Re: dynamic_cast (bug 3(?))

["Martin v. Loewis" <martin at mira dot isdn dot cs dot tu-berlin dot de>]

> I think that you have explained "v shall be an rvalue of a pointer
> to complete class type" as "if v has a virtual table (pointer), that
> pointer shall not be 0".

Well, no. A complete type is one that has been defined, not only
declared. A class type is either a struct or a class.

> (gdb) p *os
> $1 = {<ios> = {<_ios_fields> = {_strbuf = 0x40050640, _tie = 0x8049c38, 
>       _width = 0, _flags = 8209, _fill = 32, _state = 0 '\000', 
>       _exceptions = 0 '\000', _precision = 6, _arrays = 0x0}, _vptr. = 0x0}, 
>   _vb.ios = 0x8049bfc}
> 
> 
> Note that the _vptr. is 0x0.
> 
> Is this a bug in gcc? (ie, should you test if _vptr. is 0x0?), or
> is there something seriously wrong with `cerr'?

It is an error in cerr. Even though it is supposed to implement an
ostream, it does not have a virtual table. That is normally not a
problem, since ostream has only one (non-artificial) virtual function:
the destructor, which cannot be called for cout.

In your case, you manage to access the vtable of cerr, which is
constructed via

OSTREAM_DEF(cerr, CERR_SBUF,(ostream*)&cout, ios::unitbuf, )

in libio/stdstreams.cc. This, in turn, expands to 

_fake_ostream  cerr      = { {& cerr .base}, {(streambuf*)&   _IO_2_1_stderr_    ,    (ostream*)&cout  , 0, ios::skipws|ios::dec|    ios::unitbuf  , ' ',0,0,6  }};

If you count fields, you'll find that this is sufficient to initialize
_ios_fields, but not ios (which only has the vtable as extra field).

If you want to fix this, it would probably best to put the vtable of
ostream (i.e. __vt_7ostream.3ios) into the vtable of cout and
cerr. You'd also need to change _fake_ostream to actually contain a
vtable field, instead of filler.

Regards,
Martin