gcc: How to use __attribute((__may_alias__)) properly to avoid “derefencing type-punned pointer” warning

Question

I've got some code that uses type-punning to avoid having to call a member "object"'s constructor and destructor unless/until it's actually necessary to use the object.

It works fine, but under g++ 4.4.3, I get this dreaded compiler warning:

jaf@jeremy-desktop:~$ g++ -O3 -Wall puns.cpp 
puns.cpp: In instantiation of ‘Lightweight<Heavyweight>’:
puns.cpp:68:   instantiated from here 
puns.cpp:12: warning: ignoring attributes applied to ‘Heavyweight’ after definition
puns.cpp: In destructor ‘Lightweight<T>::~Lightweight() [with T = Heavyweight]’:
puns.cpp:68:   instantiated from here
puns.cpp:20: warning: dereferencing type-punned pointer will break strict-aliasing rules
puns.cpp: In member function ‘void Lightweight<T>::MethodThatGetsCalledRarely() [with T = Heavyweight]’:
puns.cpp:70:   instantiated from here
puns.cpp:36: warning: dereferencing type-punned pointer will break strict-aliasing rules

My code tries to use gcc's __attribute((__may_alias__)) to let gcc know about the potential aliasing, but gcc doesn't seem to understand what I'm trying to tell it. Am I doing something wrong, or does gcc 4.4.3 just have some problems with the __may_alias__ attribute?

Toy code to reproduce the compiler warning is below:

#include <stdio.h>
#include <memory>    // for placement new
#include <stdlib.h>  // for rand()

/** Templated class that I want to be quick to construct and destroy.
  * In particular, I don't want to have T's constructor called unless
  * I actually need it, and I also don't want to use dynamic allocation.
  **/
template<class T> class Lightweight
{
private:
   typedef T __attribute((__may_alias__)) T_may_alias;

public:
   Lightweight() : _isObjectConstructed(false) {/* empty */}

   ~Lightweight()
   {
      // call object's destructor, only if we ever constructed it
      if (_isObjectConstructed) (reinterpret_cast<T_may_alias *>(_optionalObject._buf))->~T_may_alias();
   }

   void MethodThatGetsCalledOften()
   {
      // Imagine some useful code here
   }

   void MethodThatGetsCalledRarely()
   {
      if (_isObjectConstructed == false)
      {
         // demand-construct the heavy object, since we actually need to use it now
         (void) new (reinterpret_cast<T_may_alias *>(_optionalObject._buf)) T();
         _isObjectConstructed = true;
      }
      (reinterpret_cast<T_may_alias *>(_optionalObject._buf))->DoSomething();
   }

private:
   union {
      char _buf[sizeof(T)];
      unsigned long long _thisIsOnlyHereToForceEightByteAlignment;
   } _optionalObject;

   bool _isObjectConstructed;
};

static int _iterationCounter = 0;
static int _heavyCounter     = 0;

/** Example of a class that takes (relatively) a lot of resources to construct or destroy. */
class Heavyweight
{
public:
   Heavyweight()
   {
      printf("Heavyweight constructor, this is an expensive call!\n");
      _heavyCounter++;
   }

   void DoSomething() {/* Imagine some useful code here*/}
};

static void SomeMethod()
{
   _iterationCounter++;

   Lightweight<Heavyweight> obj;
   if ((rand()%1000) != 0) obj.MethodThatGetsCalledOften();
                      else obj.MethodThatGetsCalledRarely();
}

int main(int argc, char ** argv)
{
   for (int i=0; i<1000; i++) SomeMethod();
   printf("Heavyweight ctor was executed only %i times out of %i iterations, we avoid %.1f%% of the ctor calls!.\n", _heavyCounter, _iterationCounter, 100.0f*(1.0f-(((float)_heavyCounter)/((float)_iterationCounter))));
   return 0;
}

Answer 1

I think the typedef is confusing GCC. These sorts of attributes seem to work best when applied directly to variable definitions.

This version of your class works for me (GCC 4.6.0):

template<class T> class Lightweight
{
private:
  //  typedef T __attribute((__may_alias__)) T_may_alias;

public:
  Lightweight() : _isObjectConstructed(false) {/* empty */}

  ~Lightweight()
  {
    // call object's destructor, only if we ever constructed it
    if (_isObjectConstructed) {
      T * __attribute__((__may_alias__)) p
        = (reinterpret_cast<T *>(_optionalObject._buf));
      p->~T();
    }
  }

  void MethodThatGetsCalledOften()
  {
    // Imagine some useful code here
  }

  void MethodThatGetsCalledRarely()
  {
    T * __attribute__((__may_alias__)) p
      = (reinterpret_cast<T *>(_optionalObject._buf));
    if (_isObjectConstructed == false)
      {
        // demand-construct the heavy object, since we actually need to use it now

        (void) new (p) T();
        _isObjectConstructed = true;
      }
      p->DoSomething();
  }

  [etc.]

Answer 2

I would argue for having your containing class just contain a char array of sufficient size to contain your member "object" and then using placement new to initialize on top of the char array. That has the perk of being specification-compliant as well as cross-compiler. The only problem is that you have to know the size in chars of your member object, which may get you in trouble.

Is there a reason you can't have the member be a pointer and use new and delete?

Answer 3

What if you replace _isObjectConstructed with a pointer to the object:

class Lightweight
{
public:
   Lightweight() : object(NULL) {/* empty */}

   ~Lightweight()
   {
      // call object's destructor, only if we ever constructed it
      if (object) object->~T();
   }

   void MethodThatGetsCalledOften()
   {
      // Imagine some useful code here
   }

   void MethodThatGetsCalledRarely()
   {
      if (!object)
      {
         // demand-construct the heavy object, since we actually need to use it now
         object = new (_optionalObject._buf) T();
      }
      object->DoSomething();
   }

private:
   union {
      char _buf[sizeof(T)];
      unsigned long long _thisIsOnlyHereToForceEightByteAlignment;
   } _optionalObject;

   T *object;
};

Note, no GCC extension, only pure C++ code.

Using a T* instead of a bool won't even make Lightweight any bigger!