Programs as Data Structures in λSF-Calculus

Abstract

Lambda-SF-calculus can represent programs as closed normal forms. In turn, all closed normal forms are data structures, in the sense that their internal structure is accessible through queries defined in the calculus, even to the point of constructing the Goedel number of a program. Thus, program analysis and optimisation can be performed entirely within the calculus, without requiring any meta-level process of quotation to produce a data structure.Lambda-SF-calculus is a confluent, applicative rewriting system derived from lambda-calculus, and the combinatory SF-calculus. Its superior expressive power relative to lambda-calculus is demonstrated by the ability to decide if two programs are syntactically equal, or to determine if a program uses its input. Indeed, there is no homomorphism of applicative rewriting systems from lambda-SF-calculus to lambda-calculus.Program analysis and optimisation can be illustrated by considering the conversion of a programs to combinators. Traditionally, a program p is interpreted using fixpoint constructions that do not have normal forms, but combinatory techniques can be used to block reduction until the program arguments are given. That is, p is interpreted by a closed normal form M. Then factorisation (by F) adapts the traditional account of lambda-abstraction in combinatory logic to convert M to a combinator N that is equivalent to M in the following two senses. First, N is extensionally equivalent to M where extensional equivalence is defined in terms of eta-reduction. Second, the conversion is an intensional equivalence in that it does not lose any information, and so can be reversed by another definable conversion. Further, the standard optimisations of the conversion process are all definable within lambda-SF-calculus, even those involving free variable analysis.Proofs of all theorems in the paper have been verified using the Coq theorem prover.