C++ Logo

SG16

Advanced search

Subject: Re: Bike shedding for Christmas: P1885 Naming Text Encodings
From: Tom Honermann (tom_at_[hidden])
Date: 2019-12-31 17:49:00


On 12/30/19 6:11 AM, Corentin Jabot wrote:
>
>
> On Mon, Dec 30, 2019, 05:17 Tom Honermann <tom_at_[hidden]
> <mailto:tom_at_[hidden]>> wrote:
>
> On 12/28/19 6:38 PM, Corentin Jabot via SG16 wrote:
>>
>>
>> On Sun, Dec 29, 2019, 00:32 Corentin Jabot
>> <corentinjabot_at_[hidden] <mailto:corentinjabot_at_[hidden]>> wrote:
>>
>>
>>
>> On Sat, Dec 28, 2019, 22:21 Tom Honermann <tom_at_[hidden]
>> <mailto:tom_at_[hidden]>> wrote:
>>
>> On 12/27/19 6:28 AM, Corentin Jabot via SG16 wrote:
>>> Hello
>>>
>>> In P1885, I introduce the name "text_encoding"  for the
>>> class representing the name of a text encoding.
>>> I wonder whether that might conflict or interfere with
>>> actual encoding/decoder classes and would like your opinion.
>>>
>>> Here are a few possible names:
>>> * Charset (IANA nomenclature, posix)
>>> * text_codec (Qt)
>>> * text_encoding
>>> * text_encoding_name (encoding is used by posix / python /
>>>
>>> Unicode nomenclature would favor encoding (Unicode is a
>>> charset of which utf-8 and utf-16 are both are encodings)
>>
>> I suggest text_encoding_id. I'd like to preserve
>> text_encoding for a tag type (or concept) that can be
>> used at compile time to specify a (compile-time) encoding
>> as in a template parameter to std::text.
>>
>> Tangent 1: the proposed text_encoding is not extensible,
>> at least not in a very meaningful way.  I suggest we do
>> one of the following:
>>
>> 1. Remove the text_encoding(const char*) constructor. 
>> It doesn't allow setting the MIB ID, so is
>> unsatisfactory at present.
>> 2. Allow first class extension by, for example,
>> reserving the full range of IANA MIB values, defining
>> a "private use" range of values, and modifying the
>> text_encoding(const char*) constructor to also accept
>> a MIB value (and perhaps make the name parameter
>> optional such that, if specified, it would override
>> the internally known name for the provided MIB value
>> and if not specified, name() would return a suitable
>> default).
>>
>>
>>
>> It is extensible in multiple-choice ways:
>> - implentation can provide their own aliases for existing mib
>> - the other mib + custom name can be used to use a non
>> register encoding. Hence the existence of both unknown and other
>>
>> I will not support custom mib as it is not in line with the
>> rfc - the mib being a way to standardize names.
>>
> I'm not sure whether you are referring to RFC 2978 or 3808 here. 
> Regardless, the design of both the RFCs and your paper is such
> that there are multiple possible names for any encoding.  A
> program that wishes to identify an encoding for which the
> implementation does not supply a MIB ID would have to be expected
> to recognize all aliases.  That doesn't seem realistic to me. 
> Additionally, as specified, the proposal only includes enumerators
> for a select subset of the IANA registered names with no facility
> (other than implementation provided extension) for creating a
> text_encoding object with a MIB ID for any other IANA registered
> encoding.
>
> Refined suggestions:
>
> 1. Include all IANA MIB IDs in the set of enumerators of
> text_encoding::id.  This may require either a normative
> reference to (a dated version of) the IANA registry or that we
> copy from a version of the IANA registry and update it for
> each C++ standard.  Note that support for names need not imply
> support for the named encoding; these are just identifiers.
>
>
> I did that but it poses extensibility issues (list has to be
> maintained) any offers little benefits.

I don't understand how this poses an extensibility issue.  I agree it
poses a maintenance issue, but that is true regardless (at least for
implementors that extend the set of enumerators).

The benefit is that including all of them avoids the problem of
implementors offering extensions with inconsistent or conflicting
names.  It also doesn't put us in the position of deciding which
encodings are "important".  IANA provides a good specification to
follow.  I don't think we should be subsetting, at least not without
some clear criteria for determining which encodings make the cut.  For
example, I suspect Shift-JIS gets more use than UTF-32, but the former
is not included in the proposal and the latter is.

Another tangent: I think the enumeration should have an explicit
underlying type to ensure it is able to hold all IANA assigned MIB IDs.

