Soundness of Data Flow Analyses of Weak Memory Models

Modern multi-core microprocessors implement weak memory consistency models; programming for these architectures is a challenge. This paper solves a problem open for ten years, and originally posed by Rinard: we identify sufficient conditions for a data flow analysis to be sound wrt. weak memory models. We first identify a class of analyses that are sound, and provide a formal proof of soundness at the level of trace semantics. Then we discuss how analyses unsound with respect to weak memory models can be repaired via a fixed point iteration, and provide experimental data on the runtime overhead of this method.