bug#43379: [PATCH] Double-click events can occur without preceding single-click events

[Daniel Koning <dk@danielkoning.com>]

[-- Attachment #1: Type: text/plain, Size: 4163 bytes --]

According to the documentation (21.7.7 of the Lisp ref, "Repeat Events"), a
double-click binding should "assume that the single-click command has already
run," since Emacs only produces a double-click event after an equivalent
single-click event. But (so long as 'double-click-time' is set generously
enough) it's easy to trigger a double-click event without a corresponding
single-click event, because the event handling primitives aren't quite
particular enough about what they count as a repeat.

For one thing, the code ignores whether the modifier keys of one input action
differ from the next. So the sequence "M-mouse-1 mouse-1" produces <M-mouse-1>
<double-mouse-1>. This is no good: <M-mouse-1> is likely bound to something
different from <mouse-1>, so the <double-mouse-1> command will run under false
premises.

It also uses just one repeat click counter, despite the fact that down/up click
event pairs can be interleaved. The counter increments when a button-down
matches the last button-up, and it carries over to whatever button-up comes in
next. Therefore, the mouse sequence

    down on mouse-1
    down on mouse-3
    up on mouse-1
    down on mouse-1
    up on mouse-3

produces a <double-mouse-3> event with no preceding <mouse-3>.

The function 'make_lispy_event', which translates raw events to Lisp objects, is
responsible for tagging a mouse event (button-down, click, drag, or wheel) as a
double. This patch against master revises that function to avoid creating "bare"
double mouse events. It does so by consolidating the scattered repeat-tracking
logic into its own function, where more relevant state can be maintained without
any extra global cruft. (With less cruft, in fact: 'button_down_time' is no
longer punned as an interrupt flag, wheel event codes need not be coerced to
negative to stuff them into the same variable as clicks, etc.)

Just as before, a repeat mouse event has to use the same button, happen in
roughly the same place, happen at roughly the same time, and so on. Aside from
the specific deliberate narrowing I've described, I believe this code duplicates
the logic of the old code exactly. That includes nuances that strike me as
strange. For instance,

    down on mouse-1
    up on mouse-1
    down on mouse-1
    keystroke
    up on mouse-1
    down on mouse-1

finishes with <double-mouse-1> <down-mouse-1>, even though it would make more
sense to me if the keystroke reset the count right away. But that's how it was
before.

There is still at least one fringy situation in which a <double-mouse-1> command
could run without the <mouse-1> command running first: when the first and second
click land in different windows. I would have liked to add a test for this, but
I don't fully understand the type constraint on the 'frame_or_window' member of
struct input_event (specifically when it's a mouse event). 'make_lispy_event'
seems to assume it's a frame: it does an XFRAME without a check. But then the
comparison against 'double-click-fuzz', which I folded into the new function,
tries both possibilities. I left the more cautious code in place, and I haven't
added any more logic related to windows.

This also has no effect on xterm-mouse-mode, which doesn't interact with the C
event translator at all. It just takes key sequences which have already passed
through 'make_lispy_event' and translates them to mouse-event lists.
xterm-mouse-mode has some major inconsistencies with the GUI repeat-handling
behavior, but it doesn't seem to exhibit the bug this report describes.

This patch rewords the docs slightly to say explicitly that modifier keys make a
difference. It also makes a small tweak to the criteria for drag events, and
another tweak to the Cocoa-specific trackpad code to fix a conspicuous bug
caused by a similar oversight. Namely, newly pressed modifier keys can attach to
invisible so-called "momentum" events generated by a swipe on the trackpad that
has already ended. One result is that if a Cocoa user scrolls up to the top of
the buffer and immediately presses (say) C-e, Emacs gets barraged with
<C-wheel-up> events and scales the buffer text up to a ridiculous size.

Daniel

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: repeat-events.patch --]
[-- Type: text/x-patch, Size: 25998 bytes --]

From 8acd0adbaffc20cda484031723a625b361472293 Mon Sep 17 00:00:00 2001
From: Daniel Koning <dk@danielkoning.com>
Date: Tue, 1 Sep 2020 18:56:12 -0500
Subject: [PATCH 1/6] Detect repeat mouse events more stringently

If two closely spaced mouse events come in with different modifier keys,
never treat the second as a double.  Also, be stricter about the
sequencing of overlapping clicks, where one button goes down before the
other comes up.

