Program Type Recognition for Compiler Optimization

Today’s compilers have too many optimization options, and it is difficult to understand all the details and the interactions between them. Therefore, many application developers simply use the well-known compiler flags (such as -O2, -Os) to compile all their programs; at the other end of the extreme, some researchers focus on customizing the optimizations for each program. The former method is too naive as it treats all kinds of programs in the same way, and some programs might even have performance loss since the compilation options are in fact tuned for the others. The latter approach takes care of the variations of programs and tries to individually tune them to the best. However, since the optimization space is extremely huge, the scalability of this scheme is a problem and also makes this approach less popular for the emerging dynamic optimizing compilers. We attack this problem by exploiting two insights: first, it is feasible to categorize the programs into different classes according to their functionalities and characteristics; second, dynamic profiling information is useful for recognizing the type of a program. With both insights, we present a novel approach for program optimization: first, the programs are categorized into different classes and the class-level optimization decision is made by the domain experts; then, the class of an input program is recognized by the machine learning models and the optimal class-level optimizations are applied to obtain the best performance. We demonstrate that this approach can be applied to a binary translation/optimization system and great performance improvement can be achieved.