This information is going to be especially useful in the upcoming Qt
WebEngine version of GoldenDict. In the Qt WebEngine version only the
words equal to the last found result can be highlighted, not all FTS
matches as in the Qt WebKit version.
The assertion failure can be triggered by full-text-searching for a
common word, selecting a result with many large articles and quickly
selecting text in the first article while the page is still loading.
This reverts commit c936d03fa0.
A default-constructed QRegExp is passed to ArticleView::showDefinition()
when a user navigates from an FTS-result page by clicking a link or
double-clicking a word. QRegExp().pattern() returns an empty string,
which is stored in the "regexp" query item of an URL. When this URL is
loaded, ArticleView::loadFinished() calls highlightFTSResults(), which
then calls closeSearch(), performs multiple string and regular
expression operations and early-returns because the regular expression
pattern is empty. Returning earlier skips useless work in this case.
An alternative optimization is not calling highlightFTSResults() at all
when the regular expression is empty, but that would skip the
closeSearch() call and keep the FTS search frame visible on a page with
an empty regexp.
Since the update to FFmpeg 5.0, when FFmpeg+libao internal player is
selected, most sounds fail to be pronounced. Furthermore, each failed
pronunciation attempt increases GoldenDict's CPU usage. Finally,
GoldenDict continues to hang and utilize the CPU cores when the user
attempts to exit it.
The reason for the issue is: GoldenDict's readAudioData() returns 0 at
EOF but FFmpeg expects the callback to return AVERROR_EOF. As a result,
internal player's threads are busy calling readAudioData() indefinitely.
a606f27f4c
documented the expected return value of the read_packet callback.
252500a78f
removed the code that handled the deprecated 0 return value gracefully.
The relevant deprecation warning
"Invalid return value 0 for stream protocol" had been flooding
GoldenDict's output for years
Fixes #1489.
One issue with the current implementation is that a wrong-cased match
remains selected when the user checks the Case Sensitive checkbox.
Another issue is that the noResults property of searchText is not
updated right away, which means that searchText's background color
remains wrong until next search.
Handle toggling case sensitivity in the same way as editing the searched
text: restart search from the beginning of the page. This improves the
Search-in-page UI consistency.
The current article is saved to pages history user data before any
navigation to another page and before reloading the current page.
Therefore saving it each time the user activates an article is
redundant. This redundancy is not consistent, because when a user
activates an article by clicking on it, the current article changes in
the UI, but is not immediately saved in history as setCurrentArticle()
is not called. The inconsistent redundancy is a waste of CPU time and
can hide bugs.
This variable overrides history user data, which makes the current
article and position restoring code in ArticleView::loadFinished()
difficult to understand.
Encode the same logic in the history user data instead. This should make
the code more straightforward and less brittle in the face of changes.
MainWindow calls ArticleView::reload() when the group list is updated.
This updating may add/remove/reorder dictionaries in the active group.
MainWindow also calls ArticleView::reload() when the display or addon
style changes.
In both of the above scenarios uncontrolled jumps and current article
change can occur (see also the parent commit message).
Move setting articleToJump from above the only ArticleView's reload()
call into ArticleView::reload() itself.
When the first article in the list is current, expanding or collapsing
optional parts results in:
1) uncontrolled jumps due to the content height changes if the scroll
position is not (0, 0);
2) current article change if a non-first article is saved in history
user data (e.g. if a mouse click had made the first article active).
Treat the first current article in the same way as non-first ones.
This change also affects the case when the current article is not
present in the article list. If the current article is simply empty,
then there is no behavior change. If the current article is not empty,
it *must* be in the list, unless something else went wrong.
This way it is clearer that the pages history is updated just before
each navigation to a different page.
The call to ArticleView::saveHistoryUserData() now occurs slightly later
in ArticleView::showDefinition(). I don't think the intervening code can
affect the current article or window position. So the reordering most
likely does not affect application behavior.
Without the added saveHistoryUserData() call, returning back from the
enlarged picture page ("gdpicture") restores the current article and the
window position that were last saved. At this commit the "gdpicture"
behavior is consistent with regular links: returning back from the
enlarged picture page sets the article with the picture as current and
restores the window position at the time of the click on the picture.
https://doc.qt.io/archives/qt-5.5/qtwebkit-bridge.html#internet-security
Qt WebKit Bridge documentation recommends:
When exposing native objects to an open web environment, it is
important to understand the security implications. Think whether the
exposed object enables the web environment access things that
shouldn't be open, and whether the web content loaded by that web page
comes from a trusted source.
The author of Qt WebChannel has said the following in a talk that
introduced this Qt module (WebKit Bridge replacement for Qt WebEngine):
My suggestion here is to write dedicated QObjects with a slim, minimal
API that only have the signals and methods that you deem safe to be
used from the outside.
- see a comment under https://redirect.invidious.io/watch?v=KnvnTi6XafA
The default qmake build output is overly verbose. Adding the "silent"
switch to CONFIG makes it much more concise, comparable to default CMake
output. This way warnings and errors are much easier to find.
