Methodology

To verify validity of branch prediction heuristics (see chapter for description of individual algorithms) we compile program with profile feedback, dump our predictions and real program behavior to the debug output and use our analyze_branches tool to produce summary.

We tested our results on SPECint2000 suite, a representative collection of programs run on nowadays computers. Since we used the same set of programs for specifying weights of individual heuristics, we decided to test SPECfp2000 suite to cross check our predictor on different code.

Programs contained in SPECfp differ dramatically from the ones in SPECint by both purpose and language. Many SPECfp programs solve numerical problems and are written in Fortran, while SPECint programs are written in C and C++.

We tested only subset of SPECfp tests, because some of them are written in f90 Fortran dialect that is not supported by GCC yet.