An important component of efficient approaches to software model checking and systematic concurrency testing is partial order reduction, which eliminates redundant non-deterministic thread scheduling choices during the state space traversal. Thread choices have to be created only at the execution of actions that access the global state visible by multiple threads, so the key challenge is to precisely determine the set of such globally-relevant actions.
This includes accesses to object fields and array elements, and thread synchronization. However, some tools completely disable thread choices at actions that access individual array elements in order to avoid state explosion.
We show that they can miss concurrency errors in such a case. Then, as the main contribution, we present a new hybrid analysis that identifies globally-relevant actions that access arrays.
Our hybrid analysis combines static analysis with dynamic analysis, usage of information from dynamic program states, and symbolic interpretation of program statements. Results of experiments with two popular approaches to partial order reduction show that usage of the hybrid analysis (1) eliminates many additional redundant thread choices and (2) improves the performance of software model checking on programs that use arrays.