HistLock+: Precise Memory Access Maintenance Without Lockset Comparison for Complete Hybrid Data Race Detection

Abstract

Dynamic hybrid data race detectors alleviate the detection imprecision problem incurred by pure lockset-based race detectors and the thread interleaving sensitive problem incurred by pure happens-before race detectors. Nonetheless, to ensure at least one data race on every memory location to be detected, keeping all historical memory access events in the analysis state of such a detector is impractical. Existing complete hybrid race detectors perform extensive comparisons among the locksets of the memory accesses on each memory location to identify which of them to be retained in its analysis state, which incurs significant runtime overhead. In this paper, we investigate to what extent a complete hybrid data race detector able to perform such identifications without lockset comparison. We present HistLock+, which is built atop thread epoch and lock release events to infer whether two memory accesses on the same memory location from the same thread in between consecutive lock release operations have any lock subset relation without performing expensive lockset comparison. HistLock+ guarantees exactly one racy memory access event to be reported on each thread segment separated by lock releases and hard-order thread synchronizations, and it never reports false positive on lockset violation. We have validated HistLock+ using the PARSEC benchmark suite and four real-world applications. The experimental results showed that HistLock+ was 122% faster and 28% more memory-efficient than the previous state-of-the-art complete hybrid race detector. Moreover, HistLock+ achieved the highest effectiveness in race detection among all evaluated race detectors in our experiment.