On Feb 11, 2004, at 4:45 AM, Dhruv Matani wrote:
> On Wed, 2004-02-11 at 01:59, Gabriel Dos Reis wrote:
>> Matt Austern <austern@apple.com> writes:
>>
>> | On Feb 9, 2004, at 6:20 PM, Gabriel Dos Reis wrote:
>> |
>> | > Matt Austern <austern@apple.com> writes:
>> | >
>> | > | I'm afraid you're right too. Fortunately, I really do only
>> mean "a
>> | > | bit" more
>> | > | complicated. The solution is pretty simple: instead of
>> inheriting
>> | > | directly
>> | > | from the allocator, instead inherit from a thin wrapper class
>> that
>> | > | provides
>> | > | nothing but the allocator functions we care about.
>> | >
>> | >
>> | > Just to make sure I understand your suggestion. In doing so, how
>> do
>> | > we prevent the "virtual function is a sin you inherit from your
>> | > descendents" syndrom and benefit at the same time from the empty
>> base
>> | > optimization?
>> | >
>> | > The solution I'm thinking of would be:
>> | >
>> | > template<class Allocator>
>> | > WrapAllocator {
>> | > Allocator allocator;
>> | > // standard allocator interface goes here
>> | > // as forwarding functions to Allocator
>> | > // ...
>> | > };
>> | >
>> | > template<class T, class Allocator = allocator<T> >
>> | > struct vector : private WrapAllocator<Allocator> {
>> | > // ...
>> | > };
>> | >
>> | > Is that what you're alluding to?
>> |
>> | Basically, yes. The two ways in which this isn't *exactly* what I
>> | have in mind:
>> | (1) It won't be the full standard allocator interface, only the
>> | useful part. (i.e. "allocate" and "deallocate".)
>> | (2) Two template parameters, not one. As long as we have to
>> | introduce this wrapper class anyway, may as well make it
>> | do double duty by having it take care of the icky rebind
>> | stuff.
>>
>> OK, I did my homework. I think it is possible to use the inheritance
>> trick and benefit from the empty base class optimization and still not
>> get trapped by Dhruv's example.
>>
>> (1) Say don't do that. I'm not sure if this is going to be
>> acceptable answer. Certainly it is not for a testsuite.
>> But, I don't know whether outsitde of the testsuites areas, it
>> is
>> acceptable. I would think it is, but nobody knows what C++
>> programmers do.
>>
>> (2) instead of using containment followed by inheritance, use
>> inheritance+uglification followed by containment+delegation.
>
> What's inheritance+uglification? I'm mainly interested in uglification?
>
> If you have containment anywhere in the picture, won't the EBO get
> nullified? AFAICS, containment + delagation is the same as your earlier
> proposal? Or I have misunderstood something?
>
>
>> Do people have better alternatives?
>
> I have cc'd you the patch. It mainly does this:
>
> 1. A new IMPL class that encapsulates the implementation ie. _M_start,
> _M_finish and _M_end_of_storage, and also derived from the _Alloc
> parameter. Thus, the rest of the vector's implementation now refer to
> _M_start, etc... as _M_impl._M_start, etc...
>
> 2. The vector does not get derived in any way from the allocator, and
> so
> the clear() problem is avoided. Also, the allocator may have members
> push_back, pop_back, etc...
>
> 3. The extra code for delegation is not required.
>
>
>
> --
> -Dhruv Matani.
> http://www.geocities.com/dhruvbird/
>
>
>