[openaudible]

Display.initializeCallbacks looks up the realize signal for Widget.UNREALIZE:

signalIds [Widget.UNREALIZE] = OS.g_signal_lookup (OS.realize, GTK.GTK_TYPE_WIDGET ());

So Control.gtk_unrealize runs at the start of every realize, and never on a real unrealize. A Composite's IM context then keeps a pointer to the client GdkWindow after that window is destroyed. When the widget is realized again, gtk_im_context_set_client_window (im, 0) hands the freed window to GtkIMMulticontext, which crashes:

C  [libgdk-3.so.0+0x4f3fd]  gdk_window_get_screen+0x1d
J  org.eclipse.swt.internal.gtk.GTK.gtk_im_context_set_client_window(JJ)V
j  org.eclipse.swt.widgets.Control.gtk_unrealize(J)J
...
J  org.eclipse.swt.internal.gtk.GTK.gtk_widget_realize(J)V
j  org.eclipse.swt.widgets.Text.computeSizeInPixels(IIZ)
...
j  org.eclipse.swt.widgets.TabFolder.minimumSize(IIZ)
...
j  org.eclipse.jface.window.Window.getInitialSize()

It comes before the segfault as Gdk-CRITICAL: gdk_window_get_screen: assertion 'GDK_IS_WINDOW (window)' failed, then gtk_settings_get_for_screen and g_signal_handlers_disconnect_matched criticals.

Reproduction: GTK 3.24.49 (Debian 13 / Devuan 6), X11, no ibus (GTK_IM_MODULE and XMODIFIERS unset). A JFace dialog with a native TabFolder whose pages are created in the already realized shell: TabItem.setControl unrealizes each page, and the next realize (here Text.computeSize during getInitialSize) crashes. With ibus active the stale pointer goes unnoticed.

Fix: add OS.unrealize and look up "unrealize". Verified on Debian 13: the dialog crashed every time before the change and opens and closes cleanly after it.

🤖 Generated with Claude Code

[akurtakov]

Please sign ECA https://api.eclipse.org/git/eca/status/gh/eclipse-platform/eclipse.platform.swt/3754 . It's mandatory for even looking at your PR.