bug#27193: 25.2; tmm should use completing-read-default

[Ryan Thompson <rct@thompsonclan.org>]

[-- Attachment #1: Type: text/plain, Size: 9393 bytes --]

On Fri, Jun 2, 2017 at 11:26 AM Drew Adams <drew.adams@oracle.com> wrote:

> I don't understand that argument.  (But to be clear, I don't
> really care much about `tmm'.)
>

Fair enough. My post certainly assumes quite a bit familiarity with tmm.
I'm happy to elaborate below.


> I don't see that that is an argument for _preventing_ the
> use of a `completing-read-function' by tmm.  I see it
> only as a statement that it might not be helpful to use
> some particular values of `completing-read-function' with
> tmm.
>
> If there is a general problem then consider adding a
> comment in `tmm.el' (or perhaps put it in some doc
> string) that says what you are really saying: Some
> values of `completing-read-function' might not be
> helpful with `tmm', and if you find that is the case,
> consider binding that variable to nil.
>
> But now that I've taken a quick look (just now) at the
>
use by tmm of `completing-read', I don't see that there
> is a problem to be fixed in `tmm.el'.  I don't see that
> its use of `completing-read' is particularly exotic or
> problematic.  This is it:
>
> (completing-read
>  (concat gl-str " (up/down to change, PgUp to menu): ")
>  (tmm--completion-table (reverse tmm-km-list)) nil t nil
>  (cons 'tmm--history (- (* 2 history-len) index-of-default)))
>
> I don't see anything at all unusual about that.
>
> And the collection function, `tmm--completion-table',
> is likewise pretty ordinary, I think:
>
> (defun tmm--completion-table (items)
>   (lambda (string pred action)
>     (if (eq action 'metadata)
>         '(metadata (display-sort-function . identity))
>       (complete-with-action action items string pred))))
>
> You say:
>
> > tmm already pretty much relies on the assumption that
> > completing-read is actually calling completing-read-default.
>
> I don't see any evidence of that.
>

tmm's weirdness is not in the actual call to completing-read. That
completing-read call is wrapped by "(minibuffer-with-setup-hook
#'tmm-add-prompt ...)", which in turn calls tmm-define-keys, which sets up
a bunch of non-standard key bindings, which have the effect of implementing
a menu system with single-keypress activation of menu items, rather than
completion of a substring with string matching. The result is not even
recognizable as completing-read. The use of completing-read is merely an
implementation detail in a system that is *not* actually doing completion
of any kind. So pluggable completion backends don't make any sense for tmm,
and I can't imagine any other value of completing-read-function that would
make sense for tmm besides completing-read-default. Looking at the "git
blame" output for tmm, that call to completing-read has definitely not been
updated since completing-read-function was introduced except for minor
bugfixes, so it makes sense that tmm would be expecting one and only one
implementation of completing-read.


> This kind of argument could (inappropriately, IMO) be
> applied to any number of completely normal uses of
> `completing-read'.
>
> I see no reason to impose a dichotomy of either a
> `completing-read-function' similar to yours or else
> `completing-read-default'.  There are likely other
> benign values of the variable, besides just
> `completing-read-default'.
>

I'm not trying to set a general precedent here. tmm is the only code that
I'm aware of that uses completing-read in this way.

It sounds like (and no, I haven't looked into it;
> it just sounds like it) you have some particular
> `completing-read-function' values in mind, which
> you have found are incompatible with tmm's use of
> `completing-read'.
>

The alternative completing-read-function providers that I am aware of are
of are ido-ubiquitous (written by me), ivy, and helm. I'll also include
icicles, even though uses some other mechanism besides modifying
completing-read-function. ido-ubiquitous and ivy both have code to
deactivate themselves when completing-read is called by tmm because
otherwise their completion systems would break it, while icicles and helm
simply break tmm when enabled. Thus, to my knowledge there is currently no
other completing-read-function that doesn't break tmm (except for those
that make exceptions specifically for tmm).

