Re: [Boost-users] [phoenix] function with reference implementation

Subject: Re: [Boost-users] [phoenix] function with reference implementation
From: Joel de Guzman (joel_at_[hidden])
Date: 2010-09-10 23:06:28

On 9/11/2010 9:31 AM, Alfredo Correa wrote:
> On Fri, Sep 10, 2010 at 5:47 PM, Joel de Guzman
> <joel_at_[hidden]>  wrote:
>> On 9/11/10 8:31 AM, alfC wrote:
>>>
>>>
>>> On Sep 10, 4:36 pm, Joel de Guzman<j..._at_[hidden]>    wrote:
>>>>
>>>> On 9/11/10 4:37 AM, alfC wrote:
>>>>
>>>>
>>>>
>>>>> I noticed that in some contexts I have both the object class and the
>>>>> corresponding phoenix function
>>>>
>>>>>     class impl;
>>>>>     ...
>>>>
>>>>>     impl f_impl;
>>>>>     boost::phoenix::function<impl>      f(f_impl);
>>>>
>>>>> now to avoid the copying I could use:
>>>>
>>>>>     boost::phoenix::function<impl&>      f(f_impl);
>>>>
>>>>> which works. Although
>>>>>     boost::phoenix::function<impl const&>      f(f_impl); //doesn't work.
>>>>> (error below)
>>>>
>>>>> Is there any problem with using references at all, in the first place?
>>>>
>>>> I'm not sure. Could you provide a minimal cpp file I can try?
>>>
>>> Yes, see below. Althougth now that I made this example I realize that
>>> the underlying object implementation is copied many times anyway (at
>>> least 4 times in the example), I guess during the building up of the
>>> expression template. [The final objective was to make an object
>>> function that not only evaluates but also is able to take expression
>>> and build expression templates].
>>
>> Here's my suggestion: write a proxy function object that holds
>> the target function object by reference, pointer, or boost.ref
>> and use that to wrap your phoenix function.
>>
>
> I see, so once phoenix::function takes charge everything is copied by
> value and I have to provide an extra layer of reference.
> I think this is a general problem with template code, that wants to assume that
> the type is a non-reference and suddenly everything is copied by
> value, although it is
> clear from the intention that everything should be treated by reference.
> In fact template code is full of remove_reference and remove_cv, it
> make the life easier for the template coder but at the cost of certain
> loss of flexibilty. My point is that most template code should honor
> const and reference qualifiers and not just get rid of them. (If it
> can't honor it then the code should not compile).
>
> Anyway this is the final code for a funciton object that binds itself
> upon evaluation on phoenix expression. Although not related to the
> initial question of the post I would welcome any feedback:
> (I am not sure if really need the inheritance part)

Instead of going that path, I suggest you instead tweak phoenix to
allow references (I agree with your assessment in general) and
submit a patch for it. In other words, I consider this a bug
and what I just provided is a mere workaround.

Regards,
-- 
Joel de Guzman
http://www.boostpro.com
http://spirit.sf.net