Compacting points-to sets through object clustering

Inclusion-based set constraint solving is the most popular technique for whole-program points-to analysis whereby an analysis is typically formulated as repeatedly resolving constraints between points-to sets of program variables. The set union operation is central to this process. The number of points-to sets can grow as analyses become more precise and input programs become larger, resulting in more time spent performing unions and more space used storing these points-to sets. Most existing approaches focus on improving scalability of precise points-to analyses from an algorithmic perspective and there has been less research into improving the data structures behind the analyses.

Bit-vectors as one of the more popular data structures have been used in several mainstream analysis frameworks to represent points-to sets. To store memory objects in bit-vectors, objects need to mapped to integral identifiers. We observe that this object-to-identifier mapping is critical for a compact points-to set representation and the set union operation. If objects in the same points-to sets (co-pointees) are not given numerically close identifiers, points-to resolution can cost significantly more space and time. Without data on the unpredictable points-to relations which would be discovered by the analysis, an ideal mapping is extremely challenging.

In this paper, we present a new approach to inclusion-based analysis by compacting points-to sets through object clustering. Inspired by recent staged analysis where an auxiliary analysis produces results approximating a more precise main analysis, we formulate points-to set compaction as an optimisation problem solved by integer programming using constraints generated from the auxiliary analysis's results in order to produce an effective mapping. We then develop a more approximate mapping, yet much more efficiently, using hierarchical clustering to compact bit-vectors. We also develop an improved representation of bit-vectors (called core bit-vectors) to fully take advantage of the newly produced mapping. Our approach requires no algorithmic change to the points-to analysis. We evaluate our object clustering on flow-sensitive points-to analysis using 8 open-source programs (>3.1 million lines of LLVM instructions) and our results show that our approach can successfully improve the analysis with an up to 1.83× speed up and an up to 4.05× reduction in memory usage.

CCS Concepts: • Software and its engineering → Automated static analysis; • Theory of computation → Data compression.