Hierarchical Control Prediction: Support for Aggressive Predication

Predication of control edges has the potential advantages of improving fetch bandwidth and reducing branch mispredictions. However, heavily predicated code in out-of-order processors can lose significant performance by deferring resolution of the predicates until they are executed, whereas in nonpredicated code those control arcs would have remained as branches, and would be resolved immediately in the fetch stage when they are predicted. Although predicate prediction can address this problem, three problems arise when trying to predict aggressively predicated code that contains multi-path hyperblocks: (1) How to maintain a high bandwidth of branch prediction to keep the instruction window full without having the predicate predictions interfere and without increasing the branch mispredictions, (2) how to determine which predicates in the multi-path hyperblocks should be predicted, and (3) how to achieve high predicate prediction accuracies without centralizing all prediction information in a single location. To solve these problems, this paper proposes a speculation architecture called hierarchical control prediction (HCP). In HCP, the control flow speculation is partitioned into two levels. In parallel with the branch predictor, which identifies the coarsegrain execution path by predicting the next hyperblock entry, HPC identifies and predicts the chain of predicate instructions along the predicted path of execution within each hyperblock, using encoded static path approximations in each branch instruction and local per-predicate histories to achieve high accuracies. Using a 16-core composable EDGE processor as the evaluation platform, this study shows that hierarchical control prediction can address these issues comprehensively, accelerating single-threaded execution by 19% compared to no predicate prediction, and thus achieving half of the 38% performance gain that ideal predicate prediction would attain.