In GNU Emacs 27.0.60 (build 13, x86_64-w64-mingw32)
  of 2020-01-05 built on MACHNO
Repository revision: 7f01dfca5600fcd3e3548aee50734aab4b96b02f
Repository branch: emacs-27

With emacs -Q load the attached test-set-cursor-from-row.el and use
the mouse to drag the upper mode line up and down.  As soon as one of
the windows gets sufficiently small, I get an assertion failure like


Thread 1 hit Breakpoint 1, terminate_due_to_signal (sig=22, backtrace_limit=2147483647) at ../../src/emacs.c:371
371	  signal (sig, SIG_DFL);
(gdb) bt
#0  terminate_due_to_signal (sig=22, backtrace_limit=2147483647) at ../../src/emacs.c:371
#1  0x00000004002c13cd in die (msg=0x400974295 <get_next_element+3989> "!row->mode_line_p", file=0x400970642 <DEFAULT_REHASH_SIZE+4958> "../../src/xdisp.c", line=16249) at ../../src/alloc.c:7246
#2  0x000000040007ca89 in set_cursor_from_row (w=0x76c8500, row=0x76b0070, matrix=0x76cd580, delta=0, delta_bytes=0, dy=0, dvpos=0) at ../../src/xdisp.c:16249
#3  0x0000000400087177 in redisplay_window (window=XIL(0x76c8505), just_this_one_p=false) at ../../src/xdisp.c:18780
#4  0x000000040007c87c in redisplay_window_0 (window=XIL(0x76c8505)) at ../../src/xdisp.c:16190
#5  0x000000040031a182 in internal_condition_case_1 (bfun=0x40007c83d <redisplay_window_0>, arg=XIL(0x76c8505), handlers=XIL(0x6421ebb), hfun=0x40007c801 <redisplay_window_error>) at ../../src/eval.c:1379
#6  0x000000040007c7d3 in redisplay_windows (window=XIL(0x76c8505)) at ../../src/xdisp.c:16170
#7  0x000000040007c785 in redisplay_windows (window=XIL(0x76c82f5)) at ../../src/xdisp.c:16164
#8  0x000000040007b0ad in redisplay_internal () at ../../src/xdisp.c:15638
#9  0x00000004000789af in redisplay () at ../../src/xdisp.c:14865
#10 0x00000004001b6162 in read_char (commandflag=1, map=XIL(0x77dcfb3), prev_event=XIL(0), used_mouse_menu=0xbff25f, end_time=0x0) at ../../src/keyboard.c:2493
#11 0x00000004001c93dd in read_key_sequence (keybuf=0xbff490, prompt=XIL(0), dont_downcase_last=false, can_return_switch_frame=true, fix_current_buffer=true, prevent_redisplay=false) at ../../src/keyboard.c:9553
#12 0x00000004001b28fc in command_loop_1 () at ../../src/keyboard.c:1350
#13 0x000000040031a08d in internal_condition_case (bfun=0x4001b23c6 <command_loop_1>, handlers=XIL(0x90), hfun=0x4001b1807 <cmd_error>) at ../../src/eval.c:1355
#14 0x00000004001b1f3c in command_loop_2 (ignore=XIL(0)) at ../../src/keyboard.c:1091
#15 0x00000004003194e2 in internal_catch (tag=XIL(0xdfe0), func=0x4001b1f0a <command_loop_2>, arg=XIL(0)) at ../../src/eval.c:1116
#16 0x00000004001b1e92 in command_loop () at ../../src/keyboard.c:1070
#17 0x0000000000000000 in ?? ()
Backtrace stopped: previous frame inner to this frame (corrupt stack?)

Lisp Backtrace:
"redisplay_internal (C function)" (0x0)


The immediate cause is that the *scratch* window has a header _and_ a
tab line.  Disabling either of these makes the failure disappear.

The culprit is IIUC this part of xdisp.c

       /* Finally, fall back on the first row of the window after the
	 header line (if any).  This is slightly better than not
	 displaying the cursor at all.  */
       if (!row)
	{
	  row = matrix->rows;
	  if (row->mode_line_p)
	    ++row;
	}
       set_cursor_from_row (w, row, matrix, 0, 0, 0, 0);

which is no more valid because the first row is now the tab line of
the *scratch* window.  In this case the line we get by ++row is the
header line triggering the subsequent assertion failure

   /* Don't even try doing anything if called for a mode-line or
      header-line row, since the rest of the code isn't prepared to
      deal with such calamities.  */
   eassert (!row->mode_line_p);
   if (row->mode_line_p)
     return false;

If my analysis is correct, we should probably skip two lines to move
to the first text line of the window.  But I would not exclude the
possibility (I have not tried to produce such behavior yet) that that
third line does not even exist (what would the "(if any)" in the
comment otherwise allude to?) or at least is the mode line of that
window and the assertion failure would trigger again.

So I think we should do two things there: (1) skip subsequent lines
with row->mode_line_p enabled and (2) not call set_cursor_from_row
when we have consumed all rows.

BTW: I would also recommend to rename mode_line_p to something like
say "no_text_line_p".  Presently, people like me not used to hack the
display code, might consider mode_line_p verbatim and not apply its
semantics to header or tab lines.

martin