Re: [review][JSON] Review of JSON starts today: Sept 14 - Sept 23

From: Peter Dimov (pdimov_at_[hidden])
Date: 2020-09-16 15:53:35

Vinnie Falco wrote:

> That is easy to say but what do you do about this function which returns a 
> reference:
>
>     inline value& array::operator[]( std::size_t pos ) noexcept;
>
> What do you do if you have an array of int (scalar) and someone accesses 
> an element in the middle and assigns a string to it?

The only possible answer is "copy the entire thing into value[]". This may 
or may not be acceptable. I would think that if you have an array of 8044 
ints, assigning a string somewhere in the middle would be a rare occurrence, 
but who knows.

The upside is that an array of ints would consume significantly less memory 
than today; this is often important and we've already seen that for a subset 
of users, memory consumption is _the_ important metric. (It would also make 
value_to<vector<int>> infinitely faster.)

The downside of course is that when someone assigns a single non-int 
somewhere, or even just invokes the non-const op[] without assigning a 
non-int, or the non-const begin(), the whole thing reallocates. This would 
kill the noexcept, at the very minimum. There's a thing to be said here 
about returning references to internal state, but consistency has its 
benefits.

For CBOR 1.0, I'd just reject binary values on parsing. That's still much 
more useful than not having it.