If you subtract a sufficiently large number of months from a date using
g_date_subtract_months, then a critical will be issued, either from

  g_return_if_fail (d->year > years);

or from both of these:

  index = g_date_is_leap_year (d->year) ? 1 : 0;
  g_return_if_fail (g_date_valid (d));

Since this is a data error, not a programming error, and since criticals
will cause crashes with certain settings, g_return_if_fail should not be
used for this purpose.  Instead, an error should be signaled by making
the result invalid.

Think of it this way: How can I use g_date_subtract_months without
risking a crash?  The answer is that I can perform the checks myself,
but at that point I have already done the job of the function.

g_date_subtract_days, g_return_if_fail, and g_date_subtract_years seem to
have the same problem.