Binary-compatible C++ Interfaces

文章目錄

• Chad Austin, 2002.02.15

 

http://chadaustin.me/cppinterface.html

 

Chad Austin, 2002.02.15Updates2003.02.21

Clarified some of my comments thanks to feedback and suggestions from
Razvan Surdulescu. For actual COM compatibility, I added __stdcall to
the method declarations as well. (Apparently, if you follow the
guidelines of this article, it's easy to add bindings from your DLL to
Delphi and VB!)

2002.04.03

Somebody (unfortunately, I lost his e-mail address and name before I
could write it down. if you're reading this or you know who sent it,
please drop me a line) sent me another guideline. The gist of it is
that you shouldn't use overloaded methods in your interfaces.
Different compilers will order them in the vtable differently.

2002.03.27

Ben Scott (bscott at iastate dot edu) submitted some excellent classes
that simplify usage of the concepts presented in this article. Simply
derive your interface classes from DLLInterface and your
implementations from DLLImpl! Download the source file here
.

Overview

This article explains how to create C++ DLL APIs that will work across several
compilers and configuration settings (Release, Debug, etc.).

Background

Many platforms have an
ABI
for their preferred
programming language. For example, BeOS's primary language is C++, so the C++
compiler must be able to generate code that remains binary compatible with the
operating system's C++ system calls (and classes, etc.).

The Windows API
and
ABI were defined for C, so C++ compiler writers had free reign to implement the
C++ ABI however they felt. Eventually, however, Microsoft created an
object-oriented ABI for Windows called COM. To simplify COM usage, they made
the vtables of their C++ ABI match the vtables required in COM interface.
Since a Windows compiler that can't use COM is pretty limited, other compiler
vendors enforced the mapping between COM vtables and C++ vtables.

There are several aspects to an ABI. This article only discusses the issues
with using C++ in Windows. Other platforms have different requirements.
(Fortunately, since most other platforms aren't as popular as Windows, they
have only one or two compilers, and thus there isn't much of a problem.)

Concepts

• ABI
  -
  Application Binary Interface.
  The binary interface between systems. If a binary interface changes, both
  sides of the interface (the user and the implementation) must be recompiled.
• API
  -
  Application Program Interface.
  The source interface between systems. If a source interface changes, code
  that uses that interface must be modified. API changes usually imply ABI
  changes.
• Interface
  -
  A class where every method is pure virtual, and thus has no inherent
  implementation. An interface is merely a protocol for communication between
  objects.
• Factory
  -
  Something that creates objects. In this article, we'll use a single global
  function as our factory.
• DLL Boundary
  - The line between code instantiated
  in a DLL and code in a calling process is called the DLL boundary.
  In some cases, code can be on both sides of the boundary: Consider
  an inline function in a header file that gets used in the DLL and
  the executable. The function is actually instantiated on both sides
  of the boundary. Therefore, if the inline function has a static
  variable, two variables will be created, one in the executable and
  one in the DLL, and which is used depends on whether the code in the
  DLL or the executable is calling the function.

Initial Attempt

Let's say you want to create a portable windowing API and you want to stick
the implementation in a DLL. I'm going to create a class called Window
which can represent a window in several different windowing systems: Win32,
MFC, wxWindows, Qt, Gtk, Aqua, X11, Swing (*gasp*), etc... We'll walk through
several attempts at creating an interface until it works across different
implementations, compilers, and compiler settings.

// Window.h

#include <string>

#ifdef WIN32
  #ifdef EXPORTING
    #define DLLIMPORT __declspec(dllexport)
  #else
    #define DLLIMPORT __declspec(dllimport)
  #endif
  #define CALL __stdcall
#else
  #define DLLIMPORT
  #define CALL
#endif

class DLLIMPORT Window {
public:
  Window(std::string title);
  ~Window();

  void setTitle(std::string title);
  std::string getTitle();

  // ...

private:
  HWND m_window;
};

I'm not going to show the implementation, as I'm assuming you already know how
to do that. There is one glaring problem with this interface: It assumes
you're using the basic Win32 API. That is, it holds an HWND as a
private member, which introduces a dependency between our Window class
and the Win32 SDK. One possible solution is to use the pImpl idiom to
remove the class's private members from the class definition.
You can read more about that elsewhere
[1]
,
[2]
,
[3]
, and
[4]
.
Also, you cannot add new members to the
class without breaking binary compatibility, as the size of the class
changes.

