[ruby-core:40442] Re: JIT development for MRI

On Wed, Oct 26, 2011 at 11:56 PM, Carter Cheng <[e-mail]> wrote:
> Thanks Charles. I am still somewhat curious about this- I did look at the
> link for jruby you sent me but I am not sure how JRuby is exploiting the JVM
> framework. Do you have specific JVM's that you feel may be interesting to
> study? Are most of these OSS in some form or another nowadays? I am curious
> where you do draw the line between the academic theoretical and practical
> (since many pieces showing up in places like PLDI are descriptions of actual
> projects- like Mozilla SpiderMonkey).

For JVMs I'd recommend reading about Hotspot, the VM inside OpenJDK.
It is GPL and there's lots of documents about how it optimizes code
available online. Here's a good place to start:

http://wikis.sun.com/display/HotSpotInternals/PerformanceTechniques

There are certainly many others out there, including formal academic
papers about Hotspot and other VMs.

JRuby exploits the JVM mostly by simply targeting it. We try to
produce the lightest-weight bytecode possible, and the JVM takes over
from there. So JRuby itself isn't *quite* as interesting as the JVM
would be for your purposes.

The new IR compiler in JRuby that I mentioned may be of more interest,
since we'll be able to apply traditional compiler optimization
techniques there.

And as far as academic, theoretical, and practical...I consider
"practical" to be "something I can use right now." There's lots of
interesting papers on static optimization for Ruby, runtime type
inference, and so on...but as far as I know no working implementation
that utilizes any of those techniques. That's not to say they
shouldn't be explored...I'm just saying if you want to make an impact
on MRI performance in the short term, going with a simpler approach at
first would pay dividends much more quickly.

Of course I don't know what your goals are. You may be interested in
trying to put the more "academic" approaches to practical use.

- Charlie