Getting Started with Javassist

Shigeru Chiba

Next page


1. Reading and writing bytecode

Javassist is a class library for dealing with Java bytecode. Java bytecode is stored in a binary file called a class file. Each class file contains one Java class or interface.

The class Javassist.CtClass is an abstract representation of a class file. A CtClass (compile-time class) object is a handle for dealing with a class file. The following program is a very simple example:

This program first obtains a ClassPool object, which controls bytecode modification with Javassist. The ClassPool object is a container of CtClass object representing a class file. It reads a class file on demand for constructing a CtClass object and records the constructed object for responding later accesses. To modify the definition of a class, the users must first obtain from a ClassPool object a reference to a CtClass object representing that class. get() in ClassPool is used for this purpose. In the case of the program shown above, the CtClass object representing a class test.Rectangle is obtained from the ClassPool object and it is assigned to a variable cc. The ClassPool object returned by getDefault() searches the default system search path.

From the implementation viewpoint, ClassPool is a hash table of CtClass objects, which uses the class names as keys. get() in ClassPool searches this hash table to find a CtClass object associated with the specified key. If such a CtClass object is not found, get() reads a class file to construct a new CtClass object, which is recorded in the hash table and then returned as the resulting value of get().

The CtClass object obtained from a ClassPool object can be modified (details of how to modify a CtClass will be presented later). In the example above, it is modified so that the superclass of test.Rectangle is changed into a class test.Point. This change is reflected on the original class file when writeFile() in CtClass() is finally called.

writeFile() translates the CtClass object into a class file and writes it on a local disk. Javassist also provides a method for directly obtaining the modified bytecode. To obtain the bytecode, call toBytecode():

You can directly load the CtClass as well:

toClass() requests the context class loader for the current thread to load the class file represented by the CtClass. It returns a java.lang.Class object representing the loaded class. For more details, please see this section below.

Defining a new class

To define a new class from scratch, makeClass() must be called on a ClassPool.

This program defines a class Point including no members. Member methods of Point can be created with factory methods declared in CtNewMethod and appended to Point with addMethod() in CtClass.

makeClass() cannot create a new interface; makeInterface() in ClassPool can do. Member methods in an interface can be created with abstractMethod() in CtNewMethod. Note that an interface method is an abstract method.

Frozen classes

If a CtClass object is converted into a class file by writeFile(), toClass(), or toBytecode(), Javassist freezes that CtClass object. Further modifications of that CtClass object are not permitted. This is for warning the developers when they attempt to modify a class file that has been already loaded since the JVM does not allow reloading a class.

A frozen CtClass can be defrost so that modifications of the class definition will be permitted. For example,

After defrost() is called, the CtClass object can be modified again.

If ClassPool.doPruning is set to true, then Javassist prunes the data structure contained in a CtClass object when Javassist freezes that object. To reduce memory consumption, pruning discards unnecessary attributes (attribute_info structures) in that object. For example, Code_attribute structures (method bodies) are discarded. Thus, after a CtClass object is pruned, the bytecode of a method is not accessible except method names, signatures, and annotations. The pruned CtClass object cannot be defrost again. The default value of ClassPool.doPruning is false.

To disallow pruning a particular CtClass, stopPruning() must be called on that object in advance:

The CtClass object cc is not pruned. Thus it can be defrost after writeFile() is called.

Class search path

The default ClassPool returned by a static method ClassPool.getDefault() searches the same path that the underlying JVM (Java virtual machine) has. If a program is running on a web application server such as JBoss and Tomcat, the ClassPool object may not be able to find user classes since such a web application server uses multiple class loaders as well as the system class loader. In that case, an additional class path must be registered to the ClassPool. Suppose that pool refers to a ClassPool object:

This statement registers the class path that was used for loading the class of the object that this refers to. You can use any Class object as an argument instead of this.getClass(). The class path used for loading the class represented by that Class object is registered.

You can register a directory name as the class search path. For example, the following code adds a directory /usr/local/javalib to the search path:

The search path that the users can add is not only a directory but also a URL:

This program adds "http://www.javassist.org:80/java/" to the class search path. This URL is used only for searching classes belonging to a package org.javassist. For example, to load a class org.javassist.test.Main, its class file will be obtained from:

Furthermore, you can directly give a byte array to a ClassPool object and construct a CtClass object from that array. To do this, use ByteArrayClassPath. For example,

The obtained CtClass object represents a class defined by the class file specified by b. The ClassPool reads a class file from the given ByteArrayClassPath if get() is called and the class name given to get() is equal to one specified by name.

