This is the mail archive of the libstdc++@gcc.gnu.org mailing list for the libstdc++ project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Sizes when sizeof(value_type) == 1 (Was: Re: vector<> can...)


On Tue, 2004-10-19 at 16:21, Paolo Carlini wrote:
> Dhruv Matani wrote:
> 
> >I don;t think so. It seems fine to me.
> >
> Let's consider just a series of push_back: every time you reallocate,
> the size doubles (in _M_insert_aux), and this gives problems as soon
> as you reach ~ half max_size.
> 
> Ok, a simple example with an hypothetical 16-bit machine: in this case,
> max_size = 65535 (we are always working with sizeof(vale_type) == 1).
> 
> When __old_size = 32768, _M_insert_aux computes __old_size * 2, which
> is *exactly zero*, since overflows the max rapresentable number by one!
> 
> You could easily ask the allocator whatever in the range [__old_size + 1,
> 2 * __old_size - 1], and instead you are asking exactly zero!
> 
> Stra-bummer!

Yes, this pesky case *does* remain as a bug, and I have mentioned in my
previous email:

"Even otherwise IMHO.
We *should* have a check that __len > __old_size everywhere."

However, regarding the policy for reallocating wherein we have different
forms like __old_size*2 and __old_size+max(...) that is perfectly fine.
Only the extra check needs to go there.

This problem would be solved partially IMHO with the patch I'm about to
send.


> 
> Paolo.
-- 
        -Dhruv Matani.
http://www.geocities.com/dhruvbird/

The price of freedom is responsibility, but it's a bargain, because
freedom is priceless. ~ Hugh Downs


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]