sqrt with negative arg.

sqrt with negative arg.

[Luc Teirlinck]

A few hours ago, the documentation of sqrt in numbers.texi was
changed from saying that it returns a NaN to saying that it signals a
domain-error.  It definitely returns a NaN on my machine.  What about
the following, which I could install if we agree on it:

===File ~/numbers-d31-diff==================================
*** numbers.texi.~1.32.~	Wed Dec 31 08:53:21 2003
--- numbers.texi	Wed Dec 31 15:20:59 2003
***************
*** 1146,1152 ****
  
  @defun sqrt arg
  This returns the square root of @var{arg}.  If @var{arg} is negative,
! it signals a @code{domain-error} error.
  @end defun
  
  @node Random Numbers
--- 1146,1153 ----
  
  @defun sqrt arg
  This returns the square root of @var{arg}.  If @var{arg} is negative,
! @code{sqrt} could return a NaN or signal a @code{domain-error} error,
! depending on the implementation.
  @end defun
  
  @node Random Numbers
============================================================

Re: sqrt with negative arg.

[Eli Zaretskii]

> Date: Wed, 31 Dec 2003 15:36:31 -0600 (CST)
> From: Luc Teirlinck <teirllm@dms.auburn.edu>
> 
> A few hours ago, the documentation of sqrt in numbers.texi was
> changed from saying that it returns a NaN to saying that it signals a
> domain-error.  It definitely returns a NaN on my machine.

Really?  What system is that?  Can you debug Fsqrt and see why it
happens?

The way it's supposed to work on most modern systems is that the call
to `sqrt' in the underlying C code causes errno to be set, and Fsqrt
is supposed to check for that and signal an appropriate error.  See
the macro IN_FLOAT defined on floatfns.c.

(The last time a similar problem was reported and discussed here, we
found it was due to a bug in a particular version of glibc that was
fixed in later versions.)

Re: sqrt with negative arg.

[Richard Stallman]

    A few hours ago, the documentation of sqrt in numbers.texi was
    changed from saying that it returns a NaN to saying that it signals a
    domain-error.  It definitely returns a NaN on my machine.

I think that is a bug--why does it not signal an error?

Re: sqrt with negative arg.

[Luc Teirlinck]

Richard Stallman wrote:

       A few hours ago, the documentation of sqrt in numbers.texi was
       changed from saying that it returns a NaN to saying that it signals a
       domain-error.  It definitely returns a NaN on my machine.

   I think that is a bug--why does it not signal an error?

As Eli pointed out, it is probably due to my glibc version.  So I will
definitely _not_ install the change to numbers.texi which I proposed.

Sincerely,

Luc.

Re: sqrt with negative arg.

[Andreas Schwab]

Luc Teirlinck <teirllm@dms.auburn.edu> writes:

> A few hours ago, the documentation of sqrt in numbers.texi was
> changed from saying that it returns a NaN to saying that it signals a
> domain-error.  It definitely returns a NaN on my machine.

This is a bug either in sqrt or in matherr on your system.  ISO C
requires that a domain error occurs when sqrt is called with a
negative argument, and Emacs depends on matherr (when it is supported)
to be called.  Apparently, the availability of matherr is not enough
to guarantee this.

Andreas.

-- 
Andreas Schwab, SuSE Labs, schwab@suse.de
SuSE Linux AG, Maxfeldstraße 5, 90409 Nürnberg, Germany
Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."

Re: sqrt with negative arg.

[Eli Zaretskii]

> From: Andreas Schwab <schwab@suse.de>
> Date: Thu, 01 Jan 2004 18:22:22 +0100
> 
> This is a bug either in sqrt or in matherr on your system.  ISO C
> requires that a domain error occurs when sqrt is called with a
> negative argument, and Emacs depends on matherr (when it is supported)
> to be called.  Apparently, the availability of matherr is not enough
> to guarantee this.

Actually, Emacs will do fine even if the math function only sets errno
when such problems happen.  If matherr is available, it will be used,
but it isn't necessary.  The macros on floatfns.c tell the details.