| Age | Commit message (Collapse) | Author |
|---|
|
This is a follow-up to commit 0e9f4a398735cfc19ae783d2ab054d2400d95416,
which tries to speed up sorting the items naturally by their name using
the idea that a fast non-natural pre-sorting already sorts the items
mostly correctly and thus reduces the number of expensive natural
comparisons.
This change only makes sense if the view is really sorted by "Name". In
other cases, the pre-sorting will most likely not be useful.
Thanks to Christoph Feck for pointing this out!
|
|
Since Dolphin 2.0, we have stored the selected items in a QSet<int>,
which is neither space-efficient nor particularly fast when inserting
many items which are in a consecutive range.
This commit replaces the QSet<int> by a new class "KItemSet", which
stores the items in a sorted list of ranges. For each range, we only
store the first index and the length of the range, so we need a lot
less memory for most common selection patterns, and we also save quite
a few CPU cycles in many situations, because adding an item to the
KItemSet will in many cases not need a memory allocation at all, and
it's particularly easy when inserting sorted items into the KItemSet in
a row.
KItemSet contains a minimal subset of QSet's API which makes it
suitable as a drop-in replacement for our needs. It also has iterators,
such that the items can be iterated through easily, also with foreach.
One advantage of KItemSet compared to QSet<int> is that the items are
always iterated through in ascending order.
REVIEW: 113488
|
|
Sort the items in a folder first according to their name, without doing
a natural/locale-aware sorting. This is very fast, but the order of the
items is then already close to the final order in most cases.
The number of expensive natural comparisons required to sort the items
is thus greatly reduced.
In my experiments with a folder with 100,000 items, the time required
to sort the files was reduced by 63% with this patch.
REVIEW: 113485
|
|
This patch actually does two things when collapsing an expanded folder
with expanded children:
(a) It removes all expanded children (which are removed from the model)
from m_expandedDirs.
(b) It remembers the expanded children and restores them if the
top-level folder is re-expanded.
BUG: 304363
FIXED-IN: 4.12.0
REVIEW: 113293
|
|
When we remove items from the model, we called the function
KFileItemModel::removeItems(const KFileItemList&, RemoveItemsBehavior).
This function then looked up the indexes of the items using the hash
m_items. This is wasteful in the situations when the indexes of the
removed items are known in advance (like when an expanded folder is
collapsed in Details View), and it can cause trouble if one item is
contained in the model multiple times (can happen when searching, and a
file both matches the search and is a child of a folder that matches
the search). Even if expanding folders in the search results list might
not be particularly useful most of the time, it makes sense to make the
model more robust to prevent crashes and other unexpected behavior in
such situations.
This patch makes the following changes to achieve that goal:
* Change the argument of removeItems() from KFileItemList to
KItemRangeList. To make this work, the "look the indexes up in
m_items" code is moved from that function to slotItemsDeleted(). In
the other places where removeItems() is called, the indexes are
calculated directly (which is not more difficult than determining the
removed items as a KFileItemList, if one considers that we needed the
function childItems(KFileItem) for that, which is not needed any more
with this patch).
* Also removeFilteredChildren() takes a KItemRangeList now. Rather than
putting the parent KFileItems into a QSet for O(1) lookup (which
prevents O(N^2) worst case behavior for the entire function), it uses
a QSet<ItemData*> now, which should even be more efficient (hashing a
pointer is cheaper than hashing a KFileItem/KUrl).
BUG: 324371
BUG: 325359
FIXED-IN: 4.12.0
REVIEW: 113070
|
|
To reduce unnecessary memory comsumption and CPU usage, we only fill the
QHash<QByteArray, QVariant> if the methods data(int) or setData(int) are
called for the corresponding index, or the data is necessary for sorting
the model.
According to my tests, this patch reduces the memory usage when loading
a folder with 100,000 items by 17% in Icons View, and by 26% in Details
View.
REVIEW: 112725
|
|
|
|
This fixes the problem that filtered child items in Details View may
reappear when switching the view mode and the clearing the filter.
BUG: 325344
REVIEW: 112962
FIXED-IN: 4.11.3
|
|
Also factor out the code that transforms a sorted list of ints to a
KItemRangeList. This removes some duplicated code from KFileItemModel.
Note that overriding operator<<() in KItemRangeList was necessary
because it's not a typedef for QList<KItemRange>, but a class derived
from that now, and some code fails to compile if the return type of
that function is QList<KItemRange> and not KItemRangeList.
REVIEW: 112728
|
|
|
|
When sorting by, e.g., "Size", and the name is used as a fallback
because there are multiple files with the same size, the refreshItems
signal that is received when a file's name is changed either with the
dialog or outside the current view did not cause the view to be resorted
after commit d70a4811807776966c3241a72121242f4d1eaee8. This patch fixes
it.
