Automatic static feature generation for compiler optimization problems

Modern compilers have many optimization passes which help to get a better binary code for a given program. These optimizations are NP-hard. People use different heuristics to get a near optimal solution. These heuristics are designed by a compiler expert after examining sample programs. This is a challenging task. Recently, people have used machine learning techniques instead of heuristics for compiler optimizations. Machine learning techniques have not only eliminated the human efforts but have also out-performed human made huristics. However, the human efforts have now been moved from creating heuristics to selecting good features. Selecting right set of features is important for machine learning techniques since no machine learning tool will work well with poorly choosen features. This paper introduces a noval approach to generate features for machine learning for compiler optimization problems with out any human involvement.