From: Charles Oliver Nutter <headius@...>
Date: 2011-10-26T00:12:47+09:00
Subject: [ruby-core:40357] Re: JIT development for MRI

On Tue, Oct 25, 2011 at 9:36 AM, Carter Cheng <cartercheng@gmail.com> wrote:
> Hello,
> I was recently considering some JIT development work on MRI 1.9. internals
> after spending sometime looking at the internals of rubinius and macruby. It
> would seem that various notions taken from older projects like Craig
> Chambers self thesis would make some sense when applied here or some
> developments in trace trees (older notions like in Dynamo) or the latest
> piece on Ruby in POPL 2011.
> Some of these obviously require some adaptation before being applied to MRI.
> Any comments would be appreciated.

I would love to see MRI get a JIT. Some quick rambling thoughts.

* You'll get a good perf boost just by compiling to native code and
avoiding the interpreter, but the biggest boost will come from
inlining methods. JRuby on Java 5/6 does only the former (or rather,
we don't make it possible for the JVM to inline). JRuby on Java 7 does
the latter, and performance can be many times faster.

* ruby-core folks will be sensitive to any additional dependencies.
For example, LLVM is not a small library. Of course the JIT could be
made an optional part of the build.

* I will happily share techniques we've used to optimize JRuby. We do
get to leverage the JVM a lot, but I also have a good working
knowledge of how the JVM does JIT and optimization.

- Charlie