[pooma-dev] Explicit Instantiation Problem with Pooma 2 using KCC on RH 6.2 Linux

[Jeffrey Oldham]

					Thursday, 2001May03

Dave Nyholm submitted a program equivalent to

    #include "Pooma/NewFields.h"

    template View1<Field<NoGeometry<(int)3>, int, CompressibleBrickView>,
		   Interval<(int)3> >::sv;

I fail to see why it should compile.


To complete explicit instantiation of Field<GT,T,ET>,

    View1<Field<GT,T,ET>, Domain_t>::ReadType_t
    Field<GT,T,ET>::read() const;

must be instantiated.  Assume all of the class definition upto but
excluding read() has been processed.  Here is an excerpt of the
relevant code:

    template<class GeometryTag, 
	     class T = POOMA_DEFAULT_ELEMENT_TYPE,
	     class EngineTag = POOMA_DEFAULT_ENGINE_TYPE>
    class Field {
    public:
      typedef Field<GeometryTag, T, EngineTag> This_t;

    // ...

      inline typename View1<This_t, Domain_t>::ReadType_t 
      read() const
	{
	  typedef View1<This_t, Domain_t> Ret_t;
	  return Ret_t::makeRead(*this, physicalDomain());
	}
    };


    template<class GeometryTag, class T, class EngineTag, class Sub1>
    struct View1<Field<GeometryTag, T, EngineTag>, Sub1>
    {
      // Convenience typedef for the thing we're taking a view of.

      typedef Field<GeometryTag, T, EngineTag> Subject_t;
      // ... series of typedefs; Dispatch_t relies on Subject_t
      typedef typename Dispatch_t::ReadType_t ReadType_t;

      // ...
    };

Thus, View1<Field<GT,T,ET>, Domain_t> must be instantiated at least up
to the definition of ReadType_t.  Its first typedef requires
instantiation of Field<GT,T,ET>, which is being processed.  Thus,
compilation fails.

We could try to rewrite View1<Field<GT,T,ET>, Domain_t> to just refer
to Field<GT,T,ET>'s members, avoiding the initial typedef.
Unfortunately, member function makeRead() has a Field<GT,T,ET>
argument.

How should we solve the problem?

Jeffrey D. Oldham
oldham at codesourcery.com