Re: VM: direct vs indirect threading

[Keisuke Nishida]

address@hidden (Karl M. Hegbloom) writes:

>  Keisuke?  Have you looked at `rscheme' at all?  (It's very
>  complicated, to me.)  Apparently it has some sort of automatic
>  generation of byte code interpreters.  Might be worth a study.

No.  I want to look at it, but their ftp server is not working now.
Where could I find the source code?

>  Did you find anything interesting on CiteSeer?

I briefly looked at an article "Optimizing direct threaded code by
selective inlining" (http://www-sor.inria.fr/publi/ODCSI_pldi98.html).
They says they does direct threading and inlining (combining several
instructions) dynamically.  My compiler does these at compile time.
I think direct threading has been done in my VM.

On the other hand, my VM consumes 4 times as much memory as indirect
threading interpreter.  If indirect threading is done sufficiently
fast, I will modify my VM so that it uses indirect threading, still
using computed-gotos:

  unsigned char *ip;
  void *label_table[256] = {&&l_nop, ...};

  #define NEXT   goto *label_table[*++ip]

It seems rep uses this technique, and I think this better fits for
my VM.  It's simple, uses less memory, simplifies the instruction set,
and makes the debugging engine as fast as the regular engine.  Great!

>  * Notes:
>  *      This bytecode interpreter is hand-crafted.  In the future,
>  *      BCIs can be constructed automatically using trace-driven
>  *      instruction-sequence analysis for optimization.

Sounds interesting.  I thought the same thing, but it requires
re-compilation of the interpreter.  I think hand-craft is okay.
JIT compiler is more attractive to me.  Dan is working on it with
his new QScheme.

Anyway, these optimization issues can be done any time later.
I need to complete the VM and the compiler first...

-- Kei