bug#8789: 23.3; debug backtrace buffer changes window on step-through

[martin rudalics]

> A couple of my frame window layouts cause the emacs debugger's backtrace
> buffer to cycle between several windows when stepping through code,
> making it very difficult to focus.
>
> Please follow the minimal steps below to re-produce this undesirable
> behaviour, also confirmed as present by another user (#emacs 'off-by-1')
> in a 'relatively recent git version'..

I never use debug and don't have Emacs 23 installed, so I'm probably not
of very much help here.

> emacs -q
> C-x 3
> C-x 2
> M-x debug-on-entry RET
> apropos RET
> M-x apropos RET
> x RET

IIUC this shows *backtrace* in the window on the right?

> d RET
> d RET

These steps happen without the RETs I presume?

> ..you should see the buffer alternate windows on each step through the
> code.

On my trunk the *backtrace* buffer is alternately shown in the left
lower and the right window, is that what you see?  That is, the left
upper window is never used?

I suppose it happens because `debug' contains this pretty fragile code

              (save-window-excursion
                ...
                (pop-to-buffer debugger-buffer)

which I don't understand so I can only speculate.  Your three-window
setup apparently prevents the creation of a new window so Emacs is
forced to reuse an existing one.  Now repeating "d" does apparently

(1) remove *backtrace* from the window configuration, restoring the
    previous configuration, and

(2) pop to *backtrace* in any but the selected window (which is the left
    upper one).  Now, when it reuses a window, `display-buffer' first
    tries to use the least-recently-used one, which, in your scenario,
    is alternatingly one of the two lower windows.

You can verify (2) for yourself by replacing the line

                  (or (get-lru-window frame-to-use)

in `display-buffer' with the form

                  (or (let ((window (get-lru-window frame-to-use)))
                        (when window
                          (message "%s" window)
                          (sit-for 3)
                          window))

and go through your scenario.

I don't have the slightest idea how to fix this though because I don't
understand why apparently the *backtrace* buffer is removed from display
in (1), and what the subsequent fragment

          ;; Kill or at least neuter the backtrace buffer, so that users
          ;; don't try to execute debugger commands in an invalid context.
          (if (get-buffer-window debugger-buffer 0)
              ;; Still visible despite the save-window-excursion?  Maybe it
              ;; it's in a pop-up frame.  It would be annoying to delete and
              ;; recreate it every time the debugger stops, so instead we'll
              ;; erase it (and maybe hide it) but keep it alive.
              (with-current-buffer debugger-buffer
                (erase-buffer)
                (fundamental-mode)
                (with-selected-window (get-buffer-window debugger-buffer 0)
                  (when (and (window-dedicated-p (selected-window))
                             (not debugger-will-be-back))
                    ;; If the window is not dedicated, burying the buffer
                    ;; will mean that the frame created for it is left
                    ;; around showing some random buffer, and next time we
                    ;; pop to the debugger buffer we'll create yet
                    ;; another frame.
                    ;; If debugger-will-be-back is non-nil, the frame
                    ;; would need to be de-iconified anyway immediately
                    ;; after when we re-enter the debugger, so iconifying it
                    ;; here would cause flashing.
                    ;; Drew Adams is not happy with this: he wants to frame
                    ;; to be left at the top-level, still working on how
                    ;; best to do that.
                    (bury-buffer))))
            (kill-buffer debugger-buffer))

is needed for (despite its detailed comment).  So we need help from
someone familiar with the debug code :-(

martin