[Dune] RFC 01/2011: replace row_type & by row_reference in DenseMatrix

[Carsten Gräser]

Hi Martin,
I really like this change. It allows to reflect what
is already implemented in ScaledIdentityMatrix and
DiagonalMatrix. However I used a typedef named reference
there as it's done in std.:vector<bool>.

In fact I'd prefer to call it 'reference' because that's
conforming with the stl. But DenseMatrix differs from stl
containers anyway because value_type is not the type
of the values the container contains.

We should clearly state that one should use row_reference
instead of row_type& from now on, in order to write
portable code.

Best,
Carsten


Am 14.10.2011 15:03, schrieb Martin Nolte:
> Dear all,
> 
> I have replaced the return value of the operator[] of a DenseMatrix
> (which was row_type &) by a row_reference, which is exported from the
> MatVecTraits. For the corresponding matrix implementations, i.e.,
> FieldMatrix and DynamicMatrix, I added the corresponding typedefs
> (row_reference and const_row_reference) to their MatVecTraits.
> 
> This change has very little impact on your current code, as no types are
> actually changed, but allows you to build a DenseMatrix that has a
> single FieldVector as container for its values.
> 
> Despite the heavy use of FieldMatrix in the core modules, few lines
> actually obtain the row_type from the FieldMatrix and would have to be
> changed. Actually, the word row_type does not occur in any header of
> dune-grid or dune-localfunctions, but only in dune-common and dune-istl.
> In dune-common, the only place I did not change is exprtmpl.hh, but I
> guess this is not a big problem. In dune-istl, there are only two lines
> actually using row_type &, namely:
> 
> diagonalmatrix.hh:575:
> row_ = row_type(&(mat_->diagonal(i)), i);
> 
> io.hh:441:
> const typename MatrixType::row_type& row = matrix[rowIdx];
> 
> So nearly all current code makes use of explicit knowledge of what
> row_type and, hence, row_type &, actually is. Therefore, I think the
> only files that might need to be changed are those mentioned. However,
> for dune-istl, this might break code for other kinds of matrices that do
> not, yet, support a row_reference. Maybe we have to add this to
> dune-istl, too?
> 
> Since there has already been some discussion on this, I would like to
> hear your comments, opinions, and suggestions. It would also be great to
> know whether someone objects.
> 
> As the change is not very intrusive to existing codes, I will leave the
> commit in place until the discussion is over. I hope there are no
> objections to this.
> 
> Best,
> 
> Martin