It turns out that in QML it is not unusual that during early binding
evaluations due to the undefined order, the evaluation tries to look up
properties in objects that aren't initialized yet and thus exceptions are
thrown. Eeach thrown exception saves a stack trace, which is expensive to
generate when using the JIT, as it does full stack unwinding.
This patch implements a more light-weight approach by storing the instruction
pointer in the context before leaving JIT generated code.
Change-Id: I95e1cfd01179247dfc2c1df949828f474a23161b
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
createGlyphNode() and createNativeGlyphNode() kept calling each other
on GLES whenever QML_DISABLE_DISTANCEFIELD was set.
Change-Id: Ic1c2cfe0c4c7301f82cbbcce1cb512bd515b52ef
Reviewed-by: Gunnar Sletta <gunnar.sletta@digia.com>
All the qml and image files are embedded in the QtQuick.Dialogs plugin.
If DefaultFileDialog.qml is missing from the installation directory,
loading from resources will be enabled. Otherwise, the files will be
loaded from the local path.
This is analogous to change b2e5d1acb1aca93157a6d4d0a026153134f9ad01
in Qt Quick Controls: it reduces the number of files that need to be
installed, but also preserves ease of debugging whenever the QML files
are separately installed.
Task-number: QTBUG-31565
Change-Id: Idbe6be5d818eb6a25367f2053ea52bc7ac1485bc
Reviewed-by: Caroline Chao <caroline.chao@digia.com>
Without this change I'm getting this backtrace
3 0x4025b9f2 in QQmlIncubatorPrivate::incubate (this=0x18daa78, i=...) at qml/qqmlincubator.cpp:273
4 0x4025c1c2 in QQmlIncubator::forceCompletion (this=0x1527360) at qml/qqmlincubator.cpp:592
5 0x404e1626 in QQuickVisualDataModelPrivate::object (this=this@entry=0x13909f8, group=QQuickListCompositor::Default, index=index@entry=1, asynchronous=asynchronous@entry=false) at items/qquickvisualdatamodel.cpp:900
6 0x404e1f7e in QQuickVisualDataModel::item (this=<optimized out>, index=1, asynchronous=<optimized out>) at items/qquickvisualdatamodel.cpp:968
Note: This is with patched 5.0.x, change QQuickVisualDataModel to QQmlDelegateModel for >= 5.1
and line numbers may be a bit off
What is happening:
QQmlIncubator::forceCompletion is doing
while (Loading == status()) {
while (Loading == status() && !d->waitingFor.isEmpty())
static_cast<QQmlIncubatorPrivate *>(d->waitingFor.first())->incubate(i);
if (Loading == status())
d->incubate(i);
}
Calling QQmlIncubatorPrivate::incubate on the first item of d->waitingFor
Then, that item is getting to QQmlIncubatorPrivate::incubate and happens that
progress is QQmlIncubatorPrivate::Completed and waitingFor is not empty,
so the only thing that QQmlIncubatorPrivate::incubate ends up doing is
calling a few calls over vmeGuard and returning, that way the inner
waitingFor items never finishe incubating and you end up in an inifite loop inside
while (Loading == status() && !d->waitingFor.isEmpty())
static_cast<QQmlIncubatorPrivate *>(d->waitingFor.first())->incubate(i);
This patch basically replaces this loop with a loop that does
while (QQmlIncubator::Loading == status && !waitingFor.isEmpty())
static_cast<QQmlIncubatorPrivate *>(waitingFor.first())->forceCompletion(i);
This way we make sure we incubate the waitingFor items of our waitingFor items
Change-Id: I4298efc7ba9d8af624bb138e64b92a40ed4c4dc9
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Set QReadWriteLocker to recursive. This is needed e.g. to allow
_q_changeServiceState to get a read lock when called from inside
receiveMessage.
Change-Id: I287a7c7f44e94005c0458825f8f6d1877ee914dd
Reviewed-by: Aurindam Jana <aurindam.jana@digia.com>
Keep conservative GC as a fallback for testing
Enable all tests again that were skipped due to
GC issues.
Change-Id: I8e0fa728207bdd39a96d0acf95e27841157d8402
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
When an exception is thrown and we traverse a frame that requires only cleanup
(i.e. call QV4::Scope::~Scope), control is first transferred to the generated
cleanup code. Afterwards the unwinding is resumed (on ARM) by calling
__cxa_end_cleanup, which resides in libsupc++ (libgnustl_shared).
__cxa_end_cleanup first calls __gnu_end_cleanup and then resumes the process of
stack unwinding by calling _Unwind_Resume (per specification). Given the
linking situation on Android, this will end up calling _Unwind_Resume inside
libgnustl_shared, which sidesteps our statically linked copy of the unwind
code in QtQml (libgcc.a). Therefore any further unwinding through JIT generated
frames will fail.
This patch introduces the same EABI symbol exported in libQt5Qml, which will
direct control to the correct JIT aware unwinder.
