Foundations for programming and implementing effect handlers

Abstract

First-class control operators provide programmers with an expressive and efficient means for manipulating control through reification of the current control state as a first-class object, enabling programmers to implement their own computational effects and control idioms as shareable libraries. Effect handlers provide a particularly structured approach to programming with first-class control by naming control reifying operations and separating from their handling. This thesis is composed of three strands of work in which I develop operational foundations for programming and implementing effect handlers as well as exploring the expressive power of effect handlers. The first strand develops a fine-grain call-by-value core calculus of a statically typed programming language with a structural notion of effect types, as opposed to the nominal notion of effect types that dominates the literature. With the structural approach, effects need not be declared before use. The usual safety properties of statically typed programming are retained by making crucial use of row polymorphism to build and track effect signatures. The calculus features three forms of handlers: deep, shallow, and parameterised. They each offer a different approach to manipulate the control state of programs. Traditional deep handlers are defined by folds over computation trees, and are the original con-struct proposed by Plotkin and Pretnar. Shallow handlers are defined by case splits (rather than folds) over computation trees. Parameterised handlers are deep handlers extended with a state value that is threaded through the folds over computation trees. To demonstrate the usefulness of effects and handlers as a practical programming abstraction I implement the essence of a small UNIX-style operating system complete with multi-user environment, time-sharing, and file I/O. The second strand studies continuation passing style (CPS) and abstract machine semantics, which are foundational techniques that admit a unified basis for implementing deep, shallow, and parameterised effect handlers in the same environment. The CPS translation is obtained through a series of refinements of a basic first-order CPS translation for a fine-grain call-by-value language into an untyped language. Each refinement moves toward a more intensional representation of continuations eventually arriving at the notion of generalised continuation, which admit simultaneous support for deep, shallow, and parameterised handlers. The initial refinement adds support for deep handlers by representing stacks of continuations and handlers as a curried sequence of arguments. The image of the resulting translation is not properly tail-recursive, meaning some function application terms do not appear in tail position. To rectify this the CPS translation is refined once more to obtain an uncurried representation of stacks of continuations and handlers. Finally, the translation is made higher-order in order to contract administrative redexes at translation time. The generalised continuation representation is used to construct an abstract machine that provide simultaneous support for deep, shallow, and parameterised effect handlers. kinds of effect handlers. The third strand explores the expressiveness of effect handlers. First, I show that deep, shallow, and parameterised notions of handlers are interdefinable by way of typed macro-expressiveness, which provides a syntactic notion of expressiveness that affirms the existence of encodings between handlers, but it provides no information about the computational content of the encodings. Second, using the semantic notion of expressiveness I show that for a class of programs a programming language with first-class control (e.g. effect handlers) admits asymptotically faster implementations than possible in a language without first-class control.