* src/keyboard.c (mouse_repeat_count): New function.
(interrupt_next_repeat_click): New global variable.
(make_lispy_event): Delegate to 'mouse_repeat_count'.
* doc/emacs/custom.texi (Mouse Buttons): Clarify the behavior for users.
* doc/lispref/commands.texi (Repeat Events): Clarify the behavior for
Lisp programmers.
* etc/NEWS: Describe the change.
---
 doc/emacs/custom.texi     |   8 +-
 doc/lispref/commands.texi |   9 +-
 etc/NEWS                  |  10 ++
 src/keyboard.c            | 283 ++++++++++++++++++++++----------------
 4 files changed, 186 insertions(+), 124 deletions(-)

diff --git a/doc/emacs/custom.texi b/doc/emacs/custom.texi
index a512fd14c8..a3ffd90c5b 100644
--- a/doc/emacs/custom.texi
+++ b/doc/emacs/custom.texi
@@ -2145,8 +2145,12 @@ Mouse Buttons
 
   The symbols for mouse events also indicate the status of the modifier
 keys, with the usual prefixes @samp{C-}, @samp{M-}, @samp{H-},
-@samp{s-}, @samp{A-}, and @samp{S-}.  These always precede @samp{double-}
-or @samp{triple-}, which always precede @samp{drag-} or @samp{down-}.
+@samp{s-}, @samp{A-}, and @samp{S-}.  These always precede
+@samp{double-} or @samp{triple-}, which always precede @samp{drag-} or
+@samp{down-}.  (Two clicks with different modifier keys can never
+produce a double event, no matter how close together the clicks are.
+Otherwise, Emacs could get a @code{double-mouse-1} event without getting
+a @code{mouse-1} event first.)
 
   A frame includes areas that don't show text from the buffer, such as
 the mode line and the scroll bar.  You can tell whether a mouse button
diff --git a/doc/lispref/commands.texi b/doc/lispref/commands.texi
index 25f657404f..93350927f2 100644
--- a/doc/lispref/commands.texi
+++ b/doc/lispref/commands.texi
@@ -1548,9 +1548,12 @@ Repeat Events
 @cindex triple-click events
 @cindex mouse events, repeated
 
-If you press the same mouse button more than once in quick succession
-without moving the mouse, Emacs generates special @dfn{repeat} mouse
-events for the second and subsequent presses.
+Emacs generates special @dfn{repeat} mouse events to represent actions
+like double and triple clicks.  If you press a button twice in quick
+succession without moving the mouse in between, Emacs will designate the
+second event as a repeat.  (However, if you hold down any modifier keys
+during one press and not the other, Emacs will treat the two events as
+unconnected, and the second event will not be a repeat.)
 
 The most common repeat events are @dfn{double-click} events.  Emacs
 generates a double-click event when you click a button twice; the event
diff --git a/etc/NEWS b/etc/NEWS
index ddc2fb9d60..5a45295f33 100644
--- a/etc/NEWS
+++ b/etc/NEWS
@@ -127,6 +127,16 @@ the mouse cursor is on the scroll bars, fringes, margins, header line,
 and mode line.  ('mwheel-mode' is enabled by default on most graphical
 displays.)
 
++++
+** Repeat events are now produced only when the modifier keys are the same.
+Before, when the user pressed the same mouse button repeatedly within
+the bounds specified by 'double-click-fuzz' and 'double-click-time',
+it always produced a 'double-' or 'triple-' event, even if the user
+was holding down modifier keys on one click and not another.  This
+meant that it was possible for Emacs to read a double-click event
+without reading the same kind of single-click event first.  Emacs now
+looks at modifier keys to determine if a mouse event is a repeat.
+
 \f
 * Editing Changes in Emacs 28.1
 
