Do not cause GVFS warnings from Gtk.FileChooserDialog (which may concern
unavailable monitor backends in the test environment) to fail the tests.

On 32 bit systems pygtkcompat currently fails with

  File "pygtkcompat/pygtkcompat.py", line 74, in _install_enums
    name = flag.value_names[-1].replace(modname + '_', '')
IndexError: cannot fit 'int' into an index-sized integer

on 32 bit systems as some flags in Gdk are too large to fit into a 32 bit
"long". Work around this crash until this gets fixed properly (marked as
FIXME).

We cannot use PYGLIB_PyLong_AsLong() for the range check, as on 32 bit machines
this overflows large uints. Use PyLong_AsLongLong() separately to check for
negative values, and PyLong_AsUnsignedLong() for the actual conversion.

In some test environments we do not actually get a timeout exceptions with
timeout=0, but a different error message. So only ensure that we get the right
kind of error.

In _pygi_argument_from_object() we never put the actual element GType into the
constructed array. The array contains GIArguments, or bytes in the case of
passing a string as an array.

This happened to work on 64 bit machines where GIArgument and char* have the
same size, but not on 32 bit machines (test_gi.TestPropertiesObject.test_strv
fails there).

Commit 4bfe7972 introduced a bug where
a Python subclass of a gi-provided base class overrides a property from the
base class.

The new behaviour in the above commit causes pygobject to seek the property
in the base class and try to read it from there (resulting in confusion)
rather than noticing that the property is overridden and present in the
Python object instance.

To provide a nicer solution here, we can exploit the fact that
g_object_class_find_property() will traverse the hierarchy in order to
find the right GParamSpec, and the returned GParamSpec can tell us exactly
which GType introduces the property. The strategy is:

 1. Find pspec with g_object_class_find_property()
 2. Find the class that owns that property (pspec->owner_type)
 3. See if girepository owns that class.
 3a. If yes, get property from there.
 3b. If not, get property "directly"

And the same for property setting.

Now that _pygi_lookup_property_from_g_type is always passed the type that
implements the property, it no longer has to go recursing through parent
classes, which was the original cause of confusion.

https://bugzilla.gnome.org/show_bug.cgi?id=686942

Fill in the pointer to the struct, not the pointer to the
array position. This makes the GdkAtom** argument in
gtk_clipboard_wait_for_targets() work.

https://bugzilla.gnome.org/show_bug.cgi?id=678620

Fix regression in commit 67136184: If an atom does not have a name, do not
recursively call our own str()/repr() methods, but just print
"Gdk.Atom<atom_id>".

Gdk.Atom is not proper GType'd class, so we cannot override the whole class.
Just override its __str__() and __repr__() methods so that printing atoms shows
something sensible. For nameless/invalid atoms, fall back to the old
<void at 0xdeadbeef> output to help with debugging.

https://bugzilla.gnome.org/show_bug.cgi?id=678620

In corner case where properties use values 2 ** 62 < ±x < 2 ** 63 for
TYPE_INT64 and 2 ** 63 < x < 2 ** 64 for TYPE_UINT64 they will raise
warnings even tough values are valid.

https://bugzilla.gnome.org/show_bug.cgi?id=688949

https://bugzilla.gnome.org/show_bug.cgi?id=688697

... as announced in 3.4.1.1.

In _pygi_marshal_from_py_array(), set the array length after copying the data
in the GI_TYPE_TAG_UINT8 case (which applies for GByteArrays), otherwise it
will always come out as zero length.

Cherry-picked from trunk commit b800a690. Do not cherry-pick the test case
updates as those tests do not yet exist in the 3.4 branch.

Cherry-picked from trunk commit fa568949. Do not cherry-pick the test case
udpate as those tests do not yet exist in the 3-4 branch.

Was sometimes crashing when the destroy notify was called.

https://bugzilla.gnome.org/show_bug.cgi?id=688067

- Fix segfault when using an IOChannel as an iterator: PyIter_Next() returns
  NULL on the last element, instead of raising a StopIteration.

- The default encoding of a stream is 'UTF-8', not NULL. NULL means that the
  stream is being used in binary mode; in that case, we should not attempt to
  do any automagic conversion to an Unicode object. As this special case is
  inconsistent and has never worked anyway, and the current buggy
  implementation breaks binary streams, just drop it without replacement.
  (Introduced in commit de9eae4d)

These bugs were uncovered by the previously committed tests.

This did not have any code coverage at all. The tests uncover a range of bugs,
which will be fixed in the next commit.

This drops the verification of the set priority in
test_add_watch_data_priority(), as find_source_by_id() is not yet available in
this pygobject version.

GLib's gir currently has wrong constants for MININT64 and MAXUINT64; explicitly
set them in an override, until this gets fixed properly.

https://bugzilla.gnome.org/show_bug.cgi?id=685022

GSource IDs are unsigned, so we must use 'I' for parsing then, not 'i'.

https://bugzilla.gnome.org/show_bug.cgi?id=684526

Fix install_signals to not use the parent classes __gsignals__
dict if one does not exist on the given class.

https://bugzilla.gnome.org/show_bug.cgi?id=686496

Cut a bugfix release to have the correct g-i requirement

This makes sure the .egg-info files end up in the same python top level
directory as the rest of the gi .py files.

https://bugzilla.gnome.org/show_bug.cgi?id=686315

Simple fix to dec ref returned values from closures wrapping
python functions.

https://bugzilla.gnome.org/show_bug.cgi?id=686140

We need cairo to build g-i's Regress library, gir, and typelib. Update
configure.ac to only require cairo if --disable-cairo was not given. With
--disable-cairo, skip building the Regress library and skip all tests which use
it.

https://bugzilla.gnome.org/show_bug.cgi?id=685094

Fix regression from commit 1bfcd5d9 (exposed by several test cases) when using
Python 2.x.

https://bugzilla.gnome.org/show_bug.cgi?id=685000

The fix adds an extra args_data list to the PyGIInvokeState
structure. This list is used to track dynamically generated
closures that wrap python callables. This allows the ffi closure
and python callable to be freed when call scope has finished.

https://bugzilla.gnome.org/show_bug.cgi?id=685598

https://bugzilla.gnome.org/show_bug.cgi?id=685947

Also split test_override() into several smaller test cases.

_pygi_hash_pointer_to_arg() needs to handle unsigned integers as well.

https://bugzilla.gnome.org/show_bug.cgi?id=685860

Casting an unsigned to signed and checking if positive was not good. Check the
unsigned 64 is below G_MAXUINT64 instead.

Side issue in https://bugzilla.gnome.org/show_bug.cgi?id=685000

This will help easy_install and pip to figure out that PyGObject
is already installed.

https://bugzilla.gnome.org/show_bug.cgi?id=680138

https://bugzilla.gnome.org/show_bug.cgi?id=685218

https://bugzilla.gnome.org/show_bug.cgi?id=684062

Length arg and type info need to be unref'ed.

https://bugzilla.gnome.org/show_bug.cgi?id=685082