[boost] Re: [serialization] add_facet question

From: Jonathan Turkanis (technews_at_[hidden])
Date: 2005-02-03 21:32:41

Robert Ramey wrote:
> Jonathan Turkanis wrote:
>> Robert Ramey wrote:
>>> add_facet got added to address a short-coming of the older
>>> dinkumware library. It would not really be needed except for this.
>>>
>>> As to why I did it they way I did - I don't remember.  I would guess
>>> that I didn't want any more copying than necessary and this
>>> implementation worked fine for me.
>>
>> Sounds reasonable. My guess is that the extra copy will typically be
>> optimized away, but I haven't verified this.
>
> Hmmm - I would have thought the way its done now ther would be no
> copying necessary at all. Also, I checked the documentation (msvc
> 7.1) and it didnt' show  a copy constructor, I suppose I presumed
> erroneously that there was one.

Yes, you're right. Dinkumware doesn't document the copy constructor. But it's
definitely there (and it's nothrow, too)

> For some reason I seem to think that
> copying a locale wasn't a good idea as I was storing a copy in the
> archive -

locales are meant to be pretty trivial to copy. The facets should be stored as
pointers and are not copied directly. Copying mainly involves updating reference
counts. As a result, you should be able to store an archive's locale by value,
rather than using a scoped pointer, as it appears you do. (But I am not
suggesting you change anything.)

> and I sort of remember some issues when archives where
> deleted re garding the sequence of destruction. Also I found a number
> of subtle difference in stream implementations that really drove me
> crazy (of course that's why add_facet exists).  On the other hand
> maybe I just did it that way without thinking.  oh well, I guess
> we'll just have to test it.

I'm sure you'll remember the problems as soon as they happen again ;-)

Jonathan