[boost] Re: copy_on_write

From: Alex Besogonov (cyberax_at_[hidden])
Date: 2005-05-31 11:30:33

It's perfectly possible to write threadsafe fast COW containers. The 
trick is to separate mutatable and immutable objects. Right now I'm 
writing const_containers library (inspired by const_string). For 
example, I have two classes for strings: string_builder and const_string.

string_builder (it is noncopyable, btw) is used to create strings which
then are passed as refcounted const_strings:
==========================================
const_string<char> doSomething()
{
        string_builder<char> str;
        str+="blah-blah"+another_string;

        return const_string<char>(str.yield());
}
string_builder str_again(doSomething().yieled());
str_again+="something";
==========================================
'yield' method is used to transfer ownership of memory block (without 
any coping if block is owned exclusively).

Robert Mathews wrote:
> Hmm .... you know that COW implementations of strings tend not to be
> thread-safe, or if they are thread-safe, deadly slow?
 >
> For example DinkumWare switched from a COW implementation of strings to a
> small string optimization in the 7.1 release of VC. We run a million-plus
> transactions through our system every day, and monitor every crash. We've
> observed two things from this change:
> - crashes due to accidently passing strings between threads have
> disappeared.
> - speed has not changed detectably. In fact, if anything, seems to have
> increased slightly.
> 
> So, I'd never use a COW implementation again in a multithreaded
> environment - it's just simply impossible to make thread safe.
> 
> "Maksym Motornyy" <mmotorny_at_[hidden]> wrote in message
> news:d7hvcu$usf$1_at_sea.gmane.org...
>> > There's a pretty nice implementation of COW objects in the Adobe Open
>> > Source library (http://opensource.adobe.com).  You might take a look
>> > at that.
>> I saw it. Honestly, after I implemented COW myself :) Do you want to say
>> it's no sence to double existing things?
>>
>> > A compiler
>> > that implements NRVO (Named Return Value Optimization) can return an
>> > object (essentially) by const reference without any special COW
>> > wrapper classes.
>> Agree, this is not the best example for such a compiler. AFAIK VC++
>> doesn't implement this optimization.
>>
>> > Also, some C++ Standard Library implementations use
>> > a reference-counted std::string implementation, so this extra effort
>> > would be wasted.
>> STLport and Dinkumware (I suppose most widely used) doesn't belong to
>> that implementations. That's why personally I require COW-wrapper.

-- 
With respect,
              Alex Besogonov (alexy_at_[hidden])