Martin Kronbichler

> In my opinion, there are way too many new test cases in this PR. I think that we should considerably increase the number of tests, given that many combinations...

This is a somewhat delicate topic. The problem is that inner faces create additional dependencies, both for the ghost value update and the scheduling of operations before/after the loop over...

> I wonder right now what periodic BCs would mean for structure problems? Wouldn't the system matrix be singular in this case so that the problem is not solvable? Isn't...

> What I do not understand is why the operator evaluation is affected by this. During initialization, I thought that MatrixFree receives both AffineConstraints objects and then evaluates the constraints...

I had a close look at the code now. I think the new approach looks good. For the simplices, we do not have the equivalent of `MappingQCache` anyway, so all...

Let me add that I was more skeptical against `dealii::PostProcessor` some years ago (because part of the infrastructure is so incredibly slow for high orders), but I would like to...

I must admit I do not yet understand the complete picture. We don't call `AffineConstraints::distribute` after solving the linear system because most of our data access is done via `FEEvaluation`...

We did not use continuous FEM very often yet; so in that sense this fix is very much appreciated.

This appears to be a tricky problem. I agree that we should not initialize the `__m128` data object because many places depend upon it in terms of performance. I was...

Thank you for the detailed list. I have another idea: What if we use the following declaration of `Tensor`: ```c++ class alignas(/* some constexpr that is either 1, 2 or...