Thread-Sensitive Data Race Detection for Java

Abstract

In this paper we present StaTS, a precise static data-race detection mechanism for Java. It analyzes applications in four phases. The first one is a novel points-to analysis that includes approximations of threads and execution contexts. The second phase uses the results of the points-to analysis to compute which fields are accessed by which threads, while considering the locks held by the threads. The third phase carries out a context-sensitive static happens-before analysis to rule out accesses in execution contexts that can never be executed in parallel. The final phase builds upon the results of the first three to determine conflicting accesses and report them to the user. Our proof-of-concept implementation does not scale for large programs, which is why it can optionally limit the number of points-to relations it considers, based on sampling. Nevertheless, our evaluation shows that, even with sampling enabled for large programs, StaTS detects more data races than existing approaches. In terms of execution time, the analysis without sampling takes in the order of seconds for smaller programs. For larger ones and with sampling enabled, analysis takes minutes, thus being practically usable in nightly build environments in all cases.