The weakening of branch predictor performance as an inevitable side effect of exploiting control independence

Many algorithms are inherently sequential and hard to explicitly parallelize. Cores designed to aggressively handle these problems exhibit deeper pipelines and wider fetch widths to exploit instruction-level parallelism via out-of-order execution. As these parameters increase, so does the amount of instructions fetched along an incorrect path when a branch is mispredicted. Many of the instructions squashed after a branch are control independent, meaning they will be fetched regardless of whether the candidate branch is taken or not. There has been much research in retaining these control independent instructions on misprediction of the candidate branch. This research shows that there is potential for exploiting control independence since under favorable circumstances many benchmarks can exhibit 30% or more speedup. Though these control independent processors are meant to lessen the damage of misprediction, an inherent side-effect of fetching out of order, branch weakening, keeps realized speedup from reaching its potential. This thesis introduces, formally defines, and identifies the types of branch weakening. Useful information is provided to develop techniques that may reduce weakening. A classification is provided that measures each type of weakening to help better determine potential speedup of control independence processors. Experimentation shows that certain applications suffer greatly from weakening. Total branch mispredictions increase by 30% in several cases. Analysis has revealed two broad causes of weakening: changes in branch predictor update times and changes in the outcome history used by branch predictors. Each of these broad causes are classified into more specific causes, one of which is due to the loss of nearby correlation data and cannot be avoided. The classification technique presented in this study measures that 45% of the weakening in the selected SPEC CPU 2000 benchmarks are of this type while 40% involve other changes in outcome history. The remaining 15% is caused by changes in predictor update times. In applying fundamental techniques that reduce weakening, the Control Independence Aware Branch Predictor is developed. This predictor reduces weakening for the majority of chosen benchmarks. In doing so, a control independence processor, snipper, to attain significantly higher speedup for 10 out of 15 studied benchmarks