Datalog for Static Analysis

:: HOPL, by Ben Greenman

Datalog is an old DSL that frequently appears in work on static analysis. This edition of HOPL 2017 explores the origins of Datalog in general, its early use in program analysis, and why Datalog remains a useful tool.

Full notes:

Datalog as a language was introduced by 1978 (its semantic foundations date back to 1976). It is predicate logic as a database query language. The traditional view of a Datalog program is a time invariant transformation over the time varying data stored in an external database.

In the early 1990’s, Uwe Aβmann designed a graph rewriting systems (EARS) that could:

  1. Uniformly express various problems in static analysis
  2. Systematically derive efficient solutions to such problems.

(Prior work had derived the same solutions with ad-hoc methods.) Aβmann’s system is equivalent to Datalog.

In 1993, Reps used the CORAL deductive database (an implementation of Datalog) to derive an on-demand (read: lazy) implementation of program slicing from a specification of the slicing problem.

Both Aβmann’s and Reps work appeared in 1994. This was the first time Datalog had been used to implement a static analysis.

Researchers continue to use Datalog because:

  • predicate logic (specifically: Horn clauses without function symbols or negation) is useful for expressing recursive relations … and static analyses are all about recursive relations
  • the language separates specifications from their implementation
  • there are many techniques for efficiently serving a Datalog query
  • these techniques have been implemented in at least one commercial Datalog engine

For an excellent description of how Datalog can benefit static analysis, see the introduction to Rep’s paper.