Browse Prior Art Database

Comprehensive Branch Prediction Mechanism for BC

IP.com Disclosure Number: IPCOM000065124D
Original Publication Date: 1985-Oct-01
Included in the Prior Art Database: 2005-Feb-19

Publishing Venue

IBM

Related People

Authors:
Emma, PG Pomerene, JH Rechtschaffen, RN Sparacio, FJ [+details]

Abstract

The state of the art in branch prediction involves guessing the outcome of a branch based on observations of past history for the branch in question. In a Branch History Table (BHT), the guess is that the branch will do the same thing that it did on its last execution. Thus, the accuracy of the BHT depends on branches remaining constant in both target and action. Since the accuracy of the BHT has been measured in the 80-85% range, this would seem (statistically) to be a fairly good guessing mechanism. However, since there are a large number of branches that are either always taken or never taken (i.e., their predictability is 100%), the predictability of the remaining branches is considerably lower than the group average. The branches having the least amount of predictability in action are conditional BCs (branch on condition).