Readings: Effective Modern C++

Title: Effective Modern C++

42 specific ways to improve your use of C++11 and C++14

Author: Scott Meyers

General Review

Chapter 1: Deducing Types

Item 1: Understand template type deduction

This item is the basis to understand how type deduction(auto) works in C++. auto type deduction is uses the same principals as template type deduction.

Template function type deduction

template<typename T>
void f(ParamType param);


f(expr)

Function templates have two main “variables”, T and ParamType. T is usually the type that of expr (except for the cases below) and ParamType is the parameter type declared by the function. ParamType usually contain const and reference qualifiers.

Trivial example

template<typename T>
void f(const T& param);

int x = 0;
f(x);

For this example T is int while ParamType is const int&

Edge Case 1: ParamType is a Reference or Pointer, but not a Universal Reference

Occurs when ParamType (function parameter declaration) is a reference or pointer.

Type deduction

• If expr type is a reference, ignore reference part. (do not care if the argument provided is a reference or value)
• Pattern-match expr’s type against ParamType to determine T
  • Will try to match expr to the declared ParamType to determine what T should be

Example

template<typename T>
void f(T& param);

int x = 27;
const int cx = x;
const int& rx = x;

• f(x)
  • T is int
  • ParamType is int&
• f(cx)
  • T is const int
  • ParamType is const int&
  • To match cx, will need to set T to const int
• f(rx)
  • T is const int
  • ParamType is const int&
  • The type of T is different from rx, this is due to reference part being ignored in function template

Edge Case 2: ParamType is a Universal reference

Universal reference looks like rvalue reference syntax(T&&) but it behaves differently when lvalue reference is passed as an argument

Type deduction for universal reference:

• If expr is an lvalue, both T and ParamType are deduced to be lvalue reference (not value).
  • Only time when T is deduced to be a reference
  • Even though ParamType is declared as rvalue reference, it is actually a lvalue reference
• If expr is an rvalue, the normal(parameter is a rvalue reference)

Example

template<typename T>
void f(T&& param); // param is a uni reference

int x = 27;
const int cx = x;
const int& rx = x;

• f(x)
  • T is int&
  • ParamType is int&
  • Due to x being lvalue
• f(cx)
  • T is const int&
  • ParamType is cosnt int&
  • Due to cx being lvalue
• f(rx)
  • T is const int&
  • ParamType is cosnt int&
  • Due to rx being lvalue
• f(27)
  • T is int
  • ParamType is int&&
  • Due to 27 being a rvalue

Notes:

• Universal references treats lvalue and rvalue differently

Edge Case 3: ParamType is neither a pointer nor a reference

template<typename T>
void f(T param)

• The parameter is passed by value

Type deduction

• If expr type is a reference, ignore the reference part
• If expr is const, ignore the const part
• If expr is volatile, ignore the volatile part
• For pass by value, we can ignore all the qualifiers as the parameter is totally independent from the argument provided.

int x = 27;
const int cx = x;
const int& rx = x;

f(x); // T and ParamType are both int
f(cx); // T and ParamType are both int
f(rx); // T and ParamType are both int

Special case of const pointer to const object

template<typename T>
void f(T param);

const char* const ptr = "Fun with pointers";

f(ptr);

• The constness of ptr (RHS const) will be ignored but the type of pointer (LHS const) will be preserved
• You can change ptr to point to another data but you cannot change the data that ptr points to.

Array Arguement

const char name[] = "J. P. Briggs"; // name type is const char[13]
cosnt char* ptrToName = name;       // array decays to ptr

template<typename T>
void f(T param);

f(name) // T deduced as const char *

• When an array is provided as an argument, the type will decay to pointer instead of array as ParamType.
• No such thing as function parameter that is array.

Caveat: the parameter is passed by reference

template<typename T>
void f(T& param);

f(name); // T deduced to const char[13]

• Useful if you want to determine the size of the array

  template<typename T, std::size_t N>
  constexpr std::size_t arraySize(T (&)[N]) noexcept {return N;}
  

Function argument

When passing a function as argument, it will decay to function pointers

void someFunc(int, double);

template<typename T>
void f1(T param);

template<typename T>
void f2(T& param);

f1(someFunc); // param deduced as ptr-to-func void (*)(int, double);
f2(someFunc); // param deduced to ref-to-func void (&)(int, double);

Item 2: Understand auto type deduction

This item covers how auto type deduction works for the majority of the time and the corner cases where it will behave differently.

Template type deduction and auto

During type deduction for auto, auto acts like T in template type deduction. Similar to T, auto can be surrounded by qualifiers (ie const, & or &&). You can see auto type deduction as wrapping a variable in a function template.

// auto
const auto& rx = 27;
// same as wrap in function template

template<typename T>
void func_for_rx(const T& rx) {

}

For universal reference, auto uses auto&& syntax to represent that it could either be lvalue or rvalue reference depending on the provided expression.

auto&& uref1 = x; // x is int and lvalue so uref1's type is int&
auto&& uref2 = cx; // cx is const int and lvalue, so uref2's type is const int&
auto&& uref3 = 27; // 27 is int and rvalue, so uref3's type is int&&

Motivation: The main reason for using auto&&, is when you want to preserve the rvalueness of a variable. This allows you to move the resource of the variable or forward it to another function.

Edge case (initializer list)

The only difference between auto type deduction and template type deduction is when braces ({}) are used.

For auto: braces are deduced as std::initializer_list. Where each element is of the same type

auto x4 = { 27 }; // type is std::initializer_list<int>, value is {27}
auto x5 = {1,2, 3.0} // error! cant deduce the type

For template type deduction: cannot deduce the type for expression surrounded by {}

template<typename T>
void f(T param);

f({1,2,3}); // error!

Note: There is no reason why there is a difference in their behaviour but it is what it is.

auto return type or lambda params

If auto is used as return value of functions or parameters of lambda functions, template type deduction is used instead of auto type deduction. Thus, it would be compilation error to use {} in function return value or lambda argument.

Item 3: Understand decltype

decltype will parrot back the type of a name (variable) or expression. It can be used to specify the type of a variable.

struct A { double x; };
const A* a;
 
decltype(a->x) y; 

Use Cases For function templates, decltype can be used to allow the return type depend on the parameters.

