Re: %union errors that shouldn't be there

[Hans Aberg]

At 14:16 +0100 2005/03/23, Laurence Finston wrote:
>  > The union does not
> >  contain any type information which field is selected. If one adds that,
> >  unions with non-trivial con-/de-Structors would be possible.
>
> Add it where?


Just add a field, invisible to the user, with the type information. 
Dynamic allocations must have such a field with the size, so that it 
can be properly deallocated. So it nothing strange as such. In fact, 
one can do thi by hand:

class Union {
public:
  typedef Type int;

  union Value {
    ...
  };

  Type type;
  Value Value;
};

Then write the class Union as to hide away the constructor stuff. 
Perhaps Bison should support such a version. Then one does not need 
to use the %destructor stuff. The drawback is the extra memory each 
int takes on the parser stack.

> I suspect that doing so in C++ would break compatibility
> to C.

Not really, as C and C++ are different languages. One would need to 
be aware that the C++ unions

> I haven't checked whether any C++ implementation has
> implemented such a facility as an extension, though.

Doubt it. The C++ paradigm is to use a class hierarchy instead.

> It might be possible to implement a way of using class types with
> constructors and destructors in a `%union' in Bison, but not with an
> unextended C or C++ union.  This is just my opinion, but I think it's
> simpler to just use `void*'.

So you have a suggestion of an extended union above.
--
  Hans Aberg