Saving strings with embedded zero-bytes (\0) as CFStrings would
sometimes fail, and only write the part of the string leading
up to the first zero-byte, instead of all the way to the final
zero-terminator. This bug was revealed by the code-path that
falls back to storing e.g. QTime as strings, via the helper
method QSettingsPrivate::variantToString().
We now use the same approach as on platforms such as Windows
and WinRT, where the string produced by variantToString() is
checked for null-bytes, and if so, stored using a binary
representation instead of as a string. For our case that
means we fall back to CFData when detecting the null-byte.
To separate strings from regular byte arrays, new logic has
been added to variantToString() that wraps the null-byte
strings in @String(). That way we can implement a fast-path
when converting back from CFData, that doesn't go via the
slow and lossy conversion via UTF8, and the resulting QVariant
will be of type QVariant::ByteArray. The reason for using
UTF-8 as the binary representation of the string is that
in the case of storing a QByteArray("@foo") we need to
still be able to convert it back to the same byte array,
which doesn't work if the on-disk format is UTF-16.
Task-number: QTBUG-56124
Change-Id: Iab2f71cf96cf3225de48dc5e71870d74b6dde1e8
Cherry-picked: 764f5bf48c
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Jani Heikkinen <jani.heikkinen@qt.io>
Added the binary compatibility test data files for Qt 5.7 for QtBase
Change-Id: I5b19571f5e266c52622027d820062afa5fd4fbf3
Reviewed-by: Sergio Ahumada <sahumada@texla.cl>
In tst_QCompleter, two completers were leaked because they
had no parent and setCompleter() calls don't reparent.
Fixed by giving them parents.
In tst_QUndo*, fix lots of leaked QActions by storing them
in a QScopedPointer. There were some half-hearted attempts
to clean them up with manual deletes, but I ported these to
scoped pointers, too, to make the code more robust in the
face of failures.
This fixes the remaining errors in GCC 6.1 Linux ASan runs of
tests/auto/widgets/util.
Change-Id: Icc5248cc9cf4514540915924df1c4d9e09c071fa
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
The generated data files for the binary compatibility test
updated for QtBase.
Change-Id: Idae703c83f55ff17cada4419db742ea12b22bf86
Reviewed-by: Sergio Ahumada <sahumada@texla.cl>
Reviewed-by: Tony Sarajärvi <tony.sarajarvi@qt.io>
The usual:
- delete styles
Either by using QScopedPointer.
This fixes the remaining errors in GCC 6.1 Linux ASan runs of
tests/auto/widgets/styles.
Change-Id: Ifba59085c057d474bf964cbb93010c408d773a61
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
GCC warned:
tst_qabstractslider.cpp:858:89: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]
Qt::Orientation orientation = *reinterpret_cast<Qt::Orientation*>(&sliderOrientation);
^
tst_qabstractslider.cpp:867:72: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]
orientation = *reinterpret_cast<Qt::Orientation*>(&wheelOrientation);
^
The solution, of course, would be to use a static_cast here, but
why go via int in the first place? Qt::Orientation can perfectly
well be used in QFETCH, as proven by tst_qmainwindow, among other
things.
Change-Id: I97916a50405e16d114837bc52580ce6666d74b17
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Restore Qt::AA_DontUseNativeDialogs that is disabled in the beginning
of the test function.
Change-Id: I4ff8eab4ecc458c478337824e66b5a59fbdd7c65
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@qt.io>
After e109b8a0f3, it is possible
that the cache will be flushed as a result of inserting a new
font rather than just when the timer event triggers. When doing
superscript and subscript text layouts, we would first get
a regular font engine, then a scaled one, and then reference
the regular font engine *after* getting the scaled one. If the
regular font engine was deleted as a result of inserting the scaled
one, we would get a dangling pointer and crash.
The situation was improved by 49926bb9ef.
You would now to switch between 256 different fonts in the layout
in order to trigger it. The test in the commit will trigger the
crash even with this change.
[ChangeLog][Qt Gui][Text] Fixed a crash that could happen if you
were doing many different text layouts with different fonts
and superscript or subscript alignment.
