What is the semantics of binding equations in records?

[casella]

Consider the following models

package P
  record R
    Real x;
    Real y = 1/x;
  end R;

  function f
    input Real t;
    output R r;
  algorithm
    r.x = t;
    r.y = 1;
  end f;

  function g
     input R r;
     output Real y;
   algorithm
     y := r.x + r.y;
  end g;

model M1
  R r1(x = 2);
  R r2;
equation
  r2.x = 1 + time;
end M1;

model M2
  R r3;
equation
  r3 = f(time);
end M2;

model M3
  Real y;
equation
  y = g(f(0));
end M3;
end P;

• It seems quite clear to me that M1 is valid and gives r1.x = 2, and r2.y = 1/(1+time)
• It is not clear to me if M2 is valid.
  Q1: can the algorithm in f override the binding of r.y? (A1:probably not)
  Q2: is it possible to assign the whole record in an equation, given that one element of the record has a binding? (A2: probably not)
• It is even less clear to me if M3 is valid. Dymola compiles it, even though the binding in the record is overridden by the assignment in the algorithm. But then it fails at runtime because of a division by zero.

This discussion is motivated by some issues we're having with the HelmholtzMedia library, see OpenModelica PR 7023.

IMHO if we allow to assign records (or parts thereof) that have binding equations within function algorithms, we need to make exceptions to the fundamental idea that a binding equation is an equation by definition, not an assignment that can be overridden. Then, we would probably get down a rabbit hole of infinite discussions that may never end.

Personally, I would simply suggest to forbid doing that.

What do you think?

[qlambert-pro]

To me M1 and M3 are valid, and M2 is overdetermined. The semantics of binding equation in functions is specified in Section 12.4.4, and clearly states that they are used for initialisation. As far as I understand the following is valid, I think there should not be a special rule for records:

model Example
   function f
       input  Real i;
       output Real o = i;
   algorithm
       if o < 0 then o:=0; end if;
   end f;
Real x = f(0);
end Example;

[sjoelund]

To me M2 is not overdetermined because the initialization rules for functions state that it's an error if the binding of y.y cannot be resolved, so it won't pass the front-end. M3 has the same problem.

I think this is missing some other example of a case where all the bindings in the record are present and thus would be valid.

[qlambert-pro]

To me M2 is not overdetermined because the initialization rules for functions state that it's an error if the binding of y.y cannot be resolved, so it won't pass the front-end.

You probably meant r.y, I didn't see that, but that looks like a valid reason to reject both M2 and M3.

[sjoelund]

You probably meant r.y, I didn't see that, but that looks like a valid reason to reject both M2 and M3.

Correct. I read the output from g instead of f.

[eshmoylova]

In MapleSim, M1 works. M2 and M3 fail because the binding equation for r.y is used to initialize it, and no initial value for r.x is given at this point. We give an error that that r.x is used with no assigned value. I believe it is consistent with Section 12.4.4.

[sjoelund]

@eshmoylova do you give an error also if r.x has a start-value?