Mircea Baja @ 16 November 2022 # History of normal type in C++ --- # Regular data as of C++20 ```cpp struct person { std::string first_name; std::string last_name; int age{}; constexpr std::strong_ordering operator<=>(const person &) const noexcept = default; }; ``` --- # Agenda - struct from C - C copy - C++ constructor, destructor and class - Default constructor - C++ copy - Exception safety, noexcept - C++ equality and order (first attempt) - Templates and STL - C++ move rationale - Rule of three and composing - Regular: syntax and semantics - History of concepts in C++ - The many relations - History of the three-way comparison, aka the spaceship operator - Pragmatic Regular structs - Regular wannabes - The future of regular - Efficiency and the things we know --- # struct from C ```cpp struct sample { int w; // 4 bytes char x; // 1 byte // 3 bytes of padding here int y; // 4 bytes char z; // 1 byte // 3 bytes of padding at the end }; // takes overall 16 bytes ``` - padding can be reduced by sorting members by size - such a type provides interpretation of memory --- # C copy motivation ```cpp // can do this if x and y are int, x = y; // what if they are some struct? ``` ```cpp // can do this if x is int, int foo(int x); // what if they are some struct? struct sample bar(struct sample y); ``` - initially the answer was: NO - eventually the answer was: YES --- # C copy ```cpp // to initialize: void * memset(void * ptr, int value, size_t num); // to copy: void * memcpy(void * dest, const void * src, size_t num); // to move: void * memmove(void * dest, const void * src, size_t num); // to compare: int memcmp(const void * ptr1, const void * ptr2, size_t num); ``` --- # C copy - uses `memcpy` to copy - but can't use `memcmp` to compare because of padding ```cpp // can do this if x and y are int, if (x < y) { // ... } // but can't do it if they are some struct ``` --- # C++ class ```cpp struct foo { // member variable int bar; // member function void buzz() { // member function body bar = 42; // or this->bar = 42; } }; ``` - `class` is the same as `struct` except default visibility which is private for the `class` --- # C++ constructor/destructor ```cpp struct foo { // constructor foo() { // constructor body } // destructor // "not constructor" ~foo() { // destructor body } }; ``` --- # C++ copy motivation ```cpp struct foo { bar * ptr; foo() { // allocate memory here } ~foo() { // deallocate memory here } }; ``` ```cpp int main() { foo x; // memory is allocated here foo y = x; // y gets a copy of the pointer value // memory freed twice: ups! } ``` --- # C++ copy ```cpp struct foo { bar * ptr; foo(const foo & other) { // copy construct here // e.g. allocate memory pointed by ptr // then copy bar from *(other.ptr) } foo & operator=(const foo & other) { // copy assignment here // needs to handle the case where // ptr already points to something return *this; } }; ``` - `const` and reference - reference vs. `this` - constructor and assignment: efficiency --- # Self assignment, identity ```cpp int main() { foo x; x = x; // self assignment } ``` Options: - don't self assign - handle it ```cpp foo & operator=(const foo & other) { if (this != &other) { // actual copy assignment here } return *this; } ``` --- # Non-copyable ```cpp foo(const foo & other) = delete; foo & operator=(const foo & other) = delete; ``` --- # Copy means copy - return value optimisations ```cpp foo buzz() { return foo(); // foo is default constructed // and then copied into the return value } int main() { foo x = buzz(); // the return value of buzz is copied into x } ``` --- # Exception safety - basic: does not leak - does not throw: noexcept - strong (rollback) guarantee --- # C++ comparison: first try ```cpp class foo { // ... public: bool operator==(const foo & other) const { // compare *this with other // return true if equal // return false otherwise } // ... }; ``` ```cpp // assuming a and b are of type foo // operator == is called here // *this is a, other is b if (a == b) { // ... } ``` - motivation: complex number class --- # Heterogenous: first try ```cpp class foo { // ... public: bool operator==(const char * other) { // compare *this with literal string } // ... }; ``` - motivation: std::string ```cpp if (a == "test") { // ... } // but this does not work if ("test" == a) { // ... } ``` --- # Friend trick ```cpp class foo { // ... public: friend bool operator==(const foo & x, const char * y) { // compare x with y // return true if equal // return false otherwise } // ... }; ``` --- # Similar Equality functions: ```cpp bool operator!