Principal type schemes for modular programs

Two of the most prominent features of ML are its expressive module system and its support for Damas-Milner type inference. However, while the foundations of both these features have been studied extensively, their interaction has never received a proper type-theoretic treatment. One consequence is that both the official Definition and the alternative Harper-Stone semantics of Standard ML are difficult to implement correctly. To bolster this claim, we offer a series of short example programs on which no existing SML typechecker follows the behavior prescribed by either formal definition. It is unclear how to amend the implementations to match the definitions or vice versa. Instead, we propose a way of defining how type inference interacts with modules that is more liberal than any existing definition or implementation of SML and, moreover, admits a provably sound and complete typechecking algorithm via a straightforward generalization of Algorithm W. In addition to being conceptually simple, our solution exhibits a novel hybrid of the Definition and Harper-Stone semantics of SML, and demonstrates the broader relevance of some type-theoretic techniques developed recently in the study of recursive modules.