machine does not know about this sort of grouping, the compiler creates a local protocol of access methods in C to allow D to read, write, or call the member m. These methods have names of the form access$0, access$1, etc. They are never public. Access methods are unique in that they may be added to enclosing classes, not just inner classes.
All generated variables and methods are declared in a class file attribute, so that the 1.1 compilers can prevent programs from referring to them directly.
If an inner class C requires access to a private member m of an enclosing class T, the inserted access method for m opens up T to illegal access by any class K in the same package. There at present are no known security problems with such access methods, since it is difficult to misuse a method with package scope. The compiler can be instructed to emit warnings when it creates access methods, to monitor the creation of possible loopholes.
If a class N is a protected member of another class C, then N's class file defines it as a public class. A class file attribute correctly records the protection mode bits. This attribute is ignored by the current virtual machine, which therefore will allow access to N by any class, and not just to subclasses of C. The compiler, of course, will correctly diagnose such errors, because it looks at the attribute. This is not a security hole, since malicious users can easily create subclasses of C and so gain access to N, protected or not.
Likewise, if a class is a private member of another class, its class file defines it as having package scope, and an attribute declares the true access protection, so that 1.1 compilers can prevent inadvertant access, even within the package.