[boost] Re: move_ptr

From: Howard Hinnant (hinnant_at_[hidden])
Date: 2004-01-22 10:43:15

On Jan 22, 2004, at 8:53 AM, Bronek Kozicki wrote:

> Peter Dimov <pdimov_at_[hidden]> wrote:
>> The main (and probably only) selling point of scoped_*, move_* and
>> auto_* is that they have no size overhead, i.e. sizeof(scoped_*<T>)
>
> The second strong difference is copy semantics. You may pass *ptr 
> object
> by value without copying whole allocated memory to new object. In fact,
> we have 3 different copy semantics for arrays:
> * deep copy (std::vector)
> * shallow copy (boost::shared_array)
> * move instead of copy (subject of current discussion)
>
> There is also difference on how memory is being released:
> * implementation defined (std::vector)
> * full deleter support (boost::shared_array)
> * limited deleter or no deleter at all (subject of current discussion)

More discussion of selling points:

*  The ability to transfer ownership of the buffer.  This is really 
just a generalization of move:

move:  transfers ownership of buffer to the same type, implicitly if 
from an rvalue.

release:  relinquishes ownership to anyone who wants to call release().

The latter is very important in the exception safety patterns: 
temporarily grab ownership of a buffer with move_ptr<T[]>, call other 
(possibly throwing) initialization functions for your class, then 
transfer ownership from your temporary move_ptr<T[]> to your class.

*  When you want to guarantee unique ownership:  For example if your 
class Foo wants to own a buffer of T, and not accidently make a copy of 
it, move_ptr<T[]> for a data member is a reasonable choice over 
vector<T>.  Foo's copy ctor will be automatically disabled instead of 
automatically generating a copy of the buffer.  Foo's author will have 
to take explicit action to do anything else.

-Howard