Re: [boost] Undefined primary templates & generative interfaces (was: Comments on crc9)

From: Doug Gregor (gregod_at_[hidden])
Date: 2001-03-27 13:30:14

On Monday 26 March 2001 11:28, you wrote:
> on 3/25/01 7:20 PM, Jens Maurer at Jens.Maurer_at_[hidden] wrote:
> > I've had a look at crc version 9 from the boost files section.
> >
> > Here are a few comments:
> >
> > boost/crc.hpp
> >
> > - The exact_uint_t<> template is once declared and only later defined.
> > That makes for a lot of repeated #ifdef and, as an implementation detail,
> > doesn't buy much clarity either.  Could we omit the declarations?
>
> Is it OK to define template specializations without declaring them,
> especially if you only have a forward declaration for the general version?

Yes, it is. It is often used when there are requirements on the template 
parameters that can be checked with (partial) specialization. Failure to meet 
the requirements results in an attempt to instantiate an undefined type. For 
instance,

template<typename Vector> struct require_vector;

template<typename T, typename Allocator>
struct require_vector< std::vector<T, Allocator> > {};

The technique 

template<
  std::size_t Bits,
  unsigned long TruncPoly = 0ul,
  ...,
  typename Implementation = 
    typename get_implementation_tag<Bits, TruncPoly, ...>::type
>
struct crc; 

template< std::size_t Bits, unsigned long TruncPoly, ...>
struct crc<Bits, TruncPoly, ..., crc_fast_tag>
{
  // the fast implementation
};

The benefits of this type of generative approach are:
        1) No '::type' to confuse users
        2) You can specialize based on "crc". With typical '::type' generative 
interfaces, you don't know what the returned type will be, so you can't 
possibly specialize for any CRC computation class. This could get really bad 
when you use multiple generic libraries that use generative interfaces and 
happen to have free functions of the same name.

        Doug