path: root/src/corelib/doc/src/qt6-changes.qdoc

// Copyright (C) 2020 The Qt Company Ltd.
// SPDX-License-Identifier: LicenseRef-Qt-Commercial OR GFDL-1.3-no-invariants-only

/*!
    \page qtcore-changes-qt6.html
    \title Changes to Qt Core
    \ingroup changes-qt-5-to-6
    \brief Changes to containers, strings, serialization and I/O classes.

    Qt 6 is a result of the conscious effort to make the framework more
    efficient and easy to use.

    We try to maintain binary and source compatibility for all the public
    APIs in each release. But some changes were inevitable in an effort to
    make Qt a better framework.

    In this topic we summarize those changes in Qt Core, and provide guidance
    to handle them.

    \section1 Container Classes

    \section2 QHash, QMultiHash, QSet

    \section3 qHash() Signature

    For custom types, QHash and QMultiHash rely on you providing
    a \l{qHash}{custom qHash() function}
    in the same namespace. In Qt 4 and Qt 5, the return
    value and optional second argument of a \c qHash function
    was of type \c uint. In Qt 6, it is \c size_t.

    That is, you need to change

    \code
    uint qHash(MyType x, uint seed);
    \endcode

    to

    \code
    size_t qHash(MyType x, size_t seed);
    \endcode

    This allows QHash, QMultiHash and QSet to hold more than 2^32 items on
    64 bit platforms.

    \section3 Stability of References

    The implementation of QHash, QMultiHash and QSet in Qt 6 got changed from
    a node based approach to a two stage lookup table. This design allows
    to keep the memory overhead of a hash instance very small, while
    at the same time giving good performance.

    One behavioral change to note is that the new implementation
    will not provide stable references to elements in the hash when the
    table needs to grow, or when entries are removed. Applications that
    rely on such stability might now run into undefined behavior.

    \section3 Removal of QHash::insertMulti

    In Qt 5, QHash could be used to create multi-valued hashes by using
    QHash::insertMulti, and QMultiHash was deriving vom QHash.

    In Qt 6, both types and use cases are distinct, and QHash::insertMulti
    got removed.

    \section2 QVector, QList

    Prior to Qt 6, QVector and QList were separate classes. In Qt 6, they are
    unified: Qt 5 QList implementation is gone and both classes use updated
    QVector implementation instead. QList is the class with the actual
    implementation and QVector is an alias (typedef) to QList.

    QList's fromVector() and toVector(), and QVector's fromList() and toList(),
    no longer involve data copying in Qt 6. They now return the object that they
    were called for.

    \section3 API Changes

    QList's (and hence QVector's) size type is changed from \c int to \c
    qsizetype. Together with the size type, all relevant methods' signatures are
    updated to use \c qsizetype. This allows QList to hold more than 2^31 items
    on 64 bit platforms.

    When upgrading the code base to Qt 6, this API change would most likely
    result in compiler warnings about narrowing type conversions. Having the
    following example code:

    \code
    void myFunction(QList<MyType> &data) {
        int size = data.size();
        // ...
        const int pos = getInsertPosition(size);
        data.insert(pos, MyType());
        // ...
    }
    \endcode

    you would need to update it to use either \c qsizetype or an auto keyword:

    \code
    void myFunction(QList<MyType> &data) {
        auto size = data.size();
        // ...
        const auto pos = getInsertPosition(size);
        data.insert(pos, MyType());
        // ...
    }
    \endcode

    Alternatively, you may use type casting and cast everything to \c int or to
    \c qsizetype.

    \note If you want to build against both Qt 5 and Qt 6, the auto keyword is a
    good solution to cover signature differences between the versions.

    \section3 Memory Layout

    QList received multiple changes related to the memory layout in Qt 6.

    In Qt 5, \c{sizeof(QList<T>)} was equal to a size of a pointer. Now, the
    extra pointer indirection is removed and QList data members are directly
    stored in the object. By default, expect \c{sizeof(QList<T>)} to be equal to
    the size of 3 pointers.

    At the same time, memory layout of the elements is also updated. QList now
    always stores its elements directly in the allocated memory region as
    opposed to Qt 5, where certain objects were separately allocated on the heap
    and pointers to the objects were placed into the QList instead.

    Note that the latter, in particular, affects large objects. To have Qt 5
    behavior, you could wrap your objects into smart pointers and store these
    smart pointers in QList directly. In this case, the type of your QList would
    be \c{QList<MySmartPointer<MyLargeObject>>} as opposed to
    \c{QList<MyLargeObject>} in Qt 5.

