Automating and Validating Program Annotations

Program annotations help to catch errors, improve program understanding, and specify invariants. Adding annotations, however, is often a manual, laborious, tedious, and error prone process especially when programs are large. We offer a novel approach for automating a part of this process. Developers first specify an initial set of annotations for a few variables and types. Our LEearning ANnnotations (Lean) system combines these annotations with run-time monitoring, program analysis, and machine-learning approaches to discover and validate these annotations on unannotated variables. We evaluate our prototype implementation on open-source software projects and our results suggest that a modest set of annotations capture many program variables, and Lean can generalize from a small set of annotated variables to annotate many other variables. After users annotate approximately 6% of the program variables and types, Lean correctly annotates an additional 69% of variables in the best case, 47% in the average, and 12% in the worst case.