Compiler Optimizations for Massively ParallelMachines : Transformations on Iterative

This paper presents a set of compiler optimizations and their application strategies for a common class of data parallel loop nests. The arrays updated in the body of the loop nests are assumed to be partitioned into blocks (rectangular, rows, or columns) where each block is assigned to a processor. These optimizations are demonstrated in the context of a FORTRAN-90 compiler with very encouraging preliminary results. In the case of solving tridiagonal systems by Gaussian Elimination, the performance of the optimized native code is two orders of magnitude better than the CM-FORTRAN compiler and approaching that of the hand-written Connection Machine Scienti c Library (CMSSL) routine.