Qt, range-based for loops and structured bindings Part 1

17.04.2020 Ivan Čukić 9 comments

Qt has a long history. The first stable version was released before the first version of C++ was standardized and long before the different C++ compiler vendors started shipping usable implementations of the C++ standard library. Because of this, Qt often followed (and still follows) some design idioms that feel unnatural to the usual C++ developer.

This can have some downsides for the Qt ecosystem. The evolution of the C++ programming language which sped up quite a bit in the last decade often brings improvements which don’t fit well with the Qt philosophy. In this blog, I offer some ways to work with this.

Range-based for loops

C++11 brought us the range-based for loop which allows easy iteration over any iterable sequence, including common containers such as std::vector, std::map, etc.

for (auto name: names) {
    // ...
}

When you write code like the above, the compiler (C++11) sees the following:

        {
            auto && __range = names;
            for (auto __begin = begin(names), __end = end(names); __begin != __end; ++__begin) {
                auto&& name = *__begin;
                // ...
            }
        }

It takes the begin and end iterators of the names collection, and iterates from one to the other.

Rather simple, but it is a very nice syntactical improvement to the language — it is a significant reduction of boilerplate code and much easier on the eyes.

But, it was not designed with Qt collection classes in mind. You can check out the “Goodbye, Q_FOREACH” blog post by Marc for one example where the range-based for loop collided with the Qt design principles, most notably with the copy-on-write feature because of which we got the qAsConst helper function.

Structured bindings

Another new feature, this one from C++17, are the structured bindings.

They add syntactic sugar for decomposing structures. For example, if we have a function that returns a pair of values (std::pair or a QPair) we can declare variables that refer to the values inside of the pair instead of having to access them through .first and .second.

        auto [x, y] = mousePosition();

This can significantly improve code readability. In the previous example, the names x and y communicate the meaning much better than .first and .second would.

This is even more the case if we use a function like QMultiMap::equal_range which returns a pair of iterators delimiting the range of values that are stored under a specified key in a QMap. Seeing participants.equal_range("kdab").second would look like we are accessing the second value in the map that is stored under the “kdab” key. If we used structured bindings, we could have sane variable names:

        auto [beginParticipant, endParticipant] = participants.equal_range("kdab");

Structured bindings can lead to quite succulent code when combined with the range-based for loops for iteration over STL-compliant map-like containers:

        for (auto [key, value]: stdMap) {
            // ...
        }

Each key-value pair in stdMap, will be stored in an invisible (unnamed) variable, and key and value will be references to its inner elements.

Sidebar: don’t be fooled by decltype(key) telling you key is a normal value, for all intents and purposes, it is a reference.

QHash, QMap and STL compatibility

Sadly, the previous code snippet can not work with Qt’s associative containers because the API is not compatible with that of std::map and similar containers.

The problem is in the API of the iterators for QHash and QMap. Instead of returning a key-value pair when you dereference the iterator, it returns only the value. This means that the range-based for loop can only iterate over the values stored inside of the collection, it can not (by default) access the keys:

        for (auto value: map) {
            // ...
        }

When we use the QHash and QMap iterators directly, we can access the key with .key() and the value with .value() but this does not help us with the range-based for loop. As we’ve seen, it is just a syntactic sugar which expands to an ordinary for loop that iterates over a collection automatically dereferencing (calling operator*()) the iterator in each iteration.

So, as far as the range-based for loop is concerned, the key and value member functions do not even exist.

This deficiency of the Qt’s associative containers’ API has been noticed, and QHash and QMap provide us with alternative, more STL-like, iterators since Qt 5.10 — the key_value_iterator and the const_key_value_iterator. When dereferenced, these iterators return a std::pair containing both the key and the value. To quote the Qt documentation:

The QMap::key_value_iterator typedef provides an STL-style iterator for QMap and QMultiMap.

QMap::key_value_iterator is essentially the same as QMap::iterator with the difference that operator*() returns a key/value pair instead of a value.

This means that we can use the structured bindings with them:

        auto [key, value] = *map.keyValueBegin();

But we still cannot use this with the range-based for loop to iterate over all key-value pairs in the collection. The range-based for loop will call .begin() and .end(), and not .keyValueBegin() and .keyValueEnd().

