Statistical program dependence approximation

Abstract

This dissertation investigates the beneficial incorporation of statistical methods into observation-based program dependence analysis. Observation-based slicing, which introduced observation-based dependence analysis, performs program slicing purely dynamically. It has been shown that the observation-based approach can successfully overcome the limitations of the traditional static approach. However, it still suffers from scalability and interpretability issues. We introduce three different statistical strategies to address either or both fronts of observation-based dependence analysis. We first exploit a lexical model built from the source code to discover dependence relations in the program efficiently. We also show how to formulate observation-based dependence analysis using statistical modeling. Further, we employ causal inference analysis to facilitate program dependence interpretability. We evaluate the proposed analyses, both quantitatively and qualitatively, using a mixture of real-world software and representative example programs in order to provide confidence in our modeling.