An effective automated approach to specialization of code

Abstract. Application performance is heavily dependent on the com-piler optimizations. Modern compilers rely largely on the information made available to them at the time of compilation. In this regard, spe-cializing the code according to input values is an effective way to com-municate necessary information to the compiler. However, the static specialization suffers from possible code explosion and dynamic specialization requires runtime compilation activities that may degrade the overall performance of the application. This article proposes an automated approach for specializing code that is able to address both the problems of code size increase and the over-head of runtime activities. We first obtain optimized code through spe-cialization performed at static compile time and then generate a template that can work for a large set of values through runtime specialization. Our experiments show significant improvement for different SPEC benchmarks on Itanium-II(IA-64) and Pentium-IV processors using icc and gcc compilers.