Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.
Comment: minor editorial changes

...

Pointer downcasting to a pointer of incomplete class type has similar caveats. Pointer upcasting (casting from a more derived type to a less derived type) is a standard implicit conversion operation. C++ allows static_cast to perform the inverse operation, pointer downcasting, via [expr.static.cast], paragraph 7. However, when the pointed-to type is incomplete, the compiler is unable to make any class offset adjustments that may be required in the presence of multiple inheritanceinheritances, resulting in a pointer that cannot be validly dereferenced.

The reinterpret_cast of a pointer type is defined by [expr.reinterpret.cast], paragraph 7, as being static_cast<cv T *>(static_cast<cv void *>(PtrValue)), meaning that reinterpret_cast is simply a sequence of static_cast operations. C-style casts of a pointer to incomplete object type are defined as using either static_cast or reinterpret_cast (which is picked is unspecified) in [expr.cast], paragraph 5.

...

In this compliant solution, the deletion of impl is moved to a part of the code where Body is defined:.

Code Block
bgColor#ccccff
langcpp
class Handle {
  class Body *impl;  // Declaration of a pointer to an incomplete class
public:
  ~Handle();
  // ...
};

// Elsewhere
class Body { /* ... */ };
 
Handle::~Handle() {
  delete impl;
}

...

In this compliant solution, a std::shared_ptr is used to own the memory to impl. Note that a  A std::shared_ptr is capable of referring to an incomplete type, but a std::unique_ptr is not.

...

Code Block
bgColor#FFcccc
langcpp
// File1.h
class B {
protected:
  double d;
public:
  B() : d(1.0) {}
};
 
// File2.h
void g(class D *);
class B *get_d(); // Returns a pointer to a D object

// File1.cpp
#include "File1.h"
#include "File2.h"

void f() {
  B *v = get_d();
  g(reinterpret_cast<class D *>(v));
}
 
// File2.cpp
#include "File2.h"
#include "File1.h"
#include <iostream>

class Hah {
protected:
  short s;
public:
  Hah() : s(12) {}
};

class D : public Hah, public B {
  float f;
public:
  D() : Hah(), B(), f(1.2f) {}
  void do_something() { std::cout << "f: " << f << ", d: " << d << ", s: " << s << std::endl; }
};

void g(D *d) {
  d->do_something();
}

B *get_d() {
  return new D;
}

Implementation Details

When compiled with Clang 3.8 and the function f() is executed, the noncompliant code example prints the following:.

Code Block
f: 1.89367e-40, d: 5.27183e-315, s: 0

...

This compliant solution assumes that the intent is to hide implementation details by using incomplete class types. Instead of requiring a D * to be passed to g(), it expects a B * type:.

Code Block
bgColor#ccccff
langcpp
// File1.h -- contents identical.
// File2.h
void g(class B *); // Accepts a B object, expects a D object
class B *get_d(); // Returns a pointer to a D object

// File1.cpp
#include "File1.h"
#include "File2.h"

void f() {
  B *v = get_d();
  g(v);
}
 
// File2.cpp
// ... all contents are identical until ...
void g(B *d) {
  D *t = dynamic_cast<D *>(d);
  if (t) {
    t->do_something();
  } else {
    // Handle error
  }
}

B *get_d() {
  return new D;
}

...