qSwap() is our wrapper around
using std::swap;
swap(lhs, rhs);
it needn't and shouldn't be overloaded.
ADL swap() should be, though, so qSwap(), std::ranges::swap() and all
the other adl_swap()s out there all find the optimized version.
Qt 5.15 has it correct, Qt 6 wrong. Fix it.
Can't pick to 6.2 because, while backwards-source-compatible, because
the generic qSwap() template provides the name for both qualified and
unqualified calls, it's not forwards-source-compatible: A new user of
ADL swap
// compile error w/o `using std::swap`, pessimization otherwise:
swap(dp1, dp2);
would break or performance-regress when going back to an older
version.
Pick-to: 6.3
Change-Id: I725949a4aa9ae438a182b4b7552ff2dced767e2f
Reviewed-by: Andrei Golubev <andrei.golubev@qt.io>
Reviewed-by: Mårten Nordheim <marten.nordheim@qt.io>
Reviewed-by: Fabian Kosmale <fabian.kosmale@qt.io>
Create macros that wrap the magic developed in
7d63efc16f and apply it to all
Q_DECLARE_METATYPE invocations that show up in Clang -ftime-trace for
a PCH'ed QtGui build.
Effects on compile times:
Clang 10 -ftme-trace:
$ ClangBuildAnalyzer --analyze qtgui-before.trace | head -n6
Analyzing build trace from 'qtgui-before.trace'...
**** Time summary:
Compilation (523 times):
Parsing (frontend): 628.3 s
Codegen & opts (backend): 304.5 s
$ ClangBuildAnalyzer --analyze qtgui-after.trace | head -n6
Analyzing build trace from 'qtgui-after.trace'...
**** Time summary:
Compilation (523 times):
Parsing (frontend): 546.0 s
Codegen & opts (backend): 304.4 s
GCC 11 time (bash builtin):
before:
$ time for ((i=0; i < 3; ++i)) do touch src/gui/painting/qpolygon.h ; ninja libQt6Gui.so; done
real 4m13,539s
user 49m24,416s
sys 3m18,177s
after:
$ time for ((i=0; i < 3; ++i)) do touch src/gui/painting/qpolygon.h ; ninja libQt6Gui.so; done
real 3m55,697s
user 45m19,941s
sys 3m7,370s
Task-number: QTBUG-97601
Pick-to: 6.3
Change-Id: Ia8e37a58937568a7ed21cfeb4b27274deca4d53b
Reviewed-by: Fabian Kosmale <fabian.kosmale@qt.io>
I can't test on Windows, so skipped the platform-specific code.
Pick-to: 6.3 6.2
Change-Id: Id13d4abc447ddd5d17fb67b670b83207877456f6
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
After QFuture continuations became non-copyable (see earlier commits),
we have to always use ContinuationWrapper to save the continuations
inside std::function, since it requires the callable to be copyable.
Optimize the wrapper, by storing the callable directly (instead of using
a ref-counted QSharedPointer) and introducing a fake copy-constructor
that makes sure that it's never called.
Pick-to: 6.3 6.2
Change-Id: I0ed5f90ad62ede3b5c6d6e56ef58eb6377122920
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
Reviewed-by: Andrei Golubev <andrei.golubev@qt.io>
Reviewed-by: Mårten Nordheim <marten.nordheim@qt.io>
This is required to ensure that the continuation attached to a
QFuture returned by QtFuture::when* methods is cleaned in the destructor
of the associated QPromise, so that it doesn't keep any ref-counted
copies to the shared data, thus preventing it from being deleted.
Task-number: QTBUG-99534
Pick-to: 6.3
Change-Id: If4e2929b2e638d6b48c95f0aef9dc886066cedbe
Reviewed-by: Andrei Golubev <andrei.golubev@qt.io>
Reviewed-by: Mårten Nordheim <marten.nordheim@qt.io>
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
Continuations were using QFutureInterface to create and return the
associated future to the user. Attaching a continuation to the returned
future could cause memory leaks (described in an earlier commit). Use a
QPromise when saving the continuation, to make sure that the attached
continuation is cleaned in the destructor of the associated QPromise, so
that it doesn't keep any ref-counted copies to the shared data, thus
preventing it from being deleted.
