atan2(0, 0) = ???

[David B. Anderson]

In article <9868 at watcgl.waterloo.edu>, fjhenigman at watcgl.waterloo.edu (Frank J. Henigman) writes:
> 
> The man page for atan2 explicitly states that atan2(0, 0) = 0.
> However the following program outputs NaN.
[ example deleted] 
> Which is wrong: the man page or atan2?
> (I tried both libm and libm43 on a Personal Iris and a 4D.)
> -- 
> fjhenigman at watcgl.uwaterloo.ca                       Computer Graphics Lab
> fjhenigman at watcgl.waterloo.edu   Frank J. Henigman   University of Waterloo
>  ...!watmath!watcgl!fjhenigman                       Waterloo, Ontario, Canada

The MIPS-supplied libm atan2(0.0,0.0) produces the default quiet NaN. 
errno is not set.

For release 3.2 the man page has been changed to say that
  a)   libm atan2(0.0,0.0) produces the default quiet NaN.
  b)   libm43 (the Berkeley libm) atan2(0.0,0.0) produces 0.0.
       This is the atan2 the discussion applies to.

The SVID specifies a third behavior: atan2(0.0,0.0) returns
     0.0 while setting errno to EDOM and emitting a message on
     standard error.

ANSI C May 1988 draft says about atan2 (section 4.5.2.4):
     ``A domain error may occur if both arguments are zero.''
     Note that on a domain error, the function returns an
     implementation-defined value by section 4.5.1.

Anyone who has strong opinions about libm's behavior should send me e-mail.

Regards,
[ David B. Anderson            Silicon Graphics      (415) 964-1459 x3060 ]
[ USENET: {decwrl!,hplabs!sun!}sgi!davea          Internet: davea at sgi.com ]