Re: [boost] [utility] new auto_buffer class --- RFC

Subject: Re: [boost] [utility] new auto_buffer class --- RFC
From: Thorsten Ottosen (thorsten.ottosen_at_[hidden])
Date: 2009-03-02 13:04:09

Matt Calabrese skrev:
> On Mon, Mar 2, 2009 at 10:29 AM, Thorsten Ottosen <
> thorsten.ottosen_at_[hidden]> wrote:
> 
>> Alternatively we might try to compute a reasoable default size
>> based on the size of the type and based on minimum and maximum default
>> sizes.
>>
> 
> Agreed. Especially in terms of providing a meaningful default, the amount of
> bytes being used up on the stack is much more of a concern than the number
> of individual objects on the stack. In retrospect, it might even make sense
> for the template parameter to be in bytes rather than object count.
> 
> Concerning your implementation there is at least one major issue I have --
> your use of "optimized_const_reference." You define
> optimized_const_reference to be a const value_type if the value_type has a
> trivial assignment operation. You then return this from const accessors such
> as front, back, and operator []. While it first this might seem like a
> clever trick, it is flawed in several subtle ways. Firstly, it is not
> necessarily optimized because the object may be considerably large, implying
> you'd be making a copy of a large object simply by calling operator [], but
> aside from that there are much more important reasons why you should avoid
> this.

Yes, it should also take into the account that the object's size should 
be <= sizeof(long double).

> Unless I am mistaken, it makes your container not a valid STL container,
> although it already has an O(n) swap which is unavoidable. But the main
> reason why I consider this a fault is that it now means that if you store a
> pointer/reference to an element in the array taken from a non-const accessor
> it will not reference the same object as what is returned from a const
> accessor which can cause obvious subtle issues (and it means you cannot get
> a pointer to an element in the array when all you have is a const reference
> to the container unless you use "data")! 

or *obj.begin()[n]. So there are plenty of ways to get a refence into the
array.

> It also means that you cannot
> directly take the address of the values returned from your const accessor
> functions in cases where the type contained has a trivial assignment
> operation since you can't directly take the address of a temporary.

True, you can only bind to a local const reference.

> All of this is avoided by simply returning a const_reference to the
> contained object, and it makes the behavior for your template consistent
> between different instantiations, not to mention with containers in the STL.
> While the optimized_const_reference trick is clever, I do not believe it is
> appropriate here.

Major use-cases are for types like char, int, wchar_t.

Maybe this is a good place to discuss this, but AFAIR, 
boost::circular_buffer does excatly this.

-Thorsten