This enables users to have several build directories like
build-clang, build-gcc5, build-debug, and still get all
of them ignored.

There are some tests, which are hard to port in the paamg directory.
They want to explicitly test _without_ direct solver. Previously,
you simply did not add the flags for the direct solver packages.
add_dune_test relies on all flags, so we need a way to disable components
again. That will be implemented as part of dune_add_executable. For
now, we build the targets ourselves and hand them to dune_add_test.

This partially fixes FS#1724.

Similar to commit e8d16384 which
added the same for UMFPack.
This avoids strange bugs when using the setSubMatrix method: that method
automatically removes rows and columns from the matrix, and expects
_the user_ to do the same for the x and b vectors.

Let FindPackageHandleStandardArgs handle missing dependencies.
Mark ARPACKPP_LIBRARY as advanced.
Remove dead code.

It's confusing and against our style-guide.

The arpack++ package from Debian comes with a shared library that we
have to link to. Otherwise the linker will complain about missing
symbols when we try to build arpackpptest.
As there seem to be other versions of arpack++ that are header-only, we
only link to the shared library if it was found.

These header are no longer used.

The commit broke compiling / linking if one package was found but
the includes / libraries were not added. We need a different
approach to hide the ENABLE_* trick from the user and keep full
testing capabilities.
This reverts commit 27613560.

This avoids strange bugs when using the setSubMatrix method: that method
automatically removes rows and columns from the matrix, and expects _the user_
to do the same for the x and b vectors.

This a littel more natural then int. Furthermore it allows
to use the index constants from dune-typetree.

This a littel more natural then int. Furthermore it allows
to use the index constants from dune-typetree.

In commit 3e405a0b a parameter 'eps'
was introduced for testAggregate(). However this parameter is not
supplied in the call (breaking the test) and it is never used
in the method.

Boost is no longer needed in dune-istl.

[release] Merge fix for handling 64 bit IDXTYPEWIDTH in parmetis>3

Tested with gcc-4.9.3 and clang-3.4.  Kinda tricky to make them both happy...

They do.

Well, at least they compile :-)  Don't know yet whether they actually work.

The new modules from the arpack feature branch were
never installed and this would have broken installed dune-istl
modules.

With this commit they now are installed for both CMAke and autotools.

the old code didn't work (at least not with clang). The reason is that constructs like
template<typename... Arg1, typename... Arg2> don't work, as the  compiler doesn't know
where to split the list of arguments. The whole algo_itsteps got reworked to allow for an
easier specialization for MultiTypeBlockMatrix. Now the struct is also parametrized by the
matrix type. By this we can easily add a specialization.

... thus we can always run the multitype examples

we have to preceed the variadic parameter by T1, in order to fix the deduction. The problem is
that (for what ever reason) otherwise the type pack can not be detected properly.

We use the same trick with std::integral_constant as in the MultiTypeBlockVector code.