bug#63861: [PATCH] pp.el: New "pretty printing" code

[Stefan Monnier via "Bug reports for GNU Emacs, the Swiss army knife of text editors" <bug-gnu-emacs@gnu.org>]

>> I find these results (mine) quite odd: they suggest that my `pp-region`
>> is *faster* than the old `pp-buffer` for `load-history` and `bookmark-alist`
>> data, which I definitely did not expect (and don't know how to explain
>> either).

I've just redone my tests a bit differently, added `pp-emacs-lisp-code`,
and also introduced a var to control whether to activate the `lisp-ppss`
patch or not.  I also fixed my `foo.el` file: its content was
accidentally already pretty printed rather than being on a single line,
which totally changes the behavior of `pp-region` and `pp-buffer`).

For reference:

    % (cd ~/tmp; l foo.el test*.el)
    -rw------- 1 monnier monnier 1125154  8 jun 11:20 test-load-history.el
    -rw------- 1 monnier monnier  163258  8 jun 11:20 test-bookmark-alist.el
    -rw-r--r-- 1 monnier monnier   77101  8 jun 17:20 foo.el
    %

Here's the code I used to run the test:

    for f in ~/tmp/foo.el ~/tmp/test-bookmark-alist.el ~/tmp/test-load-history.el; do for ppss in nil t; do for v in '(pp-buffer)' '(pp-region (point-min) (point-max))' '(tv/pp-region (point-min) (point-max))' '(let ((s (read (current-buffer)))) (erase-buffer) (pp-emacs-lisp-code s))'; do src/emacs -Q --batch -l ~/tmp/describe-variable --eval "(with-temp-buffer (emacs-lisp-mode) (insert-file-contents \"$f\") (setq pp-buffer-use-pp-region nil lisp--faster-ppss $ppss) (message \"%S %S %S %S\" (file-name-nondirectory \"$f\") (benchmark-run $v) '$v '$ppss))"&;  done; done; done

So, by file, from fastest to slowest:

    foo.el (0.859482743 0 0.0) (pp-buffer) t
    foo.el (0.890402623 0 0.0) (pp-buffer) nil
    foo.el (4.62344853 4 1.7225397670000002) (tv/pp-region (point-min) (point-max)) t
    foo.el (4.687414465 4 1.7116580980000002) (tv/pp-region (point-min) (point-max)) nil
    foo.el (7.932661181 1 0.3435169600000001) (pp-region (point-min) (point-max)) t
    foo.el (196.183345212 1 0.618591124) (pp-region (point-min) (point-max)) nil
    foo.el (2997.739238575 505 105.82851685700001) (let ((s (read (current-buffer)))) (erase-buffer) (pp-emacs-lisp-code s)) t

Here we see that my `pp-region` code is slower than `pp-buffer` by
a factor ~10x: I'm not very happy about it, but this `foo.el` file was
selected because it was the worst case I had come across (tho that was
before I found the `lisp-ppss` patch).

The last element in each line is whether we activated the `lisp-ppss`
patch.  As we can see here, the `lisp-ppss` patch makes an enormous
difference (~24x) for my code, but not for `pp-buffer` (because it
relies on `lisp-indent-region` rather than `lisp-indent-line`) and not
for `tv/pp-region` either (because it doesn't indent at all).

We also see that `pp-emacs-lisp-code` is *much* slower.  I don't include
other results for this function in this email because they're still
running :-)

    test-bookmark-alist.el (13.237991019999999 6 2.403892035) (tv/pp-region (point-min) (point-max)) nil
    test-bookmark-alist.el (14.853056353 6 2.705511935) (tv/pp-region (point-min) (point-max)) t
    test-bookmark-alist.el (28.059658418 5 2.005039257) (pp-region (point-min) (point-max)) t
    test-bookmark-alist.el (180.390204026 5 2.1182066760000002) (pp-region (point-min) (point-max)) nil
    test-bookmark-alist.el (265.95429676599997 10 4.445954908) (pp-buffer) t
    test-bookmark-alist.el (268.975666886 10 3.6774180120000004) (pp-buffer) nil

