Branch prediction using branch correlation in self history-based branch predictors

Abstract

One of the main obstacles in designing of today`s wide-issue, deeply-pipelined superscalar processors is the conditional branch instruction. This is because the outcome of a conditional branch instruction is resolved in the late stage of a pipeline, which results in wasted processor cycles due to pipeline stalls. Various approaches have been proposed to solve the branch problem, including dynamic branch prediction that is probably the most popular one, because it gives good performance and can be implemented without any modification of existing binary programs. Dynamic branch prediction predicts the outcome of a conditional branch instruction using branch history information collected at run-time and allows instructions from predicted path to be executed speculatively while the conditional branch instruction is being resolved. If a misprediction occurs, the speculatively executed instructions are thrown away and instruction fetch restarts from the correct path. Modern superscalar processors tend to have wider issue width and deeper pipeline depth to exploit more and more of the instruction level parallelism (ILP). In these superscalar processors, the performance degradation due to a misprediction becomes more severe, since the amount of speculative work that must be thrown away in the event of a misprediction increases significantly as the issue width and the pipeline depth increase. Therefore, wide-issue, deeply-pipelined modern superscalar processors require very accurate branch prediction to fully exploit their potential performance. Two-level branch predictors, introduced by Yeh & Patt in 1991, are generally considered to be good solutions to achieve the desired high prediction accuracy. Actually, novel variations of two-level branch predictors have been adopted by various commercial microprocessors, such as Intel Pentium-Pro and DEC Alpha 21264. Two-level branch predictors can be classified broadly as either global history-based branch predictors or self...