We study the equivalence between eval-readback and eval-apply big-step evaluators in the general setting of the pure lambda calculus. We study `one-step' equivalence (same strategy) and also discuss `big-step' equivalence (same final result). One-step equivalence extends for free to evaluators in other settings (calculi, programming languages, proof assistants, etc.) by restricting the terms (closed, convergent) while maintaining the strategy. We present a proof that one-step equivalence holds when (1) the `readback' stage satisfies straightforward well-formedness provisos, (2) the `eval' stage implements a `uniform' strategy, and (3) the eval-apply evaluator implements a `balanced hybrid' strategy that has `eval' as a subsidiary strategy. The proof proceeds the `lightweight fusion by fixed-point promotion' program transformation on evaluator implementations to fuse readback and eval into the balanced hybrid. The proof can be followed with no previous knowledge of the transformation. We use Haskell 2010 as the implementation language, with all evaluators written in monadic style to guarantee semantics (strategy) preservation, but the choice of implementation language is immaterial. To illustrate the large scope of the equivalence, we provide an extensive survey of the strategy space using canonical eval-apply evaluators in code and big-step `natural' operational semantics. We discuss the strategies' properties, some of their uses, and their abstract machines. We improve the formal definition of uniform and hybrid strategy, use it to structure the strategy space, and to obtain generic higher-order evaluators which are used in the equivalence proof. We introduce a systematic notation for both evaluator styles and use it to summarise strategy and evaluator equivalences, including (non-)equivalences within a style and (non-)equivalences between styles not proven by the transformation.

Abstract We exploit the idea of proving properties of an abstract machine by using a corresponding semantic artefact better suited to their proof. The abstract machine is an improved version of Pierre Crégut’s full-reducing Krivine machine KN. The original version works with closed terms of the pure lambda calculus with de Bruijn indices. The improved version reduces in similar fashion but works on closures where terms may be open. The corresponding semantic artefact is a structural operational semantics of a calculus of closures whose reduction relation is purposely a reduction strategy. As shown in previous work, improved KN and the structural operational semantics ‘correspond’, i.e. both artefacts realise the same reduction strategy. In this paper, we prove in the calculus of closures that the reduction strategy simulates in lockstep (at every reduction step) the complete and standard normal-order strategy (i.e. leftmost reduction to normal form) of the pure lambda calculus. The simulation is witnessed by a substitution function from closures of the closure calculus to pure terms of the pure lambda calculus. Thus, KN also simulates normal-order in lockstep by the correspondence. This result is stronger than the known proof that KN is complete, for in the pure lambda calculus there are complete but non-standard strategies. The lockstep simulation proof consists of straightforward structural inductions, thanks to three properties of the closure calculus we call ‘index alignment’, ‘parameters-as-levels’ and ‘balanced derivations’. The first two come from KN. Thanks to these properties, a proof in a calculus of closures involving de Bruijn indices and de Bruijn levels is unproblematic. There is no lexical adjustment at binding lookup, on-the-fly alpha-conversion or recursive traversals of the term to deal with bound and free variables as in other calculi. This paper contributes to the framework for environment machines of Biernacka and Danvy a full-reducing open-terms closure calculus, its corresponding abstract machine, and a lockstep simulation proof via a substitution function.

Siek and Garcia (2012) have explored the dynamic semantics of the gradually-typed lambda calculus by means of definitional interpreters and abstract machines. The correspondence between the calculus's mathematically described small-step reduction semantics and the implemented big-step definitional interpreters was left as a conjecture. We prove and generalise Siek and Garcia's conjectures using program transformation. We establish the correspondence between the definitional interpreters and the reduction semantics of a closure-converted gradually-typed lambda calculus that unifies and amends various versions of the calculus. We use a layered approach and two-level continuation-passing style so that the correspondence is parametric on the subsidiary coercion calculus. We have implemented the whole derivation for the eager error-detection policy and the downcast blame-tracking strategy. The correspondence can be established for other choices of error-detection policies and blame-tracking strategies, by plugging in the appropriate artefacts for the particular subsidiary coercion calculus.

Olivier Danvy and others have shown the syntactic correspondence between reduction semantics (a small-step semantics) and abstract machines, as well as the functional correspondence between reduction-free normalisers (a big-step semantics) and abstract machines. The correspondences are established by program transformation (so-called interderivation) techniques. A reduction semantics and a reduction-free normaliser are interderivable when the abstract machine obtained from them is the same. However, the correspondences fail when the underlying reduction strategy is hybrid, i.e., relies on another sub-strategy. Hybridisation is an essential structural property of full-reducing and complete strategies. Hybridisation is unproblematic in the functional correspondence. But in the syntactic correspondence the refocusing and inlining-of-iterate-function steps become context sensitive, preventing the refunctionalisation of the abstract machine. We show how to solve the problem and showcase the interderivation of normalisers for normal order, the standard, full-reducing and complete strategy of the pure lambda calculus. Our solution makes it possible to interderive, rather than contrive, full-reducing abstract machines. As expected, the machine we obtain is a variant of Pierre Crégut's full Krivine machine KN.

We derive by program transformation Pierre Crégut's full-reducing Krivine machine KN from the structural operational semantics of the normal order reduction strategy in a closure-converted pure lambda calculus. We thus establish the correspondence between the strategy and the machine, and showcase our technique for deriving full-reducing abstract machines. Actually, the machine we obtain is a slightly optimised version that can work with open terms and may be used in implementations of proof assistants.