There was some confusion between "index" as used for the model->files[] array,
and node->index as used for our 1-based row numbers. Now we use "index" only
for indices in the model->files[] array, and node->row for row numbers. Functions
and variables are renamed to clarify whether they refer to indexes or rows.
Signed-off-by: Federico Mena Quintero <federico@novell.com>
The non-standard "filechooser::mime-type" was a remnant of the recent-files code using
a hand-built GFileInfo; now we just query the file info ourselves.
Signed-off-by: Federico Mena Quintero <federico@novell.com>
g_array_new() doesn't reserve any size by default, so during the initial population
of the file array, we'll do more reallocs than strictly needed. We'll start with
a reasonable preallocated size, in this case the number of files which we can
get in a single chunk out of GIO.
Signed-off-by: Federico Mena Quintero <federico@novell.com>
There's no point in running a GtkFileSystemModel with invalid column types.
This way we can also avoid clearing the memory of the column_types array.
Signed-off-by: Federico Mena Quintero <federico@novell.com>
Previously, do_validate_rows() validated 300 rows per iteration. While
this is usually not problematic, tree views with a lot of columns or
complex cell renderers could take inacceptably long, like:
- Epiphany's location bar entry completion has multiline and marked up
text in every cell. Validating a single row took ~1.5ms here.
- In the list view in Nautilus, When enabling all columns, validating a
single row would take ~3ms.
With 300 rows per iteration, that made those examples take 500ms/1s in a
signle main loop callback, and this obviously caused responsiveness
problems.
Now the code uses a timer and limits the time for validating rows to
30ms. This can cause less lines to be invalidated per call, so the
function might be called more often, but generally results in more
responsive applications.
The previous function enumerated the whole directory and used a lot of
outdated API to decide how to show files.
The new code queries the filesystem model to decide about this.
The now unused old functions were removed.
Previously information about file sizes was not available for search
results and recent files, so the column was always hidden. As this
information is now available, we can stop the special handling and use
the same setting as in browse mode.
Use the faster _gtk_file_system_model_get_value() function instead of
gtk_tree_model_get() inside the sort functions. This gives a significant
speed-up when sorting large lists.
In a test case with 40.000 files, the sorting time went from ~5 seconds
to less than 0.5 seconds for my test case. There is 2 significant
problems with gtk_tree_model_get() that cause this:
1) The value is copied, which takes quite a bit of time for strings.
~25% of excessive time or ~1 second in my test
2) The tree model functions need to lookup the interface vfunc. And
gtk_tree_model_get() doesn't do that only once, but multiple times
(verifying column id, getting the actual value, ...)
~75% of excessive time or ~3 seconds in my test
Replace the list model code with the file system model and use all the
file system model API niceties we get from that.
Also adds the function _gtk_file_system_model_add_and_query_file() which
g_file_query_info()'s the file before adding it, so it gets added with
the right information.
Since the time taken by g_file_enumerate_children() depends a lot on the
attributes that are queried, we query the minimum attributes that need
to be queired to display the file chooser.
In particular, the attributes for loading the icon are ignored, as icons
are loaded on demand (see previous commit).
Because loading icons takes a noticable performance, this code loads the
pixbuf on demand and only loads icons for rows that are visible. There
is a few caveats to this:
- The pixbuf cell renderer must report the proer size even if the icon
is not yet loaded. This is achieved by setting a fixed size.
- On theme changes the cahced pixbufs and the cell renderer must be
updated to conform to the new theme.
Previously, there was a GtkTreeSortModel wrapped around the filesystem
model to make it sortable. As the new implementation implements the
GtkTreeSortable interface, we can use this instead.
A lot of code special cases accesses to the tree view for the different
browse modes, which was previously necessary, because the models were
different. Now that they are identical in the first columns, there is no
such need anymore, and the functions don't need to be special cased.
Previously custom functions were used to compute the data passed to the
cell renderers. Now that all this data is saved by the tree models with
compatible nodes, the usual default attribute-to-column mapping can be
used.
With this, caching of the values can happen in the tree model, which
avoids costly lookups of icons or computation of strings. Last but not
least it avoids spurious bugs that could happen when strings changed
without anyone noticing, like the mtime when a new day begins.
All tree models in browse mode now share the first 10 column types
containing all the necessary information to display the model on screen.
Therefor it is now easy to just operate on the tree model associated
with the file tree view and in most cases it isn't necessary anymore to
special case the browse modes.
The new model is mostly API-compatible with the old model (minimal
changes were required), but is a lot faster and has a lot of very
desirable features.
- the model does no longer support a tree, just a list of files in a
given directory
- the storage has been moved to a GArray as opposed to a tree
- no more dependency on GtkFileSystem
- columns are managed by the creator of the model, so any number of
nodes can be added as needed. This also makes the API more similar
to GtkListStore.
- Values are filled on demand using a function given when creating the
model.
- The function can decide to let the model cache returned values or
decide to be called again the next time the value is queried.
- implements GtkTreeSortable
- _gtk_file_system_model_get_value() was added to significantly speed
up value access, which is necessary when sorting large models.
Going from print-to-file to a real printer and back used to nuke
an app-provided uri; now we keep the initial settings around to
avoid that. Fixes bug 587086.
