@techreport{oai:ipsj.ixsq.nii.ac.jp:02002273,
 author = {水野,将成 and 小泉,透 and 前川,隼輝 and 黒木,地球 and 津邑,公暁 and 塩谷,亮太 and Masanari Mizuno and Toru Koizumi and Toshiki Maekawa and Maru Kuroki and Tomoaki Tsumura and Ryota Shioya},
 issue = {37},
 month = {Jun},
 note = {近年のCPUは高性能化の要求に応えるために投機実行の規模が拡大しており，分岐予測ミスが性能に及ぼす影響はかつてなく大きくなっている．こうした状況を踏まえ9年ぶりに開催される「Championship Branch Prediction 2025（CBP2025）」では，従来と異なりプロセッサパイプラインの情報が提供され，それをうまく活用できるかに注目が集まっている．我々がCBP2025に提出した分岐予測器RUNLTSは，TAGE-SCを基盤としつつ，より良い履歴長選択方法の提案及びレジスタ値と分岐方向の相関をとらえる新機構の導入などにより予測精度向上を図った．本稿では，TAGE-SC-Lをはじめとする既存の分岐予測器と我々のRUNLTSを比較し，CBP2025で提供されたフレームワークを用いた評価結果を報告する．, To meet growing performance demands, modern CPUs increasingly rely on large-scale speculative execution, making the cost of branch mispredictions more critical than ever. In light of this situation, the Championship Branch Prediction 2025 (CBP2025) competition will be held for the first time in nine years and provides a full out-of-order pipeline simulation environment. This enables participants to leverage pipeline-level information that was previously unavailable. In this paper, we introduce RUNLTS, a novel branch predictor based on TAGE-SC. RUNLTS employs an improved history length selection scheme and a new mechanism to capture correlations between register values and branch outcomes. We evaluate RUNLTS against existing predictors, including TAGE-SC-L, using the CBP2025 framework, and demonstrate superior prediction accuracy.},
 title = {アウトオブオーダCPU向け分岐予測アルゴリズムRUNLTSの提案とChampionship Branch Prediction 2025フレームワークを用いた評価},
 year = {2025}
}