Applying AI Techniques to Program Optimization for Parallel Computers

Abstract

TIIis paper describes an experiment of integrating expert systems technology and advanced compiler optimization teclmiques to the problem of paralIelizing programs for different classes of parallel computers. Our approach to solve the problem is to separate machine features from programming heuristics and organize the program parallelization knowledge in a hierarchical structure which we called heuristic hierarchy. The reasoning mechanism of the program restructuring system utilizes the heuristic hierarchy and features of the program and the target machine to choose appropriate sequences of transformations automatically. Theories and mechanisms for organizing and integrating the parallelism optimization knowledge are discussed. Methodologies for abstracting machine features, data management, and programming parallel computers are presented.