BUG: 324713
FIXED-IN: 4.11.2
REVIEW: 112561
|
|
The idea is that we no longer assume that the "expandedParentsCount"
for each item will be stored in the QHash. It is only accessed for
items which are expanded, and which are not top-level items (i.e.,
which have an expandedParentsCount > 1).
Some unit tests are added to improve the coverage of the affected code.
REVIEW: 112562
|
|
|
|
KFileItem::determineMimeType() not only determines the mime type, but
also the icon. For folders, it looks for a .directory file inside the
folder, where a custom icon might be stored. This can take quite a bit
of time and cause the problem that some folder's type still appears to
be "unknown" when the view is shown.
We can work around this problem by caching the folder mime type in a
static QString and applying to to all folders, which can be identified
easily with KFileItem::isDir(),
BUG: 321710
FIXED-IN: 4.11.1
REVIEW: 111830
|
|
to avoid too much expensive resorting calls, in case of many refresh items signals.
Followup to patch 111146
CCBUG: 303873
CCBUG: 299565
BUG: 323789
FIXED-IN: 4.11.1
REVIEW: 111195
|
|
|
|
KFileItemModel::setData() should not only cause a resorting when the
sort role is changed. The name is always used as a fallback if the sort
role of multiple files is equal, therefore, renaming a file can change
the correct order of the files even if the files are not sorted by
"name".
Unit test included.
BUG: 323518
FIXED-IN: 4.11.1
REVIEW: 111721
|
|
Storing values which are equivalent to default-constructed QVariants
does not make much sense because QHash::value returns the same value
even if the corresponding key is not found in the hash.
This commit reduces Dolphin's memory consumption in large folders by
up to 7.3% (tested a folder with 100,000 files in Details View) and
reduces the time required for loading a folder.
BUG: 323517
FIXED-IN: 4.11.1
REVIEW: 111922
|
|
Sometimes when items are renamed, the order of the items in the
directory is not affected, but the groups still change (simple example:
with files a, b, c, e, rename "c" to "d"). At the moment, we always emit
the itemsMoved signal in such a case to make sure that the view is
updated. However, it would be preferable if this signal was not emitted
because it can trigger some quite expensive operations which are not
needed at all.
This commit introduces a new signal groupsChanged and modifies
KFileItemModel and KItemListView such that these classes make use of it.
Some unit tests for the new functionality are included as well.
Thanks to Emmanuel Pescosta for finding a latent bug in the code which
was triggered by this change and fixed in
998954db6d53999dfa75d380cbb4ca3111589f66.
REVIEW: 111808
|
|
This small change saves a lot of CPU cycles when the items are resorted.
REVIEW: 111700
|
|
Since m_pendingItemsToInsert is a list of pointers now (and not a list
of KFileItems, as in the 4.10 branch), we have to delete all pointers
when clearing or destroying the model.
I think that no review request is necessary for this small and obvious
change.
CCMAIL: [email protected]
|
|
Added a hash table for target url to url mapping. So when the dir lister
sends us the target url as directory url, we can use the url mapping table
to get the right "Dolphin internal" directory url, which is the non-target url.
BUG: 306219
FIXED-IN: 4.11.0
REVIEW: 111252
|
|
When using Details View, only insert all pending items immediately if
new items are inserted which might be children of a pending item. Fixes
the problem that inserting the items in multiple bunches slows down the
folder loading.
Note that the cause of the slowness when inserting in multiple bunches
is that KItemListSizeHintResolver needs O(N^2) time in the worst case
for inserting N items into a model with N existing ones.
REVIEW: 111226
|
|
|
|
KFileItemModel::createMimeData.
BUG: 307336
FIXED-IN: 4.10.5
REVIEW: 111209
|
|
when the sorting role value is changed.
BUG: 299565
FIXED-IN: 4.11
REVIEW: 111146
|
|
|
|
Before this commit, switching from, e.g., "Sort by Name" to "Sort by
Type" sometimes had no effect until the view was refreshed. The problem
was that the re-sorting was triggered before the type information was
actually added to the model.
BUG: 310705
BUG: 312014
FIXED-IN: 4.10.5
REVIEW: 111004
|
|
The idea is based on
http://milianw.de/blog/katekdevelop-sprint-vienna-2012-take-1
REVIEW: 110686
|
|
|
|
This commit prevents repeated insertions of single items into the list
m_itemData, which shift all following items by one position and result
in O(N^2) worst case complexity for the entire function.
