On Fri, May 3, 2024 at 10:11 PM Tom Honermann <tom_at_[hidden]> wrote:

> On 5/2/24 6:35 PM, Corentin Jabot via SG16 wrote:
>
>
>
> On Thu, May 2, 2024 at 11:25 PM Tom Honermann <tom_at_[hidden]> wrote:
>
>> On 4/30/24 2:31 AM, Corentin Jabot via SG16 wrote:
>>
>>
>>
>> On Tue, Apr 30, 2024 at 12:45 AM Tom Honermann <tom_at_[hidden]> wrote:
>>
>>> On 4/29/24 4:11 PM, Peter Dimov via SG16 wrote:
>>> > Tom Honermann wrote:
>>> >> I'm not entirely sure that cout << std::format("{}", u8"...") is
>>> that much
>>> >> easier
>>> >> to specify and support.
>>> >>
>>> >> But I'll be glad to be proven wrong, of course. :-)
>>> >>
>>> >> There is a relevant SO comment
>>> >> <https://stackoverflow.com/questions/58878651/what-is-the-printf-
>>> >> formatting-character-for-char8-t/58895428#58895428> .
>>> >>
>>> >> std::format() and std::print(), to some extent, improve the
>>> likelihood that an
>>> >> implementation selected encoding will be a good match for the
>>> programmer's
>>> >> intent. This is because:
>>> >>
>>> >> 1. std::format() and std::print() are not implicitly locale
>>> dependent; that
>>> >> rules out selection of a locale dependent execution encoding.
>>> >> 2. std::format() returns a std::string; that rules out selection of
>>> an I/O
>>> >> dependent encoding.
>>> >> 3. std::print() writes to an I/O stream, but has special behavior
>>> for writes
>>> >> to a terminal; that rules out selection of a terminal encoding (as
>>> unnecessary,
>>> >> at least in important cases).
>>> >> 4. std::format() and std::print() are both strongly associated with
>>> the
>>> >> ordinary/wide literal encoding.
>>> >> 5. std::format() and std::print() should have the same behavior
>>> (other than
>>> >> that std::print(...) may produce a better result than std::cout <<
>>> >> std::format(...) when the output is directed to a terminal).
>>> >> 6. std::format() and std::print() have additional guarantees when
>>> the
>>> >> ordinary/wide literal encoding is a UTF encoding.
>>> >>
>>> >>
>>> >> Due to those characteristics, we have good motivation for implicit
>>> use of the
>>> >> ordinary/wide literal encoding as the target for transcoding for
>>> std::format()
>>> >> and std::print().
>>> > I'm afraid that I don't quite understand.
>>> >
>>> > What does std::format( "{}", u8"..." ) actually do? I suppose it
>>> transcodes
>>> > the UTF-8 input into the narrow literal encoding (replacing
>>> irrepresentable
>>> > characters with '?' instead of throwing, I presume, or it would be not
>>> very
>>> > usable)?
>>>
>>> We'll have to see what Corentin proposes :)
>>>
>>> But yes, something very much like that.
>>>
>>> Note that we could also support std::format("{:L}", u8"...") to enable a
>>> programmer to explicitly request transcoding to a locale dependent
>>> encoding (either now or at some future point).
>>>
>>> (Corentin, at a minimum, we should reserve the L option in your paper).
>>>
>>
>> We have an opportunity to not conflate locale and encodings here.
>>
>> As much as I would like that to be the case, I don't think it is.
>>
>> u8"" is a known quantity here, it's utf-8.
>> But the target is also a known quantity, we very clearly decided it to be
>> the literal encoding, because we need to parse it, and
>> we wisely decided to assume a literal encoding. So the target encoding is
>> also a known quantity
>>
>> Unfortunately, that isn't the case when a programmer opts in to use of a
>> locale. Consider the following when the literal encoding is any ASCII
>> derived encoding and the global locale encoding is EUC-JP (ujis).
>>
>> #include <chrono>
>> #include <format>
>> #include <iostream>
>> #include <locale>
>> int main() {
>> std::locale::global(std::locale(""));
>> std::cout << std::format("{:L}\n", std::chrono::August);
>> }
>>
>> The resulting string will be formed from the literal encoding (for the
>> '\n' character) and the name of the month provided by the *formatting
>> locale <http://eel.is/c++draft/time.format#2>*. Nothing ensures that the
>> latter is converted to the literal encoding. Further, a validly encoded
>> string is produced so long as the characters used in the format string are
>> from the basic literal character set.
>>
>> In my environment (Linux, using a pre-release build of Clang 19 and
>> libc++), compiling the above with the default literal encoding (UTF-8) and
>> running it with LANG=ja_JP.ujis produces output in EUC-jp as expected;
>> note the iconv invocation.
>>
>> $ clang++ -std=c++23 -stdlib=libc++ t.cpp -o t
>> $ LANG=ja_JP.ujis ./t | iconv -f ujis -t utf-8
>> 8月
>>
>> (yes, that is the right output, it is convention for some translation of
>> month names to include the month number before the localized name).
>>
>> Long time SG16 participants will recall P2373R3 (Fixing locale handling
>> in chrono formatters) <https://wg21.link/p2372r3> and LWG 3547
>> <https://wg21.link/lwg3547>. There was relevant discussion during the 2021-04-28
>> SG16 meeting
>> <https://github.com/sg16-unicode/sg16-meetings/blob/master/README-2021.md#april-28th-2021>
>> .
>>
>> I have vague recollections of discussions about requiring that locale
>> dependent translations be provided in the literal encoding when it is a UTF
>> one, but I haven't been able to identify any such recorded discussion. I
>> don't see anything in the current WP that would require this.
>>
>> Based on the above, I think that, at a minimum, the "L" option should be
>> reserved.
>>
>
> I'm not sure what you are arguing about because "L" can only be applied to
> things that can be "localized" (i.e. mangled horribly by POSIX).
>
> You are right that I wasn't very clear about what I'm suggesting. I'll try
> to clarify.
>
> The ASCII and EBCDIC code page based locale programming model used on
> POSIX and Windows systems is not broken. It does have sharp edges. Unicode
> and its associated encodings have enabled a new programming model with
> fewer constraints and pitfalls, but that has not completely displaced the
> code page based programming model nor do I think it ever will. The code
> page based programming model requires the following:
>
The windows model is different (and matches more closely Unicode), and they
have to bend backwards to emulate posix for the sake of C and C++.