diff --git a/src/keyboard.c b/src/keyboard.c
index 590d183c4c..1214e62820 100644
--- a/src/keyboard.c
+++ b/src/keyboard.c
@@ -347,6 +347,7 @@ #define READABLE_EVENTS_IGNORE_SQUEEZABLES	(1 << 2)
 static Lisp_Object read_char_x_menu_prompt (Lisp_Object,
                                             Lisp_Object, bool *);
 static Lisp_Object read_char_minibuf_menu_prompt (int, Lisp_Object);
+static intmax_t mouse_repeat_count (const struct input_event *);
 static Lisp_Object make_lispy_event (struct input_event *);
 static Lisp_Object make_lispy_movement (struct frame *, Lisp_Object,
                                         enum scroll_bar_part,
@@ -5024,18 +5025,6 @@ #define ISO_FUNCTION_KEY_OFFSET 0xfe00
 
 static Lisp_Object button_down_location;
 
-/* Information about the most recent up-going button event:  Which
-   button, what location, and what time.  */
-
-static int last_mouse_button;
-static int last_mouse_x;
-static int last_mouse_y;
-static Time button_down_time;
-
-/* The number of clicks in this multiple-click.  */
-
-static int double_click_count;
-
 /* X and Y are frame-relative coordinates for a click or wheel event.
    Return a Lisp-style event list.  */
 
@@ -5306,6 +5295,139 @@ make_scroll_bar_position (struct input_event *ev, Lisp_Object type)
 		builtin_lisp_symbol (scroll_bar_parts[ev->part]));
 }
 
+/* Whether the next click (or wheel event) should be treated as a
+   single no matter what, even if a matching prior event would make it
+   a repeat.  This gets flipped on when a keystroke or drag event
+   comes in: double clicks can't extend across one of those.  */
+
+static bool interrupt_next_repeat_click;
+
+/* If EVENT is a repeat, return the number of consecutive mouse events
+   so far.  Return 1 otherwise.  In the background, keep a record of
+   where we are in the current sequence of repeats.
+
+   This should only receive a pointer to an event of an applicable
+   kind -- namely, a button-down, button-up, or wheel event.  */
+
+static intmax_t
+mouse_repeat_count (const struct input_event *event)
+{
+  /* Keep persistent information about any chain of repeat events we
+     might be in the middle of.  */
+
+  static ENUM_BF (event_kind) kind;
+  static unsigned code;
+  static unsigned keys;
+  static unsigned wheel_dir;
+
+  static Time timeout_start;
+  static EMACS_INT x, y;
+
+  static intmax_t count;
+
+  /* Analyze the new event that came in.  */
+
+  unsigned new_keys = event->modifiers & CHAR_MODIFIER_MASK;
+  unsigned new_wheel_dir;
+
+  Time interval;
+  EMACS_INT new_x, new_y, offset;
+
+  bool is_wheel = (event->kind == WHEEL_EVENT
+                   || event->kind == HORIZ_WHEEL_EVENT);
+  bool is_button = !is_wheel;
+  bool is_button_up = (is_button && (event->modifiers & up_modifier));
+
+  if (is_wheel)
+    new_wheel_dir = (event->modifiers & (down_modifier | up_modifier));
+  else
+    new_wheel_dir = 0;
+
+  /* We need to update timestamp and position after both repeat and
