The Flow-Insensitive Precision of Andersen’s Analysis in Practice

Abstract

We present techniques for determining the precision gap between Andersen’s points-to analysis and precise flow-insensitive points-to analysis in practice. While previous work has shown that such a gap may exist, no efficient algorithm for precise flow-insensitive analysis is known, making measurement of the gap on real-world programs difficult. We give an algorithm for precise flow-insensitive analysis of programs with finite memory, based on a novel technique for refining any points-to analysis with a search for flow-insensitive witnesses. We give a compact symbolic encoding of the technique that enables computing the search using a tuned SAT solver. We also present extensions of the algorithm that enable computing lower and upper bounds on the precision gap in the presence of dynamic memory allocation. In our experimental evaluation over a suite of small- to medium-sized C programs, we never observed a precision gap between Andersen’s analysis and the precise analysis. In other words, Andersen’s analysis computed a precise flow-insensitive result for all of our benchmarks. Hence, we conclude that while better algorithms for the precise flow-insensitive analysis are still of theoretical interest, their practical impact for C programs is likely to be negligible.