Date: Sat, 23 Aug 2003 14:26:37 -0700 From: David Schultz <das@FreeBSD.ORG> To: "Dag-Erling =?us-ascii:iso-8859-1?Q?Sm=F8rgrav?=" <des@des.no> Cc: cvs-all@FreeBSD.ORG Subject: Re: cvs commit: src/usr.sbin/route6d route6d.c Message-ID: <20030823212637.GA11742@HAL9000.homeunix.com> In-Reply-To: <xzp4r0fqrdg.fsf@dwp.des.no> References: <200308171605.h7HG5nOd095330@repoman.freebsd.org> <20030818070415.W3401@gamplex.bde.org> <xzp8yprqsqo.fsf@dwp.des.no> <20030818083502.GA71675@rot13.obsecurity.org> <xzp4r0fqrdg.fsf@dwp.des.no>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Aug 18, 2003, Dag-Erling Smrgrav wrote:
> Kris Kennaway <kris@obsecurity.org> writes:
> > On Mon, Aug 18, 2003 at 10:23:27AM +0200, Dag-Erling Sm?rgrav wrote:
> > > I have patches which allow world to build successfully with
> > > -fno-builtin...
> > What did you do about the libstdc++ breakage (some symbols are
> > undefined by our tree and only provided by gcc's builtins), or did
> > someone else fix that?
>
> I implemented fabsl(3), which is what libstdc++ complained about. We
> could also get by simply by adding
>
> #ifdef __GNUC__
> #define fabsl(x) __builtin_fabsl(x)
> #endif
>
> to src/lib/msun/src/math.h.
>
> I tried to raise interest on the -standards list a few months ago for
> implementing the "long double" versions of the standard math routines,
> as mandated by C99 and SUSv3, but nobody took my bait. I also managed
> to trash the partial implementation I did have, so all I have right
> now is fabsl(3).
I have implementations of a few of the long double math functions
lying around. fabsl(3) should be as simple as:
#include "fpmath.h"
long double
fabsl(long double e)
{
union IEEEl2bits u;
u.e = e;
u.bits.sign = 0;
return (u.e);
}
However, it would be *much* faster if implemented in assembly, and
slightly faster than that as a builtin. I've been holding off on
working on this until I can figure out a good way to add MI
support in such a way that an MD version would override, if
present. That way, we could support long double floating point
efficiently on most architectures without hampering porting efforts.
Any suggestions?
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20030823212637.GA11742>