Optimizations enabling transformations and code generation for th HP V class

Abstract

It is generally accepted that state-of-the-art restructuring compiler systems can parallelize automatically only about 25% of existing Fortran codes. This thesis presents some code transformations that improve the power of such compilers and thus increase the coverage of automatic parallelization. We present a series of techniques which we have been implemented in the Polaris compiler with the goal of bringing the form of many real Fortran 77 benchmarks to a more 'rational' or normal form. For example, we transform natural loops implemented with goto statements into normalized do loops. Loops with premature exits are transformed in normal exiting do loops. We further implement 'Loop Distribution' in order to separate usually sequential recurrences from otherwise parallel loops. All presented transformations enable other analysis passes of the compiler to detect and exploit parallelism in more code instances than it was previously possible. Finally, we present some of the more challenging and unique issues involved in the implementation of the code generation back-end for the HP V class multiprocessor.