+     non-repeat events.  Make the relevant comparisons in advance,
+     then do the update immediately.  */
+
+  interval = event->timestamp - timeout_start;
+  if (!is_button_up)
+    timeout_start = event->timestamp;
+
+  new_x = XFIXNUM (event->x);
+  new_y = XFIXNUM (event->y);
+  offset = max (eabs (new_x - x), eabs (new_y - y));
+  x = new_x;
+  y = new_y;
+
+  /* Is this a repeat?  Start checking for conditions that imply
+     otherwise.  */
+
+  if (interrupt_next_repeat_click && !is_button_up)
+    {
+      interrupt_next_repeat_click = false;
+      goto not_a_repeat;
+    }
+
+  if (event->kind != kind || event->code != code || new_keys != keys
+      || new_wheel_dir != wheel_dir)
+    goto not_a_repeat;
+
+  if (is_button_up)
+    /* That's it for button-up events.  At this point, we declare that
+       it's a repeat that should inherit the count of the preceding
+       button-down.  */
+    return count;
+
+  if (FIXNATP (Vdouble_click_time))
+    {
+      if (interval >= XFIXNAT (Vdouble_click_time))
+        goto not_a_repeat;
+    }
+  else if (!EQ (Vdouble_click_time, Qt))
+    goto not_a_repeat;
+
+  {
+    /* On window-system frames, use the value of 'double-click-fuzz'
+       as is.  On other frames, interpret it as a multiple of 1/8
+       characters.  */
+    struct frame *f;
+    intmax_t fuzz;
+
+    if (WINDOWP (event->frame_or_window))
+      f = XFRAME (XWINDOW (event->frame_or_window)->frame);
+    else if (FRAMEP (event->frame_or_window))
+      f = XFRAME (event->frame_or_window);
+    else
+      emacs_abort ();
+
+    if (FRAME_WINDOW_P (f))
+      fuzz = double_click_fuzz;
+    else
+      fuzz = double_click_fuzz / 8;
+
+    if (offset > fuzz)
+      goto not_a_repeat;
+  }
+
+  /* We've ruled out everything that could disqualify it as a repeat,
+     so treat it as one.  */
+  count++;
+  return count;
+
+ not_a_repeat:
+  /* Base a new repeat chain off of this event.  */
+  count = 1;
+  kind = event->kind;
+  code = event->code;
+  keys = new_keys;
+  wheel_dir = new_wheel_dir;
+
+  if (is_button_up)
+    /* A button-up event should cut off a chain when it doesn't match
+       the last event, but it shouldn't start its own chain.  */
+    interrupt_next_repeat_click = true;
+
+  return count;
+}
+
 /* Given a struct input_event, build the lisp event which represents
    it.  If EVENT is 0, build a mouse movement event from the mouse
    movement buffer, which should have a movement event in it.
@@ -5438,7 +5560,7 @@ make_lispy_event (struct input_event *event)
 	if ((event->code) == 040
 	    && event->modifiers & shift_modifier)
 	  c |= shift_modifier;
-	button_down_time = 0;
+	interrupt_next_repeat_click = true;
 	XSETFASTINT (lispy_c, c);
 	return lispy_c;
       }
@@ -5457,7 +5579,7 @@ make_lispy_event (struct input_event *event)
       /* A function key.  The symbol may need to have modifier prefixes
 	 tacked onto it.  */
     case NON_ASCII_KEYSTROKE_EVENT:
-      button_down_time = 0;
+      interrupt_next_repeat_click = true;
 
