针对难预测分支设计专用预测器，显著降低处理器误判率。

Branch prediction is key to the performance of out-of-order processors. While the CBP-2016 winner TAGE-SC-L combines geometric-history tables, a statistical corrector, and a loop predictor, over half of its remaining mispredictions stem from a small set of hard-to-predict (H2P) branches. These branches occur under diverse global histories, causing repeated thrashing in TAGE and eviction before usefulness counters can mature. Prior work shows that simply enlarging the tables offers only marginal improvement. We augment a 159 KB TAGE-SC-L predictor with a 28 KB H2P-targeted subsystem called the Bullseye predictor. It identifies problematic PCs using a set-associative H2P Identification Table (HIT) and steers them to one of two branch-specific perceptrons, one indexed by hashed local history and the other by folded global history. A short trial phase tracks head-to-head accuracy in an H2P cache. A branch becomes perceptron-resident only if the perceptron's sustained accuracy and output magnitude exceed dynamic thresholds, after which TAGE updates for that PC are suppressed to reduce pollution. The HIT, cache, and perceptron operate fully in parallel with TAGE-SC-L, providing higher fidelity on the H2P tail. This achieves an average MPKI of 3.4045 and CycWpPKI of 145.09.