• Use auto at the start of the function to state that trailing return type is used
• Use trailing return type -> decltype(...) to state that the return type depends on the argument.
  • If return type is used at the start, the parameters would not be available

template<typename Container, typename Index>
auto authAndAccess(Container& c, Index i) -> decltype(c[i])
{
  authenticateUser();
  return c[i];
}

• In c++14, the trailing return type will not be required as return type will be deduced with just auto

Caveats:

• For auto return type, type template deduction is used and the reference-ness will be ignored.

decltype(auto)

The combination of decltype(auto) means that type deduction is used (auto) but with decltype type rules (echo the exact type)

Widge w;
const Widget& cw = w;
auto myWidget1 = cw; // auto type deduction: myWidget1 type is Widget
decltype(auto) myWidget1 = cw; // auto type deduction: myWidget1 type is const Widget&

You can use decltype(auto) and std::forward to make sure that the return type of the function preserve the rvalue reference-ness and referenceness

template<typename Container, typename Index>
decltype(auto) authAndAccess(Container&& c, Inex i)
{
  authenticateUser();
  return std::forward<Container>(c)[i];
}

• This will allow the return type to be reference (with decltype(auto))
• Using universal reference (&&) allows both rvalue and lvalue containers to be provided
• Use std::forward to perfectly forward the type of each element.

Caveats:

Wrapping a name with () will change the type from value to reference

decltype(auto) f1() 
{
  int x = 0;
  return x; // decltype(x) is int
}

decltype(auto) f2()
{
  int x = 0;
  return (x); // decltype((x)) is int&
}

Item 4: Know how to view deduced type

• Using compiler:
  • force a compilation error using the type you want to find
• Runtime output
  • Use std::cout << typeid(x).name() << "\n". Only works if the type yields std::type_info
  • This does not work as expected for all cases. typeid will return the type with the referenceness ignored
• Boost type_id_with_cvr (best option)
  • Will keep the constness and referenceness of it

Chapter 2: auto

Item 5: Prefer auto to explicit type declaration

This item provides multiple reasons why using auto is preferred. (klement: personally I am convinced that auto should be used in a c++ project)

Pros

1. Prevent uninitialized variables.
   • With auto, variables will never be able to be declared without any initialization. This will prevent bugs such as uninitialized integers or accidentally calling the default constructor on an object

    int x; // UB
    auto x = 5;
   
    Foo f; // accidentally calling default constructor
    auto f = Foo(x,y,z);
   

2. Prevent writing types that have long name (ie typename std::iterator_traits<It>value_type)
3. Declare closures
   • Previously, closures (captures surrounding scope) cannot be defined as it is known only to the compiler.
4. lambda with auto instead of std::function
   • Using std::function involves a template instantiation but auto does not
   • auto will use the minimum memory
   • std::function has a fixed size on the heap for any given function. If the size is not enough, it would allocate memory on the heap instead. This could result in OOM.
   • Due to implementation details that results in the that restricts std::function from being inlined, auto closures will definitely be faster
   • The return type of a function might be different on different architecture.
     • std::vector::size() returns 32 bit on 32-bit windows but 64 bit on 64-bit windows
     • Using auto will prevent implicit conversion
   • Prevent accidental implicit conversion each entry of map is pair<const key, val> instead of pair<key, val>. Doing for (pair<std::string, int>& p : m) will actually create a temporary object and return a reference to it.

Item 6: Use the explicitly typed initializer idiom when auto deduces undesired types

This item covers the case where simply replacing explicit type with auto will yield different behaviours. When we use auto for types that are “invisible proxy”, the type will no longer be invisible.

proxy type: Are types that are meant to be intermediate type that represent an underlying type. (ie shared_ptr or unique_ptr)

invisible proxy type: Are proxy types that are meant to live in a single statement and be implicitly converted to the underlying type.

std::vector<bool>: returns std::vector<bool>::reference instead of the traditional bool&.

vector<bool> v(5, true);

v[4] // returns std::vector<bool>::reference

bool b1 = v[4] // returns std::vector<bool>::reference but implicityly converts to bool

auto b2 = func_that_return_bool_vec()[4]; // UB! b2 is std::vector<bool>::reference which reference to a temporary object

• Using auto will prevent the invisible proxy type from converting to the desired bool
• Why vector<bool> return object instead reference of the object: c++ treats vector<bool> differently so that the vector is represented as compactly as possible.

To solve this, we should use static_cast for invisible proxy type

auto b2 = static_cast<bool>(func_that_return_bool_vec()[4]);

Chapter 3: Moving to Modern C++

Item 7: Distinguish between () and {} when creating objects

There are 4 ways to initialize an object:

int x(0); // initializer is in parentheses
int y = 0; // initializer follows "="
int z{0}; // initializer is in braces

Note that using = on an uninitialized variable calls the copy constructor instead of the copy assignment constructor. (klement: I actually made this mistake during an interview…)

Foo foo; //default constructor
Foo bar = foo; // copy constructor instead of copy assignment

() vs {} C++11 introduced uniform initializer({}) which aims to set a standard to initialization without using the braces syntax.

Trade offs:

• implicit conversion:
  • {} does not allow implicit narrowing conversion (note that non-narrowing conversion is allowed) while () allows narrowing conversion. Bugs due to narrowing conversion could be easily avoided when we use {}
• C++ most vexing parse
  • most vexing parse: Anything that can be parsed as a declaration can be interpreted as one.
  • This means that calling default constructor with parentheses will result in it declaring a function instead of calling the default constructor (klement: this is an interesting behaviour that is bug prone)
  • This issue will not occur if you use braces initializer

    Widget w2(); // declares a function w2 that returns Widget
    Widget w3{}; // calls default constructor
    

• std::initializer_list
  • braces initializer:
    • If a class has overloaded constructor with one that takes std::initializer_list, it will always prefer to use the constructor with std::initializer_list.
      • Can occur to any constructor (ie copy constructor, move constructor, …) as long as there is a non-narrowing conversion
    • This will occur when there exist an implicit widening conversion for all elements in the braces
    • Impact: suddenly adding a new std::initializer_list constructor would break all of the callers code
    • Caveat: if the braces is empty, it would call the default constructor instead of the initializer_list constructor with no elements. To call the initializer_list constructor with no element, you will need to use Widget w4({});
  • parentheses initializer: does not suffer from this issue at all

    class Widge {
      public:
      Widget(int i, bool b);
      Widget(int i, double d);
      Widget(std::initializer_list<long double> il);
    };
    Widget w1(10, true)
    