This relies on https://codereview.qt-project.org/#change,68206 in order to
ensure that libsupc++.a is gone from all link lines (not needed) and that
gnustl_shared is after libQt5Qml.
Task-Number: QTBUG-33892
Change-Id: I6ed691db3ceb287475a70b7af8cf3cd7b4ddfdd6
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@digia.com>
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Change a9cf828559 refactored
QQmlImportDatabase::importPlugin() to be used for both dynamic
and static plugin loading. In the process, the scope of a
QWriteLocker protecting a call to registerTypes ended up to wide.
That caused a deadlock to occur for some static qml applications
since the lock remained active during a subsequent call to
initializeEngine.
So narrow the the scope down to be exactly as it wore before the
change. This will remove the deadlock.
Change-Id: Ibb15c953c0f693fe75dab24f0093c3bddb3f0cbb
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@digia.com>
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Properly protect them through Scoped values.
Change-Id: I5a0a1d5580d55ecff493419baa8959751a65f1d3
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
It complicated static linking for no good reason.
Change-Id: I8eb1548b799128bfab9451963cb80c72239b961c
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
"QML Application Developer Resources" is too long and the visible text
is usually shortened to "QML Applications".
The article in qt5/qtdoc changed title to "QML Applications".
Change-Id: I301f3b4659bd87631269b912bcc36f6f00fdeb2c
Reviewed-by: Mitch Curtis <mitch.curtis@digia.com>
Reviewed-by: Jerome Pasion <jerome.pasion@digia.com>
This fixes possible bugs and crashes where the incubator
could get deleted through GC while constructing the
component.
Change-Id: Ibe0c5d4e172f0b5505ace0c3ea0369169b8b48a5
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
QQmlComponent::statusChanged() marked the incubated object
as destructable when it was done loading. This implied
that any gc call afterwards could clean it up. To fix, push
the object on the GC stack so it lives until the method returns.
Also renamed the WrapperIncubator to QQmlComponentIncubator.
Change-Id: I5a8f478a1fd65ea73ddff310392219709a935a70
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
m_dialogWindow only needs to be set if the dialog has its own window,
because otherwise setters such as QQuickAbstractDialog::setHeight
(introduced in 37492efee0)
will result in reshaping the main application window instead of the
dialog. Also the window decoration's translucent "veil" should
expand to fit the whole application.
Change-Id: I16fdb48c54dbc5e758dfec1f18063579ac2bbb05
Reviewed-by: Liang Qi <liang.qi@digia.com>
Get the formal and local names of function object from the
compilation unit to avoid creating another set of strings.
Use a ScopedFunctionObject in eval()
Change-Id: I6693aec2e88818df9c160b3780db12d8df79c2fe
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Never use multiple inheritance with Managed subclasses,
as this can easily mess up garbage collection. In this
case the vtable from the QQmlIncubator would be added
before the start of the Managed pointer. That would
not work correctly for the memory manager that casts
void pointers to Managed pointers.
Change-Id: I1c1ebc6c44bd9cb77eea49738e86ce3374c7ef80
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Inline calls to get_element if the base is an
object with a simple array structure, and the index
is an integer number.
Implemented for 64bit only for now, saves ~25% on crypto.js
Change-Id: I3e34a6409169d90d3937f62264707d52a6c2f9f7
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Holes in arrays should be represented by an empty
value, not by creating/setting array attributes.
Reason is that the creation is irreversable, and slows
down execution. This speeds up crypto.js by 10%
Change-Id: I2e5472575479a5f2dbe53f59ecb8ed3aeab1be7a
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
The global scope is compiled as EvalCode. Because of this
we were never using global lookups anymore, slowing down
the v8 test suite by ~20%.
Change-Id: I6c47ccf90f4d9ec3bf531bbb689d3f1511f69968
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
If the truncation from double returned out of bounds,
the code was not generating the right call to the fallback
method on 64 bit.
Change-Id: I6a126dd013c5b6373762beef0c1459fae1c26ef2
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Registers containing outgoing parameters for calls are saved as Value
(so, 8 bytes long) instead of native register size. For 64 bit this is
the same, but not for 32 bits.
Change-Id: If1d55e9d552a301782816c9228390dc5611c6d00
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This is a special-case instruction, where the immediate needs to be
multiplied by 4.
Change-Id: I86e5ab9d39d65b8eab99fae859969896c6e5630c
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
The window may be deleted before any other window is make current,
and that would lead to memory corruption with Mesa's EGL.
Change-Id: I414b972fd517f60c28d194fa059bf7871e422872
Reviewed-by: Gunnar Sletta <gunnar.sletta@digia.com>
Match the algorithm used by QML, look for the most
specific version of an import first.
Change-Id: Ibf1370af227c8154f657cc2d8a1c1d1ae28d2f39
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@digia.com>
Needed for static builds. qmlimportscanner will
read the "plugin" and "classname" entries.
Change-Id: I31939451366ad3e771d516ac426525c8bcdba57d
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@digia.com>
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
The unsigned right shift operations can use a signed int as input, unifying
the handling with the other shift operations. The only difference now is
in the type of the return value.
