Incrementalizing lattice-based program analyses in Datalog

Abstract

Program analyses detect errors in code, but when code changes frequently as in an IDE, repeated re-analysis from-scratch is unnecessary: It leads to poor performance unless we give up on precision and recall. Incremental program analysis promises to deliver fast feedback without giving up on precision or recall by deriving a new analysis result from the previous one. However, Datalog and other existing frameworks for incremental program analysis are limited in expressive power: They only support the powerset lattice as representation of analysis results, whereas many practically relevant analyses require custom lattices and aggregation over lattice values. To this end, we present a novel algorithm called DRed L that supports incremental maintenance of recursive lattice-value aggregation in Datalog. The key insight of DRed L is to dynamically recognize increasing replacements of old lattice values by new ones, which allows us to avoid the expensive deletion of the old value. We integrate DRed L into the analysis framework IncA and use IncA to realize incremental implementations of strong-update points-to analysis and string analysis for Java. As our performance evaluation demonstrates, both analyses react to code changes within milliseconds. CCS Concepts: • Software and its engineering → Automated static analysis;

JavaDL: automatically incrementalizing Java bug pattern detection

“…The literature has proposed a number of different techniques for incremental updates of Datalogstyle analyses [Gupta et al 1993;Szabó et al 2018]. However, we found these techniques unsuitable for JavaDL: first, they assume that all ground facts are permanently stored in a database.…”

Concise, type-safe, and efficient structural diffing

“…To demonstrate that truediff can be used to drive incremental computing, we have reimplemented the driver of the incremental program analysis framework IncA [20]. IncA incrementally maintains a Datalog database of derived properties about a syntax tree, such as typing [15,21] or points-to information [19]. However, IncA so far required that programs were edited in a projectional editor, which issues fine-grained change notifications.…”

Demanded abstract interpretation

“…The application of incremental computation to program analysis is similarly well-studied, going back at least to the development of incremental dataflow analyses to support responsive continuous compilation [35,48]. Recent work has contributed incremental versions of several classes of program analysis, including IFD-S/IDE dataflow analyses [5,16] and analyses based on extensions to Datalog [46,47]. These specialized approaches offer effective solutions for certain classes of program analysis, but place restrictions on abstract domains that rule out arbitrary abstract interpretations in infinite-height domains.…”