The Java Memory Model

Abstract

After many years, support for multithreading has been integrated into mainstream programming languages. Inclusion of this feature brings with it a need for a clear and direct explanation of how threads interact through memory. Programmers need to be told, simply and clearly, what might happen when their programs execute. Compiler writers need to be able to work their magic without interfering with the promises that are made to programmers. Java's original threading specification, its memory model, was fundamentally flawed. Some language features, like volatile fields, were under-specified: their treatment was so weak as to render them useless. Other features, including fields without access modifiers, were over-specified: the memory model prevents almost all optimizations of code containing these "normal" fields. Finally, some features, like final fields, had no specification at all beyond that of normal fields; no additional guarantees were provided about what will happen when they are used. This work has attempted to remedy these limitations. We provide a clear and concise definition of thread interaction. It is sufficiently simple for programmers to work with, and flexible enough to take advantage of compiler and processor-level optimizations. We also provide formal and informal techniques for verifying that the model provides this balance. These issues had never been addressed for any programming language: in addressing them for Java, this dissertation provides a framework for all multithreaded languages. The work described in this dissertation has been incorporated into the version 5.0 of the Java programming language.