EXACT: explicit dynamic-branch prediction with active updates

Abstract

Branches that depend directly or indirectly on load instructions are a leading cause of mispredictions by state-of-the-art branch predictors. For a branch of this type, there is a unique dynamic instance of the branch for each unique combination of producer-load addresses. Based on this definition, a study of mispredictions reveals two related problems: (i) Global branch history often fails to distinguish between different dynamic branches. In this case, the predictor is unable to specialize predictions for different dynamic branches, causing mispredictions if their outcomes differ. Ideally, the remedy is to predict a dynamic branch using its program counter (PC) and the addresses of its producer loads, since this context uniquely identifies the dynamic branch. We call this context the identity, or ID, of the dynamic branch. In general, producer loads are unlikely to have generated their addresses when the dynamic branch is fetched. We show that the ID of a distant retired branch in the global branch stream combined with recent global branch history, is effective context for predicting the current branch. (ii) Fixing the first problem exposes another problem. A store to an address on which a dynamic branch depends may flip its outcome when it is next encountered. With conventional passive updates, the branch suffers a misprediction before the predictor is retrained. We propose that stores to the memory addresses on which a dynamic branch depends, directly update its prediction in the predictor. This novel "active update" concept avoids mispredictions that are otherwise incurred by conventional passive training. We highlight two practical features that enable large EXACT predictors: the prediction path is scalably pipelinable by virtue of its decoupled indexing strategy, and active updates are tolerant of 100s of cycles of latency making it ideal for virtualizing this component in the general-purpose memory hierarchy. We also present a compact form of the predictor that caches only dynamic instances of a static branch that differ from its overall bias.