Perhaps the most important problem with this approach is
that the methods are non-virtual. Thus, they are implemented as
specially named functions that take the 'this' pointer as their first
argument. Unfortunately, I don't know of any two compilers that mangle
method names in the same way. So don't think your DLL work with an
executable compiled with another compiler!

Attempt #2

For those of you experienced in object oriented programming, you know that
every class can be broken into two concepts: an interface and a factory.
A factory is a mechanism for creating objects, and an interface allows you to
communicate with them. The next version of Window.h will separate these
concepts. Notice that you no longer need to export the class (you
have to export the factory function though!), as it is
abstract: all method calls go through the object's vtable, not
through a direct linking to the DLL. Only the call to the factory
function calls directly into the DLL.

// Window.h

#include <string>

class Window {
public:
  virtual ~Window() { }
  virtual void setTitle(std::string title) = 0;
  virtual std::string getTitle() = 0;
};

Window* DLLIMPORT CreateWindow(std::string title);

This is much better. The code that uses window objects doesn't care
what actual type the window object is, just that it implements the
Window interface. However, there is still a problem: Different
compilers mangle symbol names differently, so the
CreateWindow
function in DLLs generated by different
compilers will have a different names. This means that if you compile
the windowing DLL with Visual C++ 6, you won't be able to use it in
Borland C++, and vice versa. Fortunately, the C++ standard lets us
disable symbol mangling on specified names, via extern
"C"
.

Some of you may have noticed another problem with this code.
Different compilers implement the standard C++ library differently.
In the less obvious case, some people replace their compiler's
implementation of the library with another (such as STLPort
). Since you can't depend on STL
objects being binary compatible across compilers, you cannot safely use them
in your DLL interfaces.

If a C++ ABI is ever created for Windows, it will need to specify
exactly how to interface with every class in the standard library, but
I don't see this happening anytime soon.

The final problem here is a minor one. By convention, COM methods and
DLL functions use the __stdcall calling convention. We can fix this
with the CALL macro I defined above. (You'll want to rename it in
your project.)

Revision 3

// Window.h

class Window {
public:
  virtual ~Window() { }
  virtual void CALL setTitle(const char* title) = 0;
  virtual const char* CALL getTitle() = 0;
};

extern "C" Window* CALL CreateWindow(const char* title);

We're almost there! This particular interface will probably work in a lot of
situations. However, the virtual destructor makes things a little
interesting... Since COM doesn't use virtual destructors, you can't depend
on different compilers to use them identically. However, you can replace the
virtual destructor with a virtual method which, in the implementation class,
is implemented by delete this;
This way, both construction
and destruction are implemented on the same side of the DLL boundary. For
example, if you try to use a VC++ 6 debug DLL alongside a release executable,
you'll either crash or run into warnings like "Value of ESP not saved
across function call". This error occurs because the debug version of
the VC++ runtime library has a different allocator than the release
version. Since the two allocators are not compatible, we cannot
allocate memory on one side of the DLL boundary and delete it on the
other.

"But how is a virtual destructor different from another virtual
method?" Virtual destructors are not responsible for deallocating the
memory used by the object: They are simply called to perform
necessary cleanup before the object is deallocated. The executable
that uses your DLL will try to free the object's memory itself. On
the other hand, the destroy()
method is
responsible for deallocating memory, so all new and delete calls stay
on the same side of the DLL boundary.

It's also a good idea to make the interface's destructor protected so that
users of the interface can't inadvertently use delete on it.

Revision 4

// Window.h

class Window {
protected:
  ~Window() { }  // use destroy()

public:
  virtual void CALL destroy() = 0;
  virtual void CALL setTitle(const char* title) = 0;
  virtual const char* CALL getTitle() = 0;
};

extern "C" Window* CreateWindow(const char* title);

Since this code doesn't use any semantics not defined by COM, it should work
flawlessly across compilers and configuration settings. Unfortunately, it's
not ideal. You have to remember to delete objects with
object->destroy();
, which isn't nearly as intuitive as
delete object;
. Perhaps more importantly, you can no
longer use std::auto_ptr
on objects of this type.
auto_ptr
wants to delete the object it owns with
delete object;
. Is there a way to make the syntax
delete object;
actually call object->destroy();
?
Yes. Here's where things get a little weird... You can overload
operator delete
for the interface and have it call destroy().
Since operator delete takes a void pointer, you'll have to assume you
never call Window::operator delete on anything that isn't a Window. This
is a pretty safe assumption. Here's the operator implementation:

...

  void operator delete(void* p) {
    if (p) {
      Window* w = static_cast<Window*>(p);
      w->destroy();
    }
   }

...

