[boost] Re: In-place construction for variant

From: David Abrahams (dave_at_[hidden])
Date: 2003-10-28 00:50:21

"E. Gladyshev" <egladysh_at_[hidden]> writes:

> --- David Abrahams <dave_at_[hidden]> wrote:
> [..]
>> 
>> Wha??  Copying the _LHS_ to heap backup??  That can't possibly result
>> in the basic guarantee unless I've completely missed something.  
>
> Yeah, we all missed something. :)
> The current variant does create a heap copy of the *lhs* and 
> switches to it if there is an exception. 

Umm.  OK, I guess that one works for the basic guarantee, though it's
weird because the heterogeneous assignment only gives the basic
guarantee when it could give the strong guarantee at similar cost.

>> I believe my suggestion has been misunderstood.  The only way to
>> get this right is to use heap storage for the new rhs.  In other
>> words:
>> 
>>    variant<A,B> internally is conceptually a union {A,B,A*,B*};
>> 
>>    direct construction of a variant with A or B constructs an A or B
>>    in-place.  Assignment from the same type assigns in-place.
>>    Assignment of the other type allocates the new RHS on the heap,
>>    destroys the LHS, and copies the pointer to the heap-allocated RHS
>>    copy into the union.
>> This was exactly my understanding
>
> There should be a special case if the variant has a non-throw
> copy-constructable type. 

Yes, yes of course.

> Other than that this was my understanding too, till I looked at the
> variant's codes.
>
> If the variant doesn't have any guarantees about 
> the storage location of the controlled object,

...like the stability of that location...

> it seems that it is possible to create a heap
> copy of the *lhs* not *rhs* and try to allocate
> *rhs* in-place. If it is true, it seems that 
> the current implementation is more efficient 
> in terms of memory fragmentation (create a temp. *lhs* 
> and if everything is fine, delete it).  

Oooh, now I understand.  Actually that's clever.  OK, sorry Eric; nice
job!

> For example provided that there were no exceptions, an array of
> variants will occupy a continuous chunk of storage.
>
> Also, I mentioned it to Eric that the current implementation doesn't
> actually provide basic guarantees.  He was going to look into that.

It sounds like it should/could be modified to do so.
-- 
Dave Abrahams
Boost Consulting
www.boost-consulting.com