Abstract
Algebraic program analyses compute information about a program's behavior by first (a) computing a valid path expressionÐi.e., a regular expression that recognizes all feasible execution paths (and usually more)Ðand then (b) interpreting the path expression in a semantic algebra that defines the analysis. There are an infinite number of different regular expressions that qualify as valid path expressions, which raises the question łWhich one should we choose?ž While any choice yields a sound result, for many analyses the choice can have a drastic effect on the precision of the results obtained. This paper investigates the following two questions: (1) What does it mean for one valid path expression to be łbetterž than another? (2) Can we compute a valid path expression that is łbetter,ž and if so, how? We show that it is not satisfactory to compare two path expressions E1 and E2 solely by means of the languages that they generate. Counter to one's intuition, it is possible for L(E2) (L(E1), yet for E2 to produce a less-precise analysis result than E1Ðand thus we would not want to perform the transformation E1 → E2. However, the exclusion of paths so as to analyze a smaller language of paths is exactly the refinement criterion used by some prior methods. In this paper, we develop an algorithm that takes as input a valid path expression E, and returns a valid path expression E′ that is guaranteed to yield analysis results that are at least as good as those obtained using E. While the algorithm sometimes returns E itself, it typically does not: (i) we prove a no-degradation result for the algorithm's base caseÐfor transforming a leaf loop (i.e., a most-deeply-nested loop); (ii) at a non-leaf loop L, the algorithm treats each loop L′ in the body of L as an indivisible atom, and applies the leaf-loop algorithm to L; the no-degradation result carries over to (ii), as well. Our experiments show that the technique has a substantial impact: the loop-refinement algorithm allows the implementation of Compositional Recurrence Analysis to prove over 25% more assertions for a collection of challenging loop micro-benchmarks.
Original language | English (US) |
---|---|
Article number | 45 |
Journal | Proceedings of the ACM on Programming Languages |
Volume | 3 |
Issue number | POPL |
DOIs | |
State | Published - Jan 2019 |
All Science Journal Classification (ASJC) codes
- Software
- Safety, Risk, Reliability and Quality
Keywords
- Abstract-interpretation precision
- Algebraic program analysis
- Control-flow refinement