>
> 1. Since C and C++ programs start with the global locale set to "C",
> it is necessary to opt-in to locale dependent behavior by calling
> std::locale::global() and/or std::setlocale().
> 2. Such programs, in order to avoid mojibake, must constrain the use
> of compile-time selected characters encoded in the ordinary literal
> encoding to those that have an invariant representation in all supported
> locale dependent encodings
>
>
Again the limit of the POSIX model is that it conflates "How to spell 4th
of July" in Japanese with "How to encode "7月4日" in Shift-JIS.
There is no saving this model.

There has been a lot of code written over the last 40 or so years that
> adheres to this model. Many such programs are effectively locale agnostic
> though full localization requires translations provided by message catalogs
> (that themselves rely on locale; GNU gettext
> <https://www.gnu.org/software/gettext/> and POSIX catopen
> <https://pubs.opengroup.org/onlinepubs/9699919799/functions/catopen.html>
> are relevant). In my opinion, these programs should continue to work and
> continue to benefit from C++ standard library enhancements.
>
I would object to "effectively". We did our best to work with the
constraint we got. (And then again, a large number of scripts and
applications are not served well with that model).
I would submit the present thread as evidence to my claims
No one is suggesting breaking existing code, but it doesn't mean that we
should either reconsider our design decision, or try to shoehorn Unicode
into POSIX.
The posix model exists out of a need to localize for multiple scripts
pre-utf8, where encoding could only represent a limited set of characters.
Anyway, we are getting side tracked!


> Let's look at that example from above again:
>
> std::cout << std::format("{:L}\n", std::chrono::August);
>
> Regardless of what the ordinary literal encoding is, if LANG is ja_JP.sjis,
> then valid Shift-JIS output will be produced. Likewise, if it is
> ja_JP.utf8, zh_CN.gb18030, or zh_TW.big5, valid output will be produced
> in those encodings. This is portable code that works on all platforms (with
> the right platform dependent locale names; those sadly are not portable).
>
Only because the format string is otherwise empty.


> Let's now assume a hypothetical message catalog of translated strings that
> works similarly to gettext, but that provides UTF-8 encoded translations in
> char8_t.
>
> std::cout << std::format("{} {:L}\n", u8msg("In the month of"),
> std::chrono::August);
>
> If we unconditionally require the char8_t argument to be transcoded to
> the ordinary literal encoding, then mojibake will be produced unless the
> ordinary literal encoding happens to match the locale encoding.
>

This is status quo: u8 has no impact there.
Consider:

std::cout << std::format("{} {:L}\n", "In the month of",
std::chrono::August);

If the literal encoding and the encoding associated to the locale don't
match, you get non sense


