Method for improving prediction accuracy of branch direction predictor

Abstract

The invention discloses a method for improving prediction accuracy of a branch direction predictor, and belongs to the field of branch predictor design of a processor. The branch direction predictor is additionally provided with a branch update backup device on the basis of the composition of a traditional branch direction predictor, and the branch update backup device comprises n paths of backup tables, an HIT check module, a CTR update module and a backup address selector; the backup table, the HIT checking module and the CTR updating module are connected in sequence, and the backup table, the HIT checking module and the CTR updating module are all connected with the backup address selector. Compared with an existing method of adding an IUM module, more historical directions of the branch instruction are considered as much as possible, so that the prediction accuracy of the branch direction predictor is more effectively improved, experiments prove that the repair degree of the prediction accuracy of the TAGE predictor exceeds 80%, and the problem of delayed updating of the TAGE predictor is effectively solved.

Description

technical field [0001] The invention relates to a method for improving the prediction accuracy of a branch direction predictor, belonging to the field of processor branch predictor design. Background technique [0002] For modern processors, the prediction accuracy of the branch predictor seriously affects the performance of the processor. [0003] There are two types of branch predictors: branch direction predictor and branch target predictor. The branch direction predictor can predict the direction of the branch instruction (jump or not), and the branch target predictor can predict the jump address of the branch instruction. Make predictions. At present, mainstream branch direction predictors include Bimode, GShare, and TAGE-based, etc., and they are all predictors based on multi-bit saturation counters. [0004] As we all know, there are three stages in the pipeline of the processor that can update the saturation counter, which are the instruction fetching stage, the ex...

AI Technical Summary

Problems solved by technology

However, when dealing with the delayed update problem, the IUM module simply takes the last historical direction of the branch instruction B as the predicted direction of the branch instruction B this time, without considering more historical directions of the branch instruction B. Therefore, the IUM module The fix for the loss of prediction accuracy caused by delayed updates is limited, about 3 / 4

Images