From: SASADA Koichi <ko1@...>
Date: 2011-10-27T14:26:54+09:00
Subject: [ruby-core:40439] Re: JIT development for MRI

Hi Carter,

(2011/10/27 8:43), Carter Cheng wrote:
> I guess when it comes to Ruby my sense is that the dynamism poses some
> complex problems for development of a good JIT system. I know that Macruby
> relies on the LLVM JIT system and this simplifies much of the coding.-
> however the LLVM JIT system is rather heavyweight. There was some paper
> describing for example the effect of substituting (not in much detail) tcg
> in QEMU using the LLVM system and the performance was significantly impeded.

In fact, YARV bytecode is not suitable (optimized) for JIT compilation.
 If I make a JIT compiler based on YARV, I'll make some tricks to
overcome it.  I have several ideas, however I don't have a spare time to
try it.

One hint is: as I said, YARV instructions are course-grain, because it
is optimized for VM loop execution.

> I am curious of what your sense is of this. I am somewhat speculating here-
> but it would seem to me that one of the main limiting factors (not having
> studied the exact sequence of instructions produced by MRI) is the extra
> overhead in primitive operations due to the inability to infer which
> operations to apply to a type statically and perhaps the overhead associated
> with invoke/sends.

Language dynamism is another one.  But it should be beaten with
profiling or some other techniques as Charlse posted [ruby-core:40437].

Cheers,
Koichi

-- 
// SASADA Koichi at atdot dot net