QuickInterp: speeding up Java interpreters with superinstructions By Lukas Miedema

We propose a method which provides a optimization for the OpenJDK interpreter without requiring a JIT or platform-specific code ("inline assembly"). By using code generation to generate an application specific interpreter it is possible to offer increased performance for those applications. The interpreter is generated by use of a profile from the running application which identifies hotspots and often-executed paths and converts common sequences of bytecode instructions to a single sequence of C++ code representing a single "superinstruction", which can then be compiled to a binary application by a C++ compiler. Java bytecode is converted to these superinstructions by a matching algorithm, and performance is improved by a reduction in the number of instruction dispatches in the interpreter loop. Various algorithms for deciding the best sequences of bytecodes to compile and various matching algorithms to -- at class load time -- map bytecode instructions to these superinstructions are discussed.