Change-Id: Ia9b83568951d1c1c77322d07cd734e2c37d34573
Reviewed-by: Erik Verbruggen <erik.verbruggen@digia.com>
Now that shifts can take a signed int32, reflect it in the check if a
binop will generate a call.
Change-Id: I3cab436bace31cdda327cf6132aa873b6c5456b1
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Handle them similar to function declarations, except that we need to synthesize
the expression into a function declaration that includes the signal parameter
names. This is done quite similar to the code path in the new compiler.
Change-Id: I751081f7f1052692da6e2ed60c7f5c017372d829
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
...instead of extracting the function body as a string and compiling it in the
GUI thread.
Change-Id: I3c3108f6e35464b5581a2d8b5799e7285858ce4d
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
This is done by re-using the JS code generator from the new compiler. A few bugs were
fixed on the way:
* The index into the compiledData->runtimeFunctions array is not the same as the function
index when they are collected (from the AST), as for example binding expressions may create
extra V4IR::Function objects that break the 1:1 mapping. Therefore the JS code gen will return
a mapping from incoming function index to V4IR::Module::Function (and thus runtimeFunction)
* Binding expressions in the old backend get usually unpacked from their ExpressionStatement node.
The reference to that node is lost, and instead of trying to preserve it, we simply synthesize it
again. This won't be necessary anymore with the new compiler in the future.
* Commit 1c29d63d60 ensured to always look up locals by name, and so
we have to do the same when initializing the closures of nested functions inside binding expressions
(in qv4codegen.cpp)
* Had to change the Qml debugger service auto-test, which does toString() on a function that is now compiled.
Even if we implemented FunctionPrototype::toString() to do what v8 does by extracting the string from the
file, it wouldn't help in this test, because it feeds the input from a string instead of a file.
* In tst_parserstress we now end up compiling all JS code, which previously was only parsed. This triggers
some bugs in the SSA handling. Those tests are skipped and tracked in QTBUG-34047
Change-Id: I44df51085510da0fd3d99eb5f1c7d4d17bcffdcf
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Testcase: (covered in parserstress)
function foo(x) {
var x = 42;
}
In variableDeclaration, the lookup for "x" with findMember will return -1, and
instead code for checking against arguments using indexOfArgument is needed.
The easiest fix is to simply use identifier(), which handles this accordingly.
Change-Id: I6a738d6196d4bff1fc987f111aebbaa83ed8f88f
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
In order to determine the type of lookup we need (name or directly in
environment members), we used Codegen::_mode, which is set to the currently
suitable mode depending on the function (parameter to defineFunction). However
that's not quite correct, the look-up mode depends on the function itself, not
where it was called from. This patch corrects that by moving the compilation
mode into the Environment itself.
This is needed by follow-up patches.
Additionally the "bool deletable" parameter to the builtin_declare_vars was
always set to false, because it used mode instead of _mode, which was never set
to Eval or QmlBinding. This will be cleaned up in a future patch.
Change-Id: I878f187945e5de091689ab5d70a0f33eb5a9e38f
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Testcase (part of parserstress in tests/auto/qml):
outer: {
do {
inner: {}
} while (false)
}
The labelled statement visitor, when hitting the outter label, would call
enterLoop(), which sets _labelledStatement back to zero. That then gets added
to the Loop object the do-while loop creates, and the duplicate labelled
statement check then for inner would unconditionally dereference
loop->labelledStatement.
In all other places where we access loop->labelledStatement we have a null
pointer check, so let's have one here as well.
Change-Id: I9d5925a2abf4db691c49c0cdec3550938ee02efa
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
This can (and does crash) when a gc gets triggered during the
linking stage of a compilation unit.
Change-Id: I06f1299adab68ff8e0a4755d02e246870797e7f2
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Android apps require packaging to be usable, so they need to be
built individually, not as part of the subdirs build. The build would
fail during make install due to this.
Task-number: QTBUG-34023
Change-Id: Id68f5fbdd4eeb6f130d20a0f4b831c6c24639471
Reviewed-by: Paul Olav Tvete <paul.tvete@digia.com>
Also accept a boolean primitive as input to
Boolean.prototype.valueOf()
Change-Id: I5b94d8d65b86e26860b9844eb4bf823577c8e924
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This makes pretty much all test cases pass with exact
garbage collection.
Change-Id: Ia874e3c17c3984afb7cfe370f9bd3ad8fe46699a
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Add 64 bit code patch and avoid some duplicated
calculation in 32 bit mode
Change-Id: I0e111de8ac4e733aa8802c49b4b15d785688d7ea
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Simon Hausmann
Laszlo Agocs
Shawn Rutledge
Albert Astals Cid
Kai Koehne
Lars Knoll
Richard Moe Gustavsen
Tor Arne Vestbø
Jerome Pasion
Mitch Curtis
Erik Verbruggen
Giulio Camuffo
Morten Johan Sørvig
Jan Arve Saether
Eskil Abrahamsen Blomfeldt