The reason we need to enforce this is that the GTypeValueTable
documentation explicitly states that memory is memset to 0 when the
value_init function is called.
https://bugzilla.gnome.org/show_bug.cgi?id=603590
Makes g_signal_emit_valist from 15% to 20% faster.
Results reported from profiling the pan newsreader which uses a variant
of simple and complex signal emissions (i.e no args or various args)
https://bugzilla.gnome.org/show_bug.cgi?id=603590
Most callers of G_VALUE_COLLECT previously had to initialize the GValue
and then G_VALUE_COLLECT would still go through a cleanup phase.
The new variant allows passing a unitialized GValue along with a GType
and speedup the initialization/collection process.
https://bugzilla.gnome.org/show_bug.cgi?id=603590
If threads are available we always enable threads in gobject, which
means all gio/gobject code can enable the unconditional thread calls.
This is a minor optimization since we avoid a bunch of unnecessary
is-threads-enabled checks.
https://bugzilla.gnome.org/show_bug.cgi?id=606775
This works around the need to take a custom mutex twice and add the
object to a GSList of objects that are currently in construction for the
common case. Only when the constructor is overwritten do we use the
previous behavior and allow things like singleton objects.
The only slightly incompatible change is that previously, it was ok to
call g_object_set() on construct-only properties while the object was
initialized. This will now fail. If that behavior is needed, setting a
custom constructor that just chains up will reenable this functionality.
https://bugzilla.gnome.org/show_bug.cgi?id=557151
g_type_default_interface_peek() and g_type_value_table_peek() don't need
to acquire read locks anymore when they test the refcount instead of
node->data.
The function returns TRUE if the type was previously initialized and can
be easily reused. It returns FALSE and does not take a reference if the
type is not referenced yet.
g_type_class_ref() uses this to avoid taking locks in the common path,
which speeds up object creation a lot - in particular in multithreaded
applications.
https://bugzilla.gnome.org/show_bug.cgi?id=585375
This does not change any locking behavior at all, it just replaces
simple getters/setters of the variable with atomic versions.
The ref_count variable was kept as unsigned, even though that requires
casting for all operations, to mirror GObject->refcount.
https://bugzilla.gnome.org/show_bug.cgi?id=585375
Currently interface lookups are do a binary search over all the interfaces
an object implements. Its possible to do this lookup in constant time using for
instance the gcj algorighm described at:
http://gcc.gnu.org/ml/java/1999-q3/msg00377.html
This is an implementation of that based on GAtomicArray.
We implement lock free interface lookup by moving the n_ifaces
counter into memory pointed to by TypeNode->iface_entries, and
then updating this in RCU-style by always copying it, modifying
the copy and then when the modification is done replace the old
pointer with g_atomic_pointer_set.
There is one additional complexity when freeing the old memory,
since the old memory region can be in use. To handle this we
don't free such memory, but put it on a free list and reuse it
later. This means that lock-free lookups must be able to
handle the memory being modified in random ways without crashing,
and at the end we verify that the memory didn't change and the
transaction is ok.
With this infrastructure the patch then implements a lock-free
version of type_lookup_iface_entry_L called type_lookup_iface_vtable_I
and use it in: g_type_interface_peek, g_type_interface_peek_parent
and type_node_check_conformities_UorL.
Using the performance tests from bug 557100 shows that the general
performance difference is negligible, but the lack of a lock for each
type check and interface vfunc call should greatly enhance threaded
scalability.
This adds supports for a lock-less a non-shrinking growable array.
You can use it to do reads using no locks, as long as your read-code
can handle that during the read transaction the object can be modified
by another writer (but it will not change size or be freed), and you
can only trust the result once the transaction has finished successfully.
This doesn't free things like RCU normally does, instead it pushes the
memory on a free list that is reused for other atomic arrays.
The n_children variable can be written when locked, while the n_supers
variable is read at any time. As they both share the same bytes,
accessing them is not threadsafe.
This patch puts them into different bytes.
Thanks to Xan Lopez and valgrind for noticing this.
Store whether the object has a toggleref before decrementing the
refcount to prevent race condition when two threads simultaneously
try to unref an object with a refcount of 2.
Patch by Antoine Tremblay.
https://bugzilla.gnome.org/show_bug.cgi?id=551706
This avoids a bunch of code and makes construction of simple objects
faster.
Object construction performance improvement:
Non-Threaded Threaded
Simple: 14% 5%
Complex: -1.1% -2.2%
Other tests stable.
https://bugzilla.gnome.org/show_bug.cgi?id=557100
If the class has no properties there could be no notification anyway.
This is an important optimization for construction of simple objects.
Object construction performance improvement:
Non-Threaded Threaded
Simple: 84% 91%
Complex: -1.4% -0.6%
Other tests stable.
https://bugzilla.gnome.org/show_bug.cgi?id=557100
This is both cleaner and faster (it avoids function calls and
zeroing the memory twice).
Object construction performance improvement:
Non-Threaded Threaded
Simple: 11% 1.3%
Complex: 8% 6%
Other tests stable.
https://bugzilla.gnome.org/show_bug.cgi?id=557100
