多面体モデルに基づく並列化コンパイラに置ける制約条件の緩和

Recently, it is general that computer has more than one processor inside and it is important to parallelize program to improve performance. Polyhedral model is one of model used on automatic parallelizer. On this model, dependence analysis, scheduling and parallelization are performed based on algebra. However, conventional scheme can optimize loops only if all dependence inside loop can be accurately represented by algebra. Therefore, only simple loops like matrix operation can be optimized. On this paper, we propose scheme which relax constraint conditions by extracting dependence conservatively, and extend applicable range of polyhedral model. We also discuss about performance of proposed scheme by experiments using benchmark.