Characterization of data value unpredictability to improve predictability

Recent research has shown that it is possible to overcome the parallelism limits imposed by dataflow by predicting instruction results based on previously produced values or a sequence thereof. Unlike branch prediction schemes where prediction accuracies of 90% and above are the norms, data value prediction schemes have been able to correctly predict only about 40-70% of the result-producing instructions. In order to further improve the performance of data value predictors, it is very important to do a thorough analysis of the instructions that are currently unpredictable. In this paper, we study the characteristics of unpredictability of data values. The major insights obtained from this stud, are: (i) An instruction often becomes unpredictable due to multiple causes. This means that tackling a single cause in isolation will not provide substantial improvements in prediction accuracy. (ii) Among the different causes, control flow fluctuations are a major cause. (iii) Load instructions are a significant component of unpredictable instructions. Future data value predictors need to tackle these causes, and provide "better contexts" so as to convert current unpredictability into predictability.