Krita/C++11

From KDE Community Wiki

General links about using C++11 in Qt

There have been a few links discussing mixing C++11 with Qt, and starting with Qt 5.6 C++11 support will be default. Note: as would be expected from those most engaged in hot new technology, the writers of most of these links demonstrate Silicon Valley levels of extreme optimism. Take this with a grain of salt... there is no such thing as a free lunch.


Here are some overviews of C++11 features.


Qt's API design principles do not always overlap with the C++ Standards Committee design principles. (Range-based for demonstrates the design clash pretty clearly.)

Particular Features

Under "drawbacks," every item should list: "Programmers will face another feature they must learn about."

Type Inference (auto)

Motivation: If a function f has a return type Type, it is redundant to write a local variable Type x = f(y). Using auto declarations is a simplification in two ways scenarios. First, it allows the programmer to write code without worrying about doing the manual type deduction, for example, KoXmlReader::const_iterator x = iter.begin() versus auto x = iter.begin() . Second, it allows the programmer to change the type of a function or class member without changing the declarations of local variables.

Drawbacks: Some uses of auto can be obfuscating. For example, auto x = 2 is not obviously an integer, and auto x = {"a", "b", "c"} returns std::initializer_list. Sometimes it is not clear what some function returns by the name of the function.

Recommendation: In general, it is fine to declare a local variable as auto, if it initialized using another variable, or as the return value from a function. Don't use it to initialize constants, and don't let it reduce the clarity of the code.

Range-based for loop

Motivation: This is something a long time coming in C++. It is a standardized replacement for Qt's foreach() construct, which works not only with Qt objects but all iterable C++ types.

for (T x : list ) { }

It will work with standard tooling and static analysis, and can be faster by defaulting to in-place access. For this reason range-based iterators should always be used for STL containers, if those are ever needed in Krita.

Drawbacks: By default, Qt's foreach rewites the code to make a shallow copy and then use const accessors, while c++11 does the opposite, avoiding copying when possible. When using const accessors, this is faster, but if you try to make changes to the data, this will slow your loop down instead.

Recommendation: Sometimes, the range-based for is faster. Sometimes the Qt iterator is faster. Personally I like the range-based for in principle, since it works better with static analysis, it has a faster best-case speed, and it is always possible to write it in a way that replicates the foreach() behavior, though the reverse is not true.

On the other hand, there is a bad, dangerous worst case performance hit when a detach/copy is triggered, and this is not easy to catch with standard syntax. It is possible to define a function const_(T) that forces the compiler to use constant iterators by writing for (T x : _const(list) ) { } .

Qt's recommendation on the other hand is to use foreach() for Qt iterators, and range-based for on STL containers, because you always know what you're getting, and you always keep your syntax easy to read. In my opinion is the most meaningful feature without a clear answer.

General Initializer Lists

Motivation: Initializer lists are intended to work in many different places to simplify the syntax for complicated initialization. For example, a list of strings could be initialized const QStringList x = {"abc", "def", "xyz" }; and this would work just fine for a std::list<std::string> as well.

A second place initializer lists are used is in creating standard initial values for class members. This takes the place of writing a lengthy constructor list like f(_q) : String1("a"), Subclass1(0), Subclass2(1).

Mixed uniform initialization is another addition. It is possible to give

Drawbacks: None I can think of. This is super simple, completely obvious to read and write, and shortens code by removing big default constructors.

Recommendation: Yes!

New-style signals/slots, lambdas

Motivation: Lambda expressions are a big new addition for C++11. Many programmers claim they start to feel like an essential part of the language very quickly. One of the biggest uses for lambdas is in the standard algorithm library <algorithm>, which is described below. In Qt5, this, along with std::function and std::bind, allow for One of the most useful C++11 integrations, a new signal/slot syntax which replaces the moc macros SIGNAL() and SLOT() with standard C++. For example:

  • connect(sender, SIGNAL (valueChanged(QString,QString)), receiver, SLOT (updateValue(QString)) );
  • connect(sender, &Sender::valueChanged, receiver, &Receiver::updateValue );

New style signals and slots provide a great benefit from the tooling perspective: now, all types for functions and function arguments can be checked statically, and you don't have to catch typos by monitoring debug messages saying "no such slot."

Another possibility is to use lambdas directly inside connect(), instead of defining a class member function which is only used once. The greatest benefit is that the function can be defined right where it is used; it also aids readability to get rid of a list of tiny helper functions from the header.


Drawbacks: The new-style syntax makes it somewhat harder to use default arguments, which requires the use of lambdas. It is also perhaps a little less pretty.

Lambdas in general are have become one of the most clunky pieces of C++11 notation. Since they allow a great deal of options (for example, capturing closure variables by reference with [&] versus capturing by value with [=]) they are a significant new addition to the C++ learning curve. Using auto F = [&] ( x ) { whatever } and slinging F around in places is confusing for everyone. Lambdas will not disconnect automatically, although there is a special syntax to make that happen.

Recommendation: Lambdas will feel strange until they're used in many places, at which point they will feel normal. For the time being, they should be limited to <algorithm>, since there is no reasonable alternative. For consistency's sake, it is better to stick to a single method for signals and slots, but fully replacing them in the codebase would require lambdas instead of default arguments in some places. My recommendation is to use the new signal/slot syntax. It is simply easier to write and maintain it correctly.

constexpr

Motivation: Performing calculations at compile time can speed things up at runtime. KDAB: speed up your Qt 5 programs using C++11

Drawbacks: Not easy to use these features.

Recommendation: This could be useful in specific places, like KoCompositeOpRegistry. Overall it is not something most programmers will run into.

