A 64-Kbytes ITTAGE indirect branch predictor∗

The ITTAGE, Indirect Target TAgged GEometric length predictor, was introduced in [5] at the same time as the TAGE conditional branch predictor. ITTAGE relies on the same principles as the TAGE predictor several predictor tables indexed through independent functions of the global branch/path history and the branch address. Like the TAGE predictor, ITTAGE uses (partially) tagged components as the PPM-like predictor [2]. It relies on (partial) match to select the predicted target of an indirect jump. TAGE also uses GEometric history length as the O-GEHL predictor [3], i.e. , the set of used global history lengths forms a geometric series. This allows to efficiently capture correlation on recent branch outcomes as well as on very old branches. Due to the huge storage budget available for the ChampionShip, we propose an ITTAGE predictor featuring 16 prediction tables. On the distributed set of traces, using a path history vector recording only information from indirect jumps and calls was found to be (slightly) more efficient than using a path/branch history vector combining information from all kind of branches. 1 The ITTAGE indirect jump target predictor Building on top of the cascaded predictor [1] and on the TAGE predictor, the ITTAGE predictor was proposed in [4]. In this section, we recall the general principles of the ITTAGE indirect target predictor.equivalent storage budget. Some implementation details from the initial ITTAGE proposition are slightly modified in order to improve the global prediction accuracy. 1.1 ITTAGE predictor principles ITTAGE relies on the same principles as the TAGE predictor several predictor tables indexed through independent functions of the global branch/path history and the branch address. ∗This work was partially supported by the European Research Counsil Advanced Grant DAL The Indirect Target TAgged GEometric length, ITTAGE, predictor features a tagless base predictor T0 in charge of providing a default prediction and a set of (partially) tagged predictor components. The tagged predictor components Ti, 1 ≤ i < M are indexed using different history lengths. The set of history lengths form an increasing series, i.e L(i) = (int)(αi−1 ∗ L(1) + 0.5). This is illustrated in Figure 1. The counters representing predictions in TAGE are replaced by the target addresses Target . A predictor table entry also features a tag, a 2-bit confidence counter Ctr allowing some hysteresis on the predictor and a useful bit U for controlling the update policy ( Figure 2).