Testing Completeness of Code Selector Specifications

1 I n t r o d u c t i o n The tree pat tern matching approach for code selection has proven very successful. I t allows to specify code selectors using declarative specifications and to derive efficient code selectors automatically. The program which translates the specification into the final code selector program is called code selector generator. Sometimes it is integrated with generators for other par ts of a code generator. However this paper deals only with the aspect of code selection. The specifications are essentially rule based. The user only writes rules specifying how code can can be selected but not the exact way the rules have to be applied in order to get good code. This is determined automatical ly by the code generator generator system. The user annotates the rules with cost values and the system selects good code according to this cost measure. The rule based approach has a lot of advantages which make it much superior to hand programming. However there is a small disadvantage: During hand programming one is usually forced to consider all possible constructs of the intermediate code as inputs for the code selector. When writing rules this is no longer the case. I t is therefore much more likely to forget a rule for a certain construct of the intermediate code. This might show up as an error, only if the corresponding construct is used in a very special context. This paper presents a method to detect statically if there are rules missing and if so to produce an appropriate error message. With such a test a lgori thm at hand * This work was carried out within the Esprit Project #5399 (COMPARE)