If you do not know the fully-qualified name of the class, then you can use makeClass() in ClassPool:

makeClass() returns the CtClass object constructed from the given input stream. You can use makeClass() for eagerly feeding class files to the ClassPool object. This might improve performance if the search path includes a large jar file. Since a ClassPool object reads a class file on demand, it might repeatedly search the whole jar file for every class file. makeClass() can be used for optimizing this search. The CtClass constructed by makeClass() is kept in the ClassPool object and the class file is never read again.

The users can extend the class search path. They can define a new class implementing ClassPath interface and give an instance of that class to insertClassPath() in ClassPool. This allows a non-standard resource to be included in the search path.


2. ClassPool

A ClassPool object is a container of CtClass objects. Once a CtClass object is created, it is recorded in a ClassPool for ever. This is because a compiler may need to access the CtClass object later when it compiles source code that refers to the class represented by that CtClass.

For example, suppose that a new method getter() is added to a CtClass object representing Point class. Later, the program attempts to compile source code including a method call to getter() in Point and use the compiled code as the body of a method, which will be added to another class Line. If the CtClass object representing Point is lost, the compiler cannot compile the method call to getter(). Note that the original class definition does not include getter(). Therefore, to correctly compile such a method call, the ClassPool must contain all the instances of CtClass all the time of program execution.

Avoid out of memory

This specification of ClassPool may cause huge memory consumption if the number of CtClass objects becomes amazingly large (this rarely happens since Javassist tries to reduce memory consumption in various ways). To avoid this problem, you can explicitly remove an unnecessary CtClass object from the ClassPool. If you call detach() on a CtClass object, then that CtClass object is removed from the ClassPool. For example,

You must not call any method on that CtClass object after detach() is called. However, you can call get() on ClassPool to make a new instance of CtClass representing the same class. If you call get(), the ClassPool reads a class file again and newly creates a CtClass object, which is returned by get().

Another idea is to occasionally replace a ClassPool with a new one and discard the old one. If an old ClassPool is garbage collected, the CtClass objects included in that ClassPool are also garbage collected. To create a new instance of ClassPool, execute the following code snippet:

This creates a ClassPool object that behaves as the default ClassPool returned by ClassPool.getDefault() does. Note that ClassPool.getDefault() is a singleton factory method provided for convenience. It creates a ClassPool object in the same way shown above although it keeps a single instance of ClassPool and reuses it. A ClassPool object returned by getDefault() does not have a special role. getDefault() is a convenience method.

Note that new ClassPool(true) is a convenient constructor, which constructs a ClassPool object and appends the system search path to it. Calling that constructor is equivalent to the following code:

Cascaded ClassPools

If a program is running on a web application server, creating multiple instances of ClassPool might be necessary; an instance of ClassPool should be created for each class loader (i.e. container). The program should create a ClassPool object by not calling getDefault() but a constructor of ClassPool.

Multiple ClassPool objects can be cascaded like java.lang.ClassLoader. For example,

If child.get() is called, the child ClassPool first delegates to the parent ClassPool. If the parent ClassPool fails to find a class file, then the child ClassPool attempts to find a class file under the ./classes directory.

If child.childFirstLookup is true, the child ClassPool attempts to find a class file before delegating to the parent ClassPool. For example,

Changing a class name for defining a new class

A new class can be defined as a copy of an existing class. The program below does that:

This program first obtains the CtClass object for class Point. Then it calls setName() to give a new name Pair to that CtClass object. After this call, all occurrences of the class name in the class definition represented by that CtClass object are changed from Point to Pair. The other part of the class definition does not change.

Note that setName() in CtClass changes a record in the ClassPool object. From the implementation viewpoint, a ClassPool object is a hash table of CtClass objects. setName() changes the key associated to the CtClass object in the hash table. The key is changed from the original class name to the new class name.

Therefore, if get("Point") is later called on the ClassPool object again, then it never returns the CtClass object that the variable cc refers to. The ClassPool object reads a class file Point.class again and it constructs a new CtClass object for class Point. This is because the CtClass object associated with the name Point does not exist any more. See the followings:

cc1 and cc2 refer to the same instance of CtClass that cc does whereas cc3 does not. Note that, after cc.setName("Pair") is executed, the CtClass object that cc and cc1 refer to represents the Pair class.

The ClassPool object is used to maintain one-to-one mapping between classes and CtClass objects. Javassist never allows two distinct CtClass objects to represent the same class unless two independent ClassPool are created. This is a significant feature for consistent program transformation.

To create another copy of the default instance of ClassPool, which is returned by ClassPool.getDefault(), execute the following code snippet (this code was already shown above):

If you have two ClassPool objects, then you can obtain, from each ClassPool, a distinct CtClass object representing the same class file. You can differently modify these CtClass objects to generate different versions of the class.