Task-number: QTBUG-99534
Pick-to: 6.3 6.2
Change-Id: I52d5501292095d41d1e060b7dd140c8e5d01335c
Reviewed-by: Andrei Golubev <andrei.golubev@qt.io>
Reviewed-by: Mårten Nordheim <marten.nordheim@qt.io>
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
Capturing a QFuture in the continuations attached to it results in
memory leaks. QFuture's ref-counted data can only be deleted when the
last copy referencing the data gets deleted. The saved continuation
that keeps a copy of the future (as in case of the lambda capture) will
prevent the data from being deleted. So we need to manually clean the
continuation after it is run. But this doesn't solve the problem if the
continuation isn't run. In that case, clean the continuation in the
destructor of the associated QPromise.
To avoid similar leaks, internally we should always create futures via
QPromise, instead of the ref-counted QFutureInterface, so that the
continuation is always cleaned in the destructor. Currently QFuture
continuations and QtFuture::when* methods use QFutureInterface directly,
which will be fixed by the follow-up commits.
Fixes: QTBUG-99534
Pick-to: 6.3 6.2
Change-Id: Ic13e7dffd8cb25bd6b87e5416fe4d1a97af74c9b
Reviewed-by: Mårten Nordheim <marten.nordheim@qt.io>
Avoid some unnecessary comparisons and add more tests.
Task-number: QTBUG-99799
Change-Id: I3aee9f0b62461d38dadbe8e969444e1cd1f94e68
Reviewed-by: Sona Kurazyan <sona.kurazyan@qt.io>
Drive-by update one more for to ranged-for and make sure we don't create
the global statics on destruction.
Change-Id: I5e52dc5b093c43a3b678fffd16b5ff674dfd17ae
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
Better code style. I need to optimize QCborValueRef::toString() to avoid
a round-trip through QCborValue.
Change-Id: I5e52dc5b093c43a3b678fffd16b5f1f99851cf5f
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
This showed up on a benchmark when the number of files in the directory
was way too big.
Change-Id: I5e52dc5b093c43a3b678fffd16b5ef9a938abc63
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
This works around mismatch in threads starting and restarting QThreads,
and is safe since we don't need to establish a binding, and objectName
access in QThreadPool is locked behind a mutex.
Pick-to: 6.3 6.2
Fixes: QTBUG-96718
Change-Id: Id3f75e4f8344796ca658899645219fe3373ddd6d
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
Unnecessary. We can't read a directory's entries if it doesn't exist or
isn't a directory.
Change-Id: I5e52dc5b093c43a3b678fffd16b5edce70eb651e
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
This is added specifically for the QPA platform and theme plugins, to
honor the QT_QPA_PLATFORM_PLUGIN_PATH environment variable and the
(inadvisable) -platformpluginpath command-line argument.
This removes the last QFactoryLoader used with an empty path (also the
only two that could be reached), which were causing a scan of the
application's binary directory whenever the platform plugin path was
set. In case of applications installed to /usr/bin, the entire /usr/bin
was scanned, which can be qualified as "not good".
Fixes: QTBUG-97950
Pick-to: 6.3
Change-Id: Ice04365c72984d07a64dfffd16b47fe1d22f26d3
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
I'm going to need to call this with a different path.
Pick-to: 6.3
Change-Id: I5e52dc5b093c43a3b678fffd16b5ef59376498ee
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
GCC is unable to emit the SEH metadata about the stack aligning that is
required to execute AVX aligned instructions (VMOVDQA, VMOVAPS, etc.),
so it just doesn't align the stack. That causes crashes on a 50/50
chance every time the compiler attempts to address a stack-aligned
variable. In a debug-mode build, because it always loads & saves
everything on the stack, the chance of a crash happening is a near
certainty.
So we hack around it by going behind the compiler's back and instructing
the assembler to emit the unaligned counterparts of the instructions
every time the compiler wished to emit the aligned one. There's no
performance penalty: if the variable is actually aligned, the unaligned
instruction executes in the exact same time.
Change-Id: Ib42b3adc93bf4d43bd55fffd16c29cac0da18972
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
The Intel compiler is now based on Clang, so it always defines the
macros like Clang and GCC do, so we don't need to worry about it any
more. We only need to define the macros that MSVC lacks.
Change-Id: Ib42b3adc93bf4d43bd55fffd16c10f0f6fef43ef
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
Since it shows up as a new enum value in the 6.3 header review, it's
reasonable to assume that it was added for 6.3.
Pick-to: 6.3
Change-Id: If766ef56f3354644fbda09088514e55b28a44f32
Reviewed-by: Andy Shaw <andy.shaw@qt.io>
It's private API, but exported, so de-inline the dtor to pin the
vtable in QtGui instead of potentially duplicating it in every library
that uses the class.
