On 11/28/20 11:33 AM, Victor Zverovich wrote:
To keep this thread manageable here are my answers to Tom and Jens (thanks for the feedback!) in one convenient wall of text.

Answers to Tom:

> My choice of Windows-1251 for an example scenario was motivated solely by the use of Russian characters in the example in the paper.

Sure but it nevertheless a great choice that clearly demonstrates that ACP is definitely the wrong thing when files are involved. That said, EBCDIC and other encodings are still supported via the non-Unicode path.
I don't feel that same level of clarity.  There is a distinction to be made regarding writing to a file vs writing to stdout.  In the former case, the programmer has wide latitude for choosing an encoding and knows that the content is being written to a file.  In the latter case, the programmer doesn't (in general) know whether the output is redirected to a file, pipe, or some other character device.

> I don't think the Notepad example is particularly relevant.

It is relevant for #2 because it shows that when a Russian user creates a text file on Windows it will most definitely be encoded in UTF-8 and not "ANSI" encoding (and definitely not the terminal encoding). This is true for Notepad and other popular editors. Same with files obtained from the Internet. We should understand the common encoding for text files in order for our text facilities to be useful and consistent.

I think this misses the concern to some degree.  For #2, it is the user that is making the choice to write the output to a text file, not the programmer.  I believe the programmer should have the ability to choose the encoding used (preferably with the ability for the user to influence the choice). but I'm (so far) uncomfortable with the behavior being tied to the execution/literal encoding chosen at compile time; that choice is historically distinct from the run-time encoding used by the environment the program runs in.

For example, I believe it would be a reasonable choice for a z/OS programmer to use UTF-8 as the execution/literal encoding and still run that program in an EBCDIC environment.  This is how Java works in such environments (using UTF-16 internally of course).  This is the Unicode sandwich model.


> There is no particular expectation that a .txt file was produced by a program running on the local machine, so the local code page isn't a particularly good default in any case.

Exactly.

> If I write a version of the Windows 'type' command as you used it above, call it 'cat', compile it without Microsoft's /utf-8 option, then I would like it to still do the right thing; not the behavior you illustrated above.

I misunderstood your suggestion. Are you suggesting for the non-Unicode path (print_nonunicode) to do the transcoding to the encoding determined by ACP and for the Unicode path (print_unicode) to produce UTF-8? Note that using ACP won't solve the mojjibake problem because the terminal encoding (CP866) is separate from the ACP encoding, at least for Russian. Confusing those two is a common misconception and source of problems (see e.g. https://stackoverflow.com/questions/49259502/windows-console-codepage-866). Using the terminal encoding would produce completely useless output for anything but interaction with legacy command-line programs via pipes (and even there the usefulness of the result of the pipeline is questionable).

No.  My suggestion was that, when writing to a stream known (e.g., via _isatty()) to be directly connected to the Windows console, that the Unicode path be taken regardless of what the execution encoding is, with transcoding to UTF-16 performed as necessary.  My expectation is that writes to the console would be performed using the WriteConsoleW() function; that is where UTF-16 comes in.  So, when the execution encoding is UTF-8, the implementation would convert to UTF-16 and then call WriteConsoleW() and, for other execution encodings, the implementation would transcode to UTF-16 and then call WriteConsoleW().  This approach bypasses the console encoding entirely (the console encoding is only relevant for the ANSI implementation of the console APIs and for reads/writes to the console via ReadFile() and WriteFile()).


> That is true only if the execution/literal encoding and the run-time encoding do not match

Yes and if we use ACP they will likely not match.
For Windows, that is true, but also a reality of that environment.  For other platforms, the likelihood of a mismatch is far lower (though not 0; the LANG environment variable is still used in POSIX environments and can still be set to select an encoding other than UTF-8).

> Assuming test.txt is UTF-8 encoded, that is correct; this is a straightforward case of mojibake.

test.txt is CP1251 encoded. This example illustrates that using ACP doesn't solve mojibake.

