Mpp Machine Code Generation Technology

A automatic parallelizer plays an essential role in multiprocess parallel systems. This thesis discusses the key technologys of code generation on MPP machines ,which are implemented in AUTOPAR—an automatic parallelizer of FORTRAN on the distributed memory machines.Data dependence analysis is the basis of the parallelization. AUTOPAR parallilize the loop without loop-carried dependence,AUTOPAR has the best loop shedule algorithm to balance the processors' load and reduce the loop shedule cost. Loop restructure is of very importance in AUTOPAR: eliminating the loop-carried dependence, reducing the communication overhead. A new method is proposed to evaluate whether a loop be parallelized.Communication optimization is crucial to the automatic parallelizer. In this paper ,we introduce some advanced optimizations which indude message vectorization, message aggregation, message pipelining, etc. In AUTOPAR, we adapt, integrate,and extend these optimizations efficiently.