Re: [boost] Re: dynamic_void_cast

From: Terje Slettebø (tslettebo_at_[hidden])
Date: 2002-12-10 09:25:12

>From: "Peter Dimov" <pdimov_at_[hidden]>

> From: "Johan Nilsson" <johan.nilsson_at_[hidden]>
> >
> > "Peter Dimov" <pdimov_at_[hidden]> wrote in message
> > news:00bb01c2a046$671b1c20$1d00a8c0_at_pdimov2...
> > > From: "Johan Nilsson" <johan.nilsson_at_[hidden]>
> > > [...]
> > > > template<typename T>
> > > > T* dynamic_void_cast(void* pv)
> > > > {
> > > >   struct rtti_obj__
> > > >   {
> > > >     virtual ~rtti_obj__() = 0;
> > > >   };
> > > >
> > > >   rtti_obj__* pro = static_cast<rtti_obj__*>(pv);
> > > >
> > > >   try
> > > >   {
> > > >     return dynamic_cast<T*>(pro);
> > > >   }
> > >
> > > This is not required to work. T's vtable pointer is not guaranteed to
be
> > at
> > > offset 0.
>
> Actually I meant *pv's vtable pointer here.
>
> > But I thought that it _might_ be required to be at the same offset for T
> as
> > for rtti_obj. For that part, IIRC, a vtable isn't required at all, or?
>
> Try it (on several compilers, as some do put the vptr at offset 0.)
>
> struct X { int n; };
> struct Y: public X { virtual ~Y() {} };
>
> Y+0: X::n
> Y+4: vptr
>
> rtti_obj__+0: vptr

I've tested it. Test code:

--- Start ---

struct X { int n; };
struct Y: public X { public: virtual ~Y() {} };

int main()
{
Y y;

Y *yptr=dynamic_void_cast<Y>(&y);
}

--- End ---

Here are the results:

MSVC 6 - vptr at 0 - It works
Intel C++ 7.0 - vptr at 0 - It works
g++ 2.95.3 - vptr _not_ at zero - boom! :) It gives an OS protection error
at the point of dynamic_cast
BCB 6 - vptr _not_ at zero - boom! :) It gives an OS protection error at the
point of dynamic_cast

Conclusion: Not recommended. :)

Besides, I can't see the point in testing for bad_cast there, since it uses
T * (not T) in the dynamic_cast, which I think would never throw bad_cast.
This also means that its signature is different from dynamic_cast, and that
it therefore can't be used for casting references (unlike dynamic_cast).

In any case, this means it is inherently implementation-dependent and
unsafe.

Regards,

Terje