bug#30186: 27.0.50; Password is not hidden in read-passwd

[Alan Mackenzie <acm@muc.de>]

Hello, Eli.

On Sat, Jan 27, 2018 at 11:40:36 +0200, Eli Zaretskii wrote:
> > Date: Fri, 26 Jan 2018 20:00:27 +0000
> > Cc: juri@linkov.net, 30186@debbugs.gnu.org
> > From: Alan Mackenzie <acm@muc.de>

> > Do you have any opinion on my suggestion:

> > > > Perhaps an alternative would be for Emacs to provide a flag which
> > > > indicates to a before/after-change-function whether the current
> > > > change is a "proper" change or merely a text property change.
> > > > Change hook functions could then test this flag and, for example,
> > > > refrain from doing anything for a text property change.

> I'm not sure it would be possible to provide such a flag.  Did you
> look at the internals involved, and if so, can you tell where do we
> know which kind of change caused the hooks to run?

I envisage adding an extra boolean argument to prepare_to_modify_buffer,
and to signal_after_change.  When called from the text property
routines, that argument would be true, otherwise it would be false.

The essence of this argument would be a guarantee that the change is not
going to alter the buffer text.  There may be other primitives besides
text properties for which we could set this argument.

> I also don't understand the problem you have in CC Mode with
> text-property changes.  Can you elaborate on that?

Yes.  In a largish CC Mode file, mark the entire buffer, kill-ring-save,
and append it after itself, with

    C-x h, M-w, M->, C-y

.  In buffer-undo-list there are no entries for text property changes.
Before the with-silent-modifications was put in, there were many such
entries.  Assume this in the next paragraph

Now undo this latest change with C-_.  Each of the entries for text
property changes wants to invoke CC Mode's
before/after-change-functions, which would make the operation slow.
(But see below.)

The workaround (currently in Emacs) for this, back in 2015, was to put
with-silent-modifications around the text property manipulations in
remove-yank-excluded-properties.  This prevents these manipulations
getting into the undo list, but also stops read-passwd from working
properly.

I now see there is a second workaround, in CC Mode itself, where
c-before-change and c-after-change use backtrace-frame to check the
primitive invoking them, and do nothing if that primitive is, e.g.,
put-text-property.  This is not an elegant workaround.

So, I'm changing my mind, after looking into it a bit more.  Removing
the with-silent-modifications from remove-yank-excluded-properties would
not slow down undo in CC Mode buffers noticeably.  It might slow down
other modes which make extensive use of before/after-change-functions.

The extra flag for the change hooks might still be a good idea.  It no
longer seems pertinent for solving the current bug, though.

-- 
Alan Mackenzie (Nuremberg, Germany).