Joe Buck <[EMAIL PROTECTED]> wrote on 05/07/2005 21:10:25:
> On Tue, Jul 05, 2005 at 08:05:39PM +0200, Gabriel Dos Reis wrote:
> > It is definitely a good thing to use the full bits of value
> > representation if we ever want to make all "interesting" bits part of
> > the hash value. For reasonable or sane representations it suffices to
> > get your hand on the object representation, e.g.:
> >
> > const int objsize = sizeof (double);
> > typedef unsigned char objrep_t[objsize];
> > double x = ....;
> > objrep_t& p = reintepret_cast<objrep_t&>(x);
> > // ...
> >
> > and let frexp and friends only for less obvious value representation.
>
> I disagree; on an ILP32 machine, we pull out only 32 bits for the hash
> value, and if you aren't careful, your approach will wind up using the
> least significant bits of the mantissa. This will cause all values that
> are exactly representable as floats to collide.
For that you can do something like (or templated equivalent):
namespace Impl
{
template <class T>
size_t floating_point_hash(T in)
{
if (sizeof(in) <= sizeof(size_t))
Use Gaby's solution, with zero padding;
else
frexp and friends using Joe Buck's ideas;
}
}
Gaby's solution should be done with care - to avoid any
aliasing issues (never go directly from double& to size_t&).
Both Gaby's and Joe Buck's solutions do not take
the strangeness of IEEE (NNN?) into account.
As I remember it (I don't have the reference at home),
IEEE FP has many bit-representations for NaN, each
containing some bit-encoding of errors.
It has been years since I last saw the standard of IEEE FP,
so I may give wrong details, but the main idea should be
correct.
"There *should* be a specialization for equal_to<double> that
provides a strict weak ordering for NaNs as well as other
values." [quoted forwarded mail from P.J. Plauger]
Doing bit-wise conversions will not address this requirement.
Michael
