Simple Driving Techniques

Driving was introduced as a program transformation technique by Valentin Turchin in some papers around 1980. It was intended for the programming language REFAL and used in metasystem transitions based on super compilation. In this paper we present one version of driving for a more conventional lisp-like language. Our aim is to extract a simple notion of driving and show that even in this tamed form it has much of the power of more general notions of driving. Our driving technique may be used to simplify functional programs which use function composition and will often be able to remove intermediate data structures used in computations. A supercompiler (= supervised compiler) is a special program transformation system which, while evaluating the program, supervises its operation and uses the information to reconstruct the program in a more efficient way. A supercompiler may work on expressions with free variables and will then drive it through the interpretation process and produce a tree of states and transitions. The tree is made into a finite graph by generalizing patterns or configurations of expressions. A configuration is an expression without free variables but with ”holes” in which one may insert other expressions. A supercompiler will have a strategy for generalizing configurations (i.e. make the holes bigger or make new holes). In order to guarantee termination the driving algorithm should generalize configurations into a finite set of simpler configurations called basic configurations. The concept of super compilation is due to Valentin Turchin [10], [11]. The work and concepts are closely tied to the language Refal and not easily rephrased for other programming languages. The strategy for generalizing configurations may be viewed as an abstraction function on expressions, or as an upper bound operation when several configurations should be described by a more general configuration. Different strategies will define different sets of possible configurations. To ensure that the driving terminates we need a strategy that does not produce infinitely many configurations for a program. As in abstract interpretation this may be achieved if the set of configurations is finite or has some finite-height property. Our aim is to use driving to optimize functional programs where composition of user-defined functions plays a central role. We want to specialize functions ? This research was partly sponsored by the IT-University in Copenhagen when we know that the argument values are produced by certain other functions. We will use the driving technique and define a simple set of basic configurations. The result is a transformation system which is easy to implement and easy to use.