Pruning False Positives of Static Data-Race Detection via Thread Specialization

摘要: Static data-race detection is a powerful tool by providing clues for dynamic approaches to only instrument certain memory accesses. However, static analysis suffers from high false positive rate. A key reason that overestimates the set of shared objects thread can access. We propose specialization distinguish threads statically. By fixing number as well ID assigned each thread, program be transformed simplified version. on this infer range addresses accessed more accurately. Our approach prunes positives an average 89.2% and reduces instrumentation 63.4% in seven benchmarks.