Function level parallelism driven by data dependencies

摘要: With the rise of Chip multiprocessors (CMPs), amount parallel computing power will increase significantly in near future. However, most programs are sequential nature and have not been explicitly parallelized, so they cannot exploit these resources. Automatic parallelization sequential, non-regular codes is very hard, as illustrated by lack solutions after more than 30 years research on topic. The question remains if there parallelism that can be detected automatically so, how much is.In this paper, we propose a framework for extracting potential from programs. Applying to teach us present program, but also tells what appropriate construct program is, e.g. pipeline, master/slave work distribution, etc.Our profile-based, implying it safe. It builds two new graph representations profile-data: interprocedural data flow sharing graph. This graphs show data-flow between functions structures facilitating data-flow, respectively.We apply our SPECcpu2000 bzip2 benchmark, achieving speedup 3.74 compression part global 2.45 quad processor system.