This is easily remedied by creating a simple wrapper over QHash and QMap that will rename .keyValueBegin() and .keyValueEnd() to .begin() and .end() like so:

        template <typename T>
        class asKeyValueRange
        {
        public:
            asKeyValueRange(T &data)
                : m_data{data}
            {
            }

            auto begin() { return m_data.keyValueBegin(); }

            auto end() { return m_data.keyValueEnd(); }

        private:
            T &m_data;
        };

Instead of passing a Qt associative container directly to the range-based for loop, we just need to pass it wrapped inside of the asKeyValueRange class, and we will finally be able to use the idiomatic C++17 way of iterating over key-value pairs with QHash and QMap:

        for (auto [key, value]: asKeyValueRange(map)) {
            // ...
        }

It is important to note that this is a somewhat simplified implementation of the asKeyValueRange wrapper that takes the reference to the original QHash/QMap, so it might lead to a dangling reference if used improperly.

Iteration over Qt SQL results

Another part of Qt that hasn’t been designed with common C++ idioms in mind is the Qt SQL module. Its API is designed as if it were a Java library and not a C++ one. Instead of having normal iterators over the results in the QSqlQuery, we need to use the .next() member function to fetch each result. The .next() member function returns a Boolean value denoting if there is another result, and the result itself is accessible through other QSqlQuery member functions.

Apart from the API, iteration over the QSqlQuery results has another difference to usual C++ iterators. The QSqlQuery does not need an end iterator to know whether there is another result available. If .next() or .isValid() return false, this means we have reached the end of the result set.

Having the iterator itself know when it has reached the end of the collection it belongs to is uncommon in STL, but it still is present. The input stream iterator (std::istream_iterator) is one such beast. We cannot know in advance where the end of an input stream is, we can just read values from the stream until the read fails. We have reached the end when the read failed. Similar to to the QSqlQuery — we have reached the end when .next() fails — when it returns false.

In these cases, the end iterator is only needed to fulfil the STL iterator API. It is just a dummy object (usually called a sentinel) that will trigger the validity check on the main iterator when it != end is evaluated.

C++17 brought a small change to the range-based for loop meant for this exact use-case. When C++17 compiler sees the range-based for loop from the first code snippet we had, it will convert it to the following code:

        {
            auto && __range = names;
            auto __begin = begin(names);
            auto __end = end(names);
            for ( ; __begin != __end; ++__begin) {
                auto&& name = *__begin;
                // ...
            }
        }

The difference is easy to overlook.

        auto __begin = begin(names), __end = end(names)

is replaced by

        auto __begin = begin(names);
        auto __end = end(names);

While these two snippets look mostly the same, there is an important difference — in the first snippet, both __begin and __end need to have the same type whereas they can be different in the second.

This allows us to create a begin function that returns a powerful iterator that knows when it has reached the end of a collection, while the end function returns a dummy value — a sentinel.

The sentinel can be just a simple empty type:

        struct QSqlResultSentinel {};

        QSqlResultSentinel end(QSqlQuery& query)
        {
            return {};
        }

The iterator is where the main logic lies. For it to work with the range-based for loop, it needs to know how to do the following operations:

– It needs to have the operator++ that moves to the next result; – It needs to have the operator!= which takes a sentinel object and returns whether iterator is valid or not; – It needs to have the operator* that dereferences the iterator.

The basic implementation of the operator++ and operator!= would look like this:

        class QSqlResultIterator {
        public:
            QSqlResultIterator(QSqlQuery& query)
                : m_query(query)
            {
                m_query.next();
            }

            QSqlResultIterator& operator++()
            {
                m_query.next();
                return *this;
            }

            bool operator!=(QSqlResultSentinel sentinel) const
            {
                Q_UNUSED(sentinel);
                return m_query.isValid();
            }

            // ...

        private:
            QSqlQuery& m_query;
        };

        QSqlResultIterator begin(QSqlQuery& query)
        {
            return QSqlResultIterator(query);
        }

The only thing that remains is for us to decide what the dereference operator should do.

Since each record in the result can contain several values, it would be nice if we provided the operator[] to access those values. Therefore, when we dereference the QSqlResultIterator, we want to get a type that has operator[] defined on it which returns values from the current record.

We have several options on how to do this. The simplest, though not the cleanest, is to have the dereference operator return the reference to the iterator itself, and then define the operator[] in the iterator class.

        class QSqlResultIterator {
        public:
            // ...

            QSqlResultIterator& operator*()
            {
                return *this;
            }

            QVariant operator[] (int index) const
            {
                return m_query.value(index);
            }

            QVariant operator[] (const QString& name) const
            {
                return m_query.value(name);
            }
        };

If we weren’t interested in having the operator[] to access the values, then we could just return the current QSqlRecord with m_query.record() from the operator*().

While this is not a complete iterator implementation (we are missing iterator type traits and similar), it is sufficient for us to be able to use it with a range-based for loop:

        for (auto result: query) {
            qDebug() << result["name"] << result["surname"];
        }

We’ll see what we need to do in order to be able to use structured bindings with Qt SQL in the next part of this post.