Explicit and symbolic techniques for fast and scalable points-to analysis

Abstract

Points-to analysis that scales to large programs is still an open area of research and there are several trade-offs between speed and precision. In this paper, we report advances in achieving extremely fast and scalable analysis for field-sensitive inclusion-based points-to analysis. The first algorithm is based on an explicit representation, sparse bit-vector set representation. The second algorithm is a refinement of the first using symbolic set representations using binary decision diagrams. The first algorithm scales extremely well when compared with the state-of-the-art points-to analysis problems solving the same problem, while the second reduces the memory footprint tremendously, using, on average, 4.6x less memory than the first, and even performs slightly faster points-to set propagation. The techniques that we introduce are a judicious combination of several heuristics involving sparse bit-vector set representations, prioritized processing of worklists for efficient iteration while computing fixed-points, and usage of binary decision diagrams for storing (but not propagating) points-to sets. The implementation of our approaches scales to large real life Java applications. We evaluated our implementation on benchmark applications from two recent releases of DaCapo benchmark suite using a recent version of Java Standard Library (JRE 1.7_03). Using our techniques we can propagate points-to information on all the benchmarks in less than a minute, using at most 2GB of memory for explicit representation and, at most 600MB for symbolic representation. Comparison with the fastest and the most closely related explicit and symbolic approaches reveals that our techniques are more scalable than related explicit approaches, and in terms of time, on average 4x times faster than the current state-of-the-art.