Safe: A Semantic Technique for Transforming Programs in the Presence of Errors

Abstract

Language designers and implementors have avoided specifying and preserving the meaning of programs that produce errors. This is apparently because being forced to preserve error behavior limits severely the scope of program optimization, even for correct programs. However, error behavior preservation is desirable for debugging, and error behavior must be preserved in any language that permits user-generated errors 1995. This article presents a technique for expressing general program transformations for languages that possess a rich collection of distinguishable error values. This is accomplished by defining a higher-order function called Safe, which can be used to annotate those portions of a program that are guaranteed not to produce errors. It is shown that this facilitates the expression of very general program transformations, effectively giving program transformations in a language with many error values the same power and generality as program transformations in a language with only a single error value. Using the semantic properties of Safe, it is possible to provide some useful sufficient conditions for establishing the correctness of transformations in the presence of errors. In particular, a Substitutability theorem is proven, which can be used to justify “in-context” optimizations: transformations that alter the meanings of subexpressions without changing the meaning of the whole program. Finally, the effectiveness of the technique is demonstrated by some examples of its use in an optimizing compiler. © 1995, ACM. All rights reserved.