A new Java coprocessor architecture called JVXtreme takes a different approach to Java application acceleration. It works in cooperation and in parallel with the main processor, generating better performance than in-line conversion of Java bytecodes to native processor instructions (see the figure).
Produced by inSilicon, JVXtreme supercedes the company's earlier JVX architecture. Both speed up bytecode execution. JVXtreme, though, reduces instruction fetch time so most Java bytecodes execute in a single cycle. The Java Virtual Machine (JVM) is implemented on the native processor, but inSilicon's coprocessor handles substantial portions.
This architecture provides the speed and compact code advantages of a native Java processor while working with a native processor. It competes with Just-in-Time (JIT) native-code compilers and other Java hardware accelerators. Efficient fetch and execution of Java byte codes provide this performance. So do two key features.
The first is a large hardware stack. Its 64-entry default size is significantly larger than in-line bytecode conversion architectures that use a portion of the native processor's register block, like ARM's Jazelle. JVXtreme handles stack overflow and underflow automatically. The JVM uses this feature during a task switch.
The second, instruction folding, takes advantage of Java's stack architecture. Architectures usually push a constant onto the stack and then store the new top-of-stack entry in the variable. Instead, JVXtreme recognizes this pair of instructions and stores the constant in the variable without using the stack. Also, JVXtreme pushes the constant on the stack and then compares it to the variable in one step.
JVX and JVXtreme currently support the ARM7 and ARM9 processors. InSilicon's coprocessor has a small footprint at only 35 kgates and operates at clock speeds up to 200 MHz.
More information is available at www.insilicon.com.