       for (i = 0; i < ARRAYELTS (lispy_accent_codes); i++)
 	if (event->code == lispy_accent_codes[i])
@@ -5546,10 +5668,10 @@ make_lispy_event (struct input_event *event)
 #endif
       {
 	int button = event->code;
-	bool is_double;
 	Lisp_Object position;
 	Lisp_Object *start_pos_ptr;
 	Lisp_Object start_pos;
+	int repeat_count;
 
 	position = Qnil;
 
@@ -5637,57 +5759,20 @@ make_lispy_event (struct input_event *event)
 	    mouse_syms = larger_vector (mouse_syms, incr, -1);
 	  }
 
+	repeat_count = mouse_repeat_count (event);
+	if (repeat_count == 2)
+	  event->modifiers |= double_modifier;
+	else if (repeat_count >= 3)
+	  event->modifiers |= triple_modifier;
+
 	start_pos_ptr = aref_addr (button_down_location, button);
 	start_pos = *start_pos_ptr;
 	*start_pos_ptr = Qnil;
 
-	{
-	  /* On window-system frames, use the value of
-	     double-click-fuzz as is.  On other frames, interpret it
-	     as a multiple of 1/8 characters.  */
-	  struct frame *f;
-	  intmax_t fuzz;
-
-	  if (WINDOWP (event->frame_or_window))
-	    f = XFRAME (XWINDOW (event->frame_or_window)->frame);
-	  else if (FRAMEP (event->frame_or_window))
-	    f = XFRAME (event->frame_or_window);
-	  else
-	    emacs_abort ();
-
-	  if (FRAME_WINDOW_P (f))
-	    fuzz = double_click_fuzz;
-	  else
-	    fuzz = double_click_fuzz / 8;
-
-	  is_double = (button == last_mouse_button
-		       && (eabs (XFIXNUM (event->x) - last_mouse_x) <= fuzz)
-		       && (eabs (XFIXNUM (event->y) - last_mouse_y) <= fuzz)
-		       && button_down_time != 0
-		       && (EQ (Vdouble_click_time, Qt)
-			   || (FIXNATP (Vdouble_click_time)
-			       && (event->timestamp - button_down_time
-				   < XFIXNAT (Vdouble_click_time)))));
-	}
-
-	last_mouse_button = button;
-	last_mouse_x = XFIXNUM (event->x);
-	last_mouse_y = XFIXNUM (event->y);
-
 	/* If this is a button press, squirrel away the location, so
            we can decide later whether it was a click or a drag.  */
 	if (event->modifiers & down_modifier)
 	  {
-	    if (is_double)
-	      {
-		double_click_count++;
-		event->modifiers |= ((double_click_count > 2)
-				     ? triple_modifier
-				     : double_modifier);
-	      }
-	    else
-	      double_click_count = 1;
-	    button_down_time = event->timestamp;
 	    *start_pos_ptr = Fcopy_alist (position);
 	    ignore_mouse_drag_p = false;
 	  }
@@ -5744,19 +5829,13 @@ make_lispy_event (struct input_event *event)
 		       && EQ (Fcar (Fcdr (start_pos)), Fcar (Fcdr (position)))))
 		  {
 		    /* Mouse has moved enough.  */
-		    button_down_time = 0;
+		    interrupt_next_repeat_click = true;
 		    click_or_drag_modifier = drag_modifier;
 		  }
 	      }
 
-	    /* Don't check is_double; treat this as multiple if the
-	       down-event was multiple.  */
-	    event->modifiers
-	      = ((event->modifiers & ~up_modifier)
-		 | click_or_drag_modifier
-		 | (double_click_count < 2 ? 0
-		    : double_click_count == 2 ? double_modifier
-		    : triple_modifier));
+	    event->modifiers = ((event->modifiers & ~up_modifier)
+				| click_or_drag_modifier);
 	  }
 	else
 	  /* Every mouse event should either have the down_modifier or
@@ -5776,7 +5855,7 @@ make_lispy_event (struct input_event *event)
 	  if (event->modifiers & drag_modifier)
 	    return list3 (head, start_pos, position);
 	  else if (event->modifiers & (double_modifier | triple_modifier))
-	    return list3 (head, position, make_fixnum (double_click_count));
+	    return list3 (head, position, make_fixnum (repeat_count));
 	  else
 	    return list2 (head, position);
 	}
@@ -5787,6 +5866,7 @@ make_lispy_event (struct input_event *event)
       {
 	Lisp_Object position;
 	Lisp_Object head;
+	int repeat_count;
 
 	/* Build the position as appropriate for this mouse click.  */
 	struct frame *f = XFRAME (event->frame_or_window);
@@ -5799,38 +5879,15 @@ make_lispy_event (struct input_event *event)
 	position = make_lispy_position (f, event->x, event->y,
 					event->timestamp);
 
-	/* Set double or triple modifiers to indicate the wheel speed.  */
 	{
-	  /* On window-system frames, use the value of
-	     double-click-fuzz as is.  On other frames, interpret it
-	     as a multiple of 1/8 characters.  */
-	  struct frame *fr;
-	  intmax_t fuzz;
 	  int symbol_num;
-	  bool is_double;
-
-	  if (WINDOWP (event->frame_or_window))
-	    fr = XFRAME (XWINDOW (event->frame_or_window)->frame);
-	  else if (FRAMEP (event->frame_or_window))
-	    fr = XFRAME (event->frame_or_window);
-	  else
-	    emacs_abort ();
-
-	  fuzz = FRAME_WINDOW_P (fr)
-	    ? double_click_fuzz : double_click_fuzz / 8;
 
 	  if (event->modifiers & up_modifier)
-	    {
 	      /* Emit a wheel-up event.  */
-	      event->modifiers &= ~up_modifier;
-	      symbol_num = 0;
-	    }
+	    symbol_num = 0;
 	  else if (event->modifiers & down_modifier)
-	    {
 	      /* Emit a wheel-down event.  */
-	      event->modifiers &= ~down_modifier;
-	      symbol_num = 1;
-	    }
+	    symbol_num = 1;
 	  else
 	    /* Every wheel event should either have the down_modifier or
 	       the up_modifier set.  */
@@ -5839,32 +5896,20 @@ make_lispy_event (struct input_event *event)
           if (event->kind == HORIZ_WHEEL_EVENT)
             symbol_num += 2;
 
