Re: [WIP PATCH] Controlling Isearch from the minibuffer

[Alan Mackenzie <acm@muc.de>]

Hello, Augusto.

On Sun, May 09, 2021 at 19:58:19 +0200, Augusto Stoffel wrote:
> On Sun,  9 May 2021 at 13:36, Alan Mackenzie <acm@muc.de> wrote:

> Hi Alan,

> > On Sat, May 08, 2021 at 12:13:52 +0200, Augusto Stoffel wrote:
> >> I've attached a draft implementation of a minibuffer-controlled mode for
> >> Isearch, as described for instance in
> >> https://lists.gnu.org/archive/html/emacs-devel/2020-01/msg00447.html
> >> To enable the feature, set `isearch-from-minibuffer' to t.

> >> The basic trade-off is that this makes it easier to edit the search
> >> string, and harder to quit the search.

> > You mean, the patch makes incompatible changes to isearch, yes?

> I wouldn't dare to suggest turning this on by default :-).

> The *slight* incompatible changes are all in `isearch-edit-string'.
> This seems like a quite unpolished command anyway, but of course I can
> leave it untouched.  I suggest discussing the details in a future
> subthread.

> >> It also drops some of the eccentricities of regular Isearch.  For
> >> instance, DEL deletes and C-g quits.

> > What to you are eccentricities, are features to other people.  I'm very
> > satisfied with the current workings of C-g, and would protest vehemently
> > were they to be changed.  I'm not sure what you mean by "DEL deletes" -
> > what does it delete, and in which circumstances?

> I am sure most of the regulars in this thread share your view.
> Otherwise, something like this proposal would have surfaced much
> earlier.

I think the commands came together the way they did in early Emacs
history through people trying things out and discarding what didn't work
very well.  What we're left with is what worked/works well.

> But this doesn't mean this feature won't be appreciated by other
> people.

Neither does it mean they will be.

> In the minbuffer-controlled Isearch mode, DEL is simply
> `delete-backward-char', not some kind of undo.  It deletes in all
> circumstances.

So, you're taking away one of the core bindings of isearch.  What are you
going to put in its place?  The ease of pressing DEL to undo the last
subcommand is one of the delights of isearch, from my point of view.

[ .... ]

> > Skimming through your patch (it is too large for me to take in every
> > detail at the moment), it seems it could make isearch.el even more
> > difficult to maintain than it currently is.  For example, the two
> > defmacros `with-...' are the sort that force somebody debugging (or
> > trying to read a patch) to look somewhere else to find out what they
> > mean.  This is irksome and tedious.  There are already macros like
> > this in isearch.el, and it would seem wise to minimise any further
> > proliferation.  Each of these macros expands the ,@body twice for
> > every invocation.  Would it not be possible to rearrange things, just
> > to expand them once?

> My difficulty in understanding Isearch is the very complicated internal
> state.  This patch is just a UI thing and doesn't interact at lot with
> the search state.

There's nothing "just" about UI things.  ;-)

> As to the duplication of ,@body, sure, once can always use a lambda to
> avoid this.

That wasn't really what I meant.  I was thinking more on the lines of
perhaps assigning the results of the ,@body to a variable, and doing
different things with that variable.  Or something like that.

> However, I would rather break up a very long function that requires
> being wrap in these macros (they will be very few, anyway) into shorter
> and simpler pieces.

As would I, provided the pieces are coherent wholes, and not random
fragments.

> > In one hunk in the patch, I see a condition-case containing a
> > funcall, containing a catch, containing a minibuffer-with-setup-hook,
> > which in its turn contains a lambda function and an unwind-protect.
> > The lambda function sets functions onto both after-change-functions
> > and post-command-hook.  Whew!  Each one of these constructs on its
> > own causes difficulty in debugging, but they have their
> > justification.  Must they really all appear together in this fashion?

> The condition-case [isn't] complicated, really.  The tag in question is
> thrown at just one place (`with-isearch-window-quitting-edit') and
> caught at just one place (`isearch-edit-string').  No spaghetti logic
> here at all.  It just passes a continuation, hence the funcall.

That's complicated difficult to follow logic.  Could it not be done more
simply with functions calling functions in the standard fashion?

> I admit the after-change function and post-command hook are hairy,
> because they interact with the Isearch state.

That looks to me like a source of future bugs.

> > The biggest question, which may have been answered somewhere in the
> > thread already, is why?  What is wrong with isearch at the moment
> > that the patch will fix, or enable to be fixed?  The emacs-devel post
> > you cite above (from 2020) doesn't seem to motivate this change.
> > Could you possibly answer this question, or cite a post which answers
> > it, please?

> I used Isearch for many years and never really understood/cared to
> figure out why sometimes trying to delete a character jumped around the
> buffer.

Maybe because DEL does not delete a character, but undoes a subcommand.

> Or in which circumstances I had to press C-g once or twice to quit (I
> still don't, but in the meanwhile I found `isearch-cancel').

I think one C-g works if the current search string has been found,
otherwise you need two.

> I would also inadvertently quit the search all the time when trying to
> edit the string.  This is too complicated to my taste.

Personally, I don't see it that way at all.  I find isearch delightfully
simple and intuitive.  But I can accept you see it differently.

> Only after an interlude using Swiper I realized that Isearch isn't even
> using the minibuffer, like any other command that reads a string does.

I didn't realise this till very recently either.  But isearch isn't
reading a string - at least not primarily - it's searching through a
buffer.  And as Drew pointed out, that buffer might well be the
minibuffer itself.

> This patch is supposed to provide a simple, intuitive mode of operation
> for Isearch.

Thanks, but that doesn't really answer my question.  You're unhappy about
some of the keybindings in isearch, which is fair enough.  But how does
switching to using the minibuffer have anything to do with that?

My feeling is that if you were to make the (optional) keybinding changes
whilst using the current echo-area manipulation, the patch would be
_much_ less that 500 lines long.  Why use the minibuffer at all?

[ .... ]

> >> What do you think?

> > The patch is ~500 lines long.  This makes me worried.  I'm afraid I
> > can only react negatively and defensively at the moment.

> That's because I didn't reindent anything yet :-)

;-)

> Now seriously, the bulk of the change is to rewrite
> `isearch-edit-string'.  Apart from that, it (1) wraps a bunch of
> commands in a macro that does nothing when `isearch-from-minibuffer' is
> nil, and (2) adds a condition at the end of `isearch-mode' to possibly
> call `isearch-edit-string'.  That's about it.  Maybe with this short
> summary the patch becomes easier to digest.

Probably.  But, again, what has using the minibuffer got to do with the
keybindings a user prefers?

-- 
Alan Mackenzie (Nuremberg, Germany).