Re: Default x86-darwin to -mfpmath=sse, and update -arch settings

[Geoffrey Keating <geoffk at apple dot com>]

On 12/04/2005, at 4:17 AM, Richard Guenther wrote:
> On Apr 12, 2005 12:38 PM, Richard Henderson <rth@redhat.com> wrote:
> > On Mon, Apr 11, 2005 at 06:44:27PM -0700, Geoffrey Keating wrote:
> > > - It merges in a patch from Apple's tree to use 'i386' for -arch
> > >   instead of 'i686'.  (The people responsible for the previous 
> > > setting
> > >   changed their mind.)
> > >
> > > - It makes darwin use SSE fpmath by default.  I've made this a
> > >   convenient macro because I expect that other platforms might want
> > >   this in the future.
> >
> > I don't see how these work together.  Seems to me you'd want ...
> >
> > > These have a common motivation, which is that Darwin only runs on a
> > > pentium-m and higher, and earlier chips are not really supported.
> >
> > -march=pentium-m as the default then.  Since it would appear that
> > you're not actually enabling SSE, and thus -mfpmath=sse.
> >
> > Or are you just expecting folks to configure --with-arch?

At the moment, I'm expecting --with-arch, but am thinking of ways to 
avoid that.

What I was thinking of was adding a TARGET_ARCH_DEFAULT macro, but 
there's a problem: if you just configure i686-apple-darwin, you get 
-mtune=pentiumpro, not what you want; if you configure 
i386-apple-darwin, you get a libstdc++ with 386 locking, also not what 
you want; and config.sub says that pentium_m-apple-darwin is not a 
valid configure triplet and it seems weird to require that anyway.

> Also, without an associated ABI change, the performance impact of
> fpmath=sse is not always positive, even so with the lost x87 intrinsics
> without appropriate replacements with tuned SSE/SSE2 code, like
> the intel compiler provides.

I understand that the purpose of the fpmath setting is not so much for 
performance, but that the libm really wants proper IEEE semantics (due 
to it being mostly intended for ppc); but I'll ask, maybe someone is 
confused.  It'd be nice if it had good performance too, of course, or 
at least as good as fpmath=387.

(If you wanted best performance, I think you'd want to use 
-mfpmath=sse,387 right?)