Re: [boost] [ptr_container] Questionable strong guarantee of ptr_map::insert()

Subject: Re: [boost] [ptr_container] Questionable strong guarantee of ptr_map::insert()
From: Thorsten Ottosen (thorsten.ottosen_at_[hidden])
Date: 2009-08-20 10:01:59

Kazutoshi Satoda skrev:
> Thanks for the follow ups.
> 
> Now I noticed that the problem is not only in ptr_map::insert(). The
> same problem seems wide spread in the whole ptr_container library.
> 
> I'll make the following 2 tickets in boost trac if nobody objects nor
> gives better suggestion.
> 
> 1. (Type: Bugs)
>  False strong guarantee in ptr_container documentation

Right, it should be replaced with the basic guarantee and a note saying 
the object is deleted when an exception is thrown.

> 2. (Type: Feature Requests)
>  Provide strong guaranteed overloads in ptr_container

Hm. Do you propose to break the existing interface?

> 
> Steven Watanabe wrote:
>> David Abrahams wrote:
>>> I think the implementation of ptr_map::insert is probably as correct as
>>> can be in doing that deletion (i.e. the docs are wrong, not the
>>> implementation), but if I had written ptr_map I'd have given it an
>>> auto_ptr parameter instead.
>>
>> There is an overload that takes an auto_ptr.
> 
> Yes, but unfortunately, it doesn't provide strong guarantee, as same as
> the raw pointer version.
> 
> And also, the current auto_ptr pass-by-value version can be a
> error-prone.
> 
> An casual user of ptr_map will likely write a code at first like this:
> 
>  map.insert(some_key(...), new value(...));
> 
> This doesn't compile by design. Then the user will look for the docs to
> find alternative ways, find the auto_ptr version, and modify the code to
> shut off the strange compile error, like this:
> 
>  map.insert(some_key(...), std::auto_ptr<value>(new value));
> 
> This does compile, with the risk of subtle memory leaks. But the user
> can hardly notice the problem, and would think, "Aha, I can guess, the
> use of auto_ptr is the fix for exception-safety problem which the FAQ
> mentions. The library taught me that with the error. So nice."

People need to initialize smart pointers in seperate statements.

-Thorsten