Consider the following function:

    int test(int a, int b)
    {
      int c = a/b;
      if (f1())
        f2(a,b,c);
    }

Should a compiler be required to compute c before calling f1, and thus have to store the value of c across the function call?

Better would be to define a set of semantics for loosely-sequenced traps, along with "causality barriers" to ensure that they only occur at tolerable times.