=(const foo & x, const foo & y) { return !(x == y); } ``` Order functions: ```cpp bool operator<=(const foo & x, const foo & y) { return !(y < x); } bool operator>(const foo & x, const foo & y) { return y < x; } bool operator>=(const foo & x, const foo & y) { return !(x < y); } ``` ??? > --- # STL ```cpp #define private public ``` - macros - templates - STL (Alex Stepanov) - "What is an iterator? Not a class. Not a type. It is a concept. […] It is something which doesn’t have a linguistic incarnation in C++. But it could." - The regular concept arouse from questions like: “if you design a vector as a resizable array container, what types should the vector template accept?” --- # C++ move motivation ```cpp // written like this swap makes unnecessary copies foo tmp = a; // copy constructor a = b; // copy assignment b = tmp; // copy assignment // end result: three deep copies ``` --- # C++ move ```cpp class foo { bar * ptr; public: // ... foo(foo && other) : ptr{ other.ptr } noexcept { other.ptr = nullptr; } foo & operator=(foo && other) noexcept { if (this != &other) { delete ptr; // asumes bar destructor does not throw ptr = other.ptr; other.ptr = nullptr; } return *this; } }; ``` - `noexcept` and move --- # C++ move ```cpp T tmp = std::move(a); // move constructor a = std::move(b); // move assignment b = std::move(tmp); // move assignment // end result: three ownership transfers ``` - "moved from" state: destroy or assign - destructive vs. non-destructive --- # Regular data concept (semantics) - It's a data class. - You can create an instance of that class - You can destroy an instance of that class - It does not leak resources, it does not leak memory in particular, that's the only kind of resource you would expect such a type to acquire, the destructor in particular takes care of releasing the resources - You can take a copy (either by assignment or construction) - You can move it (also either by assignment or construction) - Move is an optimised version of copy for when the original is no longer needed - "moved from" can at least be destroyed and assigned - You can test two instances for equality `==` - It does not matter how we compare for equality: `!=` is not equal, the opposite of equal - Equality behaves properly: reflexive, symmetric, transitive, substitutable - A copy is equal to the original - ... --- # Regular data concept (semantics) - ... - You can compare two instances for order - It does not matter how we compare for order: can use either of `<`, `>`, `<=`, `>=`, `<=>`. E.g. `a < b` is the same as `b > a` - Less than `<` behaves properly, it is strict total order: irreflexive, asymmetric, transitive. Similarly the other order functions. - Equality works properly with order. Trichotomy: either `a < b` or `a == b` or `b < a` - Equality and comparison do not change the arguments, copy does not change the source. - Of all the functions above, only copy might throw (because it might need to allocate memory) - The functions above have reasonable complexity of operations: constant or at most linear with the amount of data held - Independence: changing one value does not change other unrelated values ??? > --- # Regular data concept (syntax) - Default constructor (`noexcept`) - Destructor (implicitly `noexcept`) - Copy constructor and assignment - Move constructor and assignment (`noexcept`) - Equality: `==` and `!=` (`noexcept` and `const`) - Order: `<`, `>`, `<=`, `>=`, `<=>` (`noexcept` and `const`) --- # C++20 concepts - idea from Alex Stepanov (see STL/EOP) - but even C had such an idea e.g. `short`, `int`, `long` are all signed integers - going all the way to the genus (from the ancient Greeks) - `requires` expression ```cpp requires (T x) { x + x; } ``` --- # C++20 concepts ```cpp template<typename T> concept addable = requires(T x) { x + x; }; ``` ```cpp template<typename T> concept small_addable = addable<T> && (sizeof(T) <= 4); ``` ??? > --- # C++20 concepts - `requires` clause ```cpp template<typename T> T add(T x, T y) requires addable<T> { return x + y; } ``` ```cpp