> I strongly agree that, for std::format(), the default behavior should be
> that char8_t strings are transcoded to the ordinary literal encoding.
>
> What I am arguing for is that there should also be an option for the
> programmer to opt-in to locale based transcoding of arguments that
> potentially require transcoding. Thus:
>
> std::cout << std::format("{:L} {:L}\n", _u8("In the month of"),
> std::chrono::August);
>
>
What you want apply equally to std::cout << std::format("In the month
of {:L}\n",
std::chrono::August);
And again, status quo is the result of ""localizing"" std::chrono::August will
be transcoded back to the literal encoding


> would portably produce correct locale dependent output (and transcoding
> would be reduced to a byte copy when the locale encoding is UTF-8).
>
> For the short term however, I'm content to just reserve the 'L' option;
> actually doing the work to support this can await further motivation and
> standard transcoding facilities.
>
> std::format("{:L}\n", ""); is ill-formed, so would be std::format("{:L}\n",
> u8"");
> https://eel.is/c++draft/format#string.std-17 (it's also used in chrono)
> https://compiler-explorer.com/z/58bsTaf3o
> Beside, reservation is not necessary, users cannot write formatters for
> types that do not depend on user-defined types (or, if you prefer, it's
> already reserved)
>
> Victor can correct me if I'm mistaken, but my understanding has been that
> changes to std::formatter specializations might cause (sometimes?) an ABI
> break. The following is recorded in the 2023-11-29 SG16 meeting
> <https://github.com/sg16-unicode/sg16-meetings/blob/master/README-2023.md#november-29th-2023>
> notes from discussion of - P3045R0 (Quantities and units library)
> <https://wg21.link/p3045r0>.
>
> "Victor recommended reserving an 'L' option specifier in the format
> specification that would render the code ill-formed for now so as to allow
> extension later without an ABI break."
>
>
> But the issues with the whole scenario you are describing is that:
>
> 1.
> We keep trying to give meaning to programs where the execution encoding is
> not a superset of the literal encoding even though the encoding is
> generally not part of the type system
> So for 2 arbitrary strings a and b, concatenating them might not produce a
> good result, and we can't solve it.
>
> Such programs already have a meaning and have for the last 40 years or so.
> I agree there are sharp edges here that we can't fix.
>




> You will note that format makes the assumption that everything is in the
> literal encoding and it's working wonderfully well.
>
> std::format() does not ensure that the output produced is an any
> particular encoding. We spent a lot of time talking about whether
> std::format() produces text and eventually concluded that it is not
> required to do so.
>
Sort of, except when we do (L, width, escaping, etc)!

