Reader Q&A: How can I prevent a type from being instantiated on the stack?

« C++ and Beyond ‘Encore’ in 2014: Sep 29 – Oct 1, Stuttgart, Germany

Reader Q&A: How can I prevent a type from being instantiated on the stack?

2014-05-03 by Herb Sutter

Anubhav asked:

An interesting question has come up in our project while debating operator new as a class member function. Specifically, the question is about whether such a class should be allowed to be instantiated on stack. The understanding is that a class providing its own operator new would likely have special layout considerations which would not be met if the object of such a class is instantiated locally.

The class specific operator new is used when allocating on the heap, but objects can still be allocated on the stack. The only way I know to prevent an object from being instantiated on the stack is by forcing creation through a factory function that always allocates on the heap, for example:

#include <memory>

class X {
    // make all ctors private
    X();
    X(const X&);
    // and any other ctors go here

public:
    // Factory function (or it could be a nonmember friend)
    static auto make( /*…*/ ) { return std::unique_ptr<X>(new X( /*…*/ )); }

    // the usual public member functions go here
};

int main()
{
    X x;                  // error 
    auto x2 = X::make();  // ok
}

Posted in C++ | 9 Comments

9 Responses

on 2014-06-17 at 6:22 am

Victor

My idea is a wrapper around. Is there any issues?

template <typename Wrapped> class HeapOnly 
{
public:
	typedef std::unique_ptr<Wrapped> TPtr;
	typedef const std::unique_ptr<Wrapped> TConstPtr;
	
private:
    TPtr m_wrapped;

public:

    operator TPtr&      ()       { return m_wrapped; }
    operator TConstPtr& () const { return m_wrapped; }

    HeapOnly(/* sorry, but I'm unfamiliar with variable templates so far...*/) 
        : m_wrapped( std::make_unique<Wrapped>(/*...*/) )
    {}
};

class A {
private:
    A() {} // private ctors

	// ...
public:
    // ... others 
    friend class HeapOnly<A>;
};

void f()
{
     HeapOnly<A> compilationOk;
     A compilationError;
}

on 2014-06-04 at 7:43 pm Teddy

@Grout: Deriving from enable_shared_from_this doesn’t actually block compilation of code that instantiates the object on the stack. Nor does it block compilation of new Foo(). I have found the PrivateConstructorTag to be the best way to ensure that a class that inherits from enable_shared_from_this is only constructed with make_shared().
on 2014-05-09 at 2:51 pm Grout

If
```
make_shared()
```
is the plan anyway, then deriving from
```
std::enable_shared_from_this<>
```
may be adequate. No?
on 2014-05-04 at 10:55 pm Norbert Riedlin

Hi. Does std::make_unique have the same problems with the private constructor pattern as std::make_sharedcode> have? I.e. that it is not portably possible to make std::make_shared a friend of X in the example. What is the preferred way to solve this problem? Is the use of the PrivateConstructorTag the only way?

on 2014-05-04 at 10:41 am

Zenju

Correction:

	X(..., PrivateConstructionTag);
public:

should be

public:
	X(..., PrivateConstructionTag);

on 2014-05-04 at 10:31 am Zenju

@Good point, thanks. Fixed.

But this also gives up on the perf and memory layout advantages of std::make_shared… In library code this could be premature pessimisation. What about:

class X
{
//make all ctors private

struct PrivateConstructionTag {};

public:
static auto make( /*…*/ ) { return std::make_shared( /*…*/ PrivateConstructionTag()); }
};

I’m using this pattern in my library code. Is this the best solution?
on 2014-05-04 at 4:31 am Herb Sutter

@mttpd: The protected destructor solution also works, but it’s a C++98 idiom that’s now inferior because it requires using a custom ‘deleter’ function that is possible, but less natural and idiomatic in modern C++, to use together with unique_ptr and shared_ptr.

@Petter: Good point, thanks. Fixed.
on 2014-05-04 at 3:46 am Petter

But std::make_unique also requires access to the constructor, right? X::make could create the unique_ptr manually instead.
on 2014-05-04 at 2:15 am mttpd

How about the protected destructor solution?
For instance, as suggested here:
http://en.wikibooks.org/wiki/More_C%2B%2B_Idioms/Requiring_or_Prohibiting_Heap-based_Objects