Perhaps we are focused on different instances of mojibake.  I think you are pointing out that the output of the findstr command will fail to present properly because the console encoding doesn't match.  The mojibake I was alluding to is that findstr will fail to find a match in the file because the encoding of the pattern string (as entered from the console on the command line) doesn't match the encoding of the file (unless findstr consults the wide/UTF-16 variant of its command line).


> Perhaps a 'formatter' specialization should be provided for std::filesystem::path?  Proposing something like that is likely subject matter for a different paper, but I think it would be helpful for this paper to discuss it.

I think that providing such specialization would be useful but it is out of scope of the current paper since it has nothing to do with I/O integration.
I think the relevant question for this paper, given that it does intend to specify encoding conversions in at least some cases, is how output such as filenames that may have content that is not well-formed according to the execution encoding, can be incorporated.  I think my preference is to have some method to opt-out of implicit conversions; probably via a per-field format flag.

> What happens if the UTF-8 input is ill-formed?

Good question. The current implementation throws an exception on transcoding error but the error handling mechanism is open for discussion.
I don't find throwing an exception to be acceptable, but attempted conversion with U+FFFD substitution as suggested by Peter seems ok (perhaps with an opt-out as suggested above); I prefer a loss of precision over a loss of output.

Answers to Jens:

> Doing std::format without necessarily creating a std::string is useful functionality, but unrelated to the transcoding issues. Thus, this facility should be separate.

Such a facility already exists in C++20 (format_to, format_to_n). The current paper only integrates it with I/O without adding any new functionality on the formatting level.

> Apparently, there is some OS-dependent magic going on to determine whether output is to a console and, if so, which encoding the console might prefer.  I'm fine with such magic existing, but it should be a distinct facility.

Sure, I will extract it into a separate API in the next revision of the paper.

> And then there is the facility of converting the C++ literal encoding to the console encoding, if necessary.  Again, this should be a separate facility, preferably offering a generic transcoding facility that can be specialized for the console-only use case.

While I agree that such a transcoding facility would be useful I think it is out of scope of the current paper. The latter requires only minimal transcoding facilities for the Unicode case and only on some platforms where dedicated system APIs exist.
I agree that distinct interfaces should be provided for each of these concerns, but I also think each can be pursued separately and need not hold up the proposed feature.  We can always re-specify the proposed behavior in terms of new interfaces via as-if in the future.

Also, progress is being made on these; JeanHeyd is continuing to work on general transcoding facilities.  See WG14 N2595 for his most recent work (we'll be discussing this paper in SG16 early next year).

Tom.


Cheers,
Victor


On Fri, Nov 27, 2020 at 11:15 AM Jens Maurer via SG16 <sg16@lists.isocpp.org> wrote:
On 23/11/2020 06.33, Tom Honermann via Lib-Ext wrote:
> SG16 began reviewing P2093R2 <https://wg21.link/p2093r2> in our recent telecon <https://github.com/sg16-unicode/sg16-meetings#november-11th-2020> and will continue review in our next telecon scheduled for December 9th.
>
> The following reflects my personal thoughts on this proposal.

Ditto.

As I've already said in the SG16 review, I'd like to see
smaller bits and pieces offered, instead of or at least in
addition to hiding them behind a non-trivial "printf"-style
wrapper.

 - Doing std::format without necessarily creating a std::string
is useful functionality, but unrelated to the transcoding issues.
Thus, this facility should be separate.

 - Apparently, there is some OS-dependent magic going on to
determine whether output is to a console and, if so, which
encoding the console might prefer.  I'm fine with such magic
existing, but it should be a distinct facility.

 - And then there is the facility of converting the C++ literal
encoding to the console encoding, if necessary.  Again, this
should be a separate facility, preferably offering a generic
transcoding facility that can be specialized for the console-only
use case.  (Only supporting that single transcoding might save
binary size.)


Jens


--
SG16 mailing list
SG16@lists.isocpp.org
https://lists.isocpp.org/mailman/listinfo.cgi/sg16