Towards a New Perspective on Partial Evaluation

1 Introduction We have designed a new method for performing partial evaluation. This method divides specialization into two phases: a polyvariant analysis phase and a code generation phase. The analysis phase does not diierentiate between unfolding and specializing. Symbolic execution is the sole operation performed during the analysis phase. Symbolic execution of an expression yields a characterization of the value that would be returned by the expression if it were executed at runtime, a description of the residual operation(s) that must be performed to generate the runtime value, and a record of the information utilized in performing symbolic execution. (The third result of symbolic execution will be explained in some detail later.) Only delta-reductions are performed during symbolic execution. No beta-substitution of user functions occurs during the analysis phase and no values are in-lined. Consequently, symbolic execution yields extremely polymorphic, and highly reusable, specialized function bodies. The code generation phase constructs the residual program from the specialized function bodies. It is responsible for all beta-substitution. Delaying the beta-substitution decisions until the code generation phase allows us to construct a partial evaluator that both produces highly specialized code and makes intelligent decisions regarding code size versus lower function call overhead. Termination decisions are made in the analysis phase based on lazy use-analysis, an extension of eager use-analysis 5]. Lazy use-analysis considers both the information used by symbolic execution in performing delta-reductions and how information about the values returned by delta reductions is used. Information is only used in a fundamental sense if there is a causal chain between its use in performing some delta reduction and the production of the nal runtime answer returned by a program. The less information used in creating a specialized function body that contributes to the return value of the function, the greater the number of contexts in which the specialization can be reused. Furthermore , the less information used about the value returned by a function call, the less restrictions that are placed on the