[boost] [conversion] lexical_cast size optimization

From: Alexander Nasonov (alnsn_at_[hidden])
Date: 2006-04-04 15:13:55

I noticed that instantiated lexical_cast functions are parameterized
with array types. It's not optimal as array_to_pointer_decay is applied
to argument type immediately and original type is not user anywhere
else.
My idea is to define inlined lexical_cast that just transforms
the argument type and forward a call to detail::lexical_cast:

    template<typename Target, typename Source>
    inline Target lexical_cast(const Source &arg)
    {
        typedef typename detail::array_to_pointer_decay<Source>::type NewSource;
        return detail::lexical_cast<Target,NewSource>(arg);
    }

The detail::lexical_cast is almost an original lexical_cast except that
it doesn't transform the argument type. Additionally,
call_traits<>::value_type can be applied to the argument type:

    namespace detail
    {
        template<typename Target, typename Source>
        Target lexical_cast(
            BOOST_DEDUCED_TYPENAME boost::call_traits<Source>::value_type arg)
        {
            detail::lexical_stream<Target, Source> interpreter;
            Target result;

            if(!(interpreter << arg && interpreter >> result))
                throw_exception(bad_lexical_cast(typeid(Source), typeid(Target)));
            return result;
        }
    }

I tested lexical_cast_test.cpp with gcc 3.4.4 [FreeBSD] 20050518.
Code was generated with bjam --v2 release.

Original size:           220316
Optimized:               201940 -17.95K (  8.34% )
Optimized & call_traits: 197108 -22.66K ( 10.54% )

Additional transfomations can be applied. For example, enums can be
promoted to integral types. I'm a bit unsure about this, though.
This change could break enum types that have overloaded operator<<.
What is expected behavior in this case?
-- 
Alexander Nasonov