The NF should not evaluate final record parameters unless they only depend on structural or final parameters

[Francesco Casella]

Please consider the Modelica.Mechanics.MultiBody.Examples.Elementary.HeatLosses model. The final parameter body1.R_start is currently constant-evaluated, to avoid the bindings ending up in the initial equation sections, where the back-end doesn't handle them correctly.

However, the Body model contains the following declaration:

  final parameter Frames.Orientation R_start=
      Modelica.Mechanics.MultiBody.Frames.axesRotations(
        sequence_start,
        angles_start,
        zeros(3))
    "Orientation object from world frame to frame_a at initial time";

where angles_start is not final, nor it has an Evaluate = true annotation. Hence, R_start should not be constant evaluated, and it should be possible to get it automatically updated if angles_start is changed at run time, without the need to recompile the model from scratch. In fact, that's exactly what Dymola does.

Notice that with the current new front-end and backend, simply moving the final binding to the initial equation section produces incorrect results, see the original bug report in #5288.

[Per Östlund]

Changing to backend, since the only reason the NF evaluates the binding is to work around backend issues. All that's needed to stop evaluating it and instead move it to an initial equation is to revert the small change in ​e11c6c60 (the ​results of that commit might be of interest to anyone who wants to try).

Edit: Updated link due to repository changes.

[Francesco Casella]

@Karim, feel free to reassign it to Andreas.

[Per Östlund]

It seems I accidentally fixed HeatLosses in 632d2164, see ​report (there are some regressions, but I was expecting worse and will fix those later). body1.R_start is still evaluated though, so I guess it somehow got fixed when I improved the handling of record expressions in the backend.

So the model is fixed but the ticket is still valid in that the compiler probably shouldn't evaluate that parameter, and disabling the evaluation in the frontend still causes simulation issues for some models.