Item 8: Prefer nullptr to 0 and NULL

The issue of using 0 and NULL to represent a null pointer is that 0 underlying type is an integer and NULL underlying type is an integral type (depending on compiler). This would result in

• Overloaded function with one that accepts integral type and another that accepts a pointer, will always call the integral type function even though it might be meant to be a pointer

  void f(int); 
  void f(boo);
  void f(void *);
  f(0); // calls f(int)
  f(NULL); // depending on compiler might call f(int) 
  f(nullptr) // always call f(void *);
  

• Improved readability if auto is used
• For template functions, 0 and NULL will always be deduced as int and integral type instead of void *

Item 9: Prefer alias declarations to typedefs

typedef std::unique_ptr<std::unordered_map<std::string, std::string>> UPtrMapSS;
using UPtrMapSS = std::unique_ptr<std::unordered_map<std::string, std::string>>;

We should use alias (using) instead of typedef because of the following reasons:

• It is much readable for types that are function pointers

  typedef void (*FP)(int, const std::string&);
  using FP = void(*) (int , const std::string&);
  

• Template alias are allowed

  // alias
  template<typename T>
  using MyAllocList = std::list<T, MyAlloc<T>>;
  
  MyAllocList<Widget> lw;
  
  // typedef
  template<typename T>
  struct MyAllocList {
    typedef std::list<T, MyAllocT<T>> type
  };
  MyAllocList<Widget>::type lw;
  

• Template alias with template class
  • With typedef in template class, you will need to use typename as the templated typedef could have full specialisation. This means that it could declare type as something else (might not be a type). This will result in the typedef template in template class being a dependent type.
  • For alias template, it will definitely be a type so there is no dependencies

  template<typename T>
  class Widget {
  private:
    typename MyAllocList<T>::type list;
  }
  
  template<typename T>
  using MyAllocList = std::list<T, MyAlloc<T>>;
  
  template<typename T>
  class Widget {
  private:
    MyAllocList<T> list;
  }
  

Item 10: Prefer scoped enums to unscoped enums

Unscoped enums:

enum Color {
  black,
  white,
  red
};

Scoped enums:

enum class Color {
  black,
  white,
  red
};

Enumerators: are the entities within each enum. Each enumerator can be assigned a range of values.

Unscoped enums disadvantages:

• Implicit conversion
  • unscoped enums can be implicitly converted to integral types and then to floating point types. This will result in code that does not make sense (treating enums as floats)
• Namespace pollution:
  • For unscoped enums, the enumerator (inside the enum) share the same namespace as the enumerator.
  • This does not occur for scoped enums

    enum Color {
    black,
    white,
    red
    };
    auto white = false; // error! white already declared
    

• Forward declaration
  • For unscoped enums, the underlying integral type is determined by the compiler after viewing all of the range of values for each enumerator. This means that the compiler will not be able to know how to use the enum if it does not know the enumerator. Thus it cannot be forward declared.
  • For scoped enums, the underlying integral type is int so the enum can be forward declared
  • Advantage of forward declare: prevent the entire project from recompiling if a new enumerator is added.
  • This can be avoided if you set the default integral type for unscoped enums

    enum Color : std::uint8_t;
    

Advantage of unscoped enums:

• use unscoped enums for getters of tuple

  enum UserInfoField = { uiName, uiEmail, uiReputation };
  std::tuple<std::string, std::string, std::size_t> uInfo;
  ...
  auto val = std::get<uiEmail>(uInfo);
  

Item 11: Prefer deleted functions to private undefined

When you want to remove a function that is automatically generated (ie: copy constructor, implicit conversion, …)

c++98: If the function is a member function, declare it private and do not define the function. Only limited to member functions

c++11:

• use delete keyword

  class Foo {
  public:
    Foo(const Foo&) = delete
  }
  

• delete functions should be public instead of private. In some compiler, the error message would be private function not accessible instead of function deleted

Benefits of delete over private + undefined

• c++98 method of private + undefined will only be caught in link time while c++11 method of delete will be caught in compile time

• Use overloaded delete to prevent implicit conversion. (side note: c++ always prefer to convert float to double instead of int)

  bool isLucky(int num); // allow the arguments to be int, char, float
  
  bool isLucky(char) = delete; 
  bool isLucky(bool) = delete;
  bool isLucky(double) = delete; 
  
  // Lead to compilation error
  isLucky('a');
  isLucky(true);
  isLucky(3.5f);
  

• Prevent use of template specialisation for certain template arguments.

  • Using full specialisation to state that the function is deleted

    template<typename T>
    void processPointer(T* ptr);
    
    template<>
    void processPointer<void>(void*) = delete;
    
    template<>
    void processPointer<char>(char*) = delete
    

  • If the function template is member function, declare the deleted function in namespace scope(outside of the class)

Item 12: Declare overriding functions override

Conditions for derive class overriding base class:

• Base class function must be virtual
• Base and derived function names must be identical
• Parameter type must be identical (no implicit conversion allowed)
• The constness of the base and derived function must be the same
• The return types and exception specification of the base and derived must be compatible
• The reference qualifiers must be identical
  • Member functions can have trailing & or && to overload the function on what type of reference the object is.

    class Widget {
    public:
    using DataType = std::vector<double>;
    ...
    DataType& data() & 
    { return values; }
    DataType&& data() &&
    { 
      // move data instead of copying when it is a rvalue (temp obj)
      return std::move(values); 
    }
    private:
    DataType values;
    } 
    

In c++11, we can declare overriden functions override.

class Base {
public:
  virutal void mf1() const;
};
class Derived : public Base {
public:
  virutal void mf1() const override;
}

Benefits of override keyword:

• Prevent accidentally not overriding base virtual function if any of the condition not met. Throws compilation error.
• Compiler will show the impact when changing the signature of Base class function and recompiling it.

Item 13: Prefer const_iterator to iterator

In C++11, to easily use const_iterator you can call the member functions cbegin() and cend() or std::cbegin() and std::cend() to get the const_iterator. Use the non-member std::cbegin() and std::cend() for generic code.

Item 14: Declare functions noexcept if they won’t emit exceptions

In c++11, functions can be declared noexcept to guarantee that no exceptions will be thrown. The responsibility is on the developer (not the compiler) to ensure that all the code in the body will not throw exception.

