From abstract model to efficient compilation of patterns

Abstract

Pattern matching is a technique which is used in many application areas such as text manipulation in editors, manipulation of arithmetic formulas in computer algebra systems and in artificial intelligence applications. In programming systems it can be a built-in language facility as string matching in SNOBOL4 or a language extension for matching of list structures as in INTERLISP.

This paper presents a strategy for achieving a pattern match facility with the following properties (1) it is based on a simple model of pattern matching in order to give a good understanding for the pattern matching process (2) it is powerful, that is, it contains the union of the facilities commonly found in pattern matchers (3) it provides compilation of patterns to efficient code.

Available implementations are oriented towards providing either a clear and extensible definition of pattern matching or towards good performance at execution. The underlying problem is that good optimizing compilers are too complex to also serve as a modifiable definition of the language.

Our strategy is based on the development of a simple formal definition of the problem. Then a pattern match interpreter in Lisp is developed as an extension of the formal definition and thus serves to reach the second goal. Powerful program manipulation tools are then used to transform the interpreter according to the partial evaluation principle. The Lisp code which is the result from partial evaluation is regarded as compiled patterns.

The stepwise development of a suitable model is described and the resulting Lisp program is shown. The partial evaluation of the program with respect to one given pattern is demonstrated. Examples of optimization in the pattern language is shown and similarities between partial evaluation and lazy evaluation are pointed out. Performance results of the full implementation are given. Other approaches to pattern matching systems are briefly described and related to our approach. Finally the effects of partial evaluation on the pattern matcher are characterized and their relative importance is evaluated. The outcome of applying this strategy to similar problems is discussed.