Datalog for Static Analysis

These are post-lecture notes for my February 7th presentation for HOPL 2017. The theme is the use of Datalog as a framework for building static analyses. The lecture covered the history of Datalog, who its original application to static analysis, and its later success. These notes are a bare-minimum transcription of my on-paper notes for the lecture. Better than nothing. Datalog for Static Analysis Last year, there were two new papers about an old domain-specific language. • From Datalog to Flix [14] at PLDI 2016 • Datafun: A Functional Datalog [4] at ICFP 2016 Both have a similar motivation, essentially: Datalog has been successful because it lies at a confluence between logic programming and static analysis. And both papers suggest new extensions for Datalog. The goal today is to understand why and how Datalog has been successful. To start, we’ll trace the papers’ citations. Both cite work regarding the LogicBlox Datalog engine: Flix cites the Doop program analyzer (for Java programs) [6] and Datafun cites a paper that summarizes the engine [3]. They also both cite Whaley and Lam’s bddbddb. Finally they each cite one early work on Datalog. Flix cites Ullman’s text on Principles of database and knowledgebase systems and Datafun cites the proceedings of a 1978 workshop on Logic and Databases, hosted by Gallaire, Minker, and Nicholas. (See Figure 1 for a picture of this paragraph. The solid lines represent citations in the 2016 papers.) At this point, the earliest cited work that ues Datalog to implement a static analysis is the bddbddb paper, from 2004. The bddbddb paper in turn cites work by Reps [17] that uses the Coral deductive database [16] (a Datalog database) to implement demand-driven program slicing. But the idea of Datalog for static analysis goes (at least!) one step further: to Uwe Aβmann’s dissertation work. That’s where we’ll start today. The road map is (1) Aβmann, (2) classic Datalog, (3) bddbddb, and (4) Doop. In terms of deltas and contributions, we’ll (1) see a uniform framework for expressing three static analyses, (2) rephrase that framework as Datalog, (3) see how context-sensitive points-to analysis turned Datalog from a novelty to a necessity, and (4) the power of a mature Datalog implementation. 1Ullman’s text demonstrates how Datalog can express and solve the reaching definitions problem, that is likely the first work to present the idea of implementing a static analysis with Datalog.