Looks pretty good... You can now use auto_ptr again, and you still
have a stable binary interface. When you recompile and test your new
code (you are
testing, right??), you'll notice that there is
a stack overflow in WindowImpl::destroy
! What's going
on? If you remember how the destroy method is implemented, you'll see
that it simply executes delete this;
. Since the interface
overloads operator delete
, WindowImpl::destroy
calls Window::operator delete
which calls
WindowImpl::destroy
... ad infinitum. The solution to
this particular problem is to overload operator delete in the
implementation class to call the global operator delete:

...

  void operator delete(void* p) {
    ::operator delete(p);
  }

...

Finishing Touches

If your system has a lot of interfaces and implementations, you'll find that
you'll want some way to automate undefining operator delete. Fortunately,
this is possible too. Simply create a templated class called DefaultDelete
and instead of deriving your implementation class from interface I, derive
from class DefaultDelete<I>. Here's DefaultDelete's definition:

template<typename T>
class DefaultDelete : public T {
public:
  void operator delete(void* p) {
    ::operator delete(p);
  }
};

Final Implementation

Here is the final version of the code.

// Window.h

class Window {
public:
  virtual void CALL destroy() = 0;
  virtual void CALL setTitle(const char* title) = 0;
  virtual const char* CALL getTitle() = 0;

  void operator delete(void* p) {
    if (p) {
      Window* w = static_cast<Window*>(p);
      w->destroy();
    }
  }
};

extern "C" Window* CALL CreateWindow(const char* title);

// Window.cpp
#include <string>
#include <windows.h>
#include "DefaultDelete.h"

class WindowImpl : public DefaultDelete<Window> {
public:
  WindowImpl(HWND window) {
    m_window = window;
  }

  ~WindowImpl() {
    DestroyWindow(m_window);
  }

  void CALL destroy() {
    delete this;
  }

  void CALL setTitle(const char* title) {
    SetWindowText(m_window, title);
  }

  const char* CALL getTitle() {
    char title[512];
    GetWindowText(m_window, title, 512);
    m_title = title;  // save the title past the call
    return m_title.c_str();
  }

private:
  HWND window;
  std::string m_title;
};

Window* CALL CreateWindow(const char* title) {
  // create the Win32 window object
  HWND window = ::CreateWindow(..., title, ...);
  return (window ? new WindowImpl(window) : 0);
}

// DefaultDelete.h

template<typename T>
class DefaultDelete : public T {
public:
  void operator delete(void* p) {
    ::operator delete(p);
  }
};

Summary

That's about it. In closure, I'll enumerate guidelines to keep in mind
when creating C++ interface. You can look back on this as a reference or
use it to help solidify your knowledge.

• All interface classes should be completely abstract. Every method should
  be pure virtual. (Or inline... you could safely write inline
  convenience methods that call other methods.)
• All global functions should be extern "C"
  to prevent
  incompatible name mangling. Also, exported functions and methods
  should use the __stdcall
  calling convention, as DLL
  functions and COM
  traditionally use that calling convention. This way, if a user of the
  library is compiling with __cdecl
  by default, the calls
  into the DLL will still use the correct convention.
• Don't use the standard C++ library.
• Don't use exception handling.
• Don't use virtual destructors. Instead, create a destroy() method and an
  overloaded operator delete
  that calls destroy().
• Don't allocate memory on one side of the DLL boundary and free it
  on the other. Different DLLs and executables can be built with
  different heaps, and using different heaps to allocate and free
  chunks of memory is a sure recipe for a crash. For example, don't
  inline your memory allocation functions so that they could be built
  differently in the executable and DLL.
• Don't use overloaded methods in your interface. Different
  compilers order them within the vtable differently.

References

• STLPort
  is an alternate implementation
  of the STL.
• SGI
  has another standard C++
  library implementation.
• The Corona
  image I/O library uses the
  techniques introduced in this article.

Feedback

I would really like feedback on this article. Was any part unclear? Do you
want more details about a particular situation? Is any of the code wrong?
Send e-mail to
aegis@aegisknight.org
.