A Case for a Software-Managed Reconfigurable Branch Predictor

Workloads present diverse performance challenges. Consequently, many proposed microarchitecture techniques apply to only a subset of applications. This limits their appeal, as the added resources tend to be unused by the remaining applications. Designing reconfigurable hardware is one way to tackle this issue. In this paper, we propose EXACT-S, a software-managed reconfigurable branch predictor that targets workloads with hardto-predict load-dependent branches encountered in the context of sequencing arrays and linked-lists. EXACT-S is inspired by the previous EXACT predictor and targets its limitations. EXACT is indexed by branches’ load addresses and actively updated by stores. With EXACT-S, the idea is to have a light-weight run-time layer to convey key information directly to the fetch unit that it can use to generate branches’ load addresses in a timely manner. This approach is more accurate because it uses branches’ load addresses directly rather than prior branches’ load addresses. Moreover, with regard to active updates, there is no need for a large table to convert store addresses to predictor indices because of the direct indexing strategy. As a result, EXACT-S is simpler, less expensive and more accurate than EXACT. For applications that suffer poor prediction accuracies due to load-dependent branches, EXACT-S removes up to 50% of their mispredictions. For all other applications, the proposed reconfigurable predictor relinquishes EXACT-S storage to the base L-TAGE predictor, thus achieving the same prediction accuracy as a similarly-sized fixed L-TAGE predictor.