Renaming a frozen class for defining a new class

Once a CtClass object is converted into a class file by writeFile() or toBytecode(), Javassist rejects further modifications of that CtClass object. Hence, after the CtClass object representing Point class is converted into a class file, you cannot define Pair class as a copy of Point since executing setName() on Point is rejected. The following code snippet is wrong:

To avoid this restriction, you should call getAndRename() in ClassPool. For example,

If getAndRename() is called, the ClassPool first reads Point.class for creating a new CtClass object representing Point class. However, it renames that CtClass object from Point to Pair before it records that CtClass object in a hash table. Thus getAndRename() can be executed after writeFile() or toBytecode() is called on the the CtClass object representing Point class.


3. Class loader

If what classes must be modified is known in advance, the easiest way for modifying the classes is as follows:

If whether a class is modified or not is determined at load time, the users must make Javassist collaborate with a class loader. Javassist can be used with a class loader so that bytecode can be modified at load time. The users of Javassist can define their own version of class loader but they can also use a class loader provided by Javassist.


3.1 The toClass method in CtClass

The CtClass provides a convenience method toClass(), which requests the context class loader for the current thread to load the class represented by the CtClass object. To call this method, the caller must have appropriate permission; otherwise, a SecurityException may be thrown.

The following program shows how to use toClass():

Test.main() inserts a call to println() in the method body of say() in Hello. Then it constructs an instance of the modified Hello class and calls say() on that instance.

Note that the program above depends on the fact that the Hello class is never loaded before toClass() is invoked. If not, the JVM would load the original Hello class before toClass() requests to load the modified Hello class. Hence loading the modified Hello class would be failed (LinkageError is thrown). For example, if main() in Test is something like this:

then the original Hello class is loaded at the first line of main and the call to toClass() throws an exception since the class loader cannot load two different versions of the Hello class at the same time.

If the program is running on some application server such as JBoss and Tomcat, the context class loader used by toClass() might be inappropriate. In this case, you would see an unexpected ClassCastException. To avoid this exception, you must explicitly give an appropriate class loader to toClass(). For example, if bean is your session bean object, then the following code:

would work. You should give toClass() the class loader that has loaded your program (in the above example, the class of the bean object).

toClass() is provided for convenience. If you need more complex functionality, you should write your own class loader.


3.2 Class loading in Java

In Java, multiple class loaders can coexist and each class loader creates its own name space. Different class loaders can load different class files with the same class name. The loaded two classes are regarded as different ones. This feature enables us to run multiple application programs on a single JVM even if these programs include different classes with the same name.

If the same class file is loaded by two distinct class loaders, the JVM makes two distinct classes with the same name and definition. The two classes are regarded as different ones. Since the two classes are not identical, an instance of one class is not assignable to a variable of the other class. The cast operation between the two classes fails and throws a ClassCastException.

For example, the following code snippet throws an exception:

The Box class is loaded by two class loaders. Suppose that a class loader CL loads a class including this code snippet. Since this code snippet refers to MyClassLoader, Class, Object, and Box, CL also loads these classes (unless it delegates to another class loader). Hence the type of the variable b is the Box class loaded by CL. On the other hand, myLoader also loads the Box class. The object obj is an instance of the Box class loaded by myLoader. Therefore, the last statement always throws a ClassCastException since the class of obj is a different verison of the Box class from one used as the type of the variable b.

Multiple class loaders form a tree structure. Each class loader except the bootstrap loader has a parent class loader, which has normally loaded the class of that child class loader. Since the request to load a class can be delegated along this hierarchy of class loaders, a class may be loaded by a class loader that you do not request the class loading. Therefore, the class loader that has been requested to load a class C may be different from the loader that actually loads the class C. For distinction, we call the former loader the initiator of C and we call the latter loader the real loader of C.

Furthermore, if a class loader CL requested to load a class C (the initiator of C) delegates to the parent class loader PL, then the class loader CL is never requested to load any classes referred to in the definition of the class C. CL is not the initiator of those classes. Instead, the parent class loader PL becomes their initiators and it is requested to load them. The classes that the definition of a class C referes to are loaded by the real loader of C.

To understand this behavior, let's consider the following example.

Suppose that a class Window is loaded by a class loader L. Both the initiator and the real loader of Window are L. Since the definition of Window refers to Box, the JVM will request L to load Box. Here, suppose that L delegates this task to the parent class loader PL. The initiator of Box is L but the real loader is PL. In this case, the initiator of Point is not L but PL since it is the same as the real loader of Box. Thus L is never requested to load Point.

Next, let's consider a slightly modified example.

Now, the definition of Window also refers to Point. In this case, the class loader L must also delegate to PL if it is requested to load Point. You must avoid having two class loaders doubly load the same class. One of the two loaders must delegate to the other.

