Determining transformation sequences for loop parallelization

Abstract

Considerable research on loop parallelization for shared memory multiprocessors has focused upon developing transformations for removing loop-carried dependences. In many loops, more than one such transformation is required, and hence the choice of transformations and the order in which they are applied is critical. In this paper, we present an algorithm for selecting a sequence of transformations which, applied to a given loop, will yield an equivalent maximally parallel loop. The algorithms provided in the paper have been implemented and tested in PAT, a tool for interactive parallelization of Fortran.