Re: PR 57779 New debug check

[François Dumont <frs dot dumont at gmail dot com>]

On 07/11/2013 06:58 PM, Jonathan Wakely wrote:
> On 11 July 2013 17:43, Paolo Carlini wrote:
> > > +      if (&(*(__it._M_get_sequence()->_M_base().end()))
> > >
> > > Isn't this a dereference of a non-dereferenceable iterator?
> > Looks like this function does a very tricky work! I think that at least we
> > should use __addressof and std::less, std::greater, etc to compare pointers.
> > In general, we should probably try to be conservative in the first
> > implementation of the check, act only when we *really* know we have
> > contiguous memory, not even "blocks", like for std::deque.
> The safest way would probably be to add an __is_contiguous trait,
> specialized for the __gnu_debug versions of vector, array, and
> dynarray, then check __is_contiguous<_Sequence>.
    IMO the current check is enough. Of course it will consider a 
std::deque with only 1 block as continuous memory, and so what ? Passing 
a pointer to those continuous elements is bad and will be detected then. 
As soon as std::deque elements will be split in at least 2 blocks the 
check won't be performed anymore but it doesn't make the check useless. 
It would be worst if the check was generating false positive results.

    I am indeed dereferencing the end iterator but as long as it is 
just to get the address of the resulting element it is fine. However 
that's a good remark cause I never realized that getting the address of 
this invalid iterator is in fact giving a correct result. But the 
Standard is saying that we shall assert in debug mode so it is fine.

François