If L does not delegate to PL when Point is loaded, widthIs() would throw a ClassCastException. Since the real loader of Box is PL, Point referred to in Box is also loaded by PL. Therefore, the resulting value of getSize() is an instance of Point loaded by PL whereas the type of the variable p in widthIs() is Point loaded by L. The JVM regards them as distinct types and thus it throws an exception because of type mismatch.

This behavior is somewhat inconvenient but necessary. If the following statement:

did not throw an exception, then the programmer of Window could break the encapsulation of Point objects. For example, the field x is private in Point loaded by PL. However, the Window class could directly access the value of x if L loads Point with the following definition:

For more details of class loaders in Java, the following paper would be helpful:


3.3 Using javassist.Loader

Javassist provides a class loader javassist.Loader. This class loader uses a javassist.ClassPool object for reading a class file.

For example, javassist.Loader can be used for loading a particular class modified with Javassist.

This program modifies a class test.Rectangle. The superclass of test.Rectangle is set to a test.Point class. Then this program loads the modified class, and creates a new instance of the test.Rectangle class.

If the users want to modify a class on demand when it is loaded, the users can add an event listener to a javassist.Loader. The added event listener is notified when the class loader loads a class. The event-listener class must implement the following interface:

The method start() is called when this event listener is added to a javassist.Loader object by addTranslator() in javassist.Loader. The method onLoad() is called before javassist.Loader loads a class. onLoad() can modify the definition of the loaded class.

For example, the following event listener changes all classes to public classes just before they are loaded.

Note that onLoad() does not have to call toBytecode() or writeFile() since javassist.Loader calls these methods to obtain a class file.

To run an application class MyApp with a MyTranslator object, write a main class as following:

To run this program, do:

The class MyApp and the other application classes are translated by MyTranslator.

Note that application classes like MyApp cannot access the loader classes such as Main2, MyTranslator, and ClassPool because they are loaded by different loaders. The application classes are loaded by javassist.Loader whereas the loader classes such as Main2 are by the default Java class loader.

javassist.Loader searches for classes in a different order from java.lang.ClassLoader. ClassLoader first delegates the loading operations to the parent class loader and then attempts to load the classes only if the parent class loader cannot find them. On the other hand, javassist.Loader attempts to load the classes before delegating to the parent class loader. It delegates only if:

This search order allows loading modified classes by Javassist. However, it delegates to the parent class loader if it fails to find modified classes for some reason. Once a class is loaded by the parent class loader, the other classes referred to in that class will be also loaded by the parent class loader and thus they are never modified. Recall that all the classes referred to in a class C are loaded by the real loader of C. If your program fails to load a modified class, you should make sure whether all the classes using that class have been loaded by javassist.Loader.


3.4 Writing a class loader

A simple class loader using Javassist is as follows:

The class MyApp is an application program. To execute this program, first put the class file under the ./class directory, which must not be included in the class search path. Otherwise, MyApp.class would be loaded by the default system class loader, which is the parent loader of SampleLoader. The directory name ./class is specified by insertClassPath() in the constructor. You can choose a different name instead of ./class if you want. Then do as follows:

The class loader loads the class MyApp (./class/MyApp.class) and calls MyApp.main() with the command line parameters.

This is the simplest way of using Javassist. However, if you write a more complex class loader, you may need detailed knowledge of Java's class loading mechanism. For example, the program above puts the MyApp class in a name space separated from the name space that the class SampleLoader belongs to because the two classes are loaded by different class loaders. Hence, the MyApp class cannot directly access the class SampleLoader.


3.5 Modifying a system class

The system classes like java.lang.String cannot be loaded by a class loader other than the system class loader. Therefore, SampleLoader or javassist.Loader shown above cannot modify the system classes at loading time.

If your application needs to do that, the system classes must be statically modified. For example, the following program adds a new field hiddenValue to java.lang.String:

This program produces a file "./java/lang/String.class".

To run your program MyApp with this modified String class, do as follows:

Suppose that the definition of MyApp is as follows:

If the modified String class is correctly loaded, MyApp prints hiddenValue.

Note: Applications that use this technique for the purpose of overriding a system class in rt.jar should not be deployed as doing so would contravene the Java 2 Runtime Environment binary code license.


3.6 Reloading a class at runtime

If the JVM is launched with the JPDA (Java Platform Debugger Architecture) enabled, a class is dynamically reloadable. After the JVM loads a class, the old version of the class definition can be unloaded and a new one can be reloaded again. That is, the definition of that class can be dynamically modified during runtime. However, the new class definition must be somewhat compatible to the old one. The JVM does not allow schema changes between the two versions. They have the same set of methods and fields.

Javassist provides a convenient class for reloading a class at runtime. For more information, see the API documentation of javassist.tools.HotSwapper.


Next page


Java(TM) is a trademark of Sun Microsystems, Inc.
Copyright (C) 2000-2015 by Shigeru Chiba, All rights reserved.