Benefits of declaring noexcept:

• The compiler will not keep the runtime stack in an unwindable state. No need for stack trace if there is no exception.
• The compiler will not need to ensure that objects are destroyed in inverse order of construction (only relevant when exception is thrown).
• Standard library uses std::move_if_noexcept. If the move constructor is no except, some functions would move instead of calling copy constructor for optimisation. No except condition is to ensure that these functions still have strong safety guarantee.
  • Usually important for the member functions swap and move constructor

Note: do not false move noexcept. It could come at a cost of good code quality.

Item 15: Use constexpr whenever possible

(klement: I find this item very useful as constexpr is a very new concept to me and scott explains it very well here)

constexpr objects

Are objects that are const and known during compile time (technically translation time: compile time + link time)

• Expressions involving constexpr variables can be constexpr
• Return values of constexpr functions can be constexpr

constexpr functions

• If the arguments provided are constexpr objects, the function will be evaluated at compile time and the return value will be constexpr
• else the function would be evaluated like any normal functions
• Arguments can only be literal types
  • Types that can be determined during compilation
  • Objects that have constexpr constructors
• function body requirements
  • C++11:
    • constexpr member functions must be treated as const member function. Cannot change data members.
    • must be single line. However, ?: is permitted.
  • C++14: no restrictions to function body

constexpr benefits

• The objects may be placed in read-only memory
• Evaluated at compile time means that it will have shorter run time
• Can be used as template arguments (ie std::array size)

Item 16: Make const member functions thread safe

To a caller, const member functions should be thread safe as it only involves reading of data. However, sometimes you would wish to add a layer of caching for const member functions. You can do so by setting mutable (can be modified by const member functions) data members. Thus, it is important to make sure that it is thread safe.

How to add cache:

• atomic: use std’s atomic variables. Should only use if there is only 1 mutable data member.
• std::mutex: normal lock and unlock pattern

Item 17: Understand special member functions

Generated functions

• default constructor
  • Generated if the class does not have user declared function
• copy operations
  • copy constructor
    • Deleted if the class declares a move operation
  • copy assignment
    • Deleted if the class declares a move operation
• move operation
  • move constructor
  • move assignment
  • perform memberwise move on the current object and base class part
  • non-move enabled data members will be copied instead
  • Generated only if the class does not define any of the copy operation or move operation or destructor
• destructor

Conditions for generated functions

• Copy operations are independent: declaring one does not prevent compilers from generating the other
• Move operations are not independent: declaring one prevents compilers from generating the other

Item 18: Use std::unique_ptr for exclusive ownership resource management

Exclusive ownership:

• Use to represent exclusive ownership
• OnDly supports move semantics (do not allow copy)
• functions can return std::unique_ptr. Return in function to represent the ownership transfer from callee to caller

Extra functionallity

• Most of the time std::unique_ptr have the same size as raw pointers
• Support custom deleter
  • Accepts the raw pointer as arguments.
  • Need to manually call the delete.
  • If the deleter captures the surrounding scope, the size will not be the same

    auto delInvmt = [](Investment* pInv) {
    makeLogEntry(pInv);
    delete pInv;
    }
    std::unique_ptr<Investment, decltype(delInvmt)>pInv(nullptr, decltype)
    

• Use ::reset(T*) to reset the underlying pointer the unique pointer points to
• Can easily change to std::shared_ptr
• Allow single object and array to be wrapped
  • unique_ptr<T>
    • call delete
    • allow -> syntax
    • do not allow [] syntax
  • unique_ptr<T[]>
    • call delete[]
    • allow [] syntax
    • do not allow -> syntax

Item 19: Use std::shared_ptr for shared ownership resource management

Functionality:

• Represent a shared ownership of a resource
• When the last object stop pointing to the resource, it will destroy the resource
• Supports custom deleter like unique_ptr
  • Do not need to state the type custom deleter

    auto deleter =[](){ ... }
    std::shared_ptr<Widget> pw(new Widget, deleter)
    

  • Custom deleter does not change the size of shared_ptr
• Unlike unique_ptr, shared pointer does not support T[]

Under the hood:

• Reference counting:
  • Most constructors(non-move) will increment the reference count
  • Move constructors will not change the reference count.
    • Move the reference count from one to another
    • Faster than non-move (dont need atomic operations)
  • copy assignment will increment the rhs reference count and decrement the lhs reference count
  • when the reference count becomes 0, it will destroy the underlying resource
• Performance limitations
  • twice the size of normal raw pointers
  • reference count are dynamically allocated (additional overhead)
  • increment and decrement are atomic

std::shared_ptr control block

• Each shared_ptr should have a control block that stores metadata for it
• metadata
  • reference count
  • weak count
  • custom deleter
  • allocator
• A new control block created when:
  • std::make_shared is used
  • when std::unique_ptr or std::auto_ptr is used to construct shared_ptr
  • when raw pointer pass as argument to the shared_ptr constructor
• Limitation: Each raw pointer address should only have a unique control block.
  • Virtual function is involved which would lead to more latency

UB: multiple control block for a shared pointer

When a single address is used in different shared_ptr with different control block, there will be different reference count of different value and delete will be called twice on the same address

auto pw = new Widget;
std::shared_ptr<Widget> spw1(pw, logginDel);
std::shared_ptr<Widget> spw2(pw, logginDel);

Usually happen when using this. The member function do not know if the current instance is already in a shared pointer

class Widget {
private:
  std::vector<std::shared_ptr<Widget>> processWidget;
public:
  void process() {
    processWidget.emplace_back(this);
  }
}

Mitigation: std::enable_share_from_this<T>

• virtual function this->shared_from_this() that will make sure that there will only be one control block for this
• make all constructors private and have a factory function that returns a shared_ptr from shared_from_this(). This will make sure that external callers will not have access to this

Item 20: Use std::weak_ptr for std::shared_ptr like pointers that can dangle

Motivation: smart pointers that acts like shared pointer but does not take part in shared ownership

std::weak_ptr properties:

• Allowed to dangle: the object on the heap could be destoryed
• Can only be created from std::shared_ptr

  std::shared_ptr<Foo> sptr = make_shared();
  std::weak_ptr<Foo>wptr(sptr);
  

  • Will not increase the reference count of shared_ptr
• Can check if the underlying object on the heap has been destoryed (not possible with other smart pointers)
  • call std::weak_ptr::expired
• Converted back to shared_ptr: regain the shared owner ship of the underlying object
  • std::weak_ptr::lock: returns the corresponding shared_ptr<Foo> if object still on heap or nullptr if the object has been deleted.
  • std::shared_ptr<T>(std::weak_ptr<T>): use the weak pointer as an argument for shared pointer. Throws an exception if the underlying object that the weak pointer points to has been deleted.