-	  is_double = (last_mouse_button == - (1 + symbol_num)
-		       && (eabs (XFIXNUM (event->x) - last_mouse_x) <= fuzz)
-		       && (eabs (XFIXNUM (event->y) - last_mouse_y) <= fuzz)
-		       && button_down_time != 0
-		       && (EQ (Vdouble_click_time, Qt)
-			   || (FIXNATP (Vdouble_click_time)
-			       && (event->timestamp - button_down_time
-				   < XFIXNAT (Vdouble_click_time)))));
-	  if (is_double)
-	    {
-	      double_click_count++;
-	      event->modifiers |= ((double_click_count > 2)
-				   ? triple_modifier
-				   : double_modifier);
-	    }
+	  /* Set double or triple modifiers to indicate the wheel
+	     speed.  */
+	  repeat_count = mouse_repeat_count (event);
+	  if (repeat_count == 2)
+	    event->modifiers |= double_modifier;
+	  else if (repeat_count >= 3)
+	    event->modifiers |= triple_modifier;
 	  else
-	    {
-	      double_click_count = 1;
-	      event->modifiers |= click_modifier;
-	    }
+	    /* Non-repeat wheel events are tagged as clicks.  */
+	    event->modifiers |= click_modifier;
 
-	  button_down_time = event->timestamp;
-	  /* Use a negative value to distinguish wheel from mouse button.  */
-	  last_mouse_button = - (1 + symbol_num);
-	  last_mouse_x = XFIXNUM (event->x);
-	  last_mouse_y = XFIXNUM (event->y);
+	  /* The Lisp side expects to see direction information in
+	     'symbol_num', but not in the modifier bits.  */
+	  event->modifiers &= ~(down_modifier | up_modifier);
 
 	  /* Get the symbol we should use for the wheel event.  */
 	  head = modify_event_symbol (symbol_num,
@@ -5877,10 +5922,10 @@ make_lispy_event (struct input_event *event)
 	}
 
         if (NUMBERP (event->arg))
-          return list4 (head, position, make_fixnum (double_click_count),
+          return list4 (head, position, make_fixnum (repeat_count),
                         event->arg);
 	else if (event->modifiers & (double_modifier | triple_modifier))
-	  return list3 (head, position, make_fixnum (double_click_count));
+	  return list3 (head, position, make_fixnum (repeat_count));
 	else
 	  return list2 (head, position);
       }
-- 
2.20.1


From 3e484dbae4bac122ec7093688906484e13a81156 Mon Sep 17 00:00:00 2001
From: Daniel Koning <dk@danielkoning.com>
Date: Thu, 3 Sep 2020 16:15:16 -0500
Subject: [PATCH 2/6] Clarify a quirk of how 'double-click-fuzz' works

* doc/lispref/commands.texi (Repeat Events): Mention that if the mouse
changes screen positions with a button held down, it's always a drag.
---
 doc/lispref/commands.texi | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/doc/lispref/commands.texi b/doc/lispref/commands.texi
index 93350927f2..0adc6fc17d 100644
--- a/doc/lispref/commands.texi
+++ b/doc/lispref/commands.texi
@@ -1627,7 +1627,9 @@ Repeat Events
 clicks to make a double-click.
 
 This variable is also the threshold for motion of the mouse to count
-as a drag.
+as a drag.  (But if the mouse moves from one screen position to
+another while the button is held down, it always counts as a drag, no
+matter the value of @code{double-click-fuzz}.)
 @end defopt
 
 @defopt double-click-time
-- 
2.20.1


From f8f6fcf3cd329ded23586838249fd372ece969d6 Mon Sep 17 00:00:00 2001
From: Daniel Koning <dk@danielkoning.com>
Date: Thu, 3 Sep 2020 16:26:28 -0500
Subject: [PATCH 3/6] Use event macros with meaningful names

* src/keyboard.c (make_lispy_event): Use 'POSN_POSN' and
'POSN_WINDOW_POSN' to extract fields from a mouse event list; don't just
cadr and caddr things out.
---
 doc/lispref/commands.texi |  2 +-
 src/keyboard.c            | 11 ++++-------
 2 files changed, 5 insertions(+), 8 deletions(-)

diff --git a/doc/lispref/commands.texi b/doc/lispref/commands.texi
index 0adc6fc17d..c28e87619e 100644
--- a/doc/lispref/commands.texi
+++ b/doc/lispref/commands.texi
@@ -1627,7 +1627,7 @@ Repeat Events
 clicks to make a double-click.
 
