Global Context-Based Value Prediction

Various methods for value prediction have been proposed to overcome the limits imposed by data dependencies within programs. Using a value prediction scheme, an instruction's computed value is predicted during its fetch stage and forwarded to all dependent instructions to exploit parallelism. Value prediction schemes have focused on predicting the values generated by an instruction based on a local context related to the same instruction. These schemes predict instruction values by analyzing values generated by the instruction during prior executions. This paper proposes techniques that predict values of an instruction based on a global context where the behavior of other instructions is also used in prediction. The global context considers the path along which an instruction is executed and the values computed by other previously completed instructions. We rst present techniques that augment conventional last value and stride predictors with path information to improve prediction accuracy. We then propose a technique that predicts instructions by capturing the correlations between diierent instructions. Experiments performed using path-based techniques and for realistic table sizes resulted in an increase in prediction of 7-8.9% over the current prediction schemes. Prediction using correlations of instructions not predictable using path-based techniques suggests a further improvement of 10.6% accuracy over the best path-based predictor. 1 Introduction The recent advancements in processor technology and microarchitecture research has led to the design of modern architectures capable of executing instructions in parallel. The eeciency of these architectures is highly dependent on the amount of parallelism extracted both by the compiler and the hardware. Besides the hardware limitations, the barriers to parallelism are the dependencies existing between instructions of a program. These dependencies are classiied as 1) control dependencies, arising from the control ow of the program, and 2) data dependencies, which are constraints from the ow of data between statements. One of the techniques to achieve high parallelism is to avoid the limitations imposed by control dependencies by predicting the control ow and executing statements speculatively. By predicting the outcome of branches, the instruction stream pipeline can be prevented from stalling. Branch prediction has been shown to be highly successful in predicting the outcome of branches and avoiding instruction pipeline stalls. Current branch predictors have been shown to accurately predict more than 98% of the branches 10] and even approach similar numbers for branch intensive benchmarks 17].