Characterizing the Mispredicted Branches

Studies of branch prediction have tended to emphasize aggregate measurement of prediction performance. Instead, we have taken the approach that there are likely to be a small number of branch addresses (“sites”) in the program causing the majority of all mispredictions. Using a graphical tool we developed for analyzing program traces on-the-fly, we present three interesting results, all based on measurement of benchmarks including a subset of the SPEC92 suite of programs: More than 90% of all misses can be attributed to the top 0.5% of sites causing branch mispredictions. Additionally, the top fifty sites account for more than 86% all mispredictions. For these popular misprediction sites, the distance in dynamic instructions until the next mispredicted site is quite large on average (more than 50-200 instructions), making exotic techniques such as guarded execution seem feasible. Finally, we found that these sites exhibited larger than average miss rates, indicating that these sites are both executed many times and have high miss rates. Our study was limited to correlating, stateful hardware prediction schemes.