Ditto ctor, but that's just code hygiene: we don't want the code to be
duplicated across all users.
Pick-to: 6.3
Task-number: QTBUG-45582
Change-Id: I91ea38be20fc67795466a68ca5721837255b33a0
Reviewed-by: Volker Hilsheimer <volker.hilsheimer@qt.io>
Fix various violations of the coding style or general inconsistencies.
No claim for completeness.
* indentation and line breaks
* consistent scopes for case statements where needed
* add curly-brackets for if-statements where needed
* removed {} where not needed
* const'ify a few obvious local variables
* remove random empty lines
* use auto when type is obvious from cast
Deliberately not touching nested if-statements that could be merged into
one.
Pick-to: 6.3
Change-Id: Ie22b36568f33e18d5f15c751c7fd76e1490133b9
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@qt.io>
qSwap() is a monster that looks for ADL overloads of swap() and also
detects the noexcept of the wrapped swap() function, so it should only
be used when the argument type is unknown. In the vast majority of
cases, the type is known to be efficiently std::swap()able or to have
a member-swap. Call either of these.
For the common case of pointer types, circumvent the expensive trait
checks on std::swap() by providing a hand-rolled qt_ptr_swap()
template, the advantage being that it can be unconditionally noexcept,
removing all type traits instantiations. Don't document it, otherwise
we'd be unable to pick it to 6.2.
Effects on Clang -ftime-trace of a PCH'ed libQt6Gui.so build:
before:
**** Template sets that took longest to instantiate:
[...]
27766 ms: qSwap<$> (9073 times, avg 3 ms)
[...]
2806 ms: std::swap<$> (1229 times, avg 2 ms)
(30572ms)
after:
**** Template sets that took longest to instantiate:
[...]
5047 ms: qSwap<$> (641 times, avg 7 ms)
[...]
3371 ms: std::swap<$> (1376 times, avg 2 ms)
[qt_ptr_swap<$> does not appear in the top 400, so < 905ms]
(< 9323ms)
As a drive-by, remove superfluous inline keywords and template
ornaments.
Task-number: QTBUG-97601
Pick-to: 6.3 6.2
Change-Id: I88f9b4e3cbece268c4a1238b6d50e5712a1bab5a
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
Reviewed-by: Joerg Bornemann <joerg.bornemann@qt.io>
Reviewed-by: Fabian Kosmale <fabian.kosmale@qt.io>
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
We already have a precedent for this: the QueueCreateInfoModifier
callback. Following the pattern, add a EnabledFeaturesModifier that
can alter the VkPhysicalDeviceFeatures that is passed to
vkCreateDevice().
[ChangeLog][QtGui][QVulkanWindow] QVulkanWindow can now invoke a
callback to alter the VkPhysicalDeviceFeatures object used to create the
Vulkan device. This allows enabling 1.1, 1.2, and extension features.
Fixes: QTBUG-99803
Change-Id: I5ede0c6bc3430cbb304d4961eb9e44faad5ce4d7
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
[ChangeLog][QtGui][QVulkanWindow] QVulkanWindow is now enabling all
Vulkan 1.0 features reported as supported from the physical device.
Pick-to: 6.2 6.3 5.15
Task-number: QTBUG-99803
Change-Id: Ib9cfcd449904c67b07e0e2d4ade5bcaeb4cb0ce6
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
The current behavior for handling the angle delta of a wheel event
changes index the instant there is a change in angle delta. This works
fine for mouse wheels that send events with 120 angle delta units and
there is also already behavior defined for devices with pixel deltas,
but there is nothing good for handling events from high resolution mouse
wheels that don't have pixel deltas.
This patch makes it so that the current index doesn't change until the
accumulated angle delta for the X or Y axis reaches 120.
[ChangeLog][QtWidgets][QTabBar] Scrolling with a high resolution mouse
wheel changes the current index at a rate more like a normal mouse
wheel.
Task-number: QTBUG-97844
Pick-to: 6.3
Change-Id: I2e7fd88984a253f6ef8a0008deb7233e4cb4d84a
Reviewed-by: Shawn Rutledge <shawn.rutledge@qt.io>
Not to have warnings about invalid (nullptr) parameters.
Change-Id: I5fdfa7e99df0f3c9907055cf244efa5a56b21c11
Reviewed-by: Mårten Nordheim <marten.nordheim@qt.io>
We do not test actively test setups where a separate graphics and
present queue is used because there is no combined queue at all. (it
won't be tested because we neither want to nor have the possibility to
do so)
However, QVulkanWindow (unlike, say, QRhi's Vulkan backend) attempts
to support this. It turns out the argument passed to vkQueuePresent is
wrong: the present is to be submitted to the present queue. So fix
this up.