Uses cases:

• Caching objects on the heap:
  • Stores a containers of weak pointers to the object. Retrieve from cache by converting the weak pointer back to shared pointer
• Prevent cyclic reference:
  • Shared pointers will not be able to solve dangling cyclic reference.
  • If the leaves need a reference back to the parent, use weak pointers to prevent cyclic reference
    • If the child life time will always be shorter than the parent, it is okay for the child to have a raw pointer back to the parent. When a child is a live the raw pointer to parent will always be alive
• Observer design pattern:
  • Observer subscribe to subjects by storing a pointer to the observer as a data member in the subject. Subject will call the member of functions to notify them
  • Subjects to store a weak pointer to observer:
    • Do not do anything if the weak pointer has expired
    • Convert to shared pointer and notify if the observer is not destroyed
    • Prevent the subject from affecting the life time of the observer

Efficiency: Exactly same as shared pointer

Item 21: Prefer std::make_shared and std::make_unique to direct use of new

Resource leak from constructor and new

• When:
  1. A function is called with {unique/shared}_ptr(new T) and expcetionFunction() as argument
  2. Compilers are allowed to reorder the argument instruction if there are no dependencies
  3. Instructions could be reorder to: new T -> exceptionFunction -> smart pointer exception
  4. This will result in resource leak as new T is stored in smart pointer and will not be deleted
• Wrapping the allocating of the resource + constructor of the smart pointer in a function, will prevent bad interleaving of instructions.

make_shared is faster:

• shared pointer will need to allocate memory for both the object and control block
• Using make_shared will allocate both the object and control block together
• One allocation is required instead of allocating the object first then the control block
  • Faster and smaller code
  • Control block and object in a contiguous memory location -> spatial locality

make_* limitations:

• Cannot specify custom deleter
• Will not be able to perfectly forward braced initializer

  auto upv = make_unique<vector<int>>(10, 20); 
  // calls vector<int>(10,20) instead of vector<int>{10,20}
  

• Custom newed object do not work well make_shared
• As control block and the object are in the same chunk of memory, the chunk of memory can only be freed when the reference count (number of shared pointer) and weak count (number of weak pointer) goes to 0.
  • The chunk memory can still be around when the object has been destructed (as long as the number of dangling weak pointers).
  • Using new T on shared pointer constructor will have two different chunk of memory and the object’s chunk memory can be immediately freed when no more shared pointers points to the object
  • (klement: I was not aware that the freeing of memory might not occur immediately after the destructor is called)

Mitigation for new T and shared pointer constructor

• Construct the shared pointer and new T in an independent line -> prevent resource leak
• pass the shared pointer using move to preserve rvalueness

Item 22: When using the Pimpl Idiom, define special member function in the implementation file (.cpp file)

(klement: This items goes in-depth on how to properly utilize the pimpl idiom. This is very crucial as there are not many resources out there (including effective cpp) that properly show and explain how to do so.)

Motivation: As a library author, if a client include your library header files, they will also include the headers that your header files include. This will result in them having to recompile all of their code if the indirect headers file change.

Proper pimpl example:

// In header (.h) file
class Widget {
private:
  struct Impl;
  std::unique_ptr<Impl> pImpl;
public:
  Widget();
  ~Widget();

  Widget(const Widget& rhs);
  Widget& operator=(const Widget& rhs);

  Widget(Widget&& rhs);
  Widget& operator=(Widget&& rhs);
};

// In implementation (.cpp) file
#include "widget.h"
#include "gadget.h"
#include <string>
#include <vector>

struct Widget::Impl {
  std::string name;
  std::vector<double> data;
  Gadget g1, g2, g3;
}

Widget::Widget() : pImpl(std::make_unique<Impl>()) {}
Widget::~Widget() = default; // force compiler to generate destructor after seeing Widget::Impl

Widget(Widget&& rhs) = default; // moving of underlying impl is desired behaviour
Widget& operator=(Widget&& rhs) = default; // ditto

// perform deep copy of the underlying impl
// do so by using the impl copy operations
Widget(const Widget& rhs) : pImpl(std::make_unique<Impl>(*rhs.impl)) {}
Widget& Widget::operator=(const Widget& rhs)
{
  *pImpl = *rhs.impl;
  return *this;
}

Impl:

• Declare Impl in header file but do not define it so that you do not need to need to include the dependencies for the definitions
• Any compiler generated functions (destructor, move, copy) that require the underlying implementation should also be declared but not defined (incomplete type) in the header. This will make sure that when the compiler see these function it would have already seen the implementation type.
  • Under the hood, destructor is the only operation that does not work without declaration. Without destructor, all other generated function would not be generated

unique_ptr: the underlying data has to be allocated on the heap. Use unique_ptr to easily manage the pointer

Destructor and move operations:

• The generated function behaviour for these function are desired but we have to declare them in the header and define them as =default in implementation.
• By default compiler will inline these functions and generate the functions at the header file. By declaring them, the compiler will go to the implementation file to find the definition and default to default after seeing = default

Copy operations:

• Copy operations should be deep copy instead of shallow copy (copying the pointer but still point to the same address)
• Instead of manually copying the data member outisde of impl, call the underlying impl copy operations.

Using shared_ptr instead of unique_ptr

• Destructor for shared_ptr are not part of the shared_ptr type (stored at runtime control block)

Chapter 5: Rvalue Reference, Move Semantics and Perfect Forwarding

Item 23: Understand std::move and std::forward

Both std::move and std::forward do not perform any “moving” and do not have any executable code at runtime.

std::move

Performs unconditional cast to rvalue reference

template<typename T>
decltype(auto) move(T&& param)
{
  using ReturnType = remove_reference_t<T>&&;
  return static_cast<ReturnType>(param)
}

• T&& in function template means that it can be of any reference
• move remove all reference and cast it back to the rvalue (unconditional cast)

std::move on a const variable as overloaded function argument. Call the function in the following order

1. Calls the param with const T&&
2. Calls const T&: const lvalue reference binds to const rvalue reference
   • const rvalue reference will never bind to non-const rvalue reference
   • Converse is ok -> binding non-const rvalue reference to const rvalue reference
   • Should only move non-const variable otherwise will usually bind to const lvalue reference
   • std::move does not guarantee moving.

std::forward

Problem: All function parameters are lvalue as they exist as a variable in the parameters (eventhough might be temporary on call-site). Any function call within a function body with the parameters will be rvalue.