Adding CONFIG-=silent to the qmake command line doesn't restore the
verbosity because it is applied before the "CONFIG += silent" line in
goldendict.pro. Add a new CONFIG switch "verbose_build_output" to allow
increasing build log verbosity without editing goldendict.pro.
MainWindow::showTranslationFor() overloads disable the "Pronounce Word"
action, then call ArticleView::showDefinition(). And then immediately
update pronounce availability, Found in Dictionaries list, Back and
Forward buttons. Since ArticleView::showDefinition() loads the requested
page asynchronously, the previous page is still current. Therefore the
"Pronounce Word" action is immediately re-enabled (if the still-current
article has sounds), the other state updates have no effect whatsoever.
Once the new page is loaded, the state is updated again in
MainWindow::pageLoaded() - this time with the desired effect.
So the only effect of the state updates in
MainWindow::showTranslationFor() is to revert the intentional disabling
of the "Pronounce Word" action. Plus waste some CPU time. The
pronunciation-disabling behavior looks better to me and is consistent
with the scan popup's behavior (which immediately hides the
"Pronounce Word" button).
When the current article is set and the user expands or collapses
optional parts (e.g. via the Ctrl+* shortcut),
ArticleView::setCurrentArticle() is called twice from
ArticleView::loadFinished(). Furthermore, the window scroll position is
restored before the second jump. This is wasteful. Move the
higher-priority setCurrentArticle() call up and, if it succeeds, skip
the other call and the scrolling.
I have measured the time spent running the affected code fragment on my
GNU/Linux system before and at this commit. When the loaded articles are
not very large, the performance gain of this commit is only about 1 ms.
However, when one of the displayed articles was huge (the
"United States" English Wikipedia article), the time went from 120 ms to
5 ms.
The wall time of calls to ArticleView::highlightAllFtsOccurences() on my
GNU/Linux system before and at this commit:
allMatches.size() uniqueMatches.size() before(ms) at(ms)
79 1 277 4
98 1 380 4
267 1 16803 65
Extracting this function allows to simplify the code and facilitates
optimizing it in the next commit.
Remove `#if QT_VERSION >= 0x040600` along the way as GoldenDict does not
support Qt versions older than 4.6 for several years now.
If GoldenDict's option "Close to system tray" is checked and
GoldenDict's main window is visible when the user logs out, the logout
is canceled in latest stable versions of KDE Plasma and Xfce desktop
environments (probably in other GNU/Linux desktop environments too, but
they weren't tested). The cause of this unintended and pointless logout
cancellation is ignoring the close event.
Close events are accepted by default. main() calls
`app.setQuitOnLastWindowClosed( false );`. Thus, if the close event is
not touched, the main window is hidden as before this change.
GoldenDict's configuration, history and favorites are still committed
and saved in both KDE Plasma and Xfce when logging out first
closes/hides the main window, then quits GoldenDict.
The change is limited to GNU/Linux because @Abs62 pointed out that
closing the main window breaks global hotkeys on Windows. I have
verified that closing the main window does not break global hotkeys on
GNU/Linux with Qt5 or Qt4. No one has volunteered to test whether the
change is needed on macOS, so it is safer not to apply it there.
Closes #1421.
Separate data directory is only used if history file isn't already
present in the configuration directory.
This commit together with two previous ones allows to switch to
layout described in XDG Base Directory specification. Data is not
migrated automatically and old ~/.goldendict home dir has precedence.
If you wish to use standards-compliant locations, move files and
directories manually:
```
CONFIG="${XDG_CONFIG_HOME:-$HOME/.config}/goldendict"
CACHE="${XDG_CACHE_HOME:-$HOME/.cache}/goldendict"
DATA="${XDG_DATA_HOME:-$HOME/.local/share}/goldendict"
mkdir -p "${CACHE}"
mkdir -p "${DATA}"
mv ~/.goldendict/index "${CACHE}"
mv ~/.goldendict/history "${DATA}"
mv ~/.goldendict/ "${CONFIG}"
```
Please note that this effectively discards network cache stored in
"${XDG_CACHE_HOME:-$HOME/.cache}/GoldenDict/network". You can safely
delete the obsolete cache directory. If you wish to preserve and
keep using the existing network cache, move it into the new location
manually:
```
CACHE="${XDG_CACHE_HOME:-$HOME/.cache}"
mkdir -p "${CACHE}/goldendict"
mv "${CACHE}/GoldenDict/network" "${CACHE}/goldendict"
rmdir "${CACHE}/GoldenDict"
```
If GoldenDict with this change has been already used, new cache dir
will already have "${CACHE}/goldendict/network", and mv will refuse to
merge directories. In such case, you will need to manually remove it
before moving directory with old cache.
Most callers of these member functions should escape wildcard symbols in
the `text` argument. Yet nothing in the functions' signatures suggested
such escaping. With the added enum WildcardPolicy argument, the callers
are forced to decide whether or not the wildcards should be escaped.
When not-escaped wildcard symbols are placed in the translate box/line,
word completion can occupy the CPU for seconds. So it is safer to err on
the side of escaping than the other way around.
The missed unescaping in ScanPopup::translateInputFinished() was
inconsistent with the main window. It made escaping in the scan popup's
translate box unusable by attempting to translate e.g. "\*" verbatim.
