Path-based branch prediction using signature analysis

Abstract

In high-performance processors, the accuracy of branch prediction plays a significant role in enhancing computer execution. A new hardware approach is presented in this paper to dynamically predict branch directions using path information. As an execution path contains large information, we compress the large information using a technique based on the linear feedback shift register (LFSR) that is widely used in testing and error correction coding. A modified version of LFSR, called windowed LFSR, is developed to calculate the signature of a path in one cycle. Since the windowed LFSR has a very regular structure, it can be easily implemented. For most of the benchmarks, the proposed prediction scheme shows better prediction accuracy than the pattern-based predictions. (C) 1999 Elsevier Science B.V. All rights reserved.