Task-number: QTBUG-53911
Change-Id: Ia33108252e030eff25924ef1b7c10b9d59b5bc8c
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
We need to delete the style returned from QStyleFactory::create()
ourselves, so put them into a QScopedPointer.
The alternative would have been to create this once, as a member
of tst_QWidget, but this is the minimal approach that ensures
behavior just as the old code, but without the leak.
Change-Id: I527f1031c57be6f05942f4acc057e7dae1af2571
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Found by UBSan:
tst_qapplication.cpp:1754:48: runtime error: member access within address 0x7ffda11f2220 which does not point to an object of type 'SpontaneousEvent'
0x7ffda11f2220: note: object is of type 'QMouseEvent'
The code attempted to model the layout of a QEvent with another class
that allows public access to the memory location that (hopefully)
corresponds to QEvent::spont, gaining access by casting a QEvent
object to that specifically-crafted class.
Fix by the using the existing QSpontaneKeyEvent::setSpontaneous()
call, which, despite its name, works for all QEvent subclasses, and
which has already been fixed to not invoke UB (in bc087db).
Change-Id: I7db8b8a8a823f7d61ab17375142d19dc3874fea5
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Adds two formats that does not have optimized code-paths in qdrawhelper
to ensure the generic path has coverage.
This has already uncovered one bug fixed before this patch could go in.
Change-Id: I0e0a1a873555b27f6438f69a76982b8e06263dcf
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
QDateTimeEdit ignores the time-spec of its date-time value, using its
own time-spec instead; mostly, this works because it first conforms
the value to its own time-spec. However, during construction, before
doing this, it set up its display data, which could leave it with a
different time (rather than a different representation of the given
time) than it was asked to use.
Moved the updateTimeSpec() calls to immediately after setting value in
QDateTimeEditPrivate::init() to ensure correct handling. Added test.
Task-number: QTBUG-54781
Change-Id: I3b07c10997abb858fc0b40558bff96e3fdabbd83
Reviewed-by: Jesus Fernandez <jesus.fernandez@qt.io>
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
1. QSslSocketBackendPrivate::transmit was ignoring 'readBufferMaxSize';
as a result, we can have a user trying to set read buffer's size
to a small value (and more important - reading slowly in a small
chunks from this socket), but SSL itself socket reading 'too fast',
potentially growing its internal buffer to a huge size. This also
results in auto-tests failing - whenever we're trying to limit read
rate in some test.
2. Update qsslsocket auto-test.
Task-number: QTBUG-43388
Task-number: QTBUG-55170
Change-Id: Iedece26df0ac5b3b7cad62cc8c98aedc28e7ca5b
Reviewed-by: Richard J. Moore <rich@kde.org>
In setMinAndMaxSize(), QLayout::removeItem() doesn't actually delete
the removed item. We have to do that ourselves (RAII not necessary,
since the spacer is owned by the layout until we remove it).
In distributeMultiCell(), allocate the QStyle subclass on the stack so
the compiler cleans it up properly on all exit paths (was:
unconditional leak).
Change-Id: I24f8f11af2bfc5abf78f9aab0139dcfe0187402b
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@qt.io>
In sizeConstraints(), QLayout::takeAt(), as the name suggests, doesn't
actually delete the item. We have to do that ourselves.
Likewise, in replaceWidget(), QLayout::replaceWidget() also doesn't
delete the affected item, but returns it. That's spectacularly bad API
design, but the leak is easy to fix: just delete the return value.
Change-Id: I8dcbc59898949eabce766cda2c0edae2e1f2799e
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@qt.io>
The default should be false, meaning the application will prefer to
use a native menubar if the platform supports it. The application
author can set it to true if he wants to always use a Qt-rendered
menubar instead; or, he can call QMenuBar::setNativeMenuBar().
Qt and its plugins should not override the author's wishes.
Instead, if the platform plugin cannot create a native menubar
for whatever reason, createPlatformMenuBar() will return null,
and QMenuBar will fall back to using a Qt menubar instead.
The application can check the result via QMenuBar::isNativeMenuBar().
QMdiArea when maximized inside a QMainWindow with an empty title
does not replace the main window's title if we are using native menus.
