A safe basis for automatic loop parallelization is the polyhedron model, which represents the iteration domain of a loop nest as a polyhedron in an n-dimensional integer lattice. However, turning the parallel loop program in the model to efficient code meets with several obstacles, due to which performance may deteriorate seriously - especially on distributed memory architectures. This thesis introduces a fine-grained model of the computation performed and shows how this model can be applied to create efficient code.