Exploration of Highly Accurate Path Prediction Mechanism Using Detailed Path History

摘要

Prediction mechanism is the key to achieve high performance using speculative multithreading. High prediction accuracy can be expected to improve the effectiveness of speculative multithreading. In this research, we focus on the frequency of program paths in loops. To identify the multiple paths in loops, we introduce extensive history information. The N-bit history information represents 2 to the Nth power minus 1 paths, and the predictor selects a path for the next thread from these paths. We show that the proposed path prediction mechanism increases the prediction accuracy by up to 7.6 percent compared to the conventional prediction mechanism using 1-bit history information.