template<addable T> T add(T x, T y) { return x + y; } ``` --- # C++20 concepts ```cpp template<typename T> concept incrementable = requires (T x) { ++x; x++; }; auto add_one(incrementable auto x) { return ++x; } ``` - `requires` clause, then `requires` expression ```cpp template<typename T> T add(T x, T y) requires requires(T x) { x + x; } { return x + y; } ``` --- # C++20 concepts - they are just a tool: like `struct` (does not tell you how to use it) - e.g. used for `std::ranges` - errors not necessarily shorter - `std::regular` concept was demonetized from the "regular" concept --- # Many relationships - homogeneous/heterogeneous - predicate: returns `bool` - total/partial: domain - identity, equality, equivalence - "less than" trichotomy with equality - weak order trichotomy with equivalence - the weird case of `float`: NaN and -0.0 --- # Spaceship operator - this: `<=>` - not: `>=< >-< |=| |-o-| <o>` - three-way comparison (see `memcmp`) - `= default` - simplifies symmetric heterogeneous comparisons - debatable: ``` <- partial_ordering weak_equality ^ <- weak_ordering ^ ^ strong_equality <- strong_ordering ``` - the efficiency of equality ??? > --- # Pragmatic spaceship ```cpp struct person { std::string first_name; std::string last_name; int age{}; constexpr std::strong_ordering operator<=>(const person &) const noexcept = default; }; ``` --- # Good regulars - `int`: the dream regular type. - `std::string`: copy could throw, but that's usually OK, all dynamically allocated containers have this issue. - `std::vector`: it is regular if the `value_type` is regular (it does lexicographical compare, which is what one would expect), but not if the `value_type` is not regular. It still makes sense to be able to store contiguously items, even if not regular. - `std::pair`: similar issues with the vector, it all depends on member types. It does memberwise compare, which is what one would expect. - `std::optional`: orders the `nullopt` value ahead of the underlying `value_type` order. That's OK. - `std::shared_ptr`: does shallow copy and comparison, but that OK, it uses atomics to address expectations around threading. --- # Wannabes - Types like `std::map` in the Microsoft implementation that might throw for default constructor or move constructor, but are otherwise OK. They might also not have caught fully with `constexpr` support. - Types like `std::unordered_map` where the equality time complexity might be `N^2`, but are otherwise OK. Maybe you say that those cases are outside the domain of equality comparison - Types like `float` have multiple issues around `NaN` (domain issues) and `-0.0`/`+0.0` (`==` does not mean equality) - Types like `string_view` which are not semantically regular e.g. you take a copy, you change the original string, but then it's still equal to the original (because it's a view to the same data). Because it's a constant view into data, it just about gets away with doing a shallow copy, but a deep compare. - `span` unlike `string_view` also does a shallow copy of the pointers and determine the range, but does not get away with doing a deep compare - Then the standard relaxed the requirements further where the `std::regular` concept does not require order. E.g. C++ iterators; note though they could have had order, it's just that order for two iterators would reflect the location in memory, not in the sequence (think pointers to nodes in a linked list), and some would find that confusing. - And then further relaxed for `std::semiregular` to not even have equality - ... --- # Wannabes - ... - vtable interface based types that use `clone` methods to copy, with their own issues around comparisons - Types that can't be copied like RAII C-handle wrappers, but can be moved - Types that aren't even meant to be moved or copied like trait types: they are just some syntactic mechanism to communicate with the compiler. **Some deviations from regular are meant to be, but others are caused by poor design.** In particular using `==` and `<` to mean less than equality and "less than" (pun intended) is wrong. Define instead methods like `is_equivalent` or `preceeds` and use those to customise algorithms. ??? > --- # The future of regular ```cpp regular_struct person { std::string first_name; std::string last_name; int age; }; ``` --- # Questions?