Forward declaration & incomplete template type

[verec]

Consider a first version:

--- drawable.hpp ---
#include "gcdata.hpp"
struct drawable {
...
virtual int internal_new_GC(gcdata * gcd) = 0 ;
} ;

--- gcdata.hpp ---
#include "device.hpp"
struc gcdata {
...
device * d ;
} ;

--- device.hpp ---
#include "drawable.hpp"
struct device : public drawable {
...
}

Obviously, this doesn't compile, because the "include stack",
at the point where ``struct device : public drawable'' is
reached, looks like
...drawable.hpp
.... gcdata.hpp [drawable.hpp line 2]
..... device.hpp [gcdata.hpp] line 2]
which means that. at the point where the ``struct device'' is
reached [device.hpp line 3], the struct drawable has not been
defined yet.

The fix is to rewrite gcdata with a forward declaration:
--- gcdata.hpp ---
// do NOT #include "device.hpp"
struct device ;
struc gcdata {
...
device * d ;
} ;

So far, so good.

Now, I'm not using direct pointers, but "smart" pointers

--- drawable.hpp ---
#include "gcdata.hpp"
struct drawable {
...
virtual int internal_new_GC(GCData gcd) = 0 ;
} ;
typedef envelope<drawable> Drawable ;

--- gcdata.hpp ---
#include "device.hpp"
struc gcdata {
...
Device d ;
} ;
typedef envelope<gcdata> GCData ;

--- device.hpp ---
#include "drawable.hpp"
struct device : public drawable {
...
}
typedef envelope<device> Device ;

And now I'm stuck, because the envelope template (which
is extremely close to boost::shared_ptr/bost_instrusive_ptr)
has this requirement that T must be a complete type, not a forward
declaration.

Hence, the "fix"

--- gcdata.hpp ---
// do NOT #include "device.hpp"
struct device ;
typedef envelope<device> Device ;
struc gcdata {
...
Device d ;
} ;

does NOT compile ...

Apart from tinkering with envelope<T> so that it doesn't
require a complete type (which may or may not be possible),
anyone sees a way out of this conundrum?

[BTW: this is a port of Java code: I can change the design
in only very minor ways, so the requirement that device
inherits from drawable that uses gcdata that needs device
has to stay]

Many Thanks

 

[Victor Bazarov]

[... forward-declaration doesn't work when instantiating templates...]
Apart from tinkering with envelope<T> so that it doesn't
require a complete type (which may or may not be possible),
anyone sees a way out of this conundrum?

Nope. You're truly stuck. Let's put your code in one module:
---------------------------------------------------------------
template<class A> class TA {}; // your 'envelope'

struct Data; // forward-declare it...

struct Base { // OK, let's try to define it
void foo(TA<Data>); // OOPS! not allowed
};

struct Derived : Base { }; // OK, need it here

class Data {
TA<Base> tb; // OK, Base and Derived are defined
};
---------------------------------------------------------------
Oh, wait, maybe we could rearrange it a bit...
---------------------------------------------------------------
template<class A> class TA {}; // your 'envelope'

struct Base; // forward-declare it...

class Data {
TA<Base> tb; // OOPS! Can't do that
};

struct Base { // OK, let's try to define it
void foo(TA<Data>); // OK, Data is defined
};

struct Derived : Base { }; // OK
---------------------------------------------------------------

No matter how you move things around, you're faced with something
that would require an instantiation of a template from a type that
hasn't been defined yet.

You will need to go back to a pointer or a reference in one of
those classes. That means redesign.

Are you sure that you need to pass an "envelope" to the function
of your 'drawable' class?

[..]

V

 

[verec]

[... forward-declaration doesn't work when instantiating templates...]
Apart from tinkering with envelope<T> so that it doesn't
require a complete type (which may or may not be possible),
anyone sees a way out of this conundrum?

[...]
No matter how you move things around, you're faced with something
that would require an instantiation of a template from a type that
hasn't been defined yet.

First, thank you Victor, for taking the time to understand the issue.

Are you sure that you need to pass an "envelope" to the function
of your 'drawable' class?

Well ... yes. The choices I seem to be left with are:

1. redesign the code so as to eliminate circular dependencies.
2. use raw pointers in at least parts of the code
3. rework envelope

1. is next to impossible. I'm porting, not redesigning a whole
system from scratch. That would completely change the nature
of the project. Given that the port itself, without redesign,
is already a huge task, choosing this path would mean the
project death. I simly do not have the 6 years ahead of me
to complete it

2. If #3 fails, I will very reluctantly pursue this path.

3. Seems the least bad compromise. I tried:

--- gcdata.hpp ---
template <typename T> struct Test {
T * body ;
Test(T * q = 0) : body(q) {}
} ;

struct device ;
// typedef envelope<device> Device ;
typedef Test<device> TDevice ;
struct gcdata {
// Device dev ;
TDevice tdev ;
} ;

typedef envelope<gcdata> GCData ;

And this does compile, probably because the Test<T> template only
deals with T * and never with T's. But I fear that the conversion
of envelope<T> might not be as simple ...

Many thanks for your time