An Automatic Partial Evaluator for Full Prolog

A partial evaluator for Prolog takes a program and a query and returns a program specialized for all instances of that query. The intention is that the generated program will execute more efficiently than the original one for those instances. This thesis presents “Mixtus”, an automatic partial evaluator for full Prolog, i.e. including non-logical features such as cut, assert, var, write etc. Mixtus does not need any annotations to guide it in its process, but will automatically generate a program given only an input program and a query. A definition of partial evaluation based on the procedural semantics rather than declarative semantics is presented. A program transformed by Mixtus will behave identically as the original one, including side-effects and the order of the solutions returned. The three basic program transformations “unfolding”, “folding” and “definition” are adapted to full Prolog. The most important transformation is “unfolding” which is used by all partial evaluators to eliminate run-time procedure calls. The “folding” and “definition” transformations generate new recursive predicates specialized for the query. A loop prevention mechanism specifically tailored for partial evaluation together with a mechanism for “generalized restart”, will guarantee the termination of the partial evaluation for all programs. Special attention is given to built-in predicates, the cut in particular. To facilitate the execution and removal of cuts during partial evaluation a module for determinacy analysis is used. As negation in Prolog is based on the cut, it will also be treated correctly. Some of the other predicates that are given extra care are dif/2, findall/3 and bagof/3. Many program transformations possible in a logic programming framework will not work correctly for full Prolog. Built-in and generated predicates are therefore classified according to their declarative “purity” so that the “impure” predicates will not hinder the optimization process unnecessarily. It is a non-trivial task to implement a partial evaluator, so the basic features of the implementation of Mixtus are also presented. A number of sample programs are partially evaluated, and the effect is discussed.