> I am not arguing for a change in direction; in fact, I'm arguing with
> preserving consistent behavior with regard to its existing locale dependent
> behavior so that there is an option for *not* producing mojibake.
>
> It's certainly not perfect - i.e. we taught people to compile with /utf8
> on windows but the system is still not defaulting to UTF-8, but it's as
> good as we can reasonably get.
>
> Agreed. And for those that are able to use /utf8, that is great. I have
> no data, but I would bet a good deal of cash that the vast majority of code
> that is compiled with MSVC is not compiled with /utf-8.
>
>
> 2.
> When you ask for the name of August in Japanese, as a user, you probably
> don't expect part of your program to be encoded in some weird encoding that
> is different to the rest of the program.
> We try to patch that in format/chrono, but it's certainly not perfect
> https://eel.is/c++draft/time#format-3.sentence-3
>
> Thank you! That is the wording I was looking for with regard to my "vague
> recollections of discussions" statement above.
>
>
> Anyway, I'm not sure how that is relevant to the u8 discussion, L affects
> individual arguments, not the formatting string (the literal encoding is
> the ground truth for encoding as far as format is concerned)
>
> I hope the above better explains the relevance.
>
> Tom.
>
>
>
>
>
>
>
>>
>>
>>
>>>
>>> >
>>> > And then we just fall back to std::cout << "...", where the "..." is
>>> in the
>>> > narrow literal encoding and hence we assume works, more or less.
>>> Correct.
>>> >
>>> > And we don't want to make std::cout << u8"..." do that, because it can,
>>> > in principle, do better?
>>> Not because it can do better, but because there is more uncertainty
>>> about what the user might expect. If the user writes std::cout <<
>>> std::format(...), then that is an explicit opt in to the behavior that
>>> std::format() exhibits. But they might also want to just write UTF-8
>>> bytes unmodified regardless of what the ordinary literal encoding is. Or
>>> they might expect implicit transcoding to either the current locale or
>>> the environment locale or even the terminal locale. By not providing a
>>> default behavior, we give the programmer the opportunity to think about
>>> what they are actually trying to do.
>>>
>>
>> I don't quite buy this argument.
>> When cout << 42.0; outputs "42,0", the text nature, locale and encodings
>> were made for us.
>> If the programmer wants to be creative, one can consider io manipulators.
>>
>> Consider printing of other localized names as in the example above.
>>
>> #include <chrono>
>> #include <format>
>> #include <iostream>
>> #include <iomanip>
>> #include <locale>
>> int main() {
>> std::cout << "Default locale: '" << std::cout.getloc().name() << "'\n";
>> std::cout << std::chrono::August << "\n";
>> std::cout.imbue(std::locale(""));
>> std::cout << "Environment locale: '" << std::cout.getloc().name() <<
>> "'\n";
>> std::cout << std::chrono::August << "\n";
>> std::cout.imbue(std::locale("ja_JP.utf8"));
>> std::cout << "Explicit locale: '" << std::cout.getloc().name() << "'\n";
>> std::cout << std::chrono::August << "\n";
>> }
>>
>> I get the following output running that locally with LANG=ja_JP.ujis.
>> Note the mojibake and corresponding substitution of replacement characters.
>>
>> Default locale: 'C'
>> Aug
>> Environment locale: ''
>> 8��
>> Explicit locale: 'ja_JP.utf8'
>> 8月
>>
>> The (well recognized) problem with iostreams is the implicit use of the
>> imbued locale. The consistent behavior for iostreams would be that
>> inserters and extractors for charN_t would transcode to the encoding of
>> the imbued locale. But that doesn't work well at all in the common case
>> where no locale has been explicitly imbued.
>>
>> Making a choice for std::format() is simpler because the programmer
>> chooses the locale behavior on a per-argument basis; there is a good
>> default.
>>
>>
>>
>>> >
>>> > But let me get back to your list.
>>> >
>>> >> 1. std::format() and std::print() are not implicitly locale
>>> dependent; that
>>> >> rules out selection of a locale dependent execution encoding.
>>> > What is in a locale-dependent execution encoding in std::cout <<
>>> u8"..."?
>>> iostreams implicitly consults either an imbued locale facet or the
>>> global locale for formatting operations. Think about std::cout <<
>>> std::chrono::Sunday. Depending on the current locale, this might print
>>> "Sun" or a localized weekday name in a locale dependent encoding.
>>>
>>
>> But again, the only thing we care about for u8 is the encoding.
>> And I am not aware of std::locale ever impacting that.
>>
>> I hope the above examples are motivating.
>>
>>
>>
>>> >
>>> >> 2. std::format() returns a std::string; that rules out selection of
>>> an I/O
>>> >> dependent encoding.
>>> > Same question. Where is the I/O dependent encoding in std::cout <<
>>> u8"..."
>>> > (that is not also present in std::cout << some_std_string)?
>>> In the latter case, we have to assume that some_std_string holds text in
>>> the encoding expected on the other end of the stream. We can't do that
>>> for u8"...", so we have to transcode to something (or have some other
>>> assurance that UTF-8 is intended and expected).
>>> >
>>> >> 3. std::print() writes to an I/O stream, but has special behavior
>>> for writes
>>> >> to a terminal; that rules out selection of a terminal encoding (as
>>> unnecessary,
>>> >> at least in important cases).
>>
>> > This doesn't apply here, because we're using std::format.
>>>
>>
>> Right, this is one of the reasons I feel less compelled to pursue
>> iostream surgery.
>> Output behavior is suboptimal on windows, and unlikely to be fixed.
>>
>> I am likewise not compelled to pursue iostream support.
>>
>> Agreed with later remarks below.
>>
>> Tom.
>>
>>
>>
>>> >> 5. std::format() and std::print() should have the same behavior
>>> (other than
>>> >> that std::print(...) may produce a better result than std::cout <<
>>> >> std::format(...) when the output is directed to a terminal).
>>> > OK... but this isn't relevant.
>>> The above two are relevant because we wouldn't want to differentiate
>>> behavior for formatting a u8"..." argument for std::format() vs
>>> std::print(). The latter helps to constrain the reasonable options for
>>> the former.
>>
>>
>> Right, print just does format and output the result
>>
>>
>>> >
>>> >> 6. std::format() and std::print() have additional guarantees when
>>> the
>>> >> ordinary/wide literal encoding is a UTF encoding.
>>> > What additional guarantees, and how do they help here?
>>>
>>> We specify additional constraints for fill characters, display width
>>> (well, normative encouragement), and formatting of escaped strings. None
>>> of these are relevant for reflection purposes; they help to reinforce a
>>> choice to depend on the ordinary/wide literal encoding for behavior of
>>> these functions. We don't have such precedent for iostreams.
>>>
>>
>> And you know, the format string is parsed in the ordinary encoding and
>> copied as-it
>>
>>
>>>
>>> Tom.
>>>
>>>
>>
>