Improving Dynamic Branch Prediction Accuracy by Variable Cross-Reference Prediction and Iterative Dispatch

Abstract

This paper presents two schemes, the variable cross-reference (VCR) prediction scheme and the iterative dispatch approach, to improve branch prediction accuracy for the two-level adaptive branch predictor. The VCR prediction scheme makes predictions by variably cross-referring traces in the pattern history table (PHT); it achieves desirable prediction accuracy with low time complexity and at no extra hardware cost. The iterative dispatch approach enhances prediction accuracy using the PHT history to do dispatching for an additional layer of pattern history. The proposed VCR scheme is then joined by the optimal PPM algorithm to form a combined predictor which maintains desirable prediction accuracy at reduced cost. Extensive trace-driven simulation runs have been conducted to evaluate the performance of our proposed schemes and other predictors.