On Tue, 3 Jan 2023 at 15:25, Arthur O'Dwyer <arthur.j.odwyer_at_[hidden]>
wrote:

> On Tue, Jan 3, 2023 at 3:21 AM Edward Catmur via Std-Proposals <
> std-proposals_at_[hidden]> wrote:
>
>> On Tue, 3 Jan 2023, 07:38 Frederick Virchanza Gotham via Std-Proposals, <
>> std-proposals_at_[hidden]> wrote:
>>
>>> On Monday, January 2, 2023, Edward Catmur <ecatmur_at_[hidden]>
>>> wrote:
>>>
>>>>
>>>> You could accomplish the inplace part with an appropriate allocator.
>>>>
>>>
>>> If the allocator fails to allocate because the buffer is too small, then
>>> this problem can be detected at runtime and we can throw an exception. But
>>> with the classes I wrote, we can catch this at compile time either with
>>> 'static_assert' or a concept constraint.
>>>
>>
>> Allocators are typed. Such an allocator just needs to refuse to rebind to
>> a T that is too large or too aligned.
>>
>
> Sadly, "use a different allocator" does *not* suffice to give you the
> in-place allocation strategy. This is because "an allocator is a handle
> to a heap <https://www.youtube.com/watch?v=0MdSJsCTRkY&ab_channel=CppNow>."
> You need some way for the allocator (which is a value type, and will be
> copied around quite a bit inside the internals of the STL) to point to the
> place-where-the-memory-should-come-from (which in this case is "inside
> *this* specific polymorphic_value object, and not inside *that* one
> (unless of course the polymorphic_value object is copied or moved, in which
> case we *do* mean *that* one)." The allocator model is not designed to
> work with allocators that are, themselves, heaps (and thus not value types).
>

Well, you can do it by abusing fancy pointers, but yeah, it won't work here
since polymorphic_value erases its allocator and so uses T* as its pointer
type. Ah well.


> Strongly recommend watching my talk linked above, even though it is 90
> minutes.
>
> It's quite possible to write an inplace_any, inplace_function
> <https://github.com/WG21-SG14/SG14/blob/master/SG14/inplace_function.h>,
> inplace_unique_function, and so on, but you can't do it by just abusing the
> allocator model; that physically will not work — neither in theory, nor in
> practice on any mainstream implementation.
>
> –Arthur
>