Resource usage models for instruction scheduling

Abstract

In order to perform instruction scheduling efficiently, the scheduler needs to maintain the resource usage information in a compact and efficient form. Various resource usage models which facilitate the efficient detection of resource conflicts have been proposed in the literature. We propose two new resource usage models called group automaton and dynamic collision matrix. The group automaton is based on the finite state automaton approach, but exploits certain symmetry in the constructed automaton to reduce its size. The group automaton approach is further benefited by two associated techniques that facilitate precise and efficient modeling of resource usages. The dynamic collision matrix approach proposes the use of collision matrices rather than reservation tables to identify resource conflicts in the scheduling instruction. We report a quantitative comparison of these two resource usage models with three existing models for two target architectures, namely MIPS R8000 and the Cydra VLIW processor. We classify the various models based on the level of abstraction used by them (in modeling resource usage), their space, and their time requirements in scheduling instructions. Our classification gives the perspective that the different resource models are, in fact, a continuum of points in a spectrum of resource models