bug#7728: 24.0.50; GDB backtrace from abort

[Eli Zaretskii]

> From: Stefan Monnier <monnier@iro.umontreal.ca>
> Cc: Drew Adams <drew.adams@oracle.com>,  7728@debbugs.gnu.org
> Date: Sat, 25 Dec 2010 15:35:31 -0500
> 
> If selected_window is nil because of Fset_window_configuration, then it
> is presumably nil for all the intervening and some of the subsequent
> code, and I suspect fixing it earlier in the call chain will be
> preferable: e.g., it doesn't make sense to "display_and_set_cursor" in
> the "selected_window = nil" case.

After some looking around, I'm sorry to report that I don't see how to
do that.  Here are the details:

We set selected_window to nil as a semi-kludgey way of preventing
select-window from storing in the old selected window the value of
point of the buffer that has been restored into that window.
select-window has this code:

  sf = SELECTED_FRAME ();
  if (XFRAME (WINDOW_FRAME (w)) != sf)
    {
      XFRAME (WINDOW_FRAME (w))->selected_window = window;
      /* Use this rather than Fhandle_switch_frame
         so that FRAME_FOCUS_FRAME is moved appropriately as we
         move around in the state where a minibuffer in a separate
         frame is active.  */
      Fselect_frame (WINDOW_FRAME (w), norecord);
      /* Fselect_frame called us back so we've done all the work already.  */
      eassert (EQ (window, selected_window));
      return window;
    }
  else
    sf->selected_window = window;

  /* Store the current buffer's actual point into the
     old selected window.  It belongs to that window,
     and when the window is not selected, must be in the window.  */
  if (!NILP (selected_window))
    {
      ow = XWINDOW (selected_window);
      if (! NILP (ow->buffer))
        set_marker_both (ow->pointm, ow->buffer,
                         BUF_PT (XBUFFER (ow->buffer)),
                         BUF_PT_BYTE (XBUFFER (ow->buffer)));
    }

  selected_window = window;

The last `if' clause is what we want to bypass, when we restore window
configuration as the last part of save-window-excursion.  Note that
select-window stores the right value into selected_window right after
that, but the call to Fselect_frame that crashes happens before that,
so selected_window is still nil.

Now, the sequence of calls leading to the crash inside Fselect_frame
is as follows:

  Fselect_frame
  -> do_switch_frame
     -> Fredirect_frame_focus
        -> w32_frame_rehighlight / XTframe_rehighlight
           -> x_frame_rehighlight
              -> frame_highlight
                 -> x_update_cursor
                    -> update_cursor_in_window_tree
                       -> update_window_cursor
                          -> display_and_set_cursor
                             -> erase_phys_cursor
                                -> window_text_bottom_y
                                   -> die

My conclusion after studying this is that everything that happens
below Fselect_frame is reasonable: we switch to the frame and redraw
the cursor in all of its windows.  In particular,
update_cursor_in_window_tree simply walks the entire window tree of
the newly selected frame.  I don't see how we can avoid any of this
when selected_window is nil, because selected_window has nothing to do
with the windows that are being processed.  None of these functions
even references selected_window, which is TRT.  The first place that
does reference selected_window is the CURRENT_MODE_LINE_HEIGHT macro
used in window_text_bottom_y, and that leads to the abort.

So I see 2 ways to prevent this particular problem:

 1) Handle the case of selected_window == Qnil in
    CURRENT_MODE_LINE_FACE_ID.

 2) Change the code of Fset_window_configuration and Fselect_window,
    to have some other way of preventing the latter from storing point
    in the old selected window, without setting selected_window to
    nil.

Any other ideas are welcome.