Most portable and reliable way to get the address of variable in C++

Using & to get an address of a variable can be problematic if the variable type has overloaded operator&(). For example, _com_ptr_ has operator&() overloaded with a side effect of modifying the object.

Now I have a complicated set of templates with functions like this:

template<class T>
void process( const T* object )
{
    //whatever
}    

template<class T>
void tryProcess( T& object )
{
    process( &object )
}

In tryProcess() I need to get a T* pointer holding the address of the actual object of type T.

The above implementation of tryProcess() will only work allright if class T doesn't have operator&() overloaded. So if I call tryProcess<_com_ptr_<Interface>>() I can get unexpected results - the overloaded operator&() is triggered.

In another question the following workaround is suggested:

template<class T>
T* getAddress( T& object )
{
   return reinterpret_cast<T*>( &reinterpret_cast<char&>( object ) );
}

With such a function I can implement tryProcess() as follows:

template<class T>
void tryProcess( T& object )
{
    process( getAddress( object ) )
}

and will always get the same behavior independent of whether class T has operator&() overloaded. This introduces zero overhead with optimizations on on Visual C++ 7 - the compiler gets what to do and just gets the object address.

How portable and standard-compilant is this solution to the problem? How could it be improved?

Boost addressof is implemented with that reinterpret_cast trick so I'd say it's probably portable and standard-conforming.

Here you can see the code in question.

It is standard-complaint. The issue was brought to the attention of the ISO C++ committee in relation to problems with offsetof implementations that broke on this. Amongst the solutions considered were tightening the POD definition, or adding an extra restriction on types to be used with offsetof. Those solutions were rejected when the reinterpret_cast solution was brought up. Since this offered a standard-compliant way around the problem, the committee did not see a need to add extra requirements to the offsetof, and left fixes to the implementations.