Question

Many C++ classes in Qt use explicit and implicit data sharing
to maximize resource usage and minimize copying of data.

• Overview
• A QByteArray Example
• Explicit vs. Implicit Sharing
• Explicitly Shared Classes
• Implicitly Shared Classes
• QCString: implicit or explicit?

Overview

A shared class consists of a pointer to a shared data block that
contains a reference count and the data.

When a shared object is created, it sets the reference count to 1. The
reference count is incremented whenever a new object references the shared
data, and decremented when the object dereferences the shared data. The
shared data is deleted when the reference count becomes zero.

When dealing with shared objects, there are two ways of copying an object.
We usually speak about deep and shallow copies. A deep copy implies
duplicating an object. A shallow copy is a reference copy, we only copy a
pointer to a shared data block. Making a deep copy can be expensive in
terms of memory and CPU. Making a shallow copy is very fast, because it
only involves setting a pointer and incrementing the reference count.

Object assignment (with operator=) for implicitly and explicitly
shared objects is implemented as shallow copies. A deep copy can be
made by calling a copy() function.

The benefit of sharing is that a program does not need to duplicate data
when it is not required, which results in less memory usage and less
copying of data. Objects can easily be assigned, sent as function
arguments and returned from functions.

Now comes the distinction between explicit and implicit sharing.
Explicit sharing means that the programmer must be aware of the fact that
objects share common data. Implicit sharing means that the sharing
mechanism takes place behind the scenes and the programmer does not need
to worry about it.

A QByteArray Example

QByteArray is an example of a shared class that uses explicit sharing.
Example:

                          //        a=         b=         c=
    QByteArray a(3),b(2)  // 1)     {?,?,?}    {?,?}
    b[0] = 12; b[1] = 34; // 2)     {?,?,?}    {12,34}
    a = b;                // 3)     {12,34}    {12,34}
    a[1] = 56;            // 4)     {12,56}    {12,56}
    QByteArray c = a;     // 5)     {12,56}    {12,56}    {12,56}
    a.detach();           // 6)     {12,56}    {12,56}    {12,56}
    a[1] = 78;            // 7)     {12,78}    {12,56}    {12,56}
    b = a.copy();         // 8)     {12,78}    {12,78}    {12,56}
    a[1] = 90;            // 9)     {12,90}    {12,78}    {12,56}

The assignment a = b on line 3 throws away a's
original shared block (the reference count becomes zero), sets
a's shared block to point to b's shared block
and increments the reference count.

On line 4, the contents of a is modified. b is also modified,
because a and b refer the same data block. This is the difference
between explicit and implicit sharing (explained below).

The a object detaches from the common data on line 6. Detaching means
to copy the shared data to make sure that an object has its own private
data. Therefore, modifying a on line 7 will not affect b or c.

Finally, on line 8 we make a deep copy of a and assign it to b,
so that when a is modified on line 9, b remains unchanged.

Explicit vs. Implicit Sharing

Implicit sharing automatically detaches the object from a shared
block if the object is about to change and the reference count is
greater than one. Explicit sharing leaves this job to the
programmer. If an explicitly shared object is not detached, changing
the object will change all other objects that refer to the same data.

Implicit sharing optimizes memory usage and copying of data without this
side effect. So why didn't we implement implicit sharing for all
shared classes? The answer is that a class that allows direct access to
its internal data (for efficiency reasons), like QByteArray, cannot be
implicitly shared, because it can be changed without letting QByteArray
know.

An implicitly shared class has total control of its internal data. In any
member functions that modify its data, it automatically detaches
before modifying the data.

The QPen class, which uses implicit sharing, detaches from the shared data
in all member functions that change the internal data.

Code fragment:

    void QPen::setStyle( PenStyle s )
    {
        detach();        // detach from common data
        data->style = s; // set the style member
    }

    void QPen::detach()
    {
        if ( data->count != 1 ) // only if >1 reference
            *this = copy();
    }

This is clearly not possible for QByteArray, because the programmer can
do the following:

    QByteArray array( 10 );
    array.fill( 'a' );
    array[0] = 'f';        // will modify array
    array.data()[1] = 'i'; // will modify array

If we monitor changes in a QByteArray, the QByteArray class would
become unacceptably slow.

Explicitly Shared Classes

All classes that are instances of the QMemArray template class are explicitly
shared:

• QBitArray
• QPointArray
• QByteArray
• Any other instantiation of QMemArray<type>

These classes have a detach() function that can be called if you want your
object to get a private copy of the shared data. They also have a copy()
function that returns a deep copy with a reference count of 1.

The same is true for QImage, which does not inherit QMemArray. QMovie
is also explicitly shared, but it does not support detach() and copy().

Implicitly Shared Classes

The Qt classes that are implicitly shared are:

• QBitmap
• QBrush
• QCursor
• QFont
• QFontInfo
• QFontMetrics
• QIconSet
• QMap
• QPalette
• QPen
• QPicture
• QPixmap
• QRegion
• QRegExp
• QString
• QStringList
• QValueList
• QValueStack

These classes automatically detach from common data if an object is about
to be changed. The programmer will not even notice that the objects are
shared. Thus you should treat separate instances of them as separate
objects. They will always behave as separate objects but with the added
bonus of sharing data whenever possible. For this reason, you can pass
instances of these classes as arguments to functions by value without
concern for the copying overhead.

Example:

    QPixmap p1, p2;
    p1.load( "image.bmp" );
    p2 = p1;                    // p1 and p2 share data
    QPainter paint;
    paint.begin( &p2 );         // cuts p2 loose from p1
    paint.drawText( 0,50, "Hi" );
    paint.end();

In this example, p1 and p2 share data until QPainter::begin() is
called for p2, because painting a pixmap will modify it. The same
happens also if anything is bitBlt()'ed into p2.

QCString: implicit or explicit?

QCString uses a mixture of implicit and explicit sharing. Functions
inherited from QByteArray, such as data(), employ explicit sharing, while
those only in QCString detach automatically. Thus, QCString is rather an
"experts only" class, provided mainly to ease porting from Qt 1.x to Qt 2.0.
We recommend that you use QString, a purely implicitly shared class.