    \section3 Stability of References

    There are several changes made to the QVector/QList implementation. The
    QVector related one is: insertion at the beginning is optimized (similarly
    to QList in Qt 5). The QList related one is: memory layout for the elements
    is simplified.

    \important These changes impact the stability of references. In Qt 6, you
    should consider any size or capacity modifying method to invalidate all
    references, even when QList is not \l{Implicit Sharing}{implicitly shared}.
    Exceptions to this rule are documented explicitly.

    Applications that rely on certain reference stability might run into
    undefined behavior when upgraded to use Qt 6. You should pay extra attention
    to cases where QVector or QList with a non C-compatible array layout were
    used originally.

    \section1 View classes in Qt6

    \section2 General Overview

    There are several new \c View classes coming with Qt6. There is the already
    existing \l QStringView, now accompanied by \l QByteArrayView and followed
    by a specialized \l QUtf8StringView and a more universal \l QAnyStringView.

    \section2 Introduction to view classes on the example of QStringView

    The QStringView class provides a unified view on UTF-16 strings with a
    read-only subset of the QString API. Unlike QString, which keeps its own
    copy of the string (possibly ref-counted), QStringView provides a view of
    a string that is stored elsewhere.

    \code
        char hello[]{ "Hello." };   // narrow multi-byte string literal
        QString str{hello};         // needs to make a copy of the string literal
        QString strToStr(str);      // atomic increment involved to not create a copy of hello again

        // The above code can be re-written to avoid copying and atomic increment.

        QStringView view{ u"Hello." };  // view to UTF-16 encoded string literal
        QStringView viewToView{ view }; // view of the same UTF-16 encoded string literal
    \endcode

    The string \c "Hello." is stored in the binary and is not allocated at
    run-time. \c view is only a view onto the string \c "Hello.", therefore
    no copy has to be created. When we copy a QStringView, the \c viewToView
    observes the same string as the copied-from \c view is observing. This
    means that \c viewToView does not need to create a copy or an atomic
    increment. They are views onto the existing string \c "Hello.".

    \section2 Views as function argument

    Views should be passed by value, not by reference-to-const.

    \code
        void myfun1(QStringView sv);        // preferred
        void myfun2(const QStringView &sv); // compiles and works, but slower
    \endcode

    \section2 View manipulation functions

    QStringView supports functions that let us manipulate the view of the
    string. This allows us to change the view without creating a partial copy
    of the viewed string.

    \code
        QString pineapple = "Pineapple";
        QString pine = pineapple.left(4);

        // The above code can be re-written to avoid creating a partial copy.

        QStringView pineappleView{ pineapple };
        QStringView pineView = pineappleView.left(4);
    \endcode

    \section2 Non null-terminated strings and strings containing \c {'\0'}

    QStringView supports both null-terminated and non null-terminated strings.
    The difference comes from the way you initialize the QStringView:

    \code
        QChar aToE[]{ 'a', 'b', 'c', 'd', 'e' };

        QStringView nonNull{ aToE, std::size(aToE) }; // with length given
        QStringView nonNull{ aToE }; // automatically determines the length

        QChar fToJ[]{ 'f', 'g', 'h', '\0', 'j' };

        // uses given length, doesn't search for '\0', so '\0' at position 3
        // is considered to be a part of the string similarly to 'h' and 'j
        QStringView nonNull{ fToJ, std::size(fToJ) };
        QStringView part{ fToJ }; //stops on the first encounter of '\0'
    \endcode

    \section2 Ownership model of views

    As \c views do not own the memory they reference, care must be taken to
    ensure that the referenced data (for example, owned by a \l QString)
    outlives the \c view on all code paths.

    \badcode
        QStringView sayHello()
        {
            QString hello("Hello.");
            return QStringView{ hello }; // hello gets out of scope and destroyed
        }

        void main()
        {
            QStringView hello{ sayHello() };
            qDebug() << hello; // undefined behavior
        }
    \endcode

    \section2 Converting an QStringView to QString

    QStringView will not implicitly or explicitly convert to a QString, but can
    create a deep copy of its data:

    \code
        void print(const QString &s) { qDebug() << s; }

        void main()
        {
            QStringView string{ u"string"};

            // print(string); // invalid, no implicit conversion
            // QString str{ string }; // invalid, no explicit conversion

            print(string.toString());
            QString str = string.toString(); // create QString from view
        }
    \endcode

    \section2 Important notes

    By leveraging the new view classes, one can achieve a lot of performance
    boost in many use cases. However, it is important to know that there might
    be some caveats. Therefore it is important to remember:

    \list
        \li Views should be passed by value, not by reference-to-const.
        \li Constructing a view with a negative length is undefined behavior.
        \li Care must be taken to ensure that the referenced data (for example,
            owned by a \l QString) outlives the view on all code paths.
    \endlist

    \section1 String related classes

    \section2 The QStringView class

    Starting with Qt6 it is generally recommended to use \l QStringView over
    \c QStringRef. \l QStringView references a contiguous portion of a UTF-16
    string it does not own. It acts as an interface type to all kinds of UTF-16
    strings, without the need to construct a \l QString first. The \l QStringView
    class exposes almost all read-only methods of \l QString and the previously
    existing \c QStringRef class.

    \note Care must be taken to ensure that the referenced string data (for
    example, owned by a \l QString) outlives the \l QStringView on all code paths.

    \note If a \l QStringView wraps a \l QString, care needs to be taken since
    unlike \c QStringRef \l QStringView will not update the internal data pointer
    once the \l QString data relocates.

    \code
        QString string = ...;
        QStringView view{string};

        // Appending something very long might cause a relocation and will
        // ultimately result in a garbled QStringView.
        string += ...;
    \endcode

    \section2 The QStringRef class

    In Qt6 \l QStringRef got removed from Qt Core. To ease porting of existing
    applications without touching the whole code-base, the \c QStringRef class
    did not vanish completely and instead it got moved into the Qt5Compat module.
    If you want to use \c QStringRef further, see \l {Using the Qt5Compat module}.

    Unfortunately, some methods exposed by \l QString returning a \c QStringRef,
    could not be moved to Qt5Compat. Therefore some manual porting may be
    needed. If your code uses one or more of the following functions you need to
    port them to use \l QStringView or \l QStringTokenizer. It is also
    recommended to use \l {QStringView::tokenize} over \l {QStringView::split}
    for performance critical code.

    Change code using \c QStringRef:
    \code
        QString string = ...;
        QStringRef left = string.leftRef(n);
        QStringRef mid = string.midRef(n);
        QStringRef right = string.rightRef(n);

        QString value = ...;
        const QVector<QStringRef> refs = string.splitRef(' ');
        if (refs.contains(value))
            return true;
    \endcode

    to:

    \code
        QString string = ...;
        QStringView left = QStringView{string}.left(n);
        QStringView mid = QStringView{string}.mid(n);
        QStringView right = QStringView{string}.right(n);

        QString value = ...;
        const QList<QStringView> refs = QStringView{string}.split(u' ');
        if (refs.contains(QStringView{value}))
            return true;
        // or
        const auto refs = QStringView{string}.tokenize(u' ');
        for (auto ref : refs) {
            if (ref == value)
                return true;
        }
    \endcode

    \section1 QMutex and Related Classes

    In Qt 6, QRecursiveMutex does not inherit from QMutex anymore. This change was
    done to improve the performance of both QMutex and QRecursiveMutex.

    Due to those changes, the QMutex::RecursionMode enum has been removed, and
    QMutexLocker is now a templated class that can operate on both QMutex and
    QRecursiveMutex.

    \section1 QFuture and Related Classes

    \section2 The QFuture class

    To avoid unintended usage of QFuture, there were some changes to
    QFuture API in Qt 6, which may introduce source compatibility breaks.

    \section3 Implicit conversions between QFuture and other types

    Conversion of \c QFuture<T> to \c T has been disabled. The casting
    operator was calling QFuture::result(), which may lead to undefined
    behavior if the user has moved the results from QFuture via
    QFuture::takeResult() before trying to do the conversion. Use
    QFuture::result() or QFuture::takeResult() methods explicitly,
    where you need to convert \c QFuture<T> to \c T.

    The implicit conversion from \c QFuture<T> to \c QFuture<void> has
    been also disabled. If you really intend to do the conversion, use the
    explicit \c {QFuture<void>(const QFuture<T> &)} constructor:

    \code
    QFuture<int> future = ...
    QFuture<void> voidFuture = QFuture<void>(future);
    \endcode

    \section3 Equality operators

    The equality operators of QFuture have been removed. They were comparing
    the underlying d-pointers instead of comparing the results, which is not
    what users might expect. If you need to compare QFuture objects, use
    \c QFuture::result() or \c QFuture::takeResult() methods. For example:

    \code
    QFuture<int> future1 = ...;
    QFuture<int> future2 = ...;
    if (future1.result() == future2.result())
        // ...
    \endcode

