Prescient memory: exposing weak memory model behavior by looking into the future

Abstract

Shared-memory parallel programs are hard to get right. A major challenge is that language and hardware memory models allow unexpected, erroneous behaviors for executions containing data races. Researchers have introduced dynamic analyses that expose weak memory model behaviors, but these approaches cannot expose behaviors due to loading a "future value" -- a value written by a program store that executes after the program load that uses the value. This paper presents prescient memory (PM), a novel dynamic analysis that exposes behaviors due to future values. PM speculatively returns a future value at a program load, and tries to validate the speculative value at a later store. To enable PM to expose behaviors due to future values in real application executions, we introduce a novel approach that increases the chances of using and successfully validating future values, by profiling and predicting future values and guiding execution. Experiments show that our approach is able to uncover a few previously unknown behaviors due to future values in benchmarked versions of real applications. Overall, PM overcomes a key limitation of existing approaches, broadening the scope of program behaviors that dynamic analyses can expose.