An Experiment in Abstract Machine Design

SUMMARY In this article we present Typed Feature Structures as an extension of Prolog, and show how to come up with a compilation scheme and an abstract machine using a design methodology based on partial evaluation. First we define the transformations used by our partial evaluator. Then we present the design methodology which we will use later. Next, we clarify the notion of Typed Feature Structures that underlies our work, and formally define the unification of such structures. Based on this definition, we develop a unification procedure with explicit heap representation. By partially evaluating this procedure with respect to some example programs, we show how to come up with the machine instructions and translation schemes. Finally, we briefly address coreferences, cyclic structures and the unification of types. INTRODUCTION The more programming languages are freed from the 'Von-Neumann style', the more important are abstract machines. Abstract machines provide an intermediate target language for compilation. First, the compiler generates code for the abstract machine, then this code can be further compiled into real machine code or it can be interpreted. By dividing compilation into two stages, abstract machines increase the portability and maintainability of compilers. For compiling Prolog, and as a basis for compilers of other logic programming languages, the Warren Abstract Machine (WAM) is widely used. Peter Kursawe 1 shows how some of the WAM instructions for unification , originally invented by Warren, 2 can be reinvented by using partial evaluation. Later, Nilsson 3,4 derived some of the control instructions of the WAM in a similar way. To our knowledge, we are the first to use this methodology to 'invent' WAM-like instructions, not 'reinvent' WAM instructions. Moreover, the methodology both guides the design and relates source language constructs to machine instructions. Although the methodology has only been applied in the context of logic programming, we are convinced that it can be applied to other languages, too. For example, it is not difficult to design some of the P-machine instructions by applying the methodology to an interpreter for arithmetic expressions. In the following sections we apply the methodology to come up with machine instructions for Typed Feature Structures. These are an extension of first-order terms in Prolog. Basically, we allow structures and predicates to have typed and labelled