On Thu, 2 Aug 2007, Joshua Isom wrote:
> I'm pretty sure 0.0 always equals -0.0. I think it's part of the c
> specification. Now, on OpenBSD, you can't print -0.0, as it will print 0.0
> instead which is really annoying. This is with at least 3.8 but I don't know
> if it's been changed since then. Anyway, to test for -0.0 instead of 0.0, you
> have to use copysign.
I'm afraid copysign() is not generally suitable (though it may be on some
platforms). On Linux, the copysign(3) man page has the following caveat:
The copysign() functions may treat a negative zero as positive.
Alas, there's no further information, so it's not obvious when such
misbehavior might occur.
In a C99 system, you should use signbit(3) instead, though you may need
extra compiler flags or libraries to find it. In the absence of
signbit(3), there are various bitwise tricks one can often use. Some of
these are scattered about in config/gen/platform/*/math.h.
> Anyway, comparing equality of floats is never accurate, since the floats
> themselves are inaccurate.
I disagree. In some circumstances, comparing to 0.0 can be perfectly well
defined.
This whole discussion seems pointless. The accuracy of floating point
calculations depends sensitively on what, precisely, the programmer is
attempting to do. Guessing programmer intent doesn't seem to me to be an
appropriate job for a virtual machine.
--
Andy Dougherty [EMAIL PROTECTED]
