bug#50096: args-out-of-range in redisplay_internal

[Juri Linkov <juri@linkov.net>]

>> I don't understand how this is supposed to work.  The buffer " *Minibuf-0*"
>> is always empty at the time of calling message3_nolog, whereas the buffer
>> " *Echo Area 0*" contains the message to display.
>
> Yes, some code switches to another buffer at the wrong moment.

There are no buffer switching involved neither in the successful case
nor in the failing case.

>> 5. redisplay_internal calls hscroll_window_tree
>>    where cursor_row already contains information
>>    that was valid when " *Echo Area 0*" was temporarily
>>    displayed in mini_window:
>
> Any idea why we call hscroll_window_tree in this case?

Here are all differences in cursor_row between success and failure:

@@ -1,8 +1,8 @@
 (gdb) p *cursor_row
 {
   x = 0,
   y = 228,
-  pixel_width = 1610,
+  pixel_width = 1910,
   ascent = 15,
   height = 19,
   phys_ascent = 12,
@@ -26,8 +26,8 @@
   },
   end = {
     pos = {
-      charpos = 1897,
-      bytepos = 1897
+      charpos = 2150,
+      bytepos = 2150
     },
     overlay_string_index = -1,
     string_pos = {
@@ -41,14 +41,14 @@
     bytepos = 1737
   },
   maxpos = {
-    charpos = 1897,
-    bytepos = 1897
+    charpos = 2150,
+    bytepos = 2150
   },
   overlay_arrow_bitmap = 0,
   left_user_fringe_bitmap = 0,
   right_user_fringe_bitmap = 0,
   left_fringe_bitmap = 0,
-  right_fringe_bitmap = 0,
+  right_fringe_bitmap = 31,
   left_user_fringe_face_id = 0,
   right_user_fringe_face_id = 0,
   left_fringe_face_id = 0,
@@ -59,7 +59,7 @@
   redraw_fringe_bitmaps_p = true,
   enabled_p = true,
   truncated_on_left_p = false,
-  truncated_on_right_p = false,
+  truncated_on_right_p = true,
   continued_p = false,
   displays_text_p = true,
   ends_at_zv_p = true,

When it fails, truncated_on_right_p is true, that causes
executing code block with init_to_row_start in hscroll_window_tree,
that fails with truncated long miniwindow lines.

In the successful case truncated_on_right_p=false allows
hscroll_window_tree to skip the block with init_to_row_start,
and other hscrolling.

> I'm guessing this is due to some customizations of yours, in which
> case I'd appreciate a reproduction recipe starting from "emacs -Q".
> It is very hard to debug such problems via email.

It's 100% reproducible for me, but I'm not sure how easy would be
to create a test case for "emacs -Q".

>> > You are saying that if you remove the ":(literal)" part without
>> > changing anything else, the problem goes away?
>>
>> Indeed, it broke after the commit 3572613550.
>> But after applying this patch it works again without errors:
>
> Does this patch work by preventing hscrolling?
>
>> I guess additional ":(literal)" string increases the length
>> of the displayed message, and longer message triggers the bug.
>
> How does it trigger it? via hscrolling or some other way?

It reduces the length of the displayed message,
so there is no wrapped line and no hscrolling.