Here we see that my `pp-region` code can be faster (even significantly
so) than `pp-buffer`.  I'm not sure why, but I'll take the win :-)
We also see that the faster `lisp-ppss` again makes an important
difference in the performance of `pp-region` (~8x), tho the effect is
not as drastic as in the case of `foo.el`.

    test-load-history.el (235.134082197 8 4.440112806999999) (tv/pp-region (point-min) (point-max)) nil
    test-load-history.el (235.873981253 8 4.416064476) (tv/pp-region (point-min) (point-max)) t
    test-load-history.el (506.770548196 31 9.706665932) (pp-region (point-min) (point-max)) t
    test-load-history.el (701.991875274 43 12.390212449) (pp-buffer) t
    test-load-history.el (710.843618646 43 12.156289354) (pp-buffer) nil
    test-load-history.el (1419.268184021 36 9.260999640000001) (pp-region (point-min) (point-max)) nil

Here again, we see that `pp-region` is competitive with `pp-buffer` and
the `lisp-ppss` patch speeds it up significantly (~3x).

Another thing we see in those tests is that `pp-region` (with the
`lisp-ppss` patch) is ~2x slower than `tv/pp-region`, whereas the
performance differential with `pp-buffer` varies a lot more.  Also if we
compare the time it takes to the size of the file, we see:

      77101B /   7.932661181s = 9719 B/s
     163258B /  28.059658418s = 5818 B/s
    1125154B / 506.770548196s = 2220 B/s

`pp-region`s performance is not quite linear in the size of the file :-(
Interestingly, the same holds for `tv/pp-region`:

      77101B /   4.62344853s  = 16676 B/s
     163258B /  13.237991019s = 12332 B/s
    1125154B / 235.134082197s =  4785 B/s

even though it works in a fundamentally very different way (which, to
my naive eye should result in a linear performance), so maybe the
slowdown here is due to something external (such as the fact that
various operations on buffers get slower as the buffer gets bigger).

> hmm, don't know, I ran pp-buffer with M-: from the test-load-history.el or the
> test-bookmark-alist.el buffer. May be using emacs --batch makes a
> difference?

I don't see any significant performance difference between batch and
non-batch :-(

> is the data really printed in such case?

Yes, definitely.

> More or less the code using pp-region takes between 42 to 48s and the one
> with old pp-buffer around 6s.

I wonder why we see such wildly different performance.  In my tests on
your `test-bookmark-alist.el` I basically see the reverse ratio!

> Also sorry about your last patch I tested it too fast, it is indeed
> slightly faster, but not much: 42 vs 46s.

This is also perplexing.  In my tests, the patch has a very significant
impact on the performance of `pp-region`.
Are you sure the patch is used (`lisp-mode.el` is preloaded, so you need
to re-dump Emacs, or otherwise manually force-reload `lisp-mode.elc`
into your Emacs session)?

FWIW, I'm running my tests on Emacs's `master` branch with the native
ELisp compiler enabled (tho I don't see much difference in performance
on these tests when running my other Emacs build without the native
compiler) on an i386 Debian testing system.

>> And do I understand correctly that `tv/pp-region` does not indent its
>> output?
> No, it does indent, see function tv/pp which use pp-to-string which use pp-buffer
> and pp-buffer indent the whole sexp at the end.

AFAICT `tv/pp` uses `pp-to-string` only on "atomic" values (i.e. not
lists, vectors, or hash-tables), so there's usually not much to indent in there.
What I see in the output after calling `tv/pp-region` are non-indented lists.

>> What was the reason for this choice?
> Because indentation is done at the end by pp-buffer.

When I use `describe-variable` with your code, the printed value is
indeed indented, but that code uses only `pp-buffer` and not
`tv/pp-region` (i.e. `tv/describe-variable` does not call
`tv/pp-region`, neither directly nor indirectly).

> PS (unrelated to pp-region): About the old pp-buffer, there is
> a difficult to find bug where the same operation is running twice
> (newline is added, then removed, then added again and then the loop
> continue)
>
> You can see it by edebugging pp-buffer on a simple alist like this:
>
> (defvar bar '((1 . "un") (2 . "deux") (3 . "trois") (4 . "quatre") (5 . "cinq") (6 . "six")))

That might be part of the poor performance we see on
`test-bookmark-alist.el`?


        Stefan