Why not const for Coefficient member functions?

[j-signorelli]

Hello,

There have been many times while using MFEM that I have wanted to pass a Coefficient, VectorCoefficient, or MatrixCoefficient as const to a class such as an operator, where I want to be explicit in the code that the class cannot alter the coefficient object but may use it.

However, I have noticed that all of the member functions of those three base classes (minus some additional ones in MatrixCoefficient) are not declared as const, and was wondering if there was any particular reason for this? For example, I don't immediately see how calling Eval generally may change the state of these coefficient objects, yet it is still not const. My guess is that this is why anything that uses a coefficient such as GridFunction::Project___ and LinearForm/BilinearFormIntegrator constructors must take a non-const coefficient reference.

I was just wondering if there was any particular reason for this that I am missing?

Thank you!

[v-dobrev]

One perspective: using a Coefficient may require calling its SetTime() method which does alter the Coefficient state, so time-dependent Coefficients need to be passed as non-const.

That said, I think that this interface dates back to the first MFEM release when the const-ness of methods was not considered thoroughly.

It is probably possible to change these methods to be const -- I'm not sure if this is a simple change local to the coefficient classes or if it needs to propagate to other classes too. If you feel strongly about it and you want to explore the implication of a such change, feel free to try it, and propose a PR if the changes are not too pervasive. The main thing will be to ensure that the changes are backward-compatible.

[j-signorelli]

Got it, thank you! There definitely would be a good bit of propagation through (like with methods for projecting coefficients and coefficient arguments for integrators) but I am interested in giving it a go and seeing how pervasive it is. I can't immediately think of any issues with backward compatibility but will keep that in mind.

[paul-hilscher]

I am interested in giving it a go and seeing how pervasive it is.

that would be amazing ! Currently I am using const_cast when calling Eval in my code and would love to see this gone. Let me know if I can help in any way.

[j-signorelli]

PR submitted here: #4196 @paul-hilscher @v-dobrev