Exploring value prediction with the EVES predictor

International audienceIn this study we explore the performance limits of value prediction for small value predictors (8KB and 32KB) in the context of a processor assuming a large instruction window (256-entry ROB), a perfect branch predictor, fetching 16 instructions per cycle, an unlimited number of functional units, but a large value misprediction penalty with a complete pipeline flush at commit on a value misprediction This evaluation framework emphasizes two major difficulties that an effective hardware implementation value prediction will face. First the prediction of a value should be used only when the potential performance benefit on a correct prediction outweighs the potential performance loss on a misprediction. Second, value prediction has to be used in the context of an out-of-order execution processor with a large instruction window. In many cases the result of an instruction has to be predicted while one or several occurrences of the same instruction are still progressing speculatively in the pipeline. Many value predictors (FCM predic-tors [7], stride predictors [3]) are using the result(s) of one or more previous occurrence(s) of the instruction to predict the result of the current occurrence. In this case, a prediction cannot be used in the pipeline unless no instruction occurrence is still speculative or all the occurrences present in the pipeline have been predicted with high confidence. Our proposition EVES, for Enhanced VTAGE Enhanced Stride, combines two predictor components which do not use on the result of the last occurrence of the instruction to compute the prediction. We use an enhanced version of the VTAGE predictor [5], E-VTAGE. On VTAGE, the predicted value is the value directly read at prediction time on the predictor tables. Second, we propose a enhanced version of the stride predictor, E-Stride. E-Stride computes the prediction from the last committed occurrence of the instruction and the number of speculative inflight occurrences of the instruction in the pipeline. The prediction flowing out from E-Stride or E-VTAGE is used only when its confidence is high. The major contribution of this study is the algorithm to assign confidence to predictions depending on the expected benefit/loss of a * This work was partially supported by an Intel research grant prediction. For the predictor components, the predictor entry allocation and victim selection is also guided by this expected benefit/loss. The confidence/priority assignment algorithms use probabilistic counters defined by Riley et al. [6]. On the distributed traces, the EVES predictors with respectively 8KB, 32KB and unlimited storage budgets achieve respectively 4.026 IPC, 4.202 IPC and 4.408 IPC (geometric mean), i.e., respectively 25.3 %, 30.8 % and 37.3 % improvement over the 3.211 IPC achieved without value prediction. Most of this benefit is brought by the 48-entry E-Stride predictor with 16.1 % speedup