Overiding branch predictors have been proposed recently as the means to

correct fast and low accuracy primary predictors. Overiding predictors can

offer better performance because they are larger and or more complex and

therefore suitable for capturing distinct patterns of behavior than

smaller and simpler primary predictors.



However, for different programs, and even during the execution of the same

program, is possible for the primary predictor alone to be sufficient and

using an overiding predictor to be a pleonasm.



If such situations exist, then dynamic energy can be reduced by accessing

and updating selectively the overiding predictor. This project aims to

investigate on the fly activation-deactivation of overiding branch

predictors. Our investigation will consider overiding predictors for branch direction prediction and trace predictors.



To achieve our objectives, the run time behavior of various

combinations of primary-overiding predictors will be investigated and

on-the-fly phase detection algorithms will be developed. Our initial

aim is to establish a correlation between program behavior and

usefulness of overding predictors using phase detection algorithms

proposed in the literature for reconfiguring other types of resources

in a processor.

4500 euros to cover four short visits for technical

exchange. Each trip is expected to cost 1125euros: 225 air+ 100 ground

transpo + 450 lodging + 350 daily allowance