| This research explores an idea of how to improve branch prediction with wrong-path analysis. The usefulness of data collected on the wrong-path branches are explored, and analyzed, and a new technique of how to use this data to train a branch predictor is proposed. Mis-predicted branches are often assumed to be useless and detrimental because they consume computer resources, so they are normally discarded upon detection. When blindly thrown out, useful information could be lost that may have been useful later on the correct path. Related papers offer suggestions to use longer history lengths, anti-aliasing schemes, instruction pre-fetching, or larger hardware caches to improve performance. By incorporating what is learned on the wrong-path, a hybrid branch predictor can make more-accurate branch predictions when using both paths as training data. The experiment results do show that a simple hybrid predictor can make more accurate predictions, but requires additional resources. |