Solution: std::forward will cast the function template param to rvalue if it is rvalue on the call-site

• The rvalue-ness is encoded in the function template param T
• std::forward will conditionally cast to rvalue if the argument is rvalue.

Item 24: Distinguish universal references (forwarding reference) from rvalue references

Universal reference: references that can be rvalue or lvalue, const or non-const, volatile or non-volatile

Conditions for universal reference: type deduction must occurs specific to the function (ie function template or auto)

Function template:

template<typename T>
void f(T&& param)

• When the template is specific to the function. Type is deduced every time the function is called.
  • Function inside template class do not perform type deduction every time it is called ```cpp template<class T, class Allocator = allocator> class vector { public: void push_back(T&& x); }

  class vector { public: void push_back(Widget&& x); // no type deduction } ```

• Must be of the form T&&. No qualifiers are allowed (ie const T&&)

auto&&

auto&& var2 = var1;

auto fn = [](auto&& func, auto&&... params) {
  ...
}

• Using auto&& in statement will set the type to whatever the expression is.
• Using auto&& in a lambda param uses type deduction and will deduce the type it is.

Side Note: universal reference + string literal

• Passing string literal ("...") to a function taking universal reference will be deduced to array of characters const char[N]. This will allow the param to point to the string in wherever the literal is stored (usually read only memory) instead of creating a temporary string.
• Using std::forward on the deduced const char[N] will allow the string to only be created at where it is really needed

Item 25: Use std::move on rvalue references, std::forward on universal references

• If the parameter is an rvalue, it can and should be moved. Propagate the rvalueness using std::move for rvalue and std::forward for universal reference.
• function returning ref params by value
  • Should propagate rvalueness in the return because there would be no RVO (see below)
  • If the param is rvalue it would call the move constructor for the temp variable instead of having to copy the temp object into the return value

    Matrix operator+(Matrix&& lhs, const Matrix& rhs) {
    lhs += rhs;
    return std::move(lhs);
    }
    template<typename T>
    Fraction reduceAndCopy(T&& frac) {
    frac.reduce();
    return std::forward<T>(frac);
    }
    

  • If the object does not have move constructor, return std::move(lhs) will default to calling copy constructor (non-intrusive)

Return Value Optimisation (RVO)

RVO: compilers are allowed by constructing the return value in memory allocated (call-site) for the function’s return value. This will prevent the need to copy or move constructor.

Conditions:

• The local object is the same as returned by the function
• The local object is what’s being returned
  • Value parameters are considered local variable and returning them can be RVO’ed
  • DO NOT return std::move as it will return a reference to local object and not satisfy the condition. The copy will not be elided and will call the move constructor instead.

Note: If the compiler does not perform RVO, it will call return std::move so that the return value will be moved constructed.

Item 26: Avoid overloading on universal references

If a function is overloaded with universal reference and an explicit type, the universal reference function would be called most of the time.

• Universal reference are greedy: can bind to any type and any qualifier
  • Unless the argument function matches the other function completely (type + qualifier)
  • if the provided argument is a narrowing of the overloaded non-universal function, the compiler will choose the universal reference function instead.
• Compilers will still generate the copy move constructor even if a perfectly forwarding constructor is declared

Item 27: Familiarize yourself with alternatives to overloading on universal references

Active Recall:

• What is tag dispatch
• How do you constraint template functions
• What are the problems of std::is_same<CustomType, T>

Avoid overloading on universal reference

• Abandon overloading
• Pass by const T&
• Pass by value

Tag dispatch idiom

Overview:

• An entry non-overloaded function that takes in arguments as forwarding reference and dispatches it to overloaded implementation function.
• It dispatches to the right overloaded function by passing identifying arguments to the overloaded function.

// entry function
template<typename T>
void logAndAdd(T&& name)
{
  logAndAddImpl(
        std::forward<T>(name),
        std::is_integral<typename std::remove_reference<T>()::type>;
      );
}

// overloaded function
// calls this function if not integral
void logAndAddImpl(T&& name, std::false_type)
{
  auto now = std::chrono:;system_clock::now();
  log(now, "logAndAdd");
  names.emplace(std::forward<T>(name));
}

// calls this function if is integral
void logAndAddImpl(int idx, std::true_type)
{
  logAndAdd(nameFromIdx(idx));
}

• logAndAdd(T&& name) is the non-overloaded entry function that takes the forwarding reference as an argument.
• Pass an additional argument is_integral to the overloaded functions
  • Overloading works here as the true equivalent inherits from std::true_type while false equivalent inherits from std::false_type
  • This allows overloading on the same return type of is_integral
• Application Specific: We can reduce duplicate code having the overloaded function transform the parameter to the base overloaded function (similar to logAndAddImpl(int idx, std::true_type))

Constraining templates that take universal references

Use std::enable_if to enable functions for certain type of argument provided to forwarding reference.

• Allow for disabling perfect forwarding for certain types of arguments

Use Case:

• When we want to have a forward reference constructor for all types except for itself
• If the argument is itself of a derive class of it, the copy constructor should be called instead of forward reference constructor
• If the argument provided can be narrowed or widen to fit another constructor, the type should be narrowed/widened instead of calling the forwarding reference constructor

class Person {
public:
  template<
    typename T,
    typename = std::enable_if_t<
      !std::is_base_of<Person, std::decay_t<T>>::value
      &&
      !std::is_integral<std::remove_reference_t<T>>::value
    >
  >
  explicit Person(T&& n) : name(std::forward<T>(n)) { ... }
  explicit Person(int idx) : name(nameFromIdx(idx)) { ... }

  // default copy constructor generated by compiler
}

• Use enable_if to perform custom condition checking on the provided template argument
• Use is_base_of instead of is_same to allow the check to apply to all of its type and derived types
• Use decay_t to remove any const or reference qualifier

Trade offs:

• use static_assert on universal reference if you know the specific types you want
  • Pair with static_assert(std::is_constructible<std::string, T>::value, "Some error msg")

Item 28: Understand reference collapsing

This item goes into the details on how forwarding reference actually work.

