A Decoupled Fetch-Execute Engine with Static Branch Prediction Support

We describe a method for supporting static branch prediction on a decoupled fetch-execute pipeline. Using instruction buffers to decouple instruction fetch from the execute pipeline is an effective way to minimize instruction cache penalties by allowing instruction fetch and stall miss handling to proceed independent of the execution pipeline. Dynamic branch prediction is typically used with such architectures, but it is not necessary to assume the cost of dynamic branch hardware when static prediction is sufficient. Traditional static branch prediction approaches were designed for lock-step pipelines and do not adapt well to decoupled fetch-execute pipelines, so alternative means of support were required. We describe the requirements for achieving efficient static branch prediction on a decoupled fetch-execute architecture, and presents the design and results for an implementation on an EPIC-style target architecture.