When are Bytecodes Faster than Direct Execution?

We demonstrate the tradeoffs between interpreting code in compact forms (such as the bytecoded method formats used in Smalltalk and Java), and direct execution of code (as in dynamic-translation Smalltalk and just-in-time Java compilers). Experiments with the interpreted Smalltalk language show that the widely held belief that direct execution is always faster than compact interpretation is not correct. Specifically, it is untrue for memory architectures with high cache-miss penalties, where the larger volume of less frequently repeated instructions inherent in direct execution outweighs the reduction in the number of instructions executed. The experiments presented demonstrate that the space of cycle costs for different types of cache misses can be divided into regions where a particular option is faster, and that these regions differ for different cache configurations. This suggests that interpreter designers must consider memory architecture and language characteristics when deciding whether to use direct execution. This work has been supported in part by gifts from Sun Microsystems Laboratories. †Submission contact: Eliot Moss, Phone: (413) 545-4206, Fax: (413) 545-1249