Mailing-List:	contact cygwin-help AT sourceware DOT cygnus DOT com; run by ezmlm
Sender:	cygwin-owner AT sourceware DOT cygnus DOT com
Delivered-To:	mailing list cygwin AT sourceware DOT cygnus DOT com
X-Authentication-Warning:	modi.xraylith.wisc.edu: khan owned process doing -bs
Date:	Sun, 2 May 1999 18:36:32 -0500 (CDT)
From:	Mumit Khan <khan AT xraylith DOT wisc DOT edu>
To:	Chris Faylor <cgf AT cygnus DOT com>
cc:	cygwin AT sourceware DOT cygnus DOT com
Subject:	Re: (patch) newlib rint fix
In-Reply-To:	<19990502193202.B773@cygnus.com>
Message-ID:	<Pine.SUN.3.93.990502183235.20497F-100000@modi.xraylith.wisc.edu>
MIME-Version:	1.0

On Sun, 2 May 1999, Chris Faylor wrote:

> >The fdlibm assume correct rounding according to the type throughout
> >the entire set of functions.  I'd suggest to compile all fdlibm
> >functions with --float-store which should have the same effect.  This
> >avoids fixing one function after the other.
> 
> I'll ping the newlib maintainer again and see what he thinks about this
> issue.

This is a terrible idea, since the performance goes down the drain. Please
reconsider the use of -ffloat-store. I believe that this particular problem 
has more to do with code that was written without aggressive optimization 
in mind than rounding modes in x86 FPU. Jacob Navia had suggested an 
alternative using assembly code that does do the right thing.

There was a long discussion in egcs (or was it egcs-bugs) regarding this,
and as usual, no consensus was reached other than the fact that float-
store is a performance killer.

Regards,
Mumit



--
Want to unsubscribe from this list?
Send a message to cygwin-unsubscribe AT sourceware DOT cygnus DOT com

- Raw text -