Global software pipelining with iteration preselection

Abstract

Software pipelining loops containing multiple paths is a very difficult problem. Loop shifting offers the possibility of a close to optimal schedule with acceptable code growth. Deciding how often to shift each operation is difficult, and existing heuristics are rather ad hoc. We separate loop shifting from scheduling, and present new, non-greedy heuristics. Experimental results showtha t our approach yields better performance and less code growth.