<algorithm>

Motivation: Using something like std::replace (myvector.cbegin(), myvector.cend(), 20, 99); is more concise, safer and easier to read than a handwritten loop that looks for occurences of the number 20 and replaces it with 99. There should be no performance penalty compared to a hand-written loop on STL classes. A list of standard algorithms can be found here. Historically Qt provided its own algorithm library, but now encourages programmers to use the STL versions instead, and Qt's own algorithm library will mostly become deprecated. http://doc.qt.io/qt-5/qtalgorithms.html Unlike range-based for, since we are easily able to specify constBegin/cbegin, constEnd/cend, using standard algorithms can always be as fast, and generally will be faster, than using Qt foreach loops.

Drawbacks: Some of the standard algorithms are not completely obvious from observing the name. For example, I could not describe the five arguments of std::replace_copy off the top of my head. When values inside the container need to be modified, non-const iterators may be slower than a Qt foreach().

Recommendation: Encourage the use of <algorithm> when it improves code clarity. Speed not a big problem most of the time, don't make changes which are hard to understand just for a tiny hypothetical speed boost. However, Krita's hot paths should be reviewed and moved to <algorithm> instead of Qt foreach().

enum class

Motivation: These are a type-safe version of enums. This prevents comparison with a Another benefit is that enum classes can be forward-declared. enum class Color; void paint(Color c); enum class Color {Red, Green, Blue};

Drawbacks: Virtually none. Very small changes to the code, more type safety. The only problem is fixing code that was unsafe to begin with.

Recommendation: Use enum classes whenever possible.

Local type definitions (i.e. using)

Motivation: An easier and localized way to use typedefs. Can be at the namespace, class, or function level.

Drawbacks: Very few. These are quite intuitive, and allows you to place typedefs closer to where they're used.

Recommendation: Go for it.

nullptr

Motivation: The use of nullptr as a default pointer initializer is a very small change in C++11, and mostly an aesthetic one. Technically, there are only a few things it prevents : it cannot be converted to a numeric type like int x = nullptr;, and it cannot be used as a class type in a template, so the following is a compiler error: meta_type<class A, class B>; meta_type<C, nullptr> x;.

The most important to nullptr is simply that you are tagging your code - hey: there is a null pointer lurking around here, be careful!

Drawbacks: It takes longer to type nullptr than it takes to type 0, and it's not so visually pleasing. Converting the existing code base would be fairly laborious. Tiny benefits.

Recommendation: 100% up to the maintainer's preferences.

Deleted, default, override, final

Motivation: These are keywords used for designing inheritance patterns. They are useful for preventing accidental copy construction.

Drawbacks: Since Krita does not export libraries, most of the time we won't need to worry about these. They are limited to solving some pretty specialized problems.

Recommendation: No reason to hold back from these features if they seem useful. They are well named and fairly self-explanatory, especially for developers with a Java or C# background. If you apply them correctly, you can prevent other coders from making mistakes when they use your classes. For others, these definitions can be ignored until they cause a compile error, which tell you that you're doing something the wrong way.

unique_ptr/QScopedPointer

Motivation: Here is a glowing review of unique_ptr. This is really about a philosophy of C++ memory management, not just a particular smart pointer type. The idea is that whenever you create an object on the heap, you should always house it inside a smart pointer. The reason this philosophy is considered new to C++11 is that unique_ptr is the first time they 'got it right' designing a very nice smart pointer class. Most importantly, the class uses negligible overhead. In particular: sizeof(unique_ptr<T*>) = size_t, it can be passed as a function argument without copying, and dereferencing is inline.

QScopedPointer is essentially the same thing as unique_ptr, and perhaps it is more idiomatic to use QScopedPointer instead

"Rule of Zero": more about the C++ design philosophy behind unique_ptr.

Drawbacks: The philosophy mentioned above can be summarized like this: we should state up front what we are going to prohibit programmers from doing. Like the const keyword, unique_ptr puts restrictions on what can be done with the pointer, the main one being, it cannot be copied. Like enforcing const correctness, this can be annoying to get right throughout a codebase.

One particular limitation is that Qt container classes. For example QVector<std::unique_ptr> is invalid, because QVector requires its members can be copied. This makes converting to unique_ptr a bit slow, since QVector<T *> will have to be converted to std_array<unique_ptr<T*>>. If the owner was being copied before, it will become uncopiable. This can be a good thing, but it can also be extra work.

Moving a unique_ptr requires additional semantics.

Recommendation: Smart pointers are already prevalent in the codebase with the KisSP family, but more use of them should be encouraged. Unfortunately d_ptrs cannot be easily wrapped in a smart pointer.

Performance-related (rvalues)

Using move constructors and rvalues are very subtle and advanced features, but widely celebrated as successes of C++11. In general these should be used inside hot paths when we want to move objects instead of copy or reference. C++ programmers should already be aware that writing performant code inside hot paths will sometimes require opening a can of ampersands. These features will naturally stay confined to the corners of the codebase behind the scenes where they belong, and should be introduced when they are useful.

Move Constructors

Recommendation: Use whenever it aids performance.

Reference Qualifiers

Recommendation: Use whenever it aids performance.

C++11 features mostly for template programming

Krita makes very light use of templates. These features are useful, coming across them in the code base will add complexity for new learners, and have not been necessary so far.

  • decltype : this is the most likely of these features to be useful, for example, in KisInputManager.
  • static_assert
  • variadic templates

Other C++11 features that will not be useful

  • Threading support (Qt/KDE uses its own threading model, not C++ threading model)
  • shared_ptr and weak_ptr (Relies on C++ threading model; use KisSharedPointer)
  • New literal types (already have QString/ki18n)
  • Extended Unions (already have QVariant)