libffi & LLVM

libffi & LLVM

[Anthony Green]

The latest MacRuby version replaced libffi with LLVM for closure 
support, the result of which is 3 to 4 times performance boost:
http://www.macruby.org/blog/2009/10/07/macruby05b1.html

I've been thinking for a while that a LLVM backend to libffi makes 
sense.  Does anybody have any thoughts or opinions on this?

Is this what libffi 4 should be?

AG

Re: libffi & LLVM

[Andrew Haley]

Anthony Green wrote:
> The latest MacRuby version replaced libffi with LLVM for closure
> support, the result of which is 3 to 4 times performance boost:
> http://www.macruby.org/blog/2009/10/07/macruby05b1.html

I wonder if they measured the overhead for generating the closures.
If they didn't, perhaps we should.  Also, LLVM has a pretty big
footprint.  So, it's not all positive.

> I've been thinking for a while that a LLVM backend to libffi makes
> sense.  Does anybody have any thoughts or opinions on this?

It makes some sense, but I warn you: while working on Shark we've
found that the JIT interfaces to LLVM are rather unstable.  We'd be
setting libffi up for an eternal game of catch-up every time the
LLVM interfaces change.  (Perhaps not eternal: there's every chance
that as LLVM matures its JIT interface will stabilize.)

A better solution for libffi closures would be for libffi to generate
its own code, but I think that would be too much effort, given the
complexity of some of the ABIs we have to work with.

Andrew.

Re: libffi & LLVM

[Gary Benson]

Anthony Green wrote:
> The latest MacRuby version replaced libffi with LLVM for closure  
> support, the result of which is 3 to 4 times performance boost:
> http://www.macruby.org/blog/2009/10/07/macruby05b1.html

I'm looking at doing something similar for Shark.  In this case,
it wouldn't be a straight replacement; methods would default to
using libffi, and hot methods would be replaced with LLVM-build
wrappers (in much the same way as interpreted methods default to
using the interpreter and get replaced with LLVM-compiled code
when hot).

> I've been thinking for a while that a LLVM backend to libffi
> makes sense.  Does anybody have any thoughts or opinions on
> this?

Note that LLVM does not have JIT backends for anything like as
many platforms as libffi supports.  x86, x86_64, ppc, ppc64
(maybe), alpha and arm is I think it.

Cheers,
Gary

-- 
http://gbenson.net/