Which compiler optimization options should I use for detecting data races in multithreaded programs?

Abstract

Different compiler optimization options may produce different versions of object code. To the best of our knowledge, existing studies on concurrency bug detection in the public literature have not reported the effects of different compiler optimization options on detection effectiveness. This paper reports a preliminary but the first study in the exploratory nature to investigate this aspect. The study examines the happened-before based predictive data race detection scenarios on four benchmarks from the PARSEC 3.0 suite compiled under six different GNU GCC optimization options. We observe from the data set that the same race detection technique may produce different sets of races or different detection probabilities under different optimization scenarios. Based on the observations, we formulate two hypotheses for future investigations.