Initialization problems for Java

We exhibit a grey area in the speciication of Java and of its implementation through the Java Virtual Machine (JVM): the treatment of initialization of classes and interfaces. We report the result of our experiments with diierent implementations of Java, which connrm the theoretical prediction of our work on mathematical models for Java 4] and the JVM 3], namely that the designers of Java and the JVM have used notions of initialization which do not match and which aaict the portability of Java programs. We show also that concurrent initialization may deadlock and that various current Java compilers violate the initialization semantics through standard optimization techniques. 1 The Initialization Problem Any programming language that supports libraries of autonomous components has to determine when to initialize a library component. Examples of components are modules, packages or classes. In traditional block oriented languages, like Pascal, C or C++, a global variable is associated with the initialization status of a component. This variable is initialized in the main program. However, this does not work for modern object-oriented languages like Eiiel or Java. By deenition they do not have a main program. Additionally, Java's support for dynamic class loading requires that initialization should be lazy, that is, it should take place just before a rst use of any item declared in a class. An additional complication is caused by parallel threads that try to initialize the same class at the same time. Through two recent studies, where we provide a mathematical deenition of the meaning of Java 4] and prove a natural scheme for compilation of Java on the Java Virtual Machine (JVM) to be correct 3], we discovered the following problems with initialization in Java and its implementations: { The exact moment when initializers for static elds in interfaces are executed is underspeciied in the Java Language Speciication (JLS) 6]. They can either be initialized when an interface eld is rst used, or when a class (an interface) that implements (extends) the interface is initialized. { The JLS and the JVM Speciication (JVMS) 7] use semantically diierent and connicting concepts to de-ne when a class or interface is initialized. Whereas the JLS uses the concept of rst active use to trigger class or interface initialization, the JVM implements the initialization as a mandatory step of its (constant pool) resolution process. This process, however, is also triggered for instructions whose execution does not constitute …