forward list patch

[François Dumont]

Just some other points and I will let you report it to the LWG people.

You said that swap is expected to behave in constant time and I agree. 
But I think people are also expecting the swap operation not to 
invalidate any iterator. AFAIK this assertion is correct for all 
containers, why would it be different for the forward_list ? The only 
valid proposition would be to have the before begin iterator swapped 
like the other iterators which will impact normal implementation.

On 10/27/2010 03:15 AM, Paolo Carlini wrote:
> If the standard says it can be invalidated it depends on the
> implementation details whether it is or it is not and portable code
> certainly cannot rely on it to be valid. I agree that harmonizing the
> implementation defined behavior of normal-mode and debug-mode is in
> general a good idea, but if that means adding a lot of slow, and ad-hoc,
> and cryptic code, I become very very nervous.
>    
Have you taken a look to current _Safe_sequence _M_swap implementation ? 
Here it is:

   void
   _Safe_sequence_base::
   _M_swap(_Safe_sequence_base& __x)
   {
     __gnu_cxx::__scoped_lock sentry(_M_get_mutex());
     swap(_M_iterators, __x._M_iterators);
     swap(_M_const_iterators, __x._M_const_iterators);
     swap(_M_version, __x._M_version);
     _Safe_iterator_base* __iter;
     for (__iter = _M_iterators; __iter; __iter = __iter->_M_next)
       __iter->_M_sequence = this;
     for (__iter = __x._M_iterators; __iter; __iter = __iter->_M_next)
       __iter->_M_sequence = &__x;
     for (__iter = _M_const_iterators; __iter; __iter = __iter->_M_next)
       __iter->_M_sequence = this;
     for (__iter = __x._M_const_iterators; __iter; __iter = __iter->_M_next)
       __iter->_M_sequence = &__x;
   }

You see, there are already 4 for loops to swap sequence reference on 
each safe iterator, not really a constant operation as long you have 
safe iterators pointing to the sequence when the swap operation takes 
place. Even if my forward_list _M_swap proposition looks awful it is not 
very far from this version. If there is no before begin iterator when 
the swap operation takes place then there will be only a couple of 
additionnal pointer comparisons compared to this version.

Now I will keep on working on the contention enhancement patch avoiding 
tests regarding before begin ownership.

Bests regards