Re: Interpreters for VLIW

Anton, thanks for the pointer. I implemented a similar idea and got a
~20% improvement -- not as good as I wanted it to be, but a good
starter. It will be in version 1.95.1 of GNU Smalltalk, which I'll
put out by the end of the month; that version also has a JIT for the
x86 (of course my timings were without the JIT in both cases,
otherwise I'd be comparing apples to pears).


> For the IA-64 and PPCs you also have to consider the branch registers.


Yes. The compiler translates a "goto *t" to something like


        mov b6 = r14
        br.few b6


and does not schedule the two instructions. The register is loaded
way too late. By trial and error I found that "b4" is not used by the
compiler, so I could hard code "t" in that register using GCC's
__asm__ directive; having the destination of the branch in r14 hid
most of the performance benefits.


Some other tricks I used were to cache a few global pointers in
registers.


By the way, disabling global CSE results in comparable code and less
than 20% of the compile time when you are using computed gotos. (Have
you tried that out in Gforth?)


> In the following I'll assume you want to implement direct threaded
> code (right?).


Actually I use token threaded code to share most of the code between
the non-GCC and GCC versions of the interpreter. I load the branch
register at the start of the bytecode.


> I am currently working on a VM interpreter generator that does much of
> this automatically (it's based on Gforth's generator). It should be
> ready in a few weeks. Mail me if you are interested.


It's interesting indeed. Are you going to publish it on your web site?


Paolo Bonzini

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