Re: optimizing compilers

FROM : John Hörnkvist
DATE : Fri Feb 01 23:07:10 2002

On Friday, February 1, 2002, at 08:34 PM, Finlay Dobbie wrote:

> No, it's a combination of multiple factors. It's well known that gcc
> 2.95.2 sucks for PPC code generation.

As far as I know, gcc sucks for code generation, period. You can help it
a fair bit by using some of the attributes and language extensions that
it implements.

> I think it should be better with gcc 3.1, which Stan Shebs and the
> Apple compiler team are working overtime on bringing up to scratch on
> Darwin/OS X.

I've benchmarked a fairly recent version of gcc 3.1 and 2.95.2, and the
difference has been quite small.

To make matters worse, the PPC seems rather sensitive to bad code; when
I turn on optimizations in my research compiler, I can speed up
execution by almost a factor of three on the PPC (16.11 to 5.76 seconds)
but less than two times on an AMD Duron (7.95 to 4.32 seconds).

>  I heard a while back that they moved a lot of their ex-MrC engineers
> over to work on gcc 3.1/Darwin (MrC was Apple's compiler that was part
> of MPW, and it was widely recognised as being one of the most
> optimizing PPC compilers. see
> http://developer.apple.com/tools/mpw-
> tools/compilers/benchmarks/index.html for examples).

Well, that makes sense; no reason to have them working on a dead
compiler.

Regards,
John Hornkvist
--
ToastedMarshmallow, the perfect Cocoa companion
http://www.toastedmarshmallow.com