Methods of Fine-Grain Optimization for Parallel Computer Architectures

摘要

Research concentrating on aspects of compiler optimization techniques foruniprocessor machines that exploit implicit fine grain parallelism at the machine instruction level is presented. The fine grain optimization aims at detecting parallelism between individual machine operations and translating it to the architectural parallelism provided by the hardware of the target machine by scheduling these operations at compile time (termed static scheduling). In this context, a program is considered to consist of a set of atomic machine operations, each of which is carried out by one or more functional units. The parallel architecture processor in the consideration consists of one or more pipelined functional units that cooperate in a parallel fashion. The common hardware feature found in modern parallel computers are found and accommodated in the abstract machine model defined for the purpose of retargetability. This machine model serves as a framework within which new techniques of code parallelization are developed. Attempts are made to find new concepts and more powerful methods of program analysis in order to detect more parallelism; to find good scheduling heuristics; and to develop sophisticated techniques in order to generate efficient parallel code that can take advantage of the parallel resources.