This behavior turned out to be the same on Unity as it is on macOS,
so the autotest needed adjustment to expect that behavior whenever
the menubar is native, not only on certain platforms.
tst_QMenuBar::allowActiveAndDisabled() tests a standalone QMenuBar.
In f92f78094 it was disabled on macOS, but on Ubuntu it passes as
long as we force it to be a non-native menubar, so it should pass
that way on macOS too. Removed unused variable RESET to fix warning.
Task-number: QTBUG-54793
Change-Id: I716e40da709f96331cbbf25213bd7bc153e4dbe2
Reviewed-by: Morten Johan Sørvig <morten.sorvig@qt.io>
A check "just in case" -- we don't want leaks nor crashes
due to double deletions, and so on.
Change-Id: I24f1a486f0d438595bbe352ab780b07c5d53acbd
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
The test function tries to create a file inside the application bundle.
Change-Id: Ia429b42b102d5e98f20694058fa2633e3c7de30a
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Found by UBSan:
tst_qgraphicsitem.cpp:5066:29: runtime error: downcast of address 0x2afcb006c7f0 which does not point to an object of type 'GeometryChanger'
0x2afcb006c7f0: note: object is of type 'QGraphicsRectItem'
00 00 00 00 d8 64 ca 98 fc 2a 00 00 40 a9 0b b0 fc 2a 00 00 75 65 29 00 00 00 00 00 35 00 00 00
^~~~~~~~~~~~~~~~~~~~~~~
vptr for 'QGraphicsRectItem'
#0 0x4c5f1c in tst_QGraphicsItem::prepareGeometryChange() tst_qgraphicsitem.cpp:5066
Fix by actually instantiating a GeometryChanger, which incidentally is
the pattern used by paint() a few lines below, too.
While at it, allocate the item on the stack (as is done in paint())
and create a local QRectF variable to avoid repeating the same magic
numbers over and over again.
Change-Id: If5a3d56511000a17703d78d7dd1f0ea072b8bc11
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Allows us to detect cases where the requested vsync combination was
not possible to fulfill.
Change-Id: Ie8f3665129f7a1ab7fcefb94b2298d54520b753a
Reviewed-by: Simon Hausmann <simon.hausmann@qt.io>
The FPS is now calculated and output on the command line in a single
place. The animated fps counter has been replaced with a vertical line
which should make it easier to observe tearing issues when vsync is
disabled.
Change-Id: Id356fc1958c048d85aba48edfed59124454038d4
Reviewed-by: Simon Hausmann <simon.hausmann@qt.io>
The animated FPS counter should be enough to observe smooth animations,
so we use the color of the window to visualize frame latency.
Change-Id: I1171a1c4bdc261ca8655771290c6735357821781
Reviewed-by: Simon Hausmann <simon.hausmann@qt.io>
If the sloppy menu popups - send the leave event to the last active
menu (except Cocoa), because only currect active menu gets enter/leave
events (currently Cocoa is an exception).
Check that the menu really has a mouse before hiding the sloppy menu -
don't rely on enter events.
This patch removes some unnecessary synthetic mouse enter/leave events
from QMenu which causes event duplications with different mouse cursor
position.
Refactor sloppy menu timer handling - start or restart timers on mouse
move events. Enter/leave events are not reliable.
Fixes:
- better enter/leave events handling for native widget actions,
- reduce duplicated enter/leave events for menu actions,
- better handle torn off sloppy menus.
Partially reverts: 0ed68f3f58
Amends: 57ecd5aeeb
Task-number: QTBUG-53068
Change-Id: I7ad56ac1619db124915d373fab82d0512d44c90e
Reviewed-by: Shawn Rutledge <shawn.rutledge@qt.io>
Found by UBSan:
qstring.h:1160:44: runtime error: null pointer passed as argument 1, which is declared to never be null
qstring.h:1160:44: runtime error: null pointer passed as argument 2, which is declared to never be null
Fix by avoiding the memcmp() calls if there's a chance that they
might be called with nullptr.
While at it, also implement !=, >, <=, >= in terms of ==, <,
and add a test, because this particular UB was not fingered by
any of the QtCore test cases, but by a Qt3D one.