> 1.
>
>
> 2. Use the IANA "cs" prefixed names for the names of the
> enumerators of text_encoding::id. They may not be the
> prettiest names, but using these names will better facilitate
> automation and they are intended to be used as identifiers.
> 3. Provide a text_encoding(text_encoding::id) constructor that
> enables creation of an instance with a particular ID.  If the
> implementation doesn't have a registered name for the provided
> ID, then use a name like "MIB#42".
>
> What is the use case for that?
The set of recognized names are not necessarily portable since they are
implementation-defined.  This ensures that an encoding object for the
desired encoding can be created regardless of name.
> I am not opposed to the idea.
> QTextCodec has a fromMiB function after all.
> I am however opposed to a constructor that would accept either both a
> name and a mib or would otherwise not check the existence of said mib
I understand and agree with the goal of ensuring that names are
appropriately recognized and unique across MIB IDs.  But I also
recognize the need to use MIB IDs or names that are not known to the
implementation.  If the implementation were to reject names that were
known to be associated with a MIB ID other than what was provided, I
think that would be reasonable (though that would give the interface a
wider contract than I would prefer).
>
> 1.
>
>
>> Encoding have names, mib is very close to an implentation
>> details). For that same reason the name cannot be optional.
>> The name in parameter _always_ take precedence over the iana
>> name so it can roundtrip to iconv or similar APIs
>>
>
>>
>> To rephrase that:
>>
>> Different names may map to the same mib but two encoding with the
>> same names have to compare equal.
>
> I agree with that, but there is also present in the proposal that
> two encodings with different names that map to the same MIB ID
> compare equal.  The currently proposed design has the following
> issue:  IANA doesn't have a registration for the WTF-8 encoding
> today, but it is conceivable that it could be registered in the
> future.  As proposed, text_encoding("WTF8") !=
> text_encoding("WTF-8"), but if the encoding were to be registered
> in the future with both of those names as aliases, then they would
> compare equal despite having different names.  It seems to me that
> either 1) text_encoding shouldn't have equality comparison
> operators (though text_encoding::id should) or, 2) text_encoding
> should be split into two facilities, one that stores an integral
> ID, and another that performs name lookups and returns a value of
> the former.  The latter would also facilitate dynamic registration
> of additional names.
>
>
> Dynamic registration defeats the purpose.
How so?  I see at least two use cases for dynamic registration.  1) To
support encodings that are standardized in newer versions of the IANA
registry than the implementation is aware of, and 2) to support private
encodings that are not (yet) registered with IANA.
> I agree that a name becoming registered do change the behavior of the
> program.
> However, consider the case where you have
> text_encoding("wtf8",  68854);
> text_encoding("WTF-8");
>
> That does not solve anything.
I don't agree; it still results in IDs being created as specified. I
think the recognized names being implementation-defined is the real
problem here.
> Support for multiple names exist for legacy reasons, people using non
> registered encodings have to maintain consistency of names regardless.
I agree; and I think they should be able to use this facility to do so.
>
> Java does for example not expose the mib at all.
> There is just a isRegistered method that tells you if IANA knows about
> the encoding
>
>>
>> 1.
>>
>>
>>
>>> if text_encoding remains the name of that class,
>>> encoder/decoder can be used for the class doing the
>>> actual conversions.
>>>
>>>
>>> I will further rename "system" to "environment" to be
>>> more generic and aligned with POSIX.
>> Is text_encoding::system() intended to be equivalent to
>> text_encoding::for_locale(std::locale{})? (I think the
>> answer is, and should be, no; e.g., on Windows, this
>> would query GetACP()).
>>
>>
>>
>> It would query getacp which is equivalent to query the user
>> ("") locale at the start of the program.
>>
> Good, that sounds right.
>>
>>> (user, environment and system are, for our purpose
>>> synonym and intended to mean "the encoding assumed and
>>> expected by whatever launched our program).
>>> Environment has the added benefit that it implies
>>> neither user or systems which makes it more friendly to
>>> embedded platforms
>>
>> Since locale settings are generally determined by
>> environment (variables), use of the term "environment"
>> may be confusing.  I prefer system.
>>
>> Tangent 2: I don't recall if we discussed this in
>> Belfast, but the paper identifies three sets of encodings
>> to expose (literals, system, locale).  A fourth would be
>> terminal/console encoding. This encoding can be easily
>> queried on Windows, but not on Linux/UNIX (though
>> terminal encoding rarely differs from locale there, so it
>> would be reasonable to just return the system encoding).
>>
>> I considered that a few days ago. I am not aware of it being
>> a thing on other platforms than windows (I am probably wrong)
>> and I don't believe we can come up with a nice API in the
>> short term to query that nicely.
>>
> It is a thing on other platforms in that terminals (typically)
> have an encoding setting that specifies how characters are
> translated for input/display purposes.  What isn't common is a
> facility for querying the terminal for its encoding setting (as
> far as I can tell, there is no terminfo capability specified, nor
> have I found escape sequences that can request it though there are
> escape sequences for some terminal types to provide a character
> set map).
>
>
> My understanding is that on Linux the terminal exists prior to the
> program and cannot be changed.
Terminals (as in, their inodes) can technically be created, changed, and
associated with new processes.  In practice, this is almost never done
and isn't something I see any reason to spend time on.
> So that implies the terminal and the environment are one and the same.
In terms of encoding, I agree, the terminal encoding would have to be
assumed to match the locale.
> Which is different than the windows situation where the desktop is the
> environment and a console can be attached and modified at runtime.
Correct, the goal would be to support the Windows situation in a way
that is not inconsistent with other platforms.
>
> But maybe there is a way to do that on Linux?
>
> My point is, the proposal offers a way to query the encoding attached
> to std::cout, and anything else would implies doing something for
> consoles specifically, which seems like a larger discussion.

That is the discussion I'm trying to have here.  I think there is a
legitimate use case for Windows; GetConsoleCP() is used, presumably for
legitimate reasons.

Tom.

> Tom.
>
>> Tom.
>>
>>>
>>> Thanks for your input,
>>>
>>> Corentin
>>>
>>>
>>>
>>>
>>
>>
>



SG16 list run by herb.sutter at gmail.com