DirichletBC.dof_indices() are wrong for blocked spaces

[jorgensd]

For blocked function spaces, locate_dofs_topological and DirichletBC assumes blocked indices.
However, the private functions num_owned assumes that the dofs are unrolled, which makes the number of owned indices returned by DirichletBC::dof_indices wrong.

[chrisrichardson]

Looks like bs is being applied twice at the moment (is that right?), once in num_owned and then again in constructor

[jorgensd]

Looks like bs is being applied twice at the moment (is that right?), once in num_owned and then again in constructor

It is being applied twice, but it also assumes that the initial dofs are unrolled, meaning that it captures too many degrees of freedom as owned. So there are two different things that can cause issues at the moment.

[michalhabera]

Good catch. But what is _owned_indices0 used for apart from returning in dof_indices()? Seems to me a bit weird to return it. With your fix it will return unrolled dofs and number of not unrolled owned dofs, so a bit inconsistent.

[jorgensd]

Good catch. But what is _owned_indices0 used for apart from returning in dof_indices()? Seems to me a bit weird to return it. With your fix it will return unrolled dofs and number of not unrolled owned dofs, so a bit inconsistent.

It still returns the number of unrolled dofs, as it is multiplied by bs.
I caught this in: #3938 where I require it to set boundary conditions on the Python side of things (to the local PETSc vector)

[michalhabera]

Good catch. But what is _owned_indices0 used for apart from returning in dof_indices()? Seems to me a bit weird to return it. With your fix it will return unrolled dofs and number of not unrolled owned dofs, so a bit inconsistent.

It still returns the number of unrolled dofs, as it is multiplied by bs. I caught this in: #3938 where I require it to set boundary conditions on the Python side of things (to the local PETSc vector)

Ah right, missed the multiplication later somewhere in constructor... This is good to merge as a bugfix, but unrelated to this PR, returning an array and number of owned elements in the array is not a common dolfinx pattern, isn't it? We've always relied on user figuring this out.