Change-Id: I413792dcc8431ef14f0c79f26e89a3e9fab69465
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
This one selftest is currently blocking OS X 10.11
from entering the CI. It can't be reproduced when
run manually.
Task-number: QTBUG-55155
Change-Id: I4553ef2d7813b29f5dc8577976c4482686346504
Reviewed-by: Timur Pocheptsov <timur.pocheptsov@theqtcompany.com>
This has undesired effects when converting a QSqlRecord to JSON.
A char(0) e.g. has special semantics that are undesired when
reading a Tinyint column.
I don't think that returning bool for the special case of a
Tinyint(1) is required. This also did not happen before, and
is also not happening when not using a prepared statement.
Instead, a plain int/uint QVariant is returned.
This patch extends tst_QSqlQuery::integralTypesMysql to also
cover reading and writing booleans from/to a MySQL table column
of type Tinyint(1). Additionally, the reading is now also done
with a prepared statement and we also check the raw variant
value.
The broken behavior fixed by this patch was introduced by me in
commit 194403a348.
Change-Id: I028a3abd83fdd2b42d98d478950d205e5b6bbeb5
Task-number: QTBUG-53397
Reviewed-by: Andy Shaw <andy.shaw@qt.io>
Child widgets should get the show/hide event when the TLW changes its
state, because child widgets are also visible or invisible. This
restores the Qt4 behavior (fixes the Qt4->Qt5 regression).
Restoring/minimizing the TLW now sends the spontaneous show/hide event.
Show events are now handled also in the expose event handler in the
QWidgetWindow class, because the show event must occur before the
expose event to avoid possible flicker e.g. the OpenGL content. This
can happen e.g. on XCB platform. If the "WindowStateChange" event occur
before the expose event (e.g. Windows platform) then the code in expose
event handler will be ignored to prevent event duplications.
Added autotest.
Task-number: QTBUG-50589
Change-Id: Ie9a9329b1f29bff876de28d5948d0d5fb6bc1f05
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@qt.io>
widgetlessNativeDialog() is the only test function that creates
a native file dialog instance. GTK+ versions prior 3.15.5 have
a nasty bug (https://bugzilla.gnome.org/show_bug.cgi?id=725164)
in GtkFileChooserWidget, which makes it leak its folder change
callback, causing a crash "at some point later". Running the
native test last is enough to avoid spinning the event loop after
the test, and that way circumvent the crash (QTBUG-55276).
The crash has been fixed in GTK+ 3.15.5, but the RHEL 7.2 CI has
GTK+ 3.14.13 installed.
Change-Id: I867755969a4458693bd12f848d052adf77a2086e
Task-number: QTBUG-55276
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@qt.io>
Earlier versions of the compiler cannot default
move special member functions, even though we
also define Q_COMPILER_RVALUE_REFS for them.
Fix by retracting the less-often-used of the
two compiler feature defines.
Q_COMPILER_DEFAULT_MEMBERS is not used outside
QtBase in neither 5.6 nor 5.7 (5.8 is not
released at this time, so wasn't considered).
The same is true of the dependent macros
Q_COMPILER_DEFAULT_DELETE_MEMBERS and
Q_DECL_EQ_DEFAULT.
In QtBase, the three uses are:
1. in QAtomic*, where the user also requires
Q_COMPILER_CONSTEXPR, which is not defined
for any MSVC at this time,
2. for QEnableSharedFromThis, which is a class
template with an alternative {} implementa-
tion of the default constructor, and uncon-
ditional user-defined copy special member
functions.
3. The test of the corresponding functionality
in tst_compiler, which this commit amends.
That means that neither of these two only uses
of the macro in Qt libraries are affected by
the change.
The reason we do this change, then, is that in
the future, we want to be able to more easily
restore move special member functions for
classes for which they are suppressed due to
user-defined dtors or copy special member
functions.
Change-Id: I6f88cad66d6b87a758231f16355c3bddae697b86
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
According to https://msdn.microsoft.com/en-us/library/5dae5d43.aspx,
strncpy_s' second argument must not be 0:
> If strDest or strSource is NULL, *or numberOfElements is 0*, the
> invalid parameter handler is invoked.
