The idea of the following sample code is simple.  First, we keep track
of matrix features (square, diagonal, column, or row) as a type.
Second,  we want the type to divert minimally in the process of matrix
multiplication.  Specifically, this code tries to differentiate
multiplication of column matrix by row matrix into two cases:

1) when column matrix and row matrix are the same size the result is
   square matrix,
2) otherwise the result is general (rectangular) matrix.

This code compiles cleanly by g++ 2.96 and KAI C++ 4.0.  dmc gives
numerious errors about multiple non-identical declarations of
operator*.

/* matrix4.cpp */
#include <iostream.h>

template <class T, int n> class ColMatrix;
template <class T, int m> class RowMatrix;

template <class T, int n, int m>
class Matrix {
protected:
   T v[n][m];
public:
   template <class TT, int nn, int mm>
   friend Matrix<TT,nn,mm> operator*(const ColMatrix<TT,nn> &, const RowMatrix<TT,mm> &);
};

template <class T, int n>
class SquareMatrix: public Matrix<T,n,n> {
public:
   template <class TT, int nn>
   friend SquareMatrix<TT,nn> operator*(const ColMatrix<TT,nn> &, const RowMatrix<TT,nn> &);
};

template <class T, int n>
class ColMatrix: public Matrix<T,n,1> {
public:
   template <class TT, int nn, int mm>
   friend Matrix<TT,nn,mm> operator*(const ColMatrix<TT,nn> &, const RowMatrix<TT,mm> &);
   template <class TT, int nn>
   friend SquareMatrix<TT,nn> operator*(const ColMatrix<TT,nn> &, const RowMatrix<TT,nn> &);
};

template <class T, int m>
class RowMatrix: public Matrix<T,1,m> {
public:
   template <class TT, int nn, int mm>
   friend Matrix<TT,nn,mm> operator*(const ColMatrix<TT,nn> &, const RowMatrix<TT,mm> &);
   template <class TT, int nn>
   friend SquareMatrix<TT,nn> operator*(const ColMatrix<TT,nn> &, const RowMatrix<TT,nn> &);
};

template <class TT, int nn, int mm>
Matrix<TT,nn,mm> operator*(const ColMatrix<TT,nn> &a, const RowMatrix<TT,mm> &b)
{
   Matrix<TT,nn,mm> r;
   for (int i = 0; i < nn; ++i)
      for (int j = 0; j < mm; ++j)
         r.v[i][j] = a.v[i][0]*b.v[0][j];
   return r;
}

template <class TT, int nn>
SquareMatrix<TT,nn> operator*(const ColMatrix<TT,nn> &a, const RowMatrix<TT,nn> &b)
{
   SquareMatrix<TT,nn> r;
   for (int i = 0; i < nn; ++i)
      for (int j = 0; j < nn; ++j)
         r.v[i][j] = a.v[i][0]*b.v[0][j];
   return r;
}

int main()
{
   ColMatrix<int,2> m1;
   RowMatrix<int,2> m2;
   m1*m2;  // should give SquareMatrix

   return 0;
}

The fix is posted now. -Walter

Walter wrote:
> The fix is posted now. -Walter

Ok, but if we change main() in the previous example to

int main()
{
   ColMatrix<int,2> m1;
   RowMatrix<int,2> m2;
   RowMatrix<int,3> m3;
   m1*m2;  // should give SquareMatrix
   m1*m3;  // should give Matrix

   return 0;
}

compiler say:

   m1*m3;  // should give Matrix
        ^
matrix4.cpp(66) : Error: need explicit cast for function parameter 1 to get
from: ColMatrix<int ,2>
to  : const ColMatrix<int ,3>

It seems that compiler doesn't see operator*() that return rectangular matrix.  [Why does it want to cast the first operator*() argument?  The second one should be the choice of the very same priority...]

Just a thank you to Iliya for exploring these matters, and Walter for fixing them!

Hey Walter be sure to put even the most minor patches on c++-announce...I check it all the time...I'd like to know what goes on between X.YY.ZZ.01 and X.YY.ZZ.02.

Mark

>matrix4.cpp(66) : Error: need explicit cast for function parameter 1 to get
>from: ColMatrix<int ,2>
>to  : const ColMatrix<int ,3>

The fix for that is posted now too <g>.

"Mark Evans" <Mark_member@pathlink.com> wrote in message news:amqq01$u2i$1@digitaldaemon.com...
> Just a thank you to Iliya for exploring these matters, and Walter for
fixing
> them!

Welcs!

> Hey Walter be sure to put even the most minor patches on c++-announce...I
check
> it all the time...I'd like to know what goes on between X.YY.ZZ.01 and X.YY.ZZ.02.

Done!

> The fix for that is posted now too <g>.
>

It rocks! Excellent work, Walter!

P.S.: I'm really amazed by the how fast all of this was fixed out. Looks more like a magic!

"Iliya Peregoudov" <iliyap@mail.ru> wrote in message news:3D91F652.3060300@mail.ru...
> > The fix for that is posted now too <g>.
> It rocks! Excellent work, Walter!
> P.S.: I'm really amazed by the how fast all of this was fixed out.
> Looks more like a magic!

What was wrong was the template partial ordering code was well tested for type arguments, but apparently none of the stlport tests tested it with value arguments. So once I figured out what was going wrong, it wasn't too hard to fix.