Limited run branch prediction

Abstract

A branch prediction technique which increases the likelihood of correctly predicting the direction of a conditional branch instruction is presented. The technique is based on the observation that many branches have run lengths that are constant or slowly-varying. I.e., several consecutive runs of 1's are of the same length. The technique uses the history stored for each branch, which history is enhanced by two small counters, an up counter and a down counter. These counters operate in conjunction with a state machine branch predictor of the prior art for very accurate predictions.

Description

field of invention [0001] The present invention generally relates to a pipeline computer system, and in particular to a branch prediction method adopted by a computer system for predicting the direction of a conditional branch instruction. Background technique [0002] The time required to execute a conditional branch severely limits the performance of a pipelined processor. Processors typically fetch and execute instructions in a sequential fashion; that is, the address of instruction Ei+1 (Ei's successor) to be executed immediately after instruction Ei fetched from address n is found by adding n to the length of Ei . An unconditional branch is an instruction whose execution causes control to be transferred to an instruction at a non-sequential address. Thus a successor of a branch B is fetched from an arbitrary target address. In some computers the target address of branch instruction B is contained in the instruction, while in other computers the target address is form...

AI Technical Summary

Problems solved by technology

So less efficient for shorter runs

Images