    \section2 Behavioral Changes to QFuture and QFutureWatcher

    In Qt 6, there were some improvements to QFuture and QFutureWatcher which
    caused the following behavioral changes:

    \list

    \li After pausing QFuture or QFutureWatcher (by calling \c pause() or
    \c setPaused(true)), QFutureWatcher will not immediately stop delivering
    progress and result ready signals. At the moment of pausing there may be
    still computations that are in progress and cannot be stopped. Signals
    for such computations may be still delivered after pause, instead of being
    postponed and reported only after next resume. To get notified when pause
    actually took effect, QFutureWatcher::suspended() signal can be used. In
    addition, there are new \c isSuspending() and \c isSuspended() methods,
    to check if the QFuture is in the process of suspending or it's already in
    the suspended state. Note that for consistency reasons, for both QFuture
    and QFutureWatcher the pause-related APIs were deprecated and replaced by
    similar methods having "suspend" in the name instead.

    \li QFuture::waitForFinished() will now wait until QFuture is actually in
    the finished state, instead of exiting as soon as it is not in the running
    state. This prevents \c waitForFinished() from exiting immediately, if at
    the moment of calling it the future is not started yet. The same applies to
    QFutureWatcher::waitForFinished(). This change won't affect the behavior of
    code that was using QFuture with QtConcurrent. Only the code that was using
    it with the undocumented \c QFutureInterface may be affected.

    \li QFutureWatcher::isFinished() now reflects the finished-state of the
    QFuture rather than returning false until QFutureWatcher::finished() has
    been emitted.

    \endlist

    \section2 The QPromise class

    In Qt 6, the new QPromise class should be used instead of unofficial
    QFutureInterface as a "setter" counterpart of QFuture.

    \section1 IO Classes

    \section2 The QProcess class

    In Qt 6, the QProcess::start() overload that interprets a single command string
    by splitting it into program name and arguments is renamed to QProcess::startCommand().
    However, a QProcess::start() overload that takes a single string, as well as a QStringList
    for arguments exists. Since the QStringList parameter defaults to the empty list, existing
    code only passing a string will still compile, but will fail to execute the process if it
    is a complete command string that includes arguments.

    Qt 5.15 introduced deprecation warnings for the respective overload to make it easy to
    discover and update existing code:

    \code
        QProcess process;

        // compiles with warnings in 5.15, compiles but fails with Qt 6
        process.start("dir \"My Documents\"");

        // works with both Qt 5 and Qt 6; also see QProcess::splitCommand()
        process.start("dir", QStringList({"My Documents"});

        // works with Qt 6
        process.startCommand("dir \"My Documents\"");
    \endcode

    QProcess::pid() and the Q_PID type have been removed; use QProcess::processId() instead to
    get the native process identifier. Code using native Win32 APIs to access the data in the
    Q_PID as a Win32 \c{PROCESS_INFORMATION} struct is no longer supported.

    \section1 Meta-Type system

    \section2 The QVariant class

    \c QVariant has been rewritten to use \c QMetaType for all of its operations. This implies
    behavior changes in a few methods:

    \list

    \li \c QVariant::isNull() now only returns \c true if the \c QVariant is empty or contains a
    \c nullptr. In Qt 5, it also returned true for classes in qtbase which had an \c isNull method
    themselves if that one returned true. Code relying on the old behavior needs to check whether
    the contained value returns isNull – however such code is unlikely to occur in practice, as
    \c isNull() is rarely the property one is interested in (compare \c QString::isEmpty() / \c isNull()
    and \c QTime::isValid / \c isNull).

    \li \c QVariant::operator== uses \c QMetaType::equals in Qt 6. Therefore, some graphical
    types like \c QPixmap, \c QImage and \c QIcon will never compare equal. Moreover, floating
    point numbers stored in \c QVariant are no longer compared with \c qFuzzyCompare, but instead
    use exact comparisons.

    \endlist

    Furthermore, QVariant::operator<, QVariant::operator<=, QVariant::operator> and
    QVariant::operator>= were removed, because different variants are not always orderable. This
    also means that QVariant cannot be used anymore as a key in a QMap.

    \section2 The QMetaType class

    In Qt 6, registration of comparators, and QDebug and QDataStream streaming operators is
    done automatically. Consequently, \c QMetaType::registerEqualsComparator(),
    \c QMetaType::registerComparators(), \c qRegisterMetaTypeStreamOperators() and
    \c QMetaType::registerDebugStreamOperator() do no longer exist. Calls to those methods
    have to be removed when porting to Qt 6.

