UniFlow: A CFG-Based Framework for Pluggable Type Checking and Type Inference

Abstract

A type system is a crucial component of high-level programming languages, as it enhances program correctness by ruling out certain type errors. However, the built-in type system often adheres to a specific set of rules defined by the language's specification (e.g., Java, Kotlin and C++). Pluggable type systems were then introduced as an idea to provide customizable type rules for different scenarios.

Various approaches exist for implementing a pluggable type system. The Checker Framework is a well-known framework to facilitate the development of type checkers for Java. This framework enables developers to define their type rules and override the analysis logic. Additionally, Checker Framework Inference is a framework built upon the Checker Framework to provide constraint-based whole-program inference. It helps to reduce the burden of manually annotating the codebase when applying a new type system.

However, the complexity of these frameworks presents a steep learning curve to type system developers. This work examines some of the critical issues encountered from our previous experience in developing these frameworks. The Checker Framework performs its analysis on two different program representations: abstract syntax tree (AST) and control flow graph (CFG). The shared responsibilities of these representations in the framework cause readability and maintainability issues for developers. Checker Framework Inference suffers not only from the same problem but also from difficulty in employing the same type rules for type checking and type inference. This is because the underlying Checker Framework assumes type rules can be checked modularly and immediately at any AST. In contrast, the type inference is not modular and generates constraints to be solved in a later stage.

We propose a novel CFG-based type system framework, UniFlow, addressing the aforementioned issues by providing a unified development process for type systems supporting both type checking and type inference. It strives to resolve types and apply type rules on the program's CFGs whenever possible. This approach reduces friction in type system development, allowing developers to focus on a single flow-sensitive program representation that is simpler than ASTs. It also forces developers to express type rules as constraints, such that the same set of type rules can be implemented once, but consistently reused in type checking and type inference. Moreover, our framework supports running multiple type systems and attempts to improve error message reporting for users.

We present UniFlow's architecture and explain each crucial component and functionality in detail. We discuss the advantages and limitations of our framework. Furthermore, we explore the initial implementation of the framework and outline future research directions.