Practical Unification for Dependent Type Checking

Licentiate thesis, 2020

When using popular dependently-typed languages such as Agda, Idris or Coq to write a proof or a program, some function arguments can be omitted, both to decrease code size and to improve readability.  Type checking such a program involves inferring a combination of these implicit arguments that makes the program type-correct.

Finding such a combination of implicit arguments entails solving a higher-order unification problem.
Because higher-order unification is undecidable, our aim is to infer the omitted arguments for as many programs as possible with a reasonable use of computational resources. The extent to which
these goals are achieved affect how usable a dependently-typed proof assistant or programming language is in practice.

Current approaches to higher-order unification are in some cases too inflexible, postponing unification of terms until their types have been unified (Coq, Idris). In other cases they are too optimistic, which sometimes leads to ill-typed terms that break internal invariants (Agda).

In order to increase the flexibility of our unifier without sacrificing soundness, we use the twin types technique by Gundry and McBride. We simplify their approach so that it can be used within an existing type
theory without changes to the syntax of terms. We also extend it so that it can handle more classes of constraints. We show that the resulting solutions are correct and unique.

Finally, we implement the resulting unification algorithm on an existing type checker prototype for a smaller variant of the Agda language, developed by Mazzoli and Abel. We make a suitable choice of internal term representation, and use few, if any, example-specific optimizations. We obtain a type-checker which avoids ill-typed solutions, and is also able to handle some challenging examples in time and memory comparable to the existing Agda implementation.