 This variable is also the threshold for motion of the mouse to count
-as a drag.  (But if the mouse moves from one screen position to
+as a drag.  (But if the mouse moves from one buffer position to
 another while the button is held down, it always counts as a drag, no
 matter the value of @code{double-click-fuzz}.)
 @end defopt
diff --git a/src/keyboard.c b/src/keyboard.c
index 1214e62820..66d9b816a5 100644
--- a/src/keyboard.c
+++ b/src/keyboard.c
@@ -5796,13 +5796,10 @@ make_lispy_event (struct input_event *event)
 	      ignore_mouse_drag_p = false;
 	    else
 	      {
-		Lisp_Object new_down, down;
-		intmax_t xdiff = double_click_fuzz, ydiff = double_click_fuzz;
+		Lisp_Object down = POSN_WINDOW_POSN (start_pos);
+		Lisp_Object new_down = POSN_WINDOW_POSN (position);
 
-		/* The third element of every position
-		   should be the (x,y) pair.  */
-		down = Fcar (Fcdr (Fcdr (start_pos)));
-		new_down = Fcar (Fcdr (Fcdr (position)));
+		intmax_t xdiff = double_click_fuzz, ydiff = double_click_fuzz;
 
 		if (CONSP (down)
 		    && FIXNUMP (XCAR (down)) && FIXNUMP (XCDR (down)))
@@ -5826,7 +5823,7 @@ make_lispy_event (struct input_event *event)
 			  a click.  But mouse-drag-region completely ignores
 			  this case and it hasn't caused any real problem, so
 			  it's probably OK to ignore it as well.  */
-		       && EQ (Fcar (Fcdr (start_pos)), Fcar (Fcdr (position)))))
+		       && EQ (POSN_POSN (start_pos), POSN_POSN (position))))
 		  {
 		    /* Mouse has moved enough.  */
 		    interrupt_next_repeat_click = true;
-- 
2.20.1


From 76b195d2402b097d6c04bc809c95ec947f9f73c4 Mon Sep 17 00:00:00 2001
From: Daniel Koning <dk@danielkoning.com>
Date: Thu, 3 Sep 2020 16:55:45 -0500
Subject: [PATCH 4/6] Take windows into account when detecting drags

* src/keyboard.c (make_lispy_event): Whenever a button-up event lands in
a different window than its corresponding button-down event, always
treat it as a drag, not a click.  This worked most of the time before,
but not always: if you happened to land in the same position in the
other window's buffer, it wasn't detected.
* doc/lispref/commands.texi (Repeat Events): Be explicit about this.
---
 doc/lispref/commands.texi | 7 ++++---
 src/keyboard.c            | 2 ++
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/doc/lispref/commands.texi b/doc/lispref/commands.texi
index c28e87619e..e0137b976d 100644
--- a/doc/lispref/commands.texi
+++ b/doc/lispref/commands.texi
@@ -1627,9 +1627,10 @@ Repeat Events
 clicks to make a double-click.
 
