In addition, knowledge about the characteristics of nonzero strides may motivate future explorations of the potential for hybrid predictors, which combine both the last-value predictor and the stride predictor.
The value predictor chosen to compete the address predictor is the last-value predictor, since it has gained the best prediction accuracy for load instructions.
Figure 12 shows the ILP that can be gained by employing the last-value predictor in comparison to a machine that does not employ value prediction.
The gain of ILP available with value prediction is examined for two different value predictors, the last-value predictor and the stride predictor.
Figure 16 also illustrates that the stride predictor significantly accomplishes better performance than the last-value predictor in those benchmarks which exhibited stride value predictability (like m88ksim and vortex) in our previous experiments.
Another interesting observation shown by these experiments is that the eager mode barely improves the ILP that the last-value predictor achieves in all the benchmarks.
2) Both the last-value predictor and the stride predictor gained relatively small prediction accuracy in floating-point instructions which may also affect their achievable performance.
The last-value predictor increases the ILP of this benchmark from 28 to 55 while the same predictor operating in scalar mode gains only 42.
by employing a small prediction table for the stride predictor and a bigger table for the last-value predictor and only allowing value predictions based on strides to be generated in eager mode.