Program Intelligence, Declaratively and Symbolically

The automatic understanding of programs, in insightful, high-level terms, has long been a dream of computer science.

The area of static program analysis has made significant progress in such understanding by algorithmically modeling all possible program behaviors.

In this setting, declarative program analysis has recently demonstrated great success in capturing powerful algorithms efficiently and elegantly, in a form that bridges mathematical logic and intuitive human understanding.The PI’s research has established a world-leading program in declarative program analysis, with multiple independent signs of high recognition.

However, the dream of automatic deep program understanding remains elusive: static analysis tools are still reliant on significant human insights and extensive customization for the analysis domain. Is there hope for a giant step forward?

The PINDESYM approach posits that two emerging breakthroughs offer excellent promise to take declarative program analysis to a next level, capable of realizing the dream of automatic program understanding.

The first is the idea of combining a declarative system (e.g., a Datalog fixpoint engine) and a symbolic reasoning system, such as an SMT solver or algebraic rewrite system.

The second is the seamless integration of a machine learning approach, over large amounts of data (from past code bases), in the declarative inference process.The PINDESYM project will leverage symbolic reasoning and learning approaches to greatly advanceprogram analysis.

The challenge is dual: not only to invent powerful new techniques and algorithms, but also to capture all the diversity in symbolic, value-flow, and learning-based reasoning in a single, unified, reusable, and extensible analysis framework—a true deep program understanding engine, far beyond current approaches.