    \section2 Type registration

    Types used in \c Q_PROPERTY have their meta-type stored in the class'  \c QMetaObject. This
    requires the types to be complete when moc sees them, which can lead to compilation errors
    in code that worked in Qt 5. There are three ways to fix this issue:

    \list

    \li Include the header which defines the type.
    \li Instead of using an include, use the \c Q_MOC_INCLUDE macro. This helps if including the header
        would cause a cyclic dependency, or when it would slow down compilation.
    \li If the header is present in the cpp file which implements the class, it is also possible to include
        the moc generated file there.

    \endlist

    \section1 Regular expression classes

    \section2 The QRegularExpression class

    In Qt 6, the \c QRegExp type has been retired to the Qt5Compat module
    and all Qt APIs using it have been removed from other modules.
    Client code which used it can be ported to use \l QRegularExpression
    in its place. As \l QRegularExpression is present already in Qt 5,
    this can be done and tested before migration to Qt 6.

    \include corelib/port-from-qregexp.qdocinc porting-to-qregularexpression

    \section2 The QRegExp class

    In Qt6 \l QRegExp got removed from Qt Core. If your application cannot be
    ported right now, \c QRegExp still exists in Qt5Compat to keep these
    code-bases working. If you want to use \c QRegExp further, see
    \l {Using the Qt5Compat module}.

    \section1 QEvent and subclasses

    The QEvent class defined a copy constructor and an assignment operator,
    in spite of being a polymorphic class. Copying classes with virtual methods
    can result in slicing when assigning objects from different classes to each
    other. Since copying and assigning often happens implicilty, this could
    lead to hard-to-debug problems.

    In Qt 6, the copy constructor and assignment operator for QEvent subclasses
    have been made protected to prevent implicit copying. If you need to copy
    events, use the \l{QEvent::}{clone} method, which will return a heap-allocated
    copy of the QEvent object. Make sure you delete the clone, perhaps using
    std::unique_ptr, unless you post it (in which case Qt will delete it once it
    has been delivered).

    In your QEvent subclasses, override clone(), and declare the protected and
    default-implemented copy constructor and assignment operator like this:

    \code
    class MyEvent : public QEvent
    {
    public:
        // ...

        MyEvent *clone() const override { return new MyEvent(*this); }

    protected:
        MyEvent(const MyEvent &other) = default;
        MyEvent &operator=(const MyEvent &other) = default;
        MyEvent(MyEvent &&) = delete;
        MyEvent &operator=(MyEvent &&) = delete;
        // member data
    };
    \endcode

    Note that if your MyEvent class allocates memory (e.g. through a
    pointer-to-implementation pattern), then you will have to implement
    custom copy semantics.

    \section1 Serialization classes

    In Qt 6, QJsonDocument methods for converting it to/from Qt's legacy
    JSON binary format were removed in favor of the standardized CBOR format.
    Qt JSON types can be converted to Qt CBOR types, which can in turn be
    serialized into the CBOR binary format and vice versa. See, for example,
    \l QCborValue::fromJsonValue() and \l QCborValue::toJsonValue().

    If you still need to use the binary JSON format, you can use the
    replacements provided in the Qt5Compat module. They can be found in the
    \l QBinaryJson namespace. See \l {Using the Qt5Compat module} to find out
    how to use the module in your application.

    \section1 Other classes

    In Qt 5, QCoreApplication::quit() was equivalent to calling
    \l{QCoreApplication::exit()}. This just exited the main event loop.

    In Qt 6, the method will instead try to close all top-level windows by posting
    a close event. The windows are free to cancel the shutdown process by
    ignoring the event.

    Call QCoreApplication::exit() to keep the non-conditional behavior.

    QLibraryInfo::location() and QLibraryInfo::Location were deprecated due to inconsistent
    naming. Use the new API QLibraryInfo::path() and QLibraryInfo::LibraryPath instead.

    \section1 Qt State Machine Framework

    \l {Qt State Machine} was moved into the Qt SCXML module (soon to be renamed to Qt
    State Machines) and therefore it is no longer part of Qt Core. There were very few
    cross dependencies inside Qt Core which ultimately led to this decision.

    \section1 Using the Qt5Compat module

    To use the \l {Qt 5 Core Compatibility APIs}{Qt5Compat} module, you need
    to build with its headers in your include path and link against its library.
    If you are using \l qmake, add the following to your \c .pro file:

    \code
        QT += core5compat
    \endcode

    If you build your application or library using \l {Build with CMake}{cmake},
    add the following to your \c CMakeList.txt:
    \code
        PUBLIC_LIBRARIES
            Qt::Core5Compat
    \endcode
*/