Re: Interpretersn & Intermediate languages

>>[Discussion of instruction set simulators.]


John Levine (compilers-request) says:
>[I know many systems translate to machine code; what would astonish
> me is if you got usefully fast performance (compared to a decode and
> dispatch interpreter) by compiling to something other than machine
> code.]


I suppose it hinges around what you consider to be `usefully fast'.


  - One fast decode-and-dispatch interpreter in assembly averages
      around 40I/I and peaks around 30I/I.


  - Rob Bedicheck's `g88' simulator is written in C and does
      compilation to threaded code and even though it does much more
      simulation work (address translation, etc.) it averages better than
      30I/I, peaks at better than 20I/I and could probably be streamlined
      down to 10I/I if you removed a lot of the fancy features (address
      translation, etc.).


  - Several systems that translate to machine code peak at around 2I/I.


A factor of 2-3 improvement by dynamically compiling to threaded code
seems ``usefully fast'' to me, but it does depend on your needs.


These figures aren't exact because the host and target machines vary,
but there are several implementations in each camp, and the relative
performances of various host/target combinations and implementation
technologies seem to be fairly consistant.


;-D on ( Pidgin intermediate langauges ) Pardo
[OK, OK, I'm astonished. -John]
--

Return to the comp.compilers page.
Search the comp.compilers archives again.