If you are using gdk_window_scroll() to scroll a window and that window
happens to have child windows gdk will use guffaw scrolling instead of
copyarea scrolling.  If there's an overlapping window, areas of the window
being scrolled will also be damaged and should be repainted.  Right now
they aren't repainted and the expose_event generated has the incorrect
source coordinates set in it.

Here's what owen says:

[15:55:08] <owen> OK, I see that the gdk_window_guffaw_scroll is pretty
clearly broken
[15:55:23] <owen> If you compare it to windodw gdk_window_copy_area_scroll()
[15:55:45] <owen> copy_area_scroll() has:
[15:55:57] <owen>       gdk_window_queue_translation (window, dx, dy);
[15:56:14] <owen> But guffaw_scroll_has:
[15:56:16] <owen>   gint d_xoffset = -dx;
[15:56:16] <owen>   gint d_yoffset = -dy;
[15:56:29] <owen>   if (d_xoffset < 0 || d_yoffset < 0)
[15:56:29] <owen>     gdk_window_queue_translation (window, MIN (d_xoffset,
0), MIN (d_yoffset, 0));
[15:56:31] <owen> [...]
[15:56:42] <owen>   if (d_xoffset > 0 || d_yoffset > 0)
[15:56:42] <owen>     gdk_window_queue_translation (window, MAX (d_xoffset,
0), MAX (d_yoffset, 0));
[15:57:25] <owen> So, the signs are opposite in the two cases ... but they
should be the same

[16:24:45] <owen> If you want to open the bug, that will help me remember;
I want to write a 20-line test case though before I commit a fix

Owen, it's all up to you.