Experiment flows and microbenchmarks for reverse engineering of branch predictor structures

摘要

Insights into branch predictor organization and operation can be used in architecture-aware compiler optimizations to improve program performance. Unfortunately, such details are rarely publicly disclosed. In this paper we introduce a set of experiment flows and corresponding microbenchmarks for reverse engineering cache-like branch target and outcome predictor structures, indexed by branch address or program path information. The experiment flows are demonstrated on the Intel Pentium M branch predictor. We have been able to determine the size, organization, internal operation, and interactions between various hardware structures used in the Pentium M branch predictor, namely the branch target buffer, indirect branch target buffer, loop branch predictor buffer, global predictor, and bimodal predictor. These findings have been validated using a functional PIN model.