Pick-to: 6.3 6.2 5.15
Fixes: QTBUG-73470
Change-Id: Ic9b589aba52e3326637216b98a074e27fdc3e3b9
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
The UB that the C and C++ standards talk about do not apply if we use
intrinsics. We can rely on the processors' architectural behavior
instead.
There are two ways to detect a conversion that cannot be represented in
the result. One would be to check if the #IE bit got set in the MXCSR,
but in order to do that we'd need two issue an STMXCSR+LDMCXSR pair to
clear the bit first and then another STMXCSR at the end to see if it got
set. Those instructions are 4 uops long and necessarily target memory,
so that's a bit slow.
This commit implements the second way, which is to check if the result
of the conversion is the "undefined" value. Unfortunately, that value is
a valid, precise value that double can hold for all data types except
unsigned 64-bit, so we need to recheck if that was the actual value
stored in the original double.
This implementation targets 64-bit exclusively because that avoids
having to deal with the 64-bit intrinsics not even being defined in 32-
bit code (converting a double to 64-bit integer in 32-bit is messy). The
unsigned implementation is only implemented with AVX512F because of the
unsigned conversion instructions that were introduced then.
Change-Id: I89446ea06b5742efb194fffd16bb9f04b2014bab
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
This makes it easier to set global painter options
which affect all style painting.
Change-Id: I6a38204ed2d874255e92345e6a6a50d27939fb24
Reviewed-by: Volker Hilsheimer <volker.hilsheimer@qt.io>
Arguably, when talking about «null-string» constructor, it might be
useful to read about which strings are considered null, and which
methods one can use to test that.
Change-Id: Ie30144f33000aac53f4041cfb99da28a79dad946
Reviewed-by: Shawn Rutledge <shawn.rutledge@qt.io>
Reviewed-by: Fabian Kosmale <fabian.kosmale@qt.io>
Qt 5 uses begin() so the fix there will be to use cbegin().
Found by Clang -ftime-trace pin-pointing repeated instantiations
of QList<int>::data().
Pick-to: 6.3 6.2 5.15
Task-number: QTBUG-97601
Change-Id: I6410e5b303766fdbc7e158a9ac1263adec973099
Reviewed-by: Fabian Kosmale <fabian.kosmale@qt.io>
We know decltype(d), so we don't need to jump though the ADL-enabling
that qSwap() does. Just call QExplicitlySharedDataPointer's
member-swap directly.
Found through Clang -ftrace-time over PCH'ed libQt6Gui.so build:
**** Templates that took longest to instantiate:
[...]
4050 ms: qSwap<QExplicitlySharedDataPointer<QColorTransformPrivate> > (87 times, avg 46 ms)
which is gone afterwards.
Task-number: QTBUG-97601
Pick-to: 6.3 6.2
Change-Id: Ie054848922a50dbf746781491cb28e598c0e12bc
Reviewed-by: Fabian Kosmale <fabian.kosmale@qt.io>
Doing Present(0, 0) is not necessarily sufficient to get rid of
blocking. It may very well start blocking after a few frames.
This does not apply to a non-flip-discard swapchain (when running
with QT_D3D_NO_FLIP=1), but for flip-discard we should also try using
DXGI_SWAP_CHAIN_FLAG_ALLOW_TEARING and DXGI_PRESENT_ALLOW_TEARING in
case a swap interval of 0 is wanted.
Pick-to: 6.3
Fixes: QTBUG-99949
Change-Id: I9cb13b139ba04e41b4f25b94bcd3d1e973496414
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
We set the wakeUps atomic to prevent multiple WM_QT_SENDPOSTEDEVENTS from
being posted. However, this might happen right after the event processing
thread cleared the atomic, but before it processed the previous
WM_QT_SENDPOSTEDEVENTS message. In that case, we end up with a set
atomic and an empty event queue, resulting in the event loop to block
even though there are posted QEvents.
To prevent that, always reset the atomic when we handle the
WM_QT_SENDPOSTEDEVENTS message. In that case, we either call
sendPostedEvents, or startPostedEventsTimer. The former already resets
wakeUps; reset it in the latter as well.
