intermodule optimisation patch

[Geoffrey Keating]

On Monday, May 19, 2003, at 01:33  PM, Chris Lattner wrote:

>
>> The attached patch allows GCC to perform optimisations between
>> modules.
> ...
>> I'd like to spark some discussion of this patch.  Is it the right way
>> to go for intermodule optimisation,
>
> In my (extremely biased) opinion, no it isn't.  :)  Take a look at the
> paper I'm planning to present at the GCC summit this weekend for more
> information on what I think is the right approach:
>
> http://llvm.cs.uiuc.edu/pubs/2003-05-01-GCCSummit2003.html
>
>> at least for the moment?
>
> I'm not sure about that.  As described in the paper above, the changes 
> I
> would like to see made are fairly dramatic, and are thus are unlikely 
> to
> be accepted in the near future, if ever.  :)

I saw your paper earlier and thought it was interesting.  However, your 
proposal does have some significant disadvantages:

1. It's not ready now.
2. It's not likely to be ready in the near future, since (among other 
things) it requires
    implementation of a new GCC frontend to read LLVM files and generate 
code.
3. It permits people to write non-GPLed optimisers and code generators 
and still use GCC's
    parsers.
4. There are other issues with the implementation, for instance it's 
written in C++.
5. It seems likely it'd ultimately be slower than a single-compile 
approach because of the
    overhead of reading and writing LLVM files and converting them to 
and from trees.

It does have some advantages, most significantly the compatibility with 
existing Makefiles.  However, this could be achieved in other ways, for 
instance by writing out a 'stub' .o file that contains only compiler 
options and a pointer to the real source code.

-- 
Geoff Keating <geoffk@apple.com>