-
Dave Jones just posted the first pass of his boot/login analysis. Awesome, awesome stuff. I must admit that this got me a lot more excited than a lot of the dtracing that has been going on ;)
The systemtap dudes are at it, too.
-
Tuomas has empirically validated the famous R.M.L. conjecture on the futility of CPU-focused optimizations:
As processor speed increases, the time spent in I/O approaches 100% of total execution time, until all processes are fully I/O bound.
-
Mark: Indeed, the combination of a *priv field and GObject private data is what we ended up using in PangoFcFont.
Before my patch, the public PangoFcFont looked like this:
struct _PangoFcFont { PangoFont parent_instance; FcPattern *font_pattern; /* fully resolved pattern */ PangoFontMap *fontmap; /* associated map */ gpointer context_key; /* used internally */ PangoMatrix matrix; /* used internally */ PangoFontDescription *description; GSList *metrics_by_lang; guint is_hinted : 1; guint is_transformed : 1; };That structure had no padding fields for extensions, so it wouldn't have been posible to simply add a *priv field. But after reading the Pango code, I found that the context_key field is really only used internally, and it really doesn't make sense to access it from the outside, since it is an opaque token. So, I did this:
-
Move "gpointer context_key" to PangoFcFontPrivate. Add a setter/getter for that field; change the relevant code to use them.
-
In the public PangoFcFont, replace "gpointer context_key" with "gpointer priv".
-
Simply allocate PangoFcFontPrivate with g_new(), and stuff it into the priv field in the instance_init() method.
Some misguided souls, lured by the siren call of exotic APIs, replaced my scheme by what you describe: use g_type_class_add_private() in class_init(), use G_TYPE_INSTANCE_GET_PRIVATE() in instance_init() to get a pointer to the private area, and stuff that pointer in the priv field. Let's compare both schemes.
g_new() and priv g_type_class_add_private() and priv Code changes - Call to "self->priv = g_new()" in instance_init()
- Call to "g_free (self->priv)" in finalize()
- Call to g_type_class_add_private() in class_init().
- Call to "self->priv = G_TYPE_INSTANCE_GET_PRIVATE (...)" in instance_init()
New APIs to learn NONE g_type_class_add_private(), G_TYPE_INSTANCE_GET_PRIVATE() Summary: In this case, we got lucky because we actually had room for a priv field in the public structure. If we hadn't had that luxury, we wouldn't have been able to make the optimization to bring pango_fc_font_get_glyph() from being 8% of the benchmark's time to just over 1%.
-
-
Pvanhoof: note that blindly using G_TYPE_INSTANCE_GET_PRIVATE() can lead to performance problems if you use it in a critical code path.
Since the dawn of time, the GNOME Programming Guidelines have recommended that you use a "*priv" field in your objects to point to the private data. This is perfect when implementing new APIs, since you can plan in advance and have a *priv since the beginning.
G_TYPE_INSTANCE_GET_PRIVATE() and friends were added to make it easier to add private data to existing APIs which cannot be changed and do not have *priv fields already. When you register a new GType, you can call g_type_class_add_private() to tell GType, "when instantiating an object of this type, also allocate a chunk of a certain size, which I'll access with G_TYPE_INSTANCE_GET_PRIVATE() when I need it".
This is is pretty useful, but consider how G_TYPE_INSTANCE_GET_PRIVATE() has to be implemented. Remember that you call it like this:
MyPriv *priv = G_TYPE_INSTANCE_GET_PRIVATE (my_obj, MY_OBJECT_TYPE, MyPriv);
So, we end up doing this:
-
MY_OBJECT_TYPE is my_object_get_type(), which is a function call.
-
Get the class of the my_obj instance. Get the type from the class. Do a type lookup for the metadata of that type, so that we can find the size of MyObj without taking the private data into account.
-
Look up the type metadata for MY_OBJECT_TYPE.
-
Get the offset of the private data chunk as specified in the metadata. This offset is then added to to my_obj and the size from (2).
And *then* you can dereference the pointer to the private data. This is a lot more work than having a *priv field in the first place: g_type_instance_get_private() compiles to a few hundred instructions, while "myobj->priv>some_private_field" is trivial.
-
-
Yesterday Hans Petter and I flew to Boston. A girl was dragging a huge, freaky doll around the Mexico City airport.
And they give you trouble if you try to carry a pocket knife. Anyway, after some running in the Houston airport due to slow lines at immigration, and an uneventful flight, we arrived to the North Pole. Temperature, -5°C.
We dropped our stuff at the hotel and went for dinner. Pho Pasteur is a wonderful place in Harvard Square. We had a rich soup with noodles and meat — great for the cold — and a pot where you boil thin slices of meat along with chunks of pineapple. You then put the cooked meat in rice paper, add some fresh herbs, and a delicious peanut-based hot sauce. Eating such a Vietnamese taco with chopsticks is tricky.
How cold is it in Boston? It was about -10°C by the time we came out of the restaurant. Humidity freezes on the inside of the hotel's bedroom windows.
-
Smoke-testing the GNOME 2.13.4 release
With much help from Elijah and Vincent, today I prepared the GNOME 2.13.4 release. I did some quick smoke-testing, and some things came up.
Look and feel
Stock, blank desktop is boring. Window title is too blue. Controls are too gray.
Plain black foot in menu bar? Color/shade it somehow?
Spacing between foot and "Applications" is the same as the space between the other menu items. So, the foot looks like a different item than "Applications".
Gray-bluish background is boring. Photo contest?
Panel tooltips for menu, launchers are instantaneous. They should have the normal delay.
Nothing in notification area; it just appears as a drawing glitch.
Nothing in initial window list; handle appears as a drawing glitch.
Window menu icon has a visible break with the rest of the panel.
Clearlooks doesn't seem to have a "resize grip" in the status bar, yet I can resize from that spot.
Menu bar gradient breaks off with window frame.
Close "X" in window manager menu is too thin.
Bugs:
Clearlooks has a drawing glitch in scrollbars: bug #325818.
Gucharmap - doesn't display your own script at startup: bug #311912.
Gedit File/Quit is disabled if all documents are closed: bug #325814.
Gnome-settings-daemon didn't exit upon logout while the X server was still up. Normally it gets killed when X dies, but it should exit by itself anyway.
Gnome-screenshot's UI is plain weird. If I must pick a file, why not use a full-fledged file chooser rather than the current combination of GtkEntry+GtkFileChooserButton? Also, GtkFileChooserButton displays "Home" instead of your $HOME name, in inconsistent fashion with GtkFileChooser and Nautilus: bug #325817.
