Modular monadic meta-theory

Abstract

This paper presents 3MT, a framework for modular mechanized meta-theory of languages with effects. Using 3MT, individual language features and their corresponding definitions -- semantic functions, theorem statements and proofs-- can be built separately and then reused to create different languages with fully mechanized meta-theory. 3MT combines modular datatypes and monads to define denotational semantics with effects on a per-feature basis, without fixing the particular set of effects or language constructs. One well-established problem with type soundness proofs for denotational semantics is that they are notoriously brittle with respect to the addition of new effects. The statement of type soundness for a language depends intimately on the effects it uses, making it particularly challenging to achieve modularity. 3MT solves this long-standing problem by splitting these theorems into two separate and reusable parts: a feature theorem that captures the well-typing of denotations produced by the semantic function of an individual feature with respect to only the effects used, and an effect theorem that adapts well-typings of denotations to a fixed superset of effects. The proof of type soundness for a particular language simply combines these theorems for its features and the combination of their effects. To establish both theorems, 3MT uses two key reasoning techniques: modular induction and algebraic laws about effects. Several effectful language features, including references and errors, illustrate the capabilities of 3MT. A case study reuses these features to build fully mechanized definitions and proofs for 28 languages, including several versions of mini-ML with effects.