Collective Program Analysis

摘要

Popularity of data-driven software engineering has led to an increasing demand on the infrastructures to support efficient execution of tasks that require deeper source code analysis. While task optimization and parallelization are the adopted solutions, other research directions are less explored. We present collective program analysis (CPA), a technique for scaling large scale source code analyses, especially those that make use of control and data flow analysis, by leveraging analysis specific similarity. Analysis specific similarity is about, whether two or more programs can be considered similar for a given analysis. The key idea of collective program analysis is to cluster programs based on analysis specific similarity, such that running the analysis on one candidate in each cluster is sufficient to produce the result for others. For determining analysis specific similarity and clustering analysis-equivalent programs, we use a sparse representation and a canonical labeling scheme. Our evaluation shows that for a variety of source code analyses on a large dataset of programs, substantial reduction in the analysis time can be achieved; on average a 69% reduction when compared to a baseline and on average a 36% reduction when compared to a prior technique. We also found that a large amount of analysis-equivalent programs exists in large datasets.