Metadata-based Dynamic Optimization

Dynamic Optimizations could manipulate the binary codes dynamically according to the runtime environments. The binary codes could run effectively in various environments. Moreover, Dynamic Optimization Systems could capture and monitor every single detail for the running programs, which could be provided to satisfy users' requirements.Simulators, Dynamic Binary Translators and Dynamic Instrumentation Tools are all the typical applications to Dynamic Optimizations. Program behaviors are captured and reported by these dynamic systems according to the requirements. However, a common obstacle to these dynamic systems is their low execution efficiency. As the information extracted from the pure binary codes is limited, it is hard for the dynamic systems to apply advanced or aggressive optimizations. Take dynamic instrumentation tools and simulators as example, it usually costs dozens or hundreds of time to run programs within these dynamic systems.Therefore, it becomes a popular research hotspot to improve the execution efficiency for the Dynamic Optimization Systems. Although many strategies have been proposed to improve the efficiency, some aggressive optimizations, especially the optimizations for memory, cannot be applied. Since high-level semantics information would be lost during executable binary generation, it is difficult to extract the memory related information through pure binary-based approaches through profiling tools or on-line analysis.However, while a compiler analyzes programs, such as data flow analysis, it is easy to collect memory related information which is hard to analysis but essential for the dynamic systems. Our algorithm in this paper is based on this approach. Once the information is annotated into executable binaries as metadata, more dynamic optimization opportunities for Dynamic Optimizations Systems would be available. In this paper, we present a compiler-based metadata driven optimizing framework and two paradigms: a novel phase optimization for simulation and a register allocation approach for a dynamic binary translator. Experimental results show that it is effective in both cases.