bug#33749: 26.1; input-decode-map to empty vector should preserve echo area

[Stefan Monnier]

> Also, I cannot easily add a new condition at this particular point.
> It happens when Emacs reads each character of the sequence that will
> eventually turn out to map to an empty one, but I do not know that
> will be the case until the end.

I think The Right Thing to do is likely to move this code to
read_key_sequence, more specifically, move it to the point where we
*know* we really do have an event.

IOW, I think the patch below might be a better option (where we test
`indec.start > 0` to make sure some *decoded* event was read).

>> >> Also, who/where do you intend to set input-decode-preserve-echo?
>> Ah, so you're planning to set it once and for all globally?
> Terminal-locally, if it’s any better.

Only marginally.


        Stefan


diff --git a/src/keyboard.c b/src/keyboard.c
index baf2f51440..92ef79b09f 100644
--- a/src/keyboard.c
+++ b/src/keyboard.c
@@ -2928,22 +2928,6 @@ read_char (int commandflag, Lisp_Object map,
       Vinput_method_previous_message = previous_echo_area_message;
     }
 
-  /* Now wipe the echo area, except for help events which do their
-     own stuff with the echo area.  */
-  if (!CONSP (c)
-      || (!(EQ (Qhelp_echo, XCAR (c)))
-         && !(EQ (Qswitch_frame, XCAR (c)))
-         /* Don't wipe echo area for select window events: These might
-            get delayed via `mouse-autoselect-window' (Bug#11304).  */
-         && !(EQ (Qselect_window, XCAR (c)))))
-    {
-      if (!NILP (echo_area_buffer[0]))
-       {
-         safe_run_hooks (Qecho_area_clear_hook);
-         clear_message (1, 0);
-       }
-    }
-
  reread_for_input_method:
  from_macro:
   /* Pass this to the input method, if appropriate.  */
@@ -9070,6 +9054,23 @@ read_key_sequence (Lisp_Object *keybuf, Lisp_Object 
prompt,
       /* If not, we should actually read a character.  */
       else
        {
+          /* Now wipe the echo area, except for help events which do their
+             own stuff with the echo area.  */
+          /* FIXME: This used to happen at the end of read_char (i.e. after
+             we read the first event of a key sequence), but we now do it just
+             before reading the second event, and only when we know that the
+             first event is a "real" one, rather than some internal event that
+             might be dropped altogether (e.g. help-event, switch-frame, or
+             some key that we remap to the empty sequence (bug#33749)).
+             Maybe we should even make sure that `fkey.star > 0` or maybe
+             even `keytran.start > 0`!?  */
+          if (indec.start > 0
+              && !NILP (echo_area_buffer[0]))
+           {
+             safe_run_hooks (Qecho_area_clear_hook);
+             clear_message (1, 0);
+           }
+
          {
            KBOARD *interrupted_kboard = current_kboard;
            struct frame *interrupted_frame = SELECTED_FRAME ();