 This variable is also the threshold for motion of the mouse to count
-as a drag.  (But if the mouse moves from one buffer position to
-another while the button is held down, it always counts as a drag, no
-matter the value of @code{double-click-fuzz}.)
+as a drag.  (But if the mouse moves from one window to another while
+the button is held down, or from one screen position to another, it
+always counts as a drag, no matter the value of
+@code{double-click-fuzz}.)
 @end defopt
 
 @defopt double-click-time
diff --git a/src/keyboard.c b/src/keyboard.c
index 66d9b816a5..cc38cd5ce0 100644
--- a/src/keyboard.c
+++ b/src/keyboard.c
@@ -5813,6 +5813,8 @@ make_lispy_event (struct input_event *event)
 		       && xdiff < double_click_fuzz
 		       && - double_click_fuzz < ydiff
 		       && ydiff < double_click_fuzz
+		       /* If we jumped windows, it has to be a drag.  */
+		       && EQ (POSN_WINDOW (start_pos), POSN_WINDOW (position))
 		       /* Maybe the mouse has moved a lot, caused scrolling, and
 			  eventually ended up at the same screen position (but
 			  not buffer position) in which case it is a drag, not
-- 
2.20.1


From e760b108efc51f17b23298d3271697bd278e25ee Mon Sep 17 00:00:00 2001
From: Daniel Koning <dk@danielkoning.com>
Date: Fri, 21 Aug 2020 19:39:25 -0500
Subject: [PATCH 5/6] Ignore changes in modifier keys during momentum

* src/nsterm.m ([EmacsView mouseDown:]): Treat a momentum-phase mouse
wheel event as having the same modifier keys as the non-momentum event
that triggered it.
---
 src/nsterm.m | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/src/nsterm.m b/src/nsterm.m
index 26059ab67c..9dd5249b97 100644
--- a/src/nsterm.m
+++ b/src/nsterm.m
@@ -313,6 +313,7 @@ - (NSColor *)colorUsingDefaultColorSpace
 static NSAutoreleasePool *outerpool;
 static struct input_event *emacs_event = NULL;
 static struct input_event *q_event_ptr = NULL;
+static int last_real_wheel_modifier_keys = 0;
 static int n_emacs_events_pending = 0;
 static NSMutableArray *ns_pending_files, *ns_pending_service_names,
   *ns_pending_service_args;
@@ -6984,6 +6985,21 @@ - (void)mouseDown: (NSEvent *)theEvent
           emacs_event->code = 0;
           emacs_event->modifiers = EV_MODIFIERS (theEvent) |
             (scrollUp ? up_modifier : down_modifier);
+
+          /* If this is a momentum-phase event, ignore the modifier
+           * keys it arrived with.  Inherit the modifier keys from the
+           * last non-momentum event in the sequence.  The user may be
+           * pressing or releasing modifier keys, intending to use
+           * them in the next event, while the wheel events are still
+           * firing.  */
+          if ([theEvent momentumPhase] != NSEventPhaseNone)
+            {
+              emacs_event->modifiers &= ~CHAR_MODIFIER_MASK;
+              emacs_event->modifiers |= last_real_wheel_modifier_keys;
+            }
+          else
+              last_real_wheel_modifier_keys =
+                emacs_event->modifiers & CHAR_MODIFIER_MASK;
 #if MAC_OS_X_VERSION_MIN_REQUIRED < 1070
         }
       else
-- 
2.20.1


From a66ff506f3c7dcd05a95863d724b207f02ed6329 Mon Sep 17 00:00:00 2001
From: Daniel Koning <dk@danielkoning.com>
Date: Fri, 21 Aug 2020 19:11:36 -0500
Subject: [PATCH 6/6] Give context to a FIXME comment

* src/nsterm.m ([EmacsView mouseDown:]): Provide more detail to explain
why it's difficult to adjust scrolling behavior based on wheel momentum
phase.
---
 src/nsterm.m | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/src/nsterm.m b/src/nsterm.m
index 9dd5249b97..2ddb378d75 100644
--- a/src/nsterm.m
+++ b/src/nsterm.m
@@ -6899,8 +6899,13 @@ - (void)mouseDown: (NSEvent *)theEvent
           int lines = 0;
           int scrollUp = NO;
 
-          /* FIXME: At the top or bottom of the buffer we should
-           * ignore momentum-phase events.  */
+          /* FIXME: At the top or bottom of the buffer, we should
+           * ignore momentum-phase events that are bound to scrolling
+           * the buffer down or up respectively.  But since this code
+           * doesn't know about bindings and the keymap code doesn't
+           * know about wheel momentum, that doesn't seem to be
+           * possible yet.  */
+
           if (! ns_use_mwheel_momentum
               && [theEvent momentumPhase] != NSEventPhaseNone)
             return;
-- 
2.20.1