The evidence based static branch prediction method suggests a correlation between a program's structure and its profile information. Since they believe that similar structure implies similar dynamic behavior they are able to simplify the whole process by not profiling the new programs in the testing set.
During the course of my research, I would like to study the relationship between program structure and program optimization. Instead of comparing program structures to find the appropriate optimization, I want to be able to compare program profiles to do the same. However, the kind of profile information required to decide the suitable optimization is perhaps quite a bit more than the profile information required to compare one program to another. Thus it would be possible to cut down on the expenses of profiling the programs in the testing set. This would greatly benefit the compilation time as obtaining detailed profiling information would be restricted to only the training set of programs. And, the profile information of a new program in the testing set would be sufficient to deem one program similar to another on that basis. In the process, it would be possible to deduce from the results of my experiment whether similar profiles imply similar optimization. Once that factor is determined it would be interesting to go back to the method in their paper and study the performance of the same system without profiling the new incoming program. If a transitive relation is found to exist between program structure, program profile and program optimization then it would be possible to skip the intermediate step of profiling the program and directly compare the program structures for similarity and choose the appropriate optimization.