Exploring Qt Containers

alex  07 Aug, 2018
qt  cpp  qt-containers  stl 

The following post will dive into the world of containers and explore the ones Qt provides and how they compare to the standard library's (STL). The post will explore the different types and which ones are unique to each library. API, performance and some internal implementation details will also be covered. The goal of the post is to identify which library might be preferred over the other.

According to Qt's wiki page about the Qt library:

The Qt library provides a set of general purpose template-based container classes.

These container classes are designed to be lighter, safer, and easier to use than the STL containers. If you are unfamiliar with the STL, or prefer to do things the "Qt way", you can use these classes instead of the STL classes.

In a talk by Giuseppe D'Angelo he explained that there are historical reasons why the Qt containers exist:

  • Qt supported platforms that didn't provide a standard library.
  • Qt didn't want to expose the STL symbols from its ABI.

However, since Qt5, it's required to use a working STL implementation.

Container types

Let's start by looking into the available types for each library. The following tables, list the Qt containers and the STL counterparts for each container type: sequence containers, container adaptors, associated containers (ordered and unordered) and lastly, classes which resemble containers. As you will note, some containers are only available in Qt and vice versa. Each section will cover the unique Qt types as well as some additional information.

Sequence containers

Qt STL
- std::array
QVector std::vector
- std::deque
QLinkedList std::list
QList -
- std::forward_list
Algorithmic Complexity
Type Random Access Insertion Prepending Appending
QLinkedList O(n) O(1) O(1) O(1)
QList O(1) O(n) Amortised O(1) Amortised O(1)
QVector O(1) O(n) O(n) Amortised O(1)

The corresponding STL containers have the same complexity.

QList

QList is a bit dodgy as it, at first glance, looks like a linked list, i.e. a std::list, yet it's a very different container. Bear with me here as it's rather complicated. The QList is essentially an array of void* allowing fast random access, insertion, prepending and appending (see Algorithmic Complexity above). Each void* points to the added element which is allocated on the heap. However if sizeof(T) <= sizeof(void*) and the type has been declared to be either a Q_MOVABLE_TYPE or a Q_PRIMITIVE_TYPE the QList will internally use an array of T instead of void*. Because of this design the QList is a memory waster for most types compared to a QVector.

Interestingly, the official documentation previously recommended the developer to use QList as the 'default' container in Qt4, whereas from Qt5 the official documentation encourages developers to use QVector instead:

QVector should be your default first choice. QVector will usually give better performance than QList, because QVector always stores its items sequentially in memory, where QList will allocate its items on the heap...

Container adaptors

Qt STL
QStack std::stack
QQueue std::queue
- std::priority_queue
  • QStack inherits from QVector, whereas std::stack's underlying container can be any container that meets certain requirements. By default std::stack uses std::deque as underlying container.

  • QQueue inherits from QList. Similar to std::stack the std::queue as well as std::priority_queue can use any underlying container that meets some criteria, see std::queue and std::priority_queue. By default std::queue's underlying container is std::deque and std::priority_queue uses a std::vector.

Associative containers

Qt STL
- std::set
QSet std::unordered_set
- std::multiset
- std::unordered_multiset
QMap std::map
QMultiMap std::multimap
QHash std::unordered_map
QMultiHash std::unordered_multimap

Note that QSet's STL counterpart is std::unordered_set and not std::set.

Algorithmic Complexity
Type Key Access Insertion
Average Worst Case Average Worst Case
QMap O(log n) O(log n) O(log n) O(log n)
QMultiMap O(log n) O(log n) O(log n) O(log n)
QHash Amortised O(1) O(n) Amortised O(1) O(n)
QSet Amortised O(1) O(n) Amortised O(1) O(n)

The corresponding STL containers have the same complexity.

Other classes which resemble containers in Qt

The following three template classes are related to containers but don't use implicit sharing:

  • QVarLengthArray
  • QCache (also doesn't expose any iterators)
  • QContiguousCache (also doesn't expose any iterators)
QVarLengthArray

The QVarLengthArray container is a low-level array with a variable length, used for memory optimisation. On construction, the container allocates a predefined array of length N (default to 256) on the stack*. If the amount of elements increases above N, the elements will be moved automatically to a heap memory and then function similarly to a QVector.

A typical use case is when a temporary array is needed to be constructed multiple times, but the element size is varying:

int func(const int n) {
  //int data[n + 1]; //won't compile
  //QVector<int> data(n + 1); //func() is called many times, 
                         //requiring a lot of heap allocations.
  QVarLengthArray<int, 1024> data(n + 1); //Most cases n < 1024, 
                                          //and only a few heap allocations.
  ...
  return data[n];
}

* as pointed out by Kuba Ober a container can never guarantee that it can allocate anything on the stack, see his comment for more information.

QCache and QContiguousCache

QCache resembles a QHash (the underlying container is even a QHash) and is used for memory optimisation. The class takes ownership of the passed in elements and deletes them automatically when a specified maximum is reached. When adding new elements and the maximum is reached, the last used elements will be deleted.

