A Research on the On-the-Fly Parallelization by a Dynamic Scalar Expansion

点击翻译按钮获取中文摘要

We propose a memory renaming as a generalization of the scalar expansion techniques which are employed in vectorizing compilers. The memory renaming expands memory area of a variable dynamically according to the number of threads which execute at the same time. This method can remove any anti-dependency and output-dependency among threads and increase the opportunities to extract thread-level parallelism. Our simulation results show that this method removes more than half of hazards caused by the dependencies among threads in most programs which we investigated.