COMP6771

Advanced C++ Programming

Week 5.1

Resource Management

Revision: Objects

  • What is an object in C++?

    • An object is a region of memory associated with a type

    • Unlike some other languages (Java), basic types such as int and bool are objects

  • For the most part, C++ objects are designed to be intuitive to use

  • What special things can we do with objects

    • Create

    • Destroy

    • Copy

    • Move

std::vector<int> - under the hood

  • When writing a class, always consider the "rule of 5"
    • Copy constructor
    • Destructor
    • Move assignment
    • Move constructor
    • Copy assignment
  • Though you should always consider it, you should rarely have to write it
    • If all data members have one of these defined, then the class should automatically define this for you
    • But this may not always be what you want
    • C++ follows the principle of "only pay for what you use"
      • Zeroing out the data for an int is extra work
      • Hence, moving an int actually just copies it
      • Same for other basic types
class my_vec {
  // Constructor
  my_vec(int size): data_{new int[size]}, size_{size}, capacity_{size} {}
  
  // Copy constructor
  my_vec(my_vec const&) = default;
  // Copy assignment
  my_vec& operator=(my_vec const&) = default;
  
  // Move constructor
  my_vec(my_vec&&) noexcept = default;
  // Move assignment
  my_vec& operator=(my_vec&&) noexcept = default;

  // Destructor
  ~my_vec() = default;

  int* data_;
  int size_;
  int capacity_;
}
// Call constructor.
auto vec_short = my_vec(2);
auto vec_long = my_vec(9);
// Doesn't do anything
auto& vec_ref = vec_long;
// Calls copy constructor.
auto vec_short2 = vec_short;
// Calls copy assignment.
vec_short2 = vec_long;
// Calls move constructor.
auto vec_long2 = std::move(vec_long);
// Calls move assignment
vec_long2 = std::move(vec_short);

Destructors

  • Called when the object goes out of scope
    • What might this be handy for?
    • Does not occur for reference objects
  • Implicitly noexcept
    • What would the consequences be if this were not the case
  • Why might destructors be handy?
    • Freeing pointers
    • Closing files
    • Unlocking mutexes (from multithreading)
    • Aborting database transactions

std::vector<int> - Destructors

  • What happens when vec_short goes out of scope?
    • Destructors are called on each member.
    • Destructing a pointer type does nothing
    • We have a memory leak
  • How do we solve this?
class my_vec {
  // Constructor
  my_vec(int size): data_{new int[size]}, size_{size}, capacity_{size} {}
  
  // Copy constructor
  my_vec(my_vec const&) = default;
  // Copy assignment
  my_vec& operator=(my_vec const&) = default;
  
  // Move constructor
  my_vec(my_vec&&) noexcept = default;
  // Move assignment
  my_vec& operator=(my_vec&&) noexcept = default;

  // Destructor
  ~my_vec() = default;

  int* data_;
  int size_;
  int capacity_;
}
// Call constructor.
auto vec_short = my_vec(2);
auto vec_long = my_vec(9);
// Doesn't do anything
auto& vec_ref = vec_long;
// Calls copy constructor.
auto vec_short2 = vec_short;
// Calls copy assignment.
vec_short2 = vec_long;
// Calls move constructor.
auto vec_long2 = std::move(vec_long);
// Calls move assignment
vec_long2 = std::move(vec_short);
my_vec::~my_vec() {
  delete[] data_;
}

std::vector<int> - Copy constructor

  • What does it mean to copy a my_vec?
  • What does the default synthesized copy constructor do?
    • It does a memberwise copy
  • What are the consequences?
    • Any modification to vec_short will also change vec_short2
    • We will perform a double free
  • How can we fix this?
class my_vec {
  // Constructor
  my_vec(int size): data_{new int[size]}, size_{size}, capacity_{size} {}
  
  // Copy constructor
  my_vec(my_vec const&) = default;
  // Copy assignment
  my_vec& operator=(my_vec const&) = default;
  
  // Move constructor
  my_vec(my_vec&&) noexcept = default;
  // Move assignment
  my_vec& operator=(my_vec&&) noexcept = default;

  // Destructor
  ~my_vec() = default;

  int* data_;
  int size_;
  int capacity_;
}
auto vec_short = my_vec(2);
auto vec_short2 = vec_short;

my_vec::my_vec(my_vec const& orig): data_{new int[orig.size_]},
                                    size_{orig.size_},
                                    capacity_{orig.size_} {
  // Should work if we also define .begin() and .end(), an exercise for the reader.
  ranges::copy(orig, data_);
}

std::vector<int> - Copy assignment

  • Assignment is the same as construction, except that there is already a constructed object in your destination
  • You need to clean up the destination first
  • The copy-and-swap idiom makes this trivial
class my_vec {
  // Constructor
  my_vec(int size): data_{new int[size]}, size_{size}, capacity_{size} {}
  
  // Copy constructor
  my_vec(my_vec const&) = default;
  // Copy assignment
  my_vec& operator=(my_vec const&) = default;
  