Reference collapsing rule: When there is a reference to another reference if either is an lvalue reference, the result is an lvalue reference. Otherwise (both rvalue reference), the result is an rvalue reference.

Forwarding reference under the hood:

• Forwarding reference actually takes an rvalue reference argument
• Type T deduction
  • When an lvalue argument is provided: T is Widget&
  • When an rvalue argument is provided: T is Widget
• The underyling type of the parameter for
  • lvalue argument => Widget& && => Widget& (one rvalue)
  • rvalue argument => Widget&& (no more because it is simple reference)

This ties nicely with std::forward under the hood:

template:

template<typename T>
T&& forward(typename remove_reference<T>::typename& param) {
  return static_cast<T&&> param;
}

When lvalue argument provided:

Without collapsing

Widget& && forward(typename remove_reference<Widget&>::type& param) {
  return static_cast<Widget& &&>(param);
}

After collapsing (collapse to lvalue as one of the reference is lvalue)

Widget& forward(Widget& param) {
  return static_cast<Widget&>(param);
}

When rvalue argument provided:

No need for collapsing (substitute T with Widget)

Widget&& forward(typename remove_reference<Widget>::type& param) {
  return static_cast<Widget&&>(param);
}

Reference collapsing application:

• Template instantiation as seen above
• auto: similar to template instantiation
• typedefs
• decltype

Item 29: Assume that move operations are not present, not cheap, and not used

Active recall:

• When will passing an rvalue as a constructor argument not trigger move semantics?
• Can data on the stack be moved?

• Which stl containers benefits from move semantics?

• Types that do not support move operations will not benefit when rvlaue is provided
  • The type disabled move operations by declaring copy operations or destructors
  • The type has data members that has deleted move operations
• Underlying data members are not on the heap
  • If the data members are on the heap, copy pointer instead of underlying data type
  • If the data members are on the stack cannot copy
  • Example:
    • std::array: moving will perform move on each elements to the new array
    • std::string: for small string (< 15 chars), the strings are stored as data members on the stack instead of the heap.
• Some stl containers will only move if no except to ensure strong safety (due to legacy reason)
  • std::vector: when resizing the data members will be move_if_noexcept to ensure strong safety

Item 30: Familiarize yourself with perfect forwarding failure cases

Active Recall:

• What are the scenarios where forwarding fails?
• Can you take a reference to a bit field?

Perfectly forwarding reference fails when:

• Compilers are unable to deduce type
• Compilers deduce the “wrong” type

Braced initializer

• Standard explicitly states that {...} are non-deduced types
  • Note: this is different from auto where {...} is deduced to be std::initializer_list

0 or NULL as null pointer

• This could result in run time error as 0 or NULL will be deduced to integral types

Declaration-only integral static const and constexpr data members

• Data members that are declared as static constexpr LiteralType v = 28; are not stored in memory if they are not defined later.
• You can pass v to any function that accepts that type
• If you would like to forward v through a template function accepting a forwarding reference, it would fail as v is not in memory and there is no reference to it.
• To resolve this, just define the static variable constexpr LiteralType Widget::v;

Overloaded function names and template names

• When overloaded function are passed as function pointers, compilers will resolve them by matching the correct function signature

  int processVal(int value);
  int processVal(int value, int priority);
  void f(int (*pf)(int));
  f(processVal); // compiler will pass the pointer to the first function
  

• When forwarding the function through a template function that accepts forwarding reference, the compiler will not know what the intermediate function wants and results in error. This problem also occurs for passing function template.

  fwd(processval); // dont know which processVal to pass as arg
  

• To resolve this, specify the needed type of the function or cast it.

  using ProcessFuncType = int(*)(int);
  ProcessFuncType processValPtr = processVal; // resolve which overloaded function
  fwd(processValPtr);
  fwd(static_cast<ProcessFuncType>(workOnVal));
  

Bitfields

• C++ allows for bitfields that states how many bits a data member should occupy
  • The compiler may or may not perform bit compaction
• similar to non-defined consexpr, we cant take reference to a bitfield which will result in compile error
• Resolve this by copying the value of bitfield. This is safe as all functions cannot accept bitfield as an argument so copying it would not result in invalid argument being forwarded

Chapter 6 Lambda Expressions

Definitions

• Lambda expression: a source code expression using the [..](...){...} syntax
• Closure
  • Closure Class: Each lambda expression will generate a Closure Class in compile time
    • Equivalent to the class of the traditional functor
  • Closure Object: The instantiation of of a closure object
    • There could be multiple closure object with the same closure class by invoking the copy ctor
    • Equivalent to a functor
  • Example:

    std::find_if(v.begin(), v.end(), [](int val){ return 0 < val && val < 10; })
    

    • The lambda expression will create a lambda class (compile time)
    • The lambda class will instantiate a lambda object at runtime and pas it as the third argument

Item 31: Avoid default capture modes

C++ provides 2 types of default capture modes:

• [&]: capture all local non-static variables or parameters. However it has the following disadvantages:
  • The captured variables will only live for the duration of the function (stored on stack)
  • However, the closure object can live beyond the stack (ie by copying it to heap with adding to vector)
  • This will result in closure object having a dangling reference
  • Using non default reference capture [&var] will also lead to dangling reference but developers will be more aware of it.
• [=]: capture by value of all local non-static variables. Disadvantages:
  • False sense of security by making the lambda look self-contained
  • If the lambda is instantiated in a member function, it will capture this (pointer to the current object)
    • Will result in dangling pointer of this when object is destructed
  • Does not copy data members: need to use [m_data = m_data] (generalized lambda) syntax to capture data members.
  • Capture static variables by reference: even though it uses capture by value, local static variables are captured by reference. Could result in different behaviour when there static variables change.

