Which pointer analysis should I use?

During the past two decades many different pointer analysis algorithms have been published. Although some descriptions include measurements of the effectiveness of the algorithm, qualitative comparisons among algorithms are difficult because of varying infrastructure, benchmarks, and performance metrics. Without such comparisons it is not only difficult for an implementor to determine which pointer analysis is appropriate for their application, but also for a researcher to know which algorithms should be used as a basis for future advances. This paper describes an empirical comparison of the effectiveness of five pointer analysis algorithms on C programs. The algorithms vary in their use of control flow information (flow-sensitivity) and alias data structure, resulting in worst-case complexity from linear to polynomial. The effectiveness of the analyses is quantified in terms of compile-time precision and efficiency. In addition to measuring the direct effects of pointer analysis, precision is also reported by determining how the information computed by the five pointer analyses affects typical client analyses of pointer information: Mod/Ref analysis, live variable analysis and dead assignment identification, reaching definitions analysis, dependence analysis, and conditional constant propagation and unreachable code identification. Efficiency is reported by measuring analysis time and memory consumption of the pointer analyses and their clients.