An Online Dynamic Analysis for Sound Predictive Data Race Detection

Dynamic analyses struggle to detect all of a program’s dataraces, since some data races do not manifest in observedexecutions. Predictive analyses detect true data races thatexist only in other, unobserved executions. Smaragdakiset al. introduce the causally-precedes (CP) relation and apolynomial-time analysis for sound (no false races) predic-tive data race detection. However, their analysis and otherexisting sound predictive analyses cannot scale beyond ana-lyzing bounded windows of execution traces.This work introduces a novel dynamic analysis calledRaptor that computes CP soundly and completely. Unlikeexisting analyses, Raptor is inherently an online analysis thatanalyzes an execution trace in its entirety and finds all of itsCP-races. An evaluation of a prototype implementation ofRaptor shows that it scales to program executions that exist-ing analyses cannot handle, finding data races that existinganalyses cannot find.