Created attachment 19246 [details] [review]
allow hash tables to contain complex types

Previously, dbus-glib has provided destroy functions for the keys and
values when constructing hash tables, so that any hash tables it
constructed could be entirely freed (along with their contents) by
destroying/unreffing. Unfortunately this meant that any "complex" types,
where you need to know the GType in order to free them, were not allowed
in hash tables. In real terms, this was anything which dbus-glib
marshalled to a GPtrArray, so any array of arrays, variants, structures,
object paths, or other boxed types were not allowed as hash values.

This patch allows a broader range of types to be used as the values in
hash tables, including those where no simple free function is available.
Instead of relying on the key/value destroy functions, the new hash_free
function uses g_hash_table_foreach_steal to remove the keys and values
pairwise and free them when the type is known.

Unfortunately, it's part of the API assumptions that hash tables which
were produced through the current API had valid free functions, and
particularly that if the hash table was reffed by the application, that
the keys/values would persist beyond when dbus-glib had unreffed it, and
be freed when the hash table was later destroyed. So it's not sufficient
to use only this new freeing method on all hash tables from now on - we
have to behave in the old way for all of the previously allowable types
(including any hash tables which contain any hash tables which were
freeable in the old way).

However, as these new hash tables contain values which will not be freed
if you manipulate the hash table directly (removing or replacing keys,
or destroying or unreffing it directly), and g_boxed_free should be used
instead, a false free function is provided to print a critical warning
for the developer in the case that memory would be leaked.

Looks pretty good to me, thanks for the extensive comments.  

I'm a bit unsure why you changed the existing tests if the idea was to have backwards compat; maybe add new tests?

Regardless feel free to commit, thanks!

I an existing test because I added it in another patch in this series. The tests for a{gas} and a{oas} which test 'g' and 'o' as hash keys and 'as' as a hash value. But there's no need to test 'as' values twice, so one of the two got repurposed. Merging this.