Current techniques do not take advantage of information regarding high-level program structures to improve branch prediction accuracy. The existing testing tools are inadequate in discovering these relationships. A new testing tool is thus developed to generate test programs. It allows the high-level program characteristics of these test programs to be varied. The tests conducted uncover the reasons behind the success and failure of different branch predictors, with the Gap and the path-based predictors chosen as examples for illustration. They also suggest ways the compiler can assist in improving the prediction accuracy. A new prediction scheme that takes advantage of these findings is introduced.