STL-friendly pImpl class?

I am maintaining a project that can take a considerable time to build so am trying to reduce dependencies where possible. Some of the classes could make use if the pImpl idiom and I want to make sure I do this correctly and that the classes will play nicely with the STL (especially containers.) Here is a sample of what I plan to do - does this look OK? I am using std::auto_ptr for the implementation pointer - is this acceptable? Would using a boost::shared_ptr be a better idea?

Here is some code for a SampleImpl class that uses classes called Foo and Bar:

// SampleImpl.h
#ifndef SAMPLEIMPL_H
#define SAMPLEIMPL_H

#include <memory>

// Forward references
class Foo;
class Bar;

class SampleImpl
{
public:
    // Default constructor
    SampleImpl();
    // Full constructor
    SampleImpl(const Foo& foo, const Bar& bar);
    // Copy constructor
    SampleImpl(const SampleImpl& SampleImpl);
    // Required for std::auto_ptr?
    ~SampleImpl();
    // Assignment operator
    SampleImpl& operator=(const SampleImpl& rhs);
    // Equality operator
    bool operator==(const SampleImpl& rhs) const;
    // Inequality operator
    bool operator!=(const SampleImpl& rhs) const;

    // Accessors
    Foo foo() const;
    Bar bar() const;

private:
    // Implementation forward reference
    struct Impl;
    // Implementation ptr
    std::auto_ptr<Impl> impl_;
};

#endif // SAMPLEIMPL_H

// SampleImpl.cpp
#include "SampleImpl.h"
#include "Foo.h"
#include "Bar.h"

// Implementation definition
struct SampleImpl::Impl
{
    Foo foo_;
    Bar bar_;

    // Default constructor
    Impl()
    {
    }

    // Full constructor
    Impl(const Foo& foo, const Bar& bar) :
        foo_(foo),
        bar_(bar)
    {
    }
};

SampleImpl::SampleImpl() :
    impl_(new Impl)
{
}

SampleImpl::SampleImpl(const Foo& foo, const Bar& bar) :
    impl_(new Impl(foo, bar))
{
}

SampleImpl::SampleImpl(const SampleImpl& sample) :
    impl_(new Impl(*sample.impl_))
{
}

SampleImpl& SampleImpl::operator=(const SampleImpl& rhs)
{
    if (this != &rhs)
    {
        *impl_ = *rhs.impl_;
    }
    return *this;
}

bool SampleImpl::operator==(const SampleImpl& rhs) const
{
    return  开发者_如何学JAVAimpl_->foo_ == rhs.impl_->foo_ &&
        impl_->bar_ == rhs.impl_->bar_;
}

bool SampleImpl::operator!=(const SampleImpl& rhs) const
{
    return !(*this == rhs);
}

SampleImpl::~SampleImpl()
{
}

Foo SampleImpl::foo() const
{
    return impl_->foo_;
}

Bar SampleImpl::bar() const
{
    return impl_->bar_;
}

---

You should consider using copy-and-swap for assignment if it's possible that Foo or Bar might throw as they're being copied. Without seeing the definitions of those classes, it's not possible to say whether they can or not. Without seeing their published interface, it's not possible to say whether they will in future change to do so, without you realising.

As jalf says, using auto_ptr is slightly dangerous. It doesn't behave the way you want on copy or assignment. At a quick look, I don't think your code ever allows the impl_ member to be copied or assigned, so it's probably OK.

If you can use scoped_ptr, though, then the compiler will do that tricky job for you of checking that it's never wrongly modified. const might be tempting, but then you can't swap.

---

There are a couple of problems with the Pimpl.

First of all, though not evident: if you use Pimpl, you will have to define the copy constructor / assignment operator and destructor (now known as "Dreaded 3")

You can ease that by creating a nice template class with the proper semantic.

The problem is that if the compiler sets on defining one of the "Dreaded 3" for you, because you had used forward declaration, it does know how to call the "Dreaded 3" of the object forward declared...

Most surprising: it seems to work with std::auto_ptr most of the times, but you'll have unexpected memory leaks because the delete does not work. If you use a custom template class though, the compiler will complain that it cannot find the needed operator (at least, that's my experience with gcc 3.4.2).

As a bonus, my own pimpl class:

template <class T>
class pimpl
{
public:
  /**
   * Types
   */
  typedef const T const_value;
  typedef T* pointer;
  typedef const T* const_pointer;
  typedef T& reference;
  typedef const T& const_reference;

  /**
   * Gang of Four
   */
  pimpl() : m_value(new T) {}
  explicit pimpl(const_reference v) : m_value(new T(v)) {}

  pimpl(const pimpl& rhs) : m_value(new T(*(rhs.m_value))) {}

  pimpl& operator=(const pimpl& rhs)
  {
    pimpl tmp(rhs);
    swap(tmp);
    return *this;
  } // operator=

  ~pimpl() { delete m_value; }

  void swap(pimpl& rhs)
  {
    pointer temp(rhs.m_value);
    rhs.m_value = m_value;
    m_value = temp;
  } // swap

  /**
   * Data access
   */
  pointer get() { return m_value; }
  const_pointer get() const { return m_value; }

  reference operator*() { return *m_value; }
  const_reference operator*() const { return *m_value; }

  pointer operator->() { return m_value; }
  const_pointer operator->() const { return m_value; }

private:
  pointer m_value;
}; // class pimpl<T>

// Swap
template <class T>
void swap(pimpl<T>& lhs, pimpl<T>& rhs) { lhs.swap(rhs); }

Not much considering boost (especially for the cast issues), but there are some niceties:

• proper copy semantic (ie deep)
• proper const propagation

You still have to write the "Dreaded 3". but at least you can treat it with value semantic.

