Abstract
To design a good two-level predictor, we found that a low interference among branches and an even utilization of the entries in the pattern table are two key factors. In order to arrive at a balanced design and achieve the above two goals simultaneously, we first introduce the concept of branch pattern locality. Then, a new predictor design, called the Global PAttern Locality predictor (G-PAL), is introduced. The predictor is developed based on pattern locality and employs a cachelike pattern table to keep only those patterns that are referenced most frequently and recently. In this way, not only the interference among branches can be reduced, but the entries in the pattern table can be fully utilized. © 1996, Springer Verlag. All rights reserved.