  // Move constructor
  my_vec(my_vec&&) noexcept = default;
  // Move assignment
  my_vec& operator=(my_vec&&) noexcept = default;

  // Destructor
  ~my_vec() = default;

  int* data_;
  int size_;
  int capacity_;
}
auto vec_short = my_vec(2);
auto vec_long = my_vec(9);
vec_long = vec_short;

my_vec& my_vec::operator=(my_vec const& orig) {
  return my_vec(orig).swap(*this);
}

my_vec& my_vec::swap(my_vec& other) {
  ranges::swap(data_, other.data_);
  ranges::swap(size_, other.size_);
  ranges::swap(capacity_, other.capacity_);
}

// Alternate implementation, may not be as performant.
my_vec& my_vec::operator=(my_vec const& orig) {
  my_vec copy = orig;
  std::swap(copy, *this);
  return *this;
}

Lvalue references

  • There are multiple types of references
    • Lvalue references look like T&
    • Lvalue references to const look like T& const
  • A lvalue reference denotes an object whose resource cannot be reused
    • Most objects (eg. variable, variable[0])
  • Once the lvalue reference goes out of scope, it may still be needed

Rvalue references

  • Rvalue references look like T&&
  • An rvalue denotes an object whose resources can be reused
    • eg. Temporaries (my_vec object in f(my_vec()))
    • When someone passes it to you, they don't care about it once you're done with it

 

  • “The object that x binds to is YOURS. Do whatever you like with it, no one will care anyway”
  • Like giving a copy to f… but without making a copy 
void f(my_vec&& x);

Rvalue references

void inner(std::string&& value) {
  value[0] = 'H';
  std::cout << value << '\n';
}

void outer(std::string&& value) {
  inner(value); // This fails? Why?
  std::cout << value << '\n';
}

int main() {
  f1("hello"); // This works fine.
  auto s = std::string("hello");
  f2(s); // This fails because i is an lvalue.
}
  • An rvalue reference formal parameter means that the value was disposable from the caller of the function
    • If outer modified value, who would notice / care?
      • The caller (main) has promised that it won't be used anymore
    • If inner modified value, who would notice / care?
      • The caller (outer) has never made such a promise.
      • An rvalue reference parameter is an lvalue inside the function

std::move

// Looks something like this.
T&& move(T& value) {
  return static_cast<T&&>(value);
}
  • Simply converts it to an rvalue
    • This says "I don't care about this anymore"
    • All this does is allow the compiler to use rvalue reference overloads
void inner(std::string&& value) {
  value[0] = 'H';
  std::cout << value << '\n';
}

void outer(std::string&& value) {
  inner(std::move(value));
  // Value is now in a valid but unspecified state.
  // Although this isn't a compiler error, this is bad code.
  // Don't access variables that were moved from, except to reconstruct them.
  std::cout << value << '\n';
}

int main() {
  f1("hello"); // This works fine.
  auto s = std::string("hello");
  f2(s); // This fails because i is an lvalue.
}

Moving objects

  • Always declare your moves as noexcept
    • Failing to do so can make your code slower
    • Consider: push_back in a vector
  • Unless otherwise specified, objects that have been moved from are in a valid but unspecified state
  • Moving is an optimisation on copying
    • The only difference is that when moving, the moved-from object is mutable
    • Not all types can take advantage of this
      • If moving an int, mutating the moved-from int is extra work
      • If moving a vector, mutating the moved-from vector potentially saves a lot of work
  • Moved from objects must be placed in a valid state
    • Moved-from containers usually contain the default-constructed value
    • Moved-from types that are cheap to copy are usually unmodified
    • Although this is the only requirement, individual types may add their own constraints
  • Compiler-generated move constructor / assignment performs memberwise moves
class T {
  T(T&&) noexcept;
  T& operator=(T&&) noexcept;
};

std::vector<int> - Move constructor

class my_vec {
  // Constructor
  my_vec(int size)
  : data_{new int[size]}
  , size_{size}
  , capacity_{size} {}
  
  // Copy constructor
  my_vec(my_vec const&) = default;
  // Copy assignment
  my_vec& operator=(my_vec const&) = default;
  
  // Move constructor
  my_vec(my_vec&&) noexcept = default;
  // Move assignment
  my_vec& operator=(my_vec&&) noexcept = default;

  // Destructor
  ~my_vec() = default;

  int* data_;
  int size_;
  int capacity_;
}
auto vec_short = my_vec(2);
auto vec_short2 = std::move(vec_short);

my_vec::my_vec(my_vec&& orig) noexcept
: data_{std::exchange(orig.data_, nullptr)}
, size_{std::exchange(orig.size_, 0)}
, capacity_{std::exchange(orig.size_, 0)} {}

std::vector<int> - Move assignment

  • Like the move constructor, but the destination is already constructed
class my_vec {
  // Constructor
  my_vec(int size): data_{new int[size]}, size_{size}, capacity_{size} {}
  
  // Copy constructor
  my_vec(my_vec const&) = default;
  // Copy assignment
  my_vec& operator=(my_vec const&) = default;
  