Move the existing check for len > 0 up to protect the strncpy_s
call, too.
Change-Id: I70d339ea60d4b76f3038b2e4e4756f6590a9bd31
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
uic should use Q_NULLPTR instead of 0 as the default disambiguation
context.
Task-number: QTBUG-45291
Change-Id: I889182c7fe1c4be3336f3cd645aa60838863c633
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Restructured, separated the canonicalising of output lines out (into
an object that prepares the necessary regexes and replacements)
suppress more, changed the path-stripping to strip qtbase's parent
rather than os.getcwd() and took account of shadow builds (so both
source tree and build tree provide prefixes we want to strip from
paths). Also cope with $PWD potentially having symlinks in it, where
os.getcwd() is canonical.
It's possible some output might name files elsewhere in the source
tree; these won't be filtered by the prior cwd prefix removal; and, in
any case, the problem with cwd is only that the ancestry of qtbase is
apt to vary; paths relative to there should be consistent between test
runs. This change shall lead to a one-off rewrite of all expected_*
files; but it should now catch all paths. By stripping both build
root and source root (when different) it also avoids differences for
those doing out-of-source ("shadow") builds.
In our XML formats, any hyphens in root paths (e.g. I had Qt-5.6 in my
build root's path) got represented by a character entity, confounding
the replacement; so also do replacement that catches this. We may
discover other character entity subsitutions needed along with this.
Now filtering line numbers and timing information, including benchmark
results; these numbers all get replaced with 0 to avoid noisy diffs.
Also purging dangling hspace, to placate sanity-bot.
The module can now be imported - the code it runs is packaged as a
main() function that a __name__ == '__main__' stanza runs - and all
data is localised to where it's needed, rather than held in globals.
Tidied up and organized the existing regexes. There are doc-strings;
there is a short usage comment. Data is localised rather than global
and modern pythonic idioms get used where apt.
Regexes are compiled once instead of repeatedly. An object looks
after the list of patterns to apply and its construction handles all
anticipated problems. Failures are mediated by an exception.
The output file now gets written once, instead of twice (once before
editing, then over-write to edit), and Popen uses text mode, so that
write can do the same. Its command is delivered as an array, avoiding
the need to invoke a shell.
Instead of relying on qmake being in our path (which might give us a
bogus QT_VERSION if the one in path doesn't match our build tree), use
the relative path to qmake - we rely on being run in a specific
directory in the build tree, after all. Escape dots in the version
properly, so that 51730 doesn't get mistaken for 5.7.0 (for example),
and moved this check later in the sequence (matching a smaller target
makes it more likely to falsely match).
Overtly check we are in the right directory and tell the user what we
actually need, if run from the wrong place. Simplify handling of the
unsupported use-case for MS-Windows (but note what would be needed for
it).
Change-Id: Ibdff8f8cae173f6c31492648148cc345ae29022b
Reviewed-by: Mitch Curtis <mitch.curtis@qt.io>
Reviewed-by: Frederik Gladhorn <frederik.gladhorn@qt.io>
[ChangeLog][QtGui][QTextDocument] Fixed a bug that would return a wrong
position when searching backward from the end of the document.
Task-number: QTBUG-48182
Change-Id: I6e88f808a50cb840f61e7bc579e2a28c5300089d
Reviewed-by: Simon Hausmann <simon.hausmann@qt.io>
Prevent QAbstractSpinBoxPrivate::interpret() from bailing out
in focus changes after text has been entered.
Task-number: QTBUG-55249
Change-Id: I250b3c50f7db5de2e9356038df20f18ee059df11
Reviewed-by: Maurice Kalinowski <maurice.kalinowski@qt.io>
Use the correct method of QMetaEnum to convert enum values to names.
QMetaEnum::valueToKey() should be used, because it takes an enum value
as an argument, while QMetaEnum::key() takes an index of the enum value.
Change-Id: Ie895fcc935e8835e3d9c416ca34be8bfe82fd74e
Reviewed-by: Shawn Rutledge <shawn.rutledge@qt.io>
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@qt.io>
