Re: LLVM

[David Chisnall]

On 3 Mar 2008, at 12:56, Richard Frith-Macdonald wrote:

> I'd love to see support for your runtime in GNUstep (happy to add  
> patches to gnustep-base to allow that).

Good to hear :-)

> However, it's not clear to me how that can be achieved using gcc ...  
> since I was under the impression that your runtime was sufficiently  
> different  in behavior that it would need quite a bit of compiler  
> support changes.

The behaviour isn't horrendously different.  The biggest difference is  
in terms of message lookup.  I require the sender (can be nil for  
messages sent from non-objects) to be available.  It also returns a  
slot structure, rather than an IMP, allowing the IMP to be cached and  
check that it is still valid cheaply.  Most of the more advanced  
features (late-bound ivar access and so on) are optional.

> I completely understand your reluctance to work on gcc ... I've  
> tried to hack gcc code in the past and found it pretty  
> incomprehensible myself ... but would still advocate making the  
> effort if possible (certainly, if you had gcc fixes to use your  
> runtime to support objc-2.0 features I'd be keen to help lobby to  
> get those changes into gcc).  I would have hoped that Apples  
> objc-2.0 contributions to the compiler would make it relatively easy  
> to add support for objc-2.0 using your runtime ... you wouldn't have  
> to start hacking gcc from scratch.

I had a look at the GCC code and it just made my brain hurt.  I have a  
file in my runtime (capi.h) which includes C versions of most of the  
Objective-C language constructs, so it should be easy for anyone  
familiar with GCC to add support for it.  I would be more than willing  
to help anyone interested in doing this.

More advanced features (e.g. run-time inlining of methods) are only  
possible with a more modern compiler infrastructure, however.  Unless  
the GCC team radically update their approach it will never be  
possible, but we can still get relatively good performance.

> I feel we need to try to get objc-2.0 for gnustep, but if we do that  
> by using a less widely available compiler we'll have to somehow try  
> to keep support for people who need to use gcc with the current gnu  
> runtime and that would probably be difficult.  That wouldn't stop me  
> going ahead and incorporating objc-2.0 features using LLVM though.

I would not recommend using Objective-C 2.0 features in GNUstep itself  
unless absolutely necessary.  I think GNUstep should aim for the  
widest possible compatibility, which means running on old platforms  
with limited features but supporting new platforms with newer features  
if they are available.  We will probably switch Étoilé to using my  
runtime as soon as possible because certain features we want for  
CoreObject are simply not possible with the GNU runtime (even with my  
heavily patched version).

To be honest, most of the features of Objective-C 2.0 are either  
syntactic sugar, or so completely brain-dead I can't imagine anyone  
actually wanting to use them.  Garbage collection might be nice, and  
this is much easier to do with LLVM than GCC since it was designed  
with support for garbage collected languages in mind (and allows  
mixing multiple GC strategies on a per-function basis).

David