Re: [boost] Big Int

From: Richard Peters (r.a.peters_at_[hidden])
Date: 2005-03-05 13:34:22

From: "Rob Stewart" <stewart_at_[hidden]>
> From: "Richard Peters" <r.a.peters_at_[hidden]>
> >
> > The bigint type proposed for the C++ standard library does not permit
the
> > use of expression templates. Using expression templates has a positive
and a
> > negative aspect: A significant speedup can be gained using expression
> > templates. The downside is that templates cannot deduce the correct type
of
> > expressions, for example:
> > template<class T> void f(T a, T b) will not work when invoked as f(x +
y,
> > x - y), because x + y returns an object of type different from x - y.
> > Because the C++ standard library proposal specifies the result type of
the
> > operators to be of type const integer, a library implementing that
proposal
> > cannot use expression templates.
>
> PMFJI, but I wonder whether this would work:

PMFJI... I have no clue what it means, sorry.

>    template <typename T, typename U>
>    void f(T a, U b)
>    {
>       typedef typename <magical_incantation<T, U>::type V;
>       f_impl<V>(a, b);
>    }
>
> a and b can have different types, but then magical_incantation
> computes a common type to which both can be converted and the
> implementation function can be invoked with that type.
>
> I may be off base, and you may have already tried this approach,
> but I just wanted to mention it if it could help.

Oh, there are a few ways in which you can perform computations with the
expression templates, but that was not what I meant. My point was: If you
specify the result of operators + and - to be big_integer, then expression
templates cannot be made to work, because then those operators return types
other than big_integer. User code may depend on the specification that those
operators do return big_integers. Therefore, any implementation satisfying
the requirements of the C++ standard library proposal will miss the possible
speedincrease gained by the use of expression templates.

best regards,

Richard Peters