[MLton] Re: Hurd on l4 / ConcurrentML

[Duraid Madina]

On Fri, Aug 11, 2006 at 03:13:05PM -0700, Stephen Weeks wrote:
> 
> > For now, if anyone wants to do some low-level manipulation of
> > programs built with MLton, they can just drop llvm-gcc in place of
> > gcc and they will be left with LLVM IR on their disk.
> ...
> > For now, I can do all the LLVM-related work I want to do anyway
> 
> The obvious question being: did you get any better MLton code out of
> the llvm-gcc codegen than the gcc codegen?

I did, but only just, and it depended on the code. For most things, there's
not much in it (typically +/- 15%) since one of LLVM's key features, the
ability to do whole program optimizations regardless of programming language
or compilation scheme, is of little benefit to MLton output that is more or
less solid blocks of code anyway.

However, I'm confident a proper LLVM backend should perform significantly
better. While LLVM is already more expressive than C (e.g. support for
efficient exceptions and tail calls) it can certainly be extended, and
supporting MLton well is something the LLVM folk would be happy to help
with.

	Duraid