Re: Should the readline *-meta flags reset when $LANG changes?

[Koichi Murase]

2022年8月10日(水) 23:21 Chet Ramey <chet.ramey@case.edu>:
> > Does it mean custom values of these readline variables will be lost
> > every time LANG or LC_{CTYPE,ALL} is changed even if a user or program
> > intentionally sets them up?
>
> It means those settings will now mirror the locale.
>
> > We often temporarily change LANG or LC_* to perform some binary
> > operations [such as counting the number of bytes of data and safely
> > removing trailing x from the result of $(command;printf x)].
>
> Do you often do this in interactive shells?

Yes, but I don't mean I directly type the above kinds of commands in
the command line and run them, but I use them in the functions called
through `bind -x'.  Also, the above cases (counting bytes and removing
trailing x) are just examples; I set locale variables for various
purposes in the actual codes.  For example, I often type and run
commands of the form

  LANG=C some-commands-or-functions

to get the default error messages that are not locale-specific (though
I could use LC_MESSAGES=C instead, yet LANG=C is easier to type for
me).  I normally use the locale LANG=ja_JP.UTF-8 by default, so the
commands output error messages in Japanese by default.  This is not
useful when I would like to search for the solution on the internet
because there is almost no information on the Japanese error message.

> Often enough to make a difference?

My `bind -x' functions use `LC_ALL=' and `LC_CTYPE=C' for every
keystroke, for example, in combination with `builtin read'.  They also
use `LC_ALL=' for other purposes for mostly every keystroke.  Some vi
binding also uses `LC_CTYPE=C'.  My completion functions also change
`LC_ALL` and `LC_CTYPE`.  For example, `LC_CTYPE=C' is used in
calculating a PJW hash code of a given string.  I haven't carefully
checked, but there are probably other cases of changing `LC_CTYPE'.
Also, `LC_ALL=' is used everywhere.

> Across multiple calls to readline?

I think I am missing the point.  What does ``multiple calls to
readline'' mean?  Is the situation different from a single call to
readline?

Hmm, I think I first need to make it clear that the behavior of my
code, which is supposed to be sourced in an interactive session by
users, is unaffected by these readline settings.  I just do not want
to break or change the existing user settings inside the functions
that I provide.  The behavior of my functions is unaffected (except
for « bind -x '"\M-x":....'  » which is affected by `convert-meta',
for which I already implemented a workaround) because it doesn't try
to communicate with readline inside a single call of `bind -x'.  The
problem is that, with the new automatic adjustment of these readline
variables, the settings by users can be lost after using `LC_ALL=' or
`LC_CTYPE=C' inside my functions.

I believe this is a general problem for writers of Bash
configurations. `bash_completion' also uses `LC_CTYPE=C' and
`LC_ALL=C'.  The behavior of such configurations itself will be
unaffected by the change of readline settings, but they need to
implement special treatment to preserve the user settings if the user
settings will be lost by changing locales.

> And, if the change is intended to be temporary, why would you not
> want the relevant readline variables to reflect the locale when you
> were finished?

Because I would not like to break the users' settings.  In general, a
third-party Bash configuration should not overwrite the users'
settings as far as the configuration does not need the setting.

> > Also, if these readline variables would be cleared every time, it
> > seems to me that these readline variables would be effectively
> > unconfigurable and would lose the point of their existence, or we
> > could not touch LANG or LC_* at all after the initial setup.
>
> It seems to me that the scenario Alan describes is much more common.

I agree with this point because I have also faced this problem
for « bind -x '"\M-x":...' » vs « convert-meta » before.  For this
problem, I have added a partial workaround at my side [1] where I
decided to save and restore `convert-meta' before and after running
`bind -x'.  Actually, the patch [2] I posted in this list before has
been a part of the workaround to this problem.

[1] 
https://github.com/akinomyoga/ble.sh/commit/f32808070796d3978787f4491f812d06a629ab3f
[2] https://lists.gnu.org/archive/html/bug-bash/2019-02/msg00036.html

I agree that we should somehow change the current behavior that the
default values of *-meta settings are determined by the locale on the
startup of Bash, but the proposed change will break the opposite
scenario while it solves Alan's scenario.

The combination (UTF-8 & 7bit-mode) doesn't make much sense, so we
might force (UTF-8 & 8bit-mode) for UTF-8 or similar for multibyte
character encodings with 8-bit bytes.  [ Note: Here, 7bit/8bit-mode
means « convert-meta on/off » and « {input,output}-meta off/on »,
respectively. ] However, on the opposite side of the single-byte
character encoding (e.g. for C), I think combinations (C & 7bit-mode)
and (C & 8bit-mode) are both possible, so users can still set «
convert-meta off » or « {output-meta,input-meta,meta-flag} on ».

----------------------------------------

> I'm not going to make this much of a change at this point in the release
> process. I was willing to make the change I did because the changed
> behavior is a superset of the previous behavior.
>
> So, assuming we say that the scenario Alan outlined is reasonable (it is),
> it looks like there are four alternatives:
>
> 1. Do nothing; maintain the bash-5.1 behavior and force the change to the
>    user.
>
> 2. Leave the new function in place; automatically adjust to locale
>    changes.
>
> 3. Push it off to the application: introduce a new readline API that
>    applications can call when locale variables change. This is very cheap.
>
> 4. Push it onto readline: instead of checking the locale and making the
>    eight-bit variables mirror it on each call, make readline check for
>    locale changes (well, LC_CTYPE) and reset the eight-bit variables only
>    if the current value doesn't match the value from the last call.
>
> The last option is about as much of a change as I'm willing to make at
> this point.

2022年8月11日(木) 3:27 Chet Ramey <chet.ramey@case.edu>:
> There is a fifth option:
>
> 5. Make the locale-aware behavior dependent on a new readline option, which
>    would be enabled by default.

Maybe a large change should be considered for bash-5.3, but I still
think three states is one possible implementation that is a real
superset of the previous behavior:

6. Add a third state `auto' of `convert-meta', `input-meta'
  (`meta-flag'), and `output-meta' in addition to `on' and `off' and
  change the default values of these variables to `auto'.  When `auto'
  is set, the behavior always matches with the locale.  When `on' or
  `off' is set by a user or an application, the behavior is not
  affected by the current locale.

If these readline variables should always be uniquely determined by the
current locale and the users actually should never set them to the
different side, I think another option might be just to remove these
readline variables (though I'm not sure if this really makes sense):

7. Remove the readline variables `convert-meta', `input-meta'
  (`meta-flag'), and `output-meta'.

Or we can leave the readline variables but make it readonly:

8. Make `convert-meta', `input-meta' (`meta-flag'), and `output-meta'
  readonly and do not allow users to change these settings through
  inputrc or « bind 'set ...' ».  The users instead need to set
  LC_CTYPE up.

Then, the Bash configuration writers do not have to be bothered with
saving and restoring the custom user settings when touching LC_ALL and
LC_CTYPE.

--
Koichi