If so, that's not an argument for preventing the use of
> other values of `completing-read-function' with tmm.
> (Clearly the value `completing-read-default' is fine,
> for instance.)  That's not an argument for tmm to do
> something to prevent all use of `completing-read-function'.
>
> Instead, it's an argument for the code that defines and
> uses a particular `completing-read-function' to take
> care of the contexts where it makes sense, and to stop
> itself from being used in other contexts, where it might
> cause trouble.
>
> Only that code knows the kinds of context where its own
> `completing-read-function' makes sense and where it does
> not.  Code such as tmm should not try to guess what kinds
> of trouble different values of `completing-read-function'
> might cause.
>
> I don't think tmm should throw up its hands and say, "Gee,
> there might be some values of `completing-read-function'
> that are troublesome, so let's just prevent all use of
> that variable."  That doesn't make sense, to me.
>

Based on my explanation above, that is precisely what I think tmm should
do: avoid using completing-read-function entirely. To look at it another
way, tmm was originally written to use completing-read as an implementation
detail, and later the function that used to be called completing-read was
renamed to completing-read-default, but tmm was never updated to use the
new name. This patch rectifies that.


> If you want additional suggestions, maybe describe just
> what the problem is that your completion function causes
> for tmm.  It's hard to offer suggestions if you only
> state that it is incompatible, without going into any
> detail.  (Not that you must ask for input about this.
> But if you would like some then giving more detail might
> help.)
>
> Please use your own judgment (as I said, I don't really
> care much about `tmm'), but a priori this sounds like
> overkill.
>
> It sounds a bit like trying to bend Emacs to fit your
> `completing-read-function'.  I can understand such a
> motivation, believe me; I don't ascribe a bad intention
> to you.  A guess is that you are not sure what to do,
> to prevent inappropriate application of your value of
> `completing-read-function' in this or that context.
>
> If you think it causes trouble in some contexts, or it
> is not able to handle some contexts properly, then
> I'd suggest you consider walling it off from those use
> cases.  It might take some time to discover which
> contexts it causes trouble for, but that's OK - you
> could add them as you discover them.  Tmm sounds like
> a start.


> The right approach, IMO, is to teach your code when to
> use its `completing-read-function' and when not to use
> it.  Put differently, consider teaching your
> `completing-read-function' when and where to hold back
> and just punt to the default behavior.
>

This is exactly how ido-ubiquitous and ivy both currently work: they
essentially have a blacklist of callers for which they revert to standard
completion. tmm is on the blacklist for both packages. Certainly, for any
alternative completion system there will be cases where it needs to fall
back to standard completion. In my view, the completion system should be
able to determine purely based on the calling context (i.e. its arguments
and any relevant dynamically-bound variables) whether or not it needs to
punt. Making this decision based on the name of the caller instead of the
context to make this decision is admitting that not enough context was
provided. I view it as a workaround, not a desirable design pattern, and
someday in the future I hope to be able to remove the blacklist from
ido-ubiquitous.

In the case of tmm, the best heuristic I can think of would be to inspect
the key bindings of all the letters and numbers. If any of them are locally
rebound in the minibuffer to something other than self-insert-command, then
punt. However, this doesn't work in practice because the bindings happen in
minibuffer-setup-hook, so putting a check at the front of this hook is too
early, and putting it at the end is too late because (in the case of
ido-ubiquitous) an error is triggered before reaching the end of the hook.
This was partly my motivation for suggesting the change in tmm rather than
working around it in ido-ubiquitous: because the blacklist approach is the
only way for ido-ubiquitous to fix it.

It's obvious that it is possible for someone to create
> a `completing-read-function' that causes trouble here
> or there.  But such trouble is up to the causer to fix
> or tame.
>

For the reasons described above, I would be very surprised if there was
*any* alternative completion system that didn't break tmm.

The approach of preventing code like `tmm.el' from
> letting other code use `completing-read-function' does
> not look like it heads in the right direction.  But
> mine is just one opinion.  I ask only that you think
> some more about this.
>

As mentioned above, tmm is the only code I'm aware of that does anything
like this, and I don't see this as a general approach to be applied
anywhere else.

Hopefully the above clarifies my rationale in proposing this change.

[-- Attachment #2: Type: text/html, Size: 11671 bytes --]