Approximating happens-before order: interplay between static analysis and state space traversal

Abstract

Techniques and tools for verification of multi-threaded programs must cope with the huge number of possible thread interleavings. Tools based on systematic exploration of a program state space employ partial order reduction to avoid redundant thread interleavings. The key idea is to make non-deterministic thread scheduling choices only at statements that read or modify the global state shared by multiple threads. We focus on the approach to partial order reduction used in tools such as Java Pathfinder (JPF), which construct the program state space on-the-fly, and therefore can use only information available in the current program state and execution history to identify statements that may be globally-relevant. In our previous work, we developed a field access analysis that provides information about fields that may be accessed during program execution, and used it in JPF for more precise identification of globally-relevant statements. We build upon that and propose a may-happen-before analysis that computes a sound approximation of the happens-before ordering. Partial order reduction techniques can use the happens-before ordering to detect pairs of globally-relevant field access statements that cannot be interleaved arbitrarily (due to thread synchronization), and based on that avoid making unnecessary thread scheduling choices. The may-happen-before analysis combines static analysis with knowledge of information available from the dynamic program state. Results of experiments with several Java programs show that usage of the may-happen-before analysis further improves the performance of JPF.