Re: [boost] [any] Divergence between Boost.Any and standard library proposal

Subject: Re: [boost] [any] Divergence between Boost.Any and standard library proposal
From: Antony Polukhin (antoshkka_at_[hidden])
Date: 2013-05-04 12:38:22

2013/5/4 Beman Dawes <bdawes_at_[hidden]>:
> On Fri, May 3, 2013 at 12:30 PM, Antony Polukhin <antoshkka_at_[hidden]> wrote:
>> 2013/5/3 Beman Dawes <bdawes_at_[hidden]>:
>>>...
>>>
>>> AFAICS, the Adobe (and the proposed standard) behavior is preferred.
>>> The type does need to decay.
>>>
>>> Comments?
>>
>> Looks like I broke it in https://svn.boost.org/trac/boost/ticket/6999 .
>
> Should I create a new ticket:-)

I'll fix it in #6999 and add some tests for that case. I'll need a day or two.

>> There are also some other differences:
>> * boost implementation has perfect forwarding of ValueType
>> (constructor and operator=)
>
> I created the attached little test program to see the effects of that.
> For both VC++ 2012 and GCC 4.7.0 -std=c++11, your implementation is
> more efficient:
>
> Adobe:
>
> copy construct any
>   copy ctor
>   move ctor
>   move ctor
> move construct any
>   move ctor
>   move ctor
>   move ctor
> copy assign to any
>   copy ctor
>   move ctor
>   move ctor
> move assign to any
>   move ctor
>   move ctor
>   move ctor
>
> Boost:
>
> copy construct any
>   copy ctor
> move construct any
>   move ctor
> copy assign to any
>   copy ctor
> move assign to any
>   move ctor
>
> Only the tests with differences are shown. The results are the same
> for both debug and release builds for both compilers.
>
> Please look at the test program and verify the tests are OK.

They are OK, almost the same tests are part of Boost.Any test case:
http://svn.boost.org/svn/boost/trunk/libs/any/test/any_test_rv.cpp

> I'll email Sean Parent to see what he has to say about those test results:-)
>
>> * ValueType requirement is MoveConstructible or CopyConstructible (it
>> can be used with move only types)
>
> Interesting point. IIRC, there has been some discussion in the LWG
> about such types, but I can't remember the conclusions. I'll check
> with Daniel Krügler.

Idea was to allow usage with unique_ptr. Something like:

boost:any a = std::unique_ptr<Handle>(ConstructWindowHndle(params),
[](Handle* p){ FreeHandle(p); });

>> * ~any() is noexcept
>
> By convention, the standard library doesn't add noexcept to
> destructors. I've forgotten the rationale for that.
>
>> * proposal misses noexcept in for bad_any_cast: virtual const char*
>> what() const;
>
> Ha! Thanks for spotting that. Proposal fixed.
>
>> * Boosts version does not accept Allocator
>
> The standard library is tentatively planning to add type erased
> allocator support to vocabulary types like any if they typically
> allocate heap memory, so it was added at the request of the LWG.

I may patch Boost.Any to work with type erased allocators. This shall
not break users code.

>> A few things that I think would be helpful:
>> * clear() method
>
> Interesting. I'll give it some thought for the proposal.
>
>> * ??? assign() method with Allocator
>
> I'll do whatever the committee wants on that one.
>
> Thanks for your comments! Much appreciated!

I'll need your help later with Boost.LexicalCast proposal. It requires
some updates and C++14 Concepts for better detection of character
types.

--
Best regards,
Antony Polukhin