  // Move constructor
  my_vec(my_vec&&) noexcept = default;
  // Move assignment
  my_vec& operator=(my_vec&&) noexcept = default;

  // Destructor
  ~my_vec() = default;

  int* data_;
  int size_;
  int capacity_;
}
auto vec_short = my_vec(2);
auto vec_long = my_vec(9);
vec_long = std::move(vec_short);

my_vec& my_vec::operator=(my_vec&& orig) noexcept {
  // The easiest way to write a move assignment is generally to do
  // memberwise swaps, then clean up the orig object.
  // Doing so may mean some redundant code, but it means you don't
  // need to deal with mixed state between objects.
  ranges::swap(data_, orig.data_);
  ranges::swap(size_, orig.size_);
  ranges::swap(capacity_, orig.capacity_);
  
  // The following line may or may not be nessecary, depending on
  // if you decide to add additional constraints to your moved-from object.
  orig.clear();
  return *this;
}

void my_vec::clear() noexcept {
  delete[] data_
  data_ = nullptr;
  size_ = 0;
  capacity = 0;
}

Passing references to be copied

struct S {
  // modernize-pass-by-value error here
  S(std::string const& x) : x_{x} {}
  std::string x_;
};

auto str = std::string("hello world");
auto a = S(str);
auto b = S(std::move(str));

Consider the following code

  • When we construct "a"
    • We create a const reference
    • We copy it into x_
  • When we construct "b"
    • We create a const reference
    • Since we have a const reference, we cannot move from it
    • We copy it into x_

Passing references to be copied

struct S {
  // modernize-pass-by-value error no longer here
  S(std::string x) : x_{std::move(x)} {}
  std::string x_;
};

auto str = std::string("hello world");
auto a = S(str);
auto b = S(std::move(str));

Now consider the following

  • When we construct "a"
    • We create a temporary object by copying
    • We then move that temporary copy
  • When we construct "b"
    • We create a temporary object by moving (since the argument is an rvalue)
    • We then move that temporary copy
  • It turns out that moving from temporary objects is something the compiler can pretty trivially optimise out
  • This should be the same performance for lvalues, but allow moving instead of copying for rvalues

Explicitly deleted copies and moves

  • We may not want a type to be copyable / moveable
  • If so, we can declare fn() = delete
class T {
  T(const T&) = delete;
  T(T&&) = delete;
  T& operator=(const T&) = delete;
  T& operator=(T&&) = delete;
};

Implicitly deleted copies and moves

  • Under certain conditions, the compiler will not generate copies and moves
  • The implicitly defined copy constructor calls the copy constructor memberwise
    • If one of its members doesn't have a copy constructor, the compiler can't generate one for you
    • Same applies for copy assignment, move constructor, and move assignment
  • Under certain conditions, the compiler will not automatically generate copy / move assignment / constructors
    • eg. If you have manually defined a destructor, the copy constructor isn't generated
  • If you define one of the rule of five, you should explictly delete, default, or define all five
    • If the default behaviour isn't sufficient for one of them, it likely isn't sufficient for others
    • Explicitly doing this tells the reader of your code that you have carefully considered this
    • This also means you don't need to remember all of the rules about "if I write X, then is Y generated"

RAII (Resource Acquisition Is Initialization)

  • A concept where we encapsulate resources inside objects

    • Acquire the resource in the constructor​
    • Release the resource in the destructor

    • eg. Memory, locks, files

  • Every resource should be owned by either:

    • Another resource (eg. smart pointer, data member)

    • The stack

    • A nameless temporary variable

Object lifetimes

To create safe object lifetimes in C++, we always attach the lifetime of one object to that of something else

  • Named objects:
    • A variable in a function is tied to its scope
    • A data member is tied to the lifetime of the class instance
    • An element in a std::vector is tied to the lifetime of the vector
  • Unnamed objects:
    • A heap object should be tied to the lifetime of whatever object created it
    • Examples of bad programming practice
      • An owning raw pointer is tied to nothing
      • A C-style array is tied to nothing
  • Strongly recommend watching the first 44 minutes of Herb Sutter's cppcon talk "Leak freedom in C++... By Default"

Object lifetime with references

  • We need to be very careful when returning references.
  • The object must always outlive the reference.
  • This is undefined behaviour - if you're unlucky, the code might even work!
  • Moral of the story: Do not return references to variables local to the function returning
auto okay(int& i) -> int& {
  return i;
}

auto okay(int& i) -> int const& {
  return i;
}
auto questionable(int const& x) -> int const& {
  return i;
}

auto not_okay(int i) -> int& {
  return i;
}

auto not_okay() -> int& {
  auto i = 0;
  return i;
}

COMP6771 20T2 - 5.1 - Resource Management

By cs6771

Made with Slides.com

COMP6771 20T2 - 5.1 - Resource Management

  • 936

More from cs6771

  • COMP6771 21T2 - 1.2 - Intro & Types
    cs6771
    520
  • COMP6771 21T2 - 2.1 - C++ Standard Library
    cs6771
    568