---

EDIT: Spurred on by Frerich Raabe, here is the lazy version, when writing the Big Three (now Four) is a hassle.

The idea is to "capture" information where the full type is available and use an abstract interface to make it manipulable.

struct Holder {
    virtual ~Holder() {}
    virtual Holder* clone() const = 0;
};

template <typename T>
struct HolderT: Holder {
    HolderT(): _value() {}
    HolderT(T const& t): _value(t) {}

    virtual HolderT* clone() const { return new HolderT(*this); }
    T _value;
};

And using this, a true compilation firewall:

template <typename T>
class pimpl {
public:
    /// Types
    typedef T value;
    typedef T const const_value;
    typedef T* pointer;
    typedef T const* const_pointer;
    typedef T& reference;
    typedef T const& const_reference;

    /// Gang of Five (and swap)
    pimpl(): _holder(new HolderT<T>()), _p(this->from_holder()) {}

    pimpl(const_reference t): _holder(new HolderT<T>(t)), _p(this->from_holder()) {}

    pimpl(pimpl const& other): _holder(other->_holder->clone()),
                               _p(this->from_holder())
    {}

    pimpl(pimpl&& other) = default;

    pimpl& operator=(pimpl t) { this->swap(t); return *this; }

    ~pimpl() = default;

    void swap(pimpl& other) {
        using std::swap;
        swap(_holder, other._holder);
        swap(_p, other._p)
    }

    /// Accessors
    pointer get() { return _p; }
    const_pointer get() const { return _p; }

    reference operator*() { return *_p; }
    const_reference operator*() const { return *_p; }

    pointer operator->() { return _p; }
    const_pointer operator->() const { return _p; }

private:
    T* from_holder() { return &static_cast< HolderT<T>& >(*_holder)._value; }

    std::unique_ptr<Holder> _holder;
    T* _p;           // local cache, not strictly necessary but avoids indirections
}; // class pimpl<T>

template <typename T>
void swap(pimpl<T>& left, pimpl<T>& right) { left.swap(right); }

---

I've been struggling with the same question. Here's what I think the answer is:

You can do what you are suggesting, so long as you define the copy and assignment operators to do sensible things.

It's important to understand that the STL containers create copies of things. So:

class Sample {
public:
    Sample() : m_Int(5) {}
    void Incr() { m_Int++; }
    void Print() { std::cout << m_Int << std::endl; }
private:
    int m_Int;
};

std::vector<Sample> v;
Sample c;
v.push_back(c);
c.Incr();
c.Print();
v[0].Print();

The output from this is:

6
5

That is, the vector has stored a copy of c, not c itself.

So, when you rewrite it as a PIMPL class, you get this:

class SampleImpl {
public:
    SampleImpl() : pimpl(new Impl()) {}
    void Incr() { pimpl->m_Int++; }
    void Print() { std::cout << m_Int << std::endl; }
private:
    struct Impl {
        int m_Int;
        Impl() : m_Int(5) {}
    };
    std::auto_ptr<Impl> pimpl;
};

Note I've mangled the PIMPL idiom a bit for brevity. If you try to push this into a vector, it still tries to create a copy of the SampleImpl class. But this doesn't work, because std::vector requires that the things it store provide a copy constructor that doesn't modify the thing it's copying.

An auto_ptr points to something that is owned by exactly one auto_ptr. So when you create a copy of an auto_ptr, which one now owns the underlying pointer? The old auto_ptr or the new one? Which one is responsible for cleaning up the underlying object? The answer is that ownership moves to the copy and the original is left as a pointer to nullptr.

What auto_ptr is missing that prevents its use in a vector is copy constructor taking a const reference to the thing being copied:

auto_ptr<T>(const auto_ptr<T>& other);

(Or something similar - can't remember all the template parameters). If auto_ptr did provide this, and you tried to use the SampleImpl class above in the main() function from the first example, it would crash, because when you push c into the vector, the auto_ptr would transfer ownership of pimpl to the object in the vector and c would no longer own it. So when you called c.Incr(), the process would crash with a segmentation fault on the nullptr dereference.

So you need to decide what the underlying semantics of your class are. If you still want the 'copy everything' behaviour, then you need to provide a copy constructor that implements that correctly:

    SampleImpl(const SampleImpl& other) : pimpl(new Impl(*(other.pimpl))) {}
    SampleImpl& operator=(const SampleImpl& other) { pimpl.reset(new Impl(*(other.pimpl))); return *this; }

Now when you try to take a copy of a SampleImpl, you also get a copy of its Impl struct, owned by the copy SampleImpl. If you're taking an object that had lots of private data members and was used in STL containers and turning it into a PIMPL class, then this is probably what you want, as it provides the same semantics as the original. But note that pushing the object into a vector will be considerably slower as there is now dynamic memory allocation involved in copying the object.

If you decide you don't want this copy behaviour, then the alternative is for the copies of SampleImpl to share the underlying Impl object. In this case, it's not longer clear (or even well-defined) which SampleImpl object owns the underlying Impl. If ownership doesn't clearly belong in one place, then std::auto_ptr is the wrong choice for storing it and you need to use something else, probably a boost template.

Edit: I think the above copy constructor and assignment operator are exception-safe so long as ~Impl doesn't throw an exception. This should always be true of your code anyway.