Modular Normalization with Types

Doctoral thesis, 2023

With the increasing use of software in today’s digital world, software is becoming more and more complex and the cost of developing and maintaining software has skyrocketed. It has become pressing to develop software using effective tools that reduce this cost. Programming language research aims to develop such tools using mathematically rigorous foundations. A recurring and central concept in programming language research is normalization: the process of transforming a complex expression in a language to a canonical form while preserving its meaning. Normalization has compelling benefits in theory and practice, but is extremely difficult to achieve. Several program transformations that are used to optimise programs, prove properties of languages and check program equivalence, for instance, are after all instances of normalization, but they are seldom viewed as such.

Viewed through the lens of current methods, normalization lacks the ability to be broken into sub-problems and solved independently, i.e., lacks modularity. To make matters worse, such methods rely excessively on the syntax of the language, making the resulting normalization algorithms brittle and sensitive to changes in the syntax. When the syntax of the language evolves due to modification or extension, as it almost always does in practice, the normalization algorithm may need to be revisited entirely. To circumvent these problems, normalization is currently either abandoned entirely or concrete instances of normalization are achieved using ad hoc means specific to a particular language. Continuing this trend in programming language research poses the risk of building on a weak foundation where languages either lack fundamental properties that follow from normalization or several concrete instances end up repeated in an ad hoc manner that lacks reusability.

This thesis advocates for the use of type-directed Normalization by Evaluation (NbE) to develop normalization algorithms. NbE is a technique that provides an opportunity for a modular implementation of normalization algorithms by allowing us to disentangle the syntax of a language from its semantics. Types further this opportunity by allowing us to dissect a language into isolated fragments, such as functions and products, with an individual specification of syntax and semantics. To illustrate type-directed NbE in context, we develop NbE algorithms and show their applicability for typed programming language calculi in three different domains (modal types, static information-flow control and categorical combinators) and for a family of embedded-domain specific languages in Haskell.