Date: Tue, 18 Mar 2003 13:30:41 +0200
From: "Eli Zaretskii" <eliz AT elta DOT co DOT il>
Sender: halo1 AT zahav DOT net DOT il
To: djgpp-workers AT delorie DOT com
Message-Id: <9003-Tue18Mar2003133041+0200-eliz@elta.co.il>
X-Mailer: emacs 21.3.50 (via feedmail 8 I) and Blat ver 1.8.9
In-reply-to: <200303180858.JAA03661@lws256.lu.erisoft.se> (message from Martin
	Stromberg on Tue, 18 Mar 2003 09:58:34 +0100 (MET))
Subject: Re: isnan and isinf
References:  <200303180858 DOT JAA03661 AT lws256 DOT lu DOT erisoft DOT se>
Reply-To: djgpp-workers AT delorie DOT com
Errors-To: nobody AT delorie DOT com
X-Mailing-List: djgpp-workers AT delorie DOT com
X-Unsubscribes-To: listserv AT delorie DOT com
Precedence: bulk

> From: Martin Stromberg <Martin DOT Stromberg AT epl DOT ericsson DOT se>
> Date: Tue, 18 Mar 2003 09:58:34 +0100 (MET)
> > 
> > IMHO, we should indeed have these functions in libc.a, but perhaps a
> > simpler code could be written to implement them.  While at that, how
> > about supporting IA32-specific bit patterns for Inf's and NaN's (see
> > the Intel manual for details)?
> 
> Huh!? 
> 
> 1. Are you saying isnan() doesn't detect all possible nans?

I don't know.  Intel's manual documents an SNaN, a QNaN, and a Real
Indefinite (which is a variant of a QNaN).  They all have different
bit pattens in the mantissa, and the two first ones may have the sign
bit set.  Do we know for sure that `isnan' supports all of those?

> 2. The code in src/libm/math/sf_isnan.c doesn't look overly
> complex. In fact it looks like it will be hard to improve it.

The libm code tries to be portable, so it fiddles with bits in
floating-point numbers by using a union of an FP number and one or
more unsigned int's.  By contrast, since we _know_ we are writing for
an x87, we can access those bits directly and avoid the overhead of
the macros such as GET_FLOAT_WORD and EXTRACT_WORDS.  For example, I
could write a version of `isnan' that tests that the exponent bits are
all set (1 AND and 1 test), and then that at least one of the bits in
the mantissa is set (1 more AND and 1 more test).  I'm not sure, but
this might use less clocks than the libm version.

> 3. I only know about one bit pattern for inf and another one for
> -inf. Are there other ones?

No, I don't think so.