Experience Report: Speeding up a Nested VM Using Bytecode Comp

The focus of this paper is Weld, a bytecode compiler that compiles a given Java class into a new Java class with added compiled methods. The Java bytecode to Java bytecode compilation that Weld performs is key to accelerating Rivet, a Java virtual machine written in Java that serves as an extensible tool platform. In accelerating Rivet, Weld employs several bytecode compilation techniques. Some of these techniques resulted in significant speedup, while other techniques were less successful. The Rivet JVM was designed and implemented at MIT from the summer of 1998 through the spring of 1999. Rivet is a platform for debugging and testing tools that is structured as a Java Virtual Machine. The Java platform was chosen for Rivet's implementation because of its characteristic clarity, facility, and portability. Unlike the Java VM developed in the Jalapeno project [1], which was written in Java but designed to run on top of bare hardware, Rivet was designed to run on top of an underlying Java VM. The Java VM on top of which Rivet runs is referred to in this paper as the lower-level JVM, while the programs that run on top of Rivet are referred to as client programs. The nested virtual machine design enabled Rivet to be portable and to reuse the Java APIs implemented by the lower-level JVM. Because Rivet's goal was to support tools, clarity and extensibility initially took priority over performance in Rivet’s design. Thus, Rivet started out as a pure interpreter, effectively built around a bytecode dispatching case statement. Performance tests revealed that Rivet spent 70% of its time in this dispatch loop. This result indicated that bytecode compilation offered the possibility of dramatically improving Rivet’s performance. Weld was designed in pursuit of this goal, and also with careful attention to clean and simple integration with Rivet.