Counteracting Bank Misprediction in Sliced First-Level Caches

Future processors having sliced memory pipelines will rely on bank prediction to schedule memory instructions to a first-level cache split into banks. In a deeply pipelined processor, even a small bank misprediction rate may degrade performance severely. The goal of this paper is to counteract the bank misprediction penalty, so that in spite of such bank misprediction, performance suffers little. Our contribution is twofold: a new recovery scheme for latency misprediction, and two policies for selectively replicating loads to all banks. The proposals have been evaluated for 4 and 8-way superscalar processors and a wide range of pipeline depths. The best combination of our mechanisms improves IPC of an 8-way baseline processor up to 11%, removing up to two thirds of the bank misprediction penalty.