Moreover, the hash m_items is updated only for the items starting from
the first inserted/removed item to save some superfluous calculations
of hash values.
REVIEW: 110355
|
|
If an item is moved out of an expanded folder, the model receives the
dir lister's refreshItems signal for the folder. The method
retrieveData() then updates the folder's properties. This commit makes
sure that the 'isExpanded' state is not touched by retrieveData(). A
side-effect is that the 'isExpanded' role is not initialized to 'false',
but this does not matter because trying to read a non-existing role from
the QHash<QByteArray, QVariant> yields a default-constructed QVariant,
which evaluates to 'false'.
BUG: 299675
FIXED-IN: 4.10.4
REVIEW: 110401
|
|
Conflicts:
plasma/applets/folderview/folderview.cpp
|
|
This commit prevents that all non-ASCII letters which are not umlauts
are grouped in a single group "Others", and that the joint group of
an ASCII letter and the associated umlaut in some locales is called
"Others" if there is only the umlaut in the group.
BUG: 315569
REVIEW: 109457
FIXED-IN: 4.10.2
|
|
@emmanuel: this merge includes your fix from
https://git.reviewboard.kde.org/r/109488/ . I assume this is not a
problem.
CCMAIL: [email protected]
|
|
BUG: 295300
FIXED-IN: 4.10.2
REVIEW: 109488
|
|
If an expanded folder with filtered children is collapsed or removed,
and the parent-child relationship cannot be determined by parsing the
URLs, this patch makes sure that the filtered children do not reappear
when the filter bar is cleared.
REVIEW: 109455
|
|
This commit ensures that the mime types and icons are determined
synchronously for 200 ms when "Sort by Type" is enabled.
REVIEW: 109344
|
|
Conflicts:
dolphin/src/kitemviews/kfileitemmodel.cpp
|
|
This is analogous to commit e053ecdcd57cc39fdcbc314fc8dd22c8b9dbdd4f,
which fixes the same problem for the case that the parent folder is
deleted.
BUG: 316335
FIXED-IN: 4.10.2
REVIEW: 109343
|
|
Before this change, files were grouped by calendar weeks if they were
modified in the current month and ignoring the "first week day setting",
but by the actual distance measured in multiples of seven days if the
were modified last month.
This not only fixes the "ignores first week day" bug, but also the
problem that the inconsistencies in the algorithm could lead to a broken
grouping in the first days of a month (see screenshot in the review
request).
BUG: 181337
FIXED-IN: 4.11
REVIEW: 108667
|
|
This concept is not needed anymore because of the new "determine parents and
expansion levels" approach
Side effect:
Enables treeview for trash, ...
REVIEW: 109191
|
|
changes! :)
* Fixed the "Network browser" and "timeline" issues, by using the
KDirLister's itemsAdded(KUrl,KFileItemList) signal -> Use the
given Url to define the parent-child relationship.
* Changed the name of the slot "slotNewItems" to "slotItemsAdded"
for consistency with the signal.
* Use a QHash<KFileItem, ItemData*> instead of a QSet<KFileItem> to
store the filtered data (needed to keep the O(1) lookup for filtered
KFileItems in slotItemsDeleted + needed to fix bug 311912 "After
erasing a filter, some thumbnails randomly disappear")
* Made the determination of the "expandedParentsCount" slightly
simpler - just adding 1 to the parent's level (Also needed to fix the
"Network browser" and "timeline" issues)
FIXED-IN: 4.11.0
REVIEW: 109180
BUG: 304565
BUG: 311912
BUG: 312890
BUG: 315593
|
|
Conflicts:
dolphin/src/tests/kfileitemmodeltest.cpp
lib/konq/konq_operations.cpp
|
|
Fixes the problem that filtered children of expanded deleted folders
reappear if the filter is cleared.
BUG: 315210
FIXED-IN: 4.10.1
REVIEW: 108976
|
|
The previous approach, which was based on comparing the URLs as
strings, was not only very complex, but also could lead to
inconsistencies in the model, namely, that not all children were
removed from the model when the dir lister reported the parent as
deleted. Later on, this could even lead to a crash.
BUG: 311947
FIXED-IN: 4.11
REVIEW: 108766
|
|
|
|
This prevents expensive and unnecessary repeated rehashing when many
items are inserted into the model.
|
|
1. It seems that it really can happen that KFileItems that we get from
the dir lister's itemsDeleted signal are not in the model any more,
e.g., if a folder where hidden files are shown is left and a folder
where hidden files are not shown is entered. There is no need to
output warnings then.
2. Remove the emptiness-check for the KFileItemList at the beginning.
Even in the unlikely event that we do get an empty list, we return
just a few lines later in the code.
|
|
|
