Kangaroo: Dynamic Fusion of Branch Instructions in a Pipelined Uniprocessor

Abstract

Small pipelined processors are becoming more common as a complement to superscalars in a multi-core chip. However, current uniprocessors offer little in the way of ILP. We present kangaroo, a novel approach to instruction fusion in a pipelined processor. Kangaroo dynamically fuses two adjacent instructions to create a pair that travels through the pipeline as a unit. The instructions re-enter the pipeline as a pair the next time the first instruction is fetched. Unlike in prior art, an instruction, once fused, is not fetched again. Any pair of adjacent instructions can be fused using this technique, including dependent instructions. In this thesis, we focus on fusing branches to prior instructions. We introduce the fusion mechanism and discuss its implementation. Finally, we evaluate its performance with respect to a baseline six-stage pipelined uniprocessor with several benchmarks.