Fixes: QTBUG-99323
Pick-to: 6.2 6.3 5.15
Change-Id: I931c02be9c42b02e8ca20daba5059cd8185f0a37
Reviewed-by: Alex Trotsenko <alex1973tr@gmail.com>
Recommending a Qt 4 book in Qt 6 documentation tells us something about
how much we maintain the list :)
The other books might still be good sources. Anyhow, the chances of a
customer looking exactly in this place to learn good books about icons are
not very high. So let's just ditch the page, and use external links instead.
Pick-to: 6.3
Change-Id: I8013a5ab9d3416fe795f4aaed647e26db79508a1
Reviewed-by: Leena Miettinen <riitta-leena.miettinen@qt.io>
Reviewed-by: Nicholas Bennett <nicholas.bennett@qt.io>
In Qt 5 times, the core of QList::realloc() was out-of-line by design,
because it was independent of T.
Now that QList is QVector, its equivalent detachAndGrow() function on
QArrayDataPointer is inline and instantiated for each type anew. We
therefore need to be careful to not use detach()ing QList operations
in non-generic code inline code (in public, but also private,
headers), because (common) PCH builds force this code to be compiled
over and over again. Generic code is only instantiated when used in a
TU, so that's ok. But for non-generic code, the only option is to
de-inline.
If there is an effect on compile-times, it's hidden in the run-by-run
noise of building QtGui, but at least this entry is gone afterwards
from clang -ftime-trace:
**** Templates that took longest to instantiate:
[...]
4676 ms: QList<QPoint>::operator[] (261 times, avg 17 ms)
Added 'inline' to the definition of the setPoint(int, QPoint)
overload, since MinGW used to complain about it missing.
Task-number: QTBUG-97601
Pick-to: 6.3
Change-Id: Ie6f67da7ef39a16c98a7451d37b6d96531656392
Reviewed-by: Fabian Kosmale <fabian.kosmale@qt.io>
One of them has managed to percolate up to the top of the Clang
-ftime-trace list of expensive template instantiations when building
libQt6Gui.so with -pch:
**** Templates that took longest to instantiate:
7882 ms: std::is_trivially_destructible<QPropertyBindingSourceLocation> (135 times, avg 58 ms)
The checks aren't really necessary, because the compiler would
complain about the union's deleted dtor if any of the members were not
trivially destructible. Keep it around, though, but in the .cpp file.
Task-number: QTBUG-97601
Pick-to: 6.3 6.2
Change-Id: I74a513a907735bde298e0bd9557d10abbcee5c91
Reviewed-by: Fabian Kosmale <fabian.kosmale@qt.io>
If the SIMD code has already determined that the byte content differs,
we don't need to actually subtract the bytes we loaded from memory in
vector operations to return a sorting result.
Change-Id: I0e5f6bec596a4a78bd3bfffd16c908b2902e1b1b
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
So we don't accidentally make modifications to one and not the other.
Change-Id: I0e5f6bec596a4a78bd3bfffd16c94f1025aea521
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
If the lengths aren't equal, the strings can't be equal either, so we
can skip the entire comparison. Some of the front-end functions that
call these entry points already check for this, actually.
Change-Id: Ib42b3adc93bf4d43bd55fffd16c8ceb9594512f2
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
The compareStrings() entry points take QStringView and QLatin1String,
which are both ordered [size, pointer], so match that in the ucstricmp()
parameters. This further reduces the prologue of the compareStrings()
functions before reaching the case-sensitive comparison.
There's no need to do the same for the case-sensitive functions because
they're getting inlined.
Change-Id: I0e5f6bec596a4a78bd3bfffd16c8ffc980c8af0c
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
Case-insensitive comparisons are not common, but both GCC and Clang
inlined the ucstricmp() functions into QtPrivate::compareStrings(), with
the side-effect that a lot of unnecessary setup code saving CPU
registers was executed in the prologue of those functions.
After this, Clang 13 emits both compareString() functions without any
push/pop to save registers on x86-64; GCC 11 still emits a few, but
fewer than before (it's emitting some unnecessary overhead for the
loops).
Change-Id: I0e5f6bec596a4a78bd3bfffd16c8fc2c0be9165f
Reviewed-by: Marc Mutz <marc.mutz@qt.io>
Marc Mutz
Sona Kurazyan
Volker Hilsheimer
Ulf Hermann
Thiago Macieira
Allan Sandfeld Jensen
Laszlo Agocs
Noah Davis
Timur Pocheptsov
Alexey Edelev
Samuli Piippo
Morten Johan Sørvig
Ivan Tkachenko
Kai Köhne