Efficient evaluation of normal logic programs

An accident of implementation may be responsible for the fact that Logic Programming, Deductive Databases and Non-Monotonic Reasoning are different subfields. Logic Programming views logic as a programming language--usually through Prolog or an extension of Prolog. The Deductive Database community regards logic as a database language, often using a variant of magic sets as a basis for implementation. Finally the field of Non-Monotonic Reasoning studies non-classical logics of interest to Artificial Intelligence or other applications. However, there are currently few engines powerful enough to execute practical programs in Non-Monotonic Reasoning. Tabling methods have been formulated recently that have the potential to unify these subfields. This thesis explores how to efficiently implement one such method: SLG. Previous formulations of SLG do not make explicit the search strategy for an evaluation, or consider trade-offs between alternate search strategies. To bring out these features, an operational semantics for SLG, SLG$\sb{O}$, is defined which makes explicit algorithms for completion and other operations. This formalism allows proofs of correctness of algorithms upon which the SLG-WAM is based, as well as proofs of termination for programs of bounded-term size. The modelling of search strategy is powerful enough to derive preliminary results for the cut/O operator as extended to SLG, and for combining SLG and SLDNF using Existential Negation. Based on the SLG$\sb{O}$ search strategy, Part 2 describes the SLG-WAM, an engine which evaluates SLG for stationary stratified programs, and integrates SLG with full Prolog functionality, including the cut, findall, and meta-predicates. Data structures and instructions of the SLG-WAM are described in detail. The SLG-WAM is in fact the engine for the XSB system, and has been installed in hundreds of sites around the world. Part three analyzes the SLG-WAM and compares its performance with the WAM and with other deductive databases. When executing queries using SLD, the overhead of SLG to the WAM is about 10-15% on some standard benchmarks. At the same time, the SLG-WAM is an order of magnitude faster than current deductive databases for in-memory queries. These results are analyzed in detail, and areas for further optimization are identified.