Improved composite confidence mechanisms for a perceptron branch predictor