QContiguousCache functions in a similar manner to QCache but restricts the elements in the container to be contiguous. This enables the class to be more memory efficient than the QCache and use fewer processor cycles.

API

The Qt containers provide two sets of APIs: the "Qt Way" as well as the "STL Compatible Way". Personally, I think the Qt Way is more readable, however by using the STL API it will be easier to swap out a Qt container for its STL counterpart if so is desired in the future. Although Qt mimics the STL API, there are still some inconvenient subtleties that might throw you off. For example, the size() function and the index for the QVector and QList are using int as default type, whereas STL's counterparts are usually indexed with std::size_t. This makes it somewhat awkward to mix and convert between Qt and STL containers as it will require explicit casts.

See examples below for some of the API differences using a QVector:

QVector<int> data {0, 1, 2};
data.count() //or .size()
data << 3; //or .append() or .push_back()
if (data.isEmpty()) {} // or .empty()

The Qt containers have also been extended with some convenience functions which are not available in the STL counterparts. See example below:

//Find an element in a vector
QVector<int> qt_vector {0, 1, 2};
if (qt_vector.contains(1)) {
  ...
}

std::vector<int> stl_vector {0, 1, 2};
if (std::find(std::begin(stl_vector), std::end(stl_vector), 1) != std::end(stl_vector)) {
  ...
}

In addition, the Qt containers expose two types of iterators: "java-style" iterators and "STL-style" iterators:

QVector<QString> names {"Alex", "Chris", "Andy", "Swati"};

//java-style iterator
QVectorIterator<QString> it{names};
while (it.hasNext()) {
  qDebug() << it.next();
}

//stl-style iterator
for (auto it = std::begin(names); it != std::end(list); ++it) {
  qDebug() << *it;
}

Note that java-style iterators can't be used together with STL's generic algorithms.

Although the Qt containers have quite a few additional features they also lack some fundamental functionality compared to the STL containers:

  • Qt containers don't provide range-construction nor range-insertion.
  • It's not possible to use a custom allocator in any of the Qt containers.
  • None of the C++11 (or above) features/API have been implemented such as emplace() or rvalue-push_back().
  • No exception handling.
  • The types must implement a default constructor, a copy-constructor and an assignment operator in order to be used in a Qt container.
  • Move-only types are not (and will never be) supported because of Qt containers' principle of implicit sharing (see below for details about implicit sharing).

Implicit sharing

Most of Qt's containers use implicit data sharing which is an optimisation to maximise resource usage by minimising copying. Basically the containers use reference counting and copy-on-write internally. What this means is that when a Qt container is copied, it's only a shallow copy and the "real", deep copy, is done when a non-const function is called. The best way to understand this is with an example:

QVector<int> data {1, 1}; //Reference count = 1
auto data_copy = data; //Pointing to the same payload. Reference count = 2
auto element = data_copy[0]; //Still shared payload, const function called.
data_copy[1] = 2; //Detaches and performs a deep-copy. Reference count = 1

This design results in lower memory usage and less duplication of data as the program avoids unnecessary data copies. However, because of this behaviour it might result in some difficult-to-spot bugs. Example:

QStringList names {"Alex", "Chris"};
auto &name = names[0];
auto data_copy = data;
name = "Andy"; //Even data_copy[0] has been changed to "Andy". 

Another example, which might cause some head-scratches is when using a range-based for-loop. The loop might force the Qt container to detach and create deep copies unintentionally:

QStringList getNames() {
  return QStringList{} << "Alex" << "Chris";
}

for (const auto& name : getNames()) { //non-const begin() is called 
  qDebug() << name;                 //resulting in a deep copy
}

//Preferred
const auto names = getNames();
for (const auto& name : names) { //const begin() is called 
  qDebug() << name;            //still shared payload.
}

//or in Qt 5.7
for (const auto& name : qAsConst(getNames())) {
  qDebug() << name;
}

};

However, by using compiler plugins such as clazy, it's possible to detect some of the unintended detaches at compile time. Also, avoid using Q_FOREACH (or foreach).

Conclusion... STL or Qt?

My recommendation is to use the STL containers as your default containers for the following reasons:

  • Qt itself is moving away from the Qt containers and have replaced some of its internal structures with the STL ones.
  • As mentioned previously, none of the C++11 (or beyond) features have been implemented, and some of the C++98 are still missing (custom allocators, range-construction and range-insertion).
  • Implicit sharing might introduce some subtle bugs that might be difficult to find and debug.
  • The STL containers are usually implemented by the same people who implement the C++ compiler and are therefore most likely better optimised.

However, if you prefer the Qt API and are aware of their drawbacks, I personally don't believe that they are much worse than their STL counterparts. If you decide to use the Qt containers you should also be aware of the type classification of your own types and how it affects the performance when using them in the containers.

Lastly, some of Qt's API returns Qt containers (QStringList is very common) and in those cases I recommend to avoid converting it to a STL counterpart.

Read more

Share the post

Latest