Machine Knowledge Maniputation Issues in Parallel Compilers

Parallel compilers need detail architectural knowledge about the target machines to optimize user programs. The knowledge is needed in order to realize the potential parallelism provided by the hardware and to match the program parallelism with the machine parallelism. In most parallel compilers, the architectural infonnation of the target machine are blurred into the control structures of the compilers. Consequently, these parallel compilers are inflexible and substantial efforts are needed to modify them or to port them to different machines. A solution to this problem is to separate the hardware features from the knowledge for the program optimization and describe the later based on these features. In this way, the control of parallel compilers relies explicitly on the computational model that is described by machine features rather than the hard-coded heuristics. High degree of flexibility. portability and knowledge sharing can be achieved among different target machines. In this paper. a parallel machine knowledge representation scheme that features hierarchical structure and object oriented approach is presented. Under the scheme, machine features are represented as objects and are organized into hierarchical structures based on relationship between the feature objects. An abstraction process which translates basic machine features into different levels of abstraction is presented. Optimizing compilers can select the levels that suit its objectives and tasks most to work with. The parallel machine knowledge manipulation system is implemented in Prolog and includes mechanism for interactive feature specification. feature classification, and support for reasoning based on the features. MACHINE KNOWLEDGE MANIPULATION ISSUES FOR PARALLEL COMPILERS Ko·Yang Wang Department of Computer Sciences, Purdue University, West Lafayellc, IN 47907. Dennis Gannon Department of Computer Science, Indiana University, Bloomington, IN 47405. Piyush Mehrotra leASE, NASA Langley Research Cenler. Hampton, VA 23665.