multirhstest fails to compile when arpack++ is installed

See also !102 (merged).

When arpack++ is installed, multirhstest fails to compile: the condition estimate calculation can only handle double and float but it is instantiated here with a different type (AlignedNumber<double>). It should probably suppress its instantiation using Hybrid::ifElse(), the same way it suppressed instantiation of std::real() elsewhere.

@martin.nolte, @andreas.dedner: I've set the milestone to 2.6, since CGSolver is pretty important. At least at our university arpack is installed by default, and even if it isn't it is not uncommon for people to want to do eigenvalue computations with arpack in connection with dune.

On the other hand, besides the failing test the problem can only occur if you instantiate CGSolver with a non-double non-float type, so you may want to consider ignoring this problem for 2.6.0

mentioned in issue infrastructure/issues0#48 (closed)

@linus.seelinger: Could you provide a patch?

I tried to extend the use of Hybrid::ifElse() to the actual eigenproblem. However, this does not keep Arpack from being instantiated with the incompatible type. Is there a way to prevent the relevant section from being instantiated for unsuitable types?

(Disclaimer: I don't know the specific use case.). If the condition is given statically, i.e. as std::integral_constant<bool, x>, then the non-matching branch is not instantiated but non-dependent expressions will be. To avoid this you can pass values through the identity function provided as callback argument. Since this function is only known after instantiation of the callback it allows to delay other instantiations until then:

Hybrid::ifElse(isSuitable(foo), [&](auto id) {
  id(foo).bar();
});

You're probably going to have to do this for types as well. E.g. in the Arpack-section you have a few uses of the type real_type. You can do

using dep_real_type = decltype(id(std::declval<real_type>()));

and use dep_real_type instead. You may have to wrap the decltype(...) into std::decay_t<...>.

You probably want to move the typedefs COND_MAT and COND_VEC into the Arpack-section, and use dep_real_type to define them.

You probably want to change the loop over int row to

        for (auto row = id(0); row < i; ++row) {

so sqrt(betas[row-1]) becomes a dependent expression and the compiler will defer fully resolving sqrt(). Likewise for the divisions.

@carsten.graeser , @joe Thank you very much for your help!

Weird enough, it seems to get this section to work I only need ifElse and no id's after all. I don't think anything about my setup changed though... Maybe I was thrown off by an error for having to access _verbose via this-> now.

Before merging, could someone with Arpack installed please confirm that this works fine now?

created branch 40-multirhstest-fails-to-compile-when-arpack-is-installed

mentioned in merge request !161 (merged)

I'm on holidays until mid of January, I probably won't check before I'm back.

closed via commit 9d012149

closed via merge request !161 (merged)

mentioned in commit 80a9182c

mentioned in merge request !165 (merged)

mentioned in commit 9d012149