bug#70007: [PATCH] native JSON encoder

["Mattias Engdegård" <mattias.engdegard@gmail.com>]

27 mars 2024 kl. 20.05 skrev Eli Zaretskii <eliz@gnu.org>:

>>> This rejects unibyte non-ASCII strings, AFAU, in which case I suggest
>>> to think whether we really want that.  E.g., why is it wrong to encode
>>> a string to UTF-8, and then send it to JSON?
>> 
>> The way I see it, that would break the JSON abstraction: it transports strings of Unicode characters, not strings of bytes.
> 
> What's the difference?  AFAIU, JSON expects UTF-8 encoded strings, and
> whether that is used as a sequence of bytes or a sequence of
> characters is in the eyes of the beholder: the bytestream is the same,
> only the interpretation changes.

Well no -- JSON transports Unicode strings: the JSON serialiser takes a Unicode string as input and outputs a byte sequence; the JSON parser takes a byte sequence and returns a Unicode string (assuming we are just interested in strings).

That the transport format uses UTF-8 is unrelated; if the user hands an encoded byte sequence to us then it seems more likely that it's a mistake. After all, it cannot have come from a received JSON message.

I think it was just an another artefact of the old implementation. That code incorrectly used encode_string_utf_8 even on non-ASCII unibyte strings and trusted Jansson to validate the result. That resulted in a lot of wasted work and some strange strings getting accepted.

While it's theoretically possible that there are users with code relying on this behaviour, I can't find any evidence for it in the packages that I've looked at.

> I didn't suggest to decode the input string, not at all.  I suggested
> to allow unibyte strings, and process them just like you process
> pure-ASCII strings, leaving it to the caller to make sure the string
> has only valid UTF-8 sequences.

Users of this raw-bytes-input feature (if they exist at all) previously had their input validated by Jansson. While mistakes would probably be detected at the other end I'm not sure it's a good idea.

>  Forcing callers to decode such
> strings is IMO too harsh and largely unjustified.

We usually force them to do so in most other contexts. To take a random example, `princ` doesn't work with encoded strings. But it's rarely a problem.

Let's see how testing goes. We'll find a solution no matter what, pass-through or separate slow-path validation, if it turns out that we really need to after all.