Item 32: Use init capture to move objects into closures

Active Recall:

• What is the lifetime of a lambda
• How do you move objects from surrounding scope into a lambda
  • In c++11?
  • In c++14?
• How are init capture variables stored

Init Capture (C++14)

• In C++14 you can use init capture that allows you to declare the data members of the closure class.
• This allows you to move external objects into the data member of the closure object.

auto pw = std:::make_unique<Widget>();
auto func = [pw = std::move(pw)] 
    {return pw->isValidated() && pw->isActive();}

• The lhs of = is the name of the data member of the closure class which is accessible by the lambda function body.
• The rhs of = is the scope of where the lambda is declared
• Does not allow default capture ([=] or [&]) but still allow variable capture by value or reference ([&x] or [x])

std::bind (C++11)

• C++11 does not have access to init capture
• To still move surrounding variables into the lambda, you can wrap the closure object in a std::bind object.
• The bind object takes a callable object as the first argument and a variadic args which would be forwarded to the callable object args when it is called.
  • Check std::placeholders on how to supply arguments when calling a bind object

auto func = std::bind(
  [](const std::vector<doubles>& data) 
  {/** use of data */},
  std::move(data))

• The first argument is a callable object
• Second object is passed to the std::bind which would move-constructed as a data member to the bind object
• The data member will then be passed to the lambda by lvalue reference
• As the bind object is a wrapper around the closure object, you can treat objects in the bind as if they are in the closure

Functor with init capture

You can simulate lambda with init capture by using traditional functor

class IsValAndArch {
public:
  using DataType = std::unique_ptr<Widget>;
  explicit IsValAndArch(DataType&& ptr) : pw(std::move(ptr)) {}

  bool operator()() const 
  { return pw->isValidated() && pw->isArchived(); }
private:
  DataType pw;
}

auto func = IsValAndArch(std::make_unique<Widget>());

Item 33: Use decltype on auto&& parameters to std::forward

Active Recall:

• By convention should the template arguments for std::forward be rvalue reference or lvalue reference

Generic Lambda

`auto f = [](auto x){return ...}`

Generic lambda perfect forwarding

• In c++14, lambdas can accept generic types as parameters
• std::forward will requires the types of the arguments to be supplied as template arguments (ie std::forward<T>).
• For generic lambda, the T is generated in the closure class and not accessible to the lambda expression
• By convention when forwarding forwarding reference (auto&& x):
  • If the argument is a lvalue reference, the template argument should be a lvalue
    • This is because of reference collapsing: & && -> & perfectly forward lvalue
    • Works perfectly with decltype(x) -> lvalue reference
  • If the argument is a rvalue reference, the template argument should be non-reference
    • This is because _ && -> &&
    • Does not work as expected with decltype(x) -> rvalue reference instead of non-reference

Solution:

• use std::forward<delctype(x)>(x)
• rvlaue: still works for rvlaue reference eventhough it does not follow convention
  • reference collapsing of rvalue and rvalue && && -> &&

Item 34: Prefer lambdas to std:bind

Lambdas vs std::bind with callable object

• Lambdas are more readable than std::bind
  • std::bind:
    • To allow arguments for the callable object and capture surrounding variables
    • You will need to use std::placeholders (_1,…) to state which argument should be forwarded from std::bind constructor and which one should be forwarded when the bind object is called
  • lambda: use simple capture syntax
• Arguements are evaluated immediately
  • std::bind:
    • Args for std::bind constructor are evaluated immediately instead of when they are called
• Overloaded function:
  • std::bind: does not know how to resolve overloaded functions and will require casting to the right overloaded function
• Performance:
  • std::bind: callable object is stored as a function object in the bind object data member
    • Calling bind will result in the function pointer being derefenced then called.

Chapter 7 Concurrency API

Item 35: Prefer task-based programming to thread-based

Active Recall:

• What are the ways to perform asynchrnous work in c++?
• What type of threads does c++ create?
• What happens when there are already too many threads and you try to create more
• What are the downside to having a lot of threads

Types of thread

• Hardware Threads: threads that actually perform computation. The maximum number of hardware thread usually is usually equal to the number of processing unit (normally number of cores or 2 times number of cores if it is hyper threading)
• Software Threads: threads that are managed by the operating system across all processes. OS will schedule the software thread to hardware thread. When a software thread is blocked, the OS can schedule other software threads.
• std::thread: act as a handler to the underlying software thread. Some std::thread object could be null handlers when the std::thread is default constructed, moved, joined or detached

Disadvantages of std::thread:

• Could throw std::system_error:
  • Software threads are limited resource and if you try to construct a std::thread when the limit is reach, it will throw std::system_error
  • Could solve this by executing the function on the current thread or wait for existing software threads to complete. (hard to implement)
• Over subscription:
  • Over subscription due to more software threads than hardware threads when using std::thread
  • Context switch into different CPU core:
    • A software thread could be mapped to different core when it is context switch
    • Result in cold CPU cache when context switch into a new CPU
    • CPU cache “pollution” when the new software thread evicts previous (useful) cache entries

Advantage of std::async:

• Acts as a thread manager (decides if a software thread should be created)
• Preventing out of thread exception:
  • If there too many software thread, function will execute the task on the same thread requesting the result
    • Thread calling get or wait
    • Does not create a new thread
    • Could occur when there is over subscription (but not close to limit)
• Allows a return value from the function: get the result of async work
• Use std::launch::async if you would to force a new thread to be created

Item 36: Specify std::launch::async if asynchronicity is essential

Active recall:

• What are the different types of launch policy?
• What is the default type of launch policy?
• Will a function guarantee to execute with the default launch policy?

Launch Policies:

• std::launch::async launch policy: f must be run asynchrnously (ie on a different thread)
• std::launch::deferred launch policy: f may run only when getor wait is called on the future returned by std::async
  • f execution is deferred until get or wait is called and will be executed synchronously
  • If neither wait nor get is called, f will never run
• Default launch policy:
  • Policy is async or deferred
  • f can be executed asynchrnously or synchronously
  • Will decide on which policy to be executed to minimise over subscription and perform load balancing

Implications of default policy:

• Not possible to predict whether f will run concurrently with t. Could be deferred and run on t instead
• Not possible to predict whether f runs on a thread different from the thread invoking get or wait
  • Could create a software thread and mapped to another hardware thread
• Not possible to predict whether f runs at all:
  • Could be deferred and the future is not get or wait which will result in f not being executed at all
• Spin lock with wait_for does not work
  • If f is deemed to be deferred, it could never be executed and future_status will never be ready

    auto fut = std::async(f);
    while (fut.wait