> One problem with this approach was that the mechanism to count
> the bytecodes slowed down the bytecode dispatch considerably--by a
> factor of two, as I recall! A better approach might be to use a timer
> interrupt to tell the VM when to return control to the browser.
> 
> Do you have any thoughts on how to handle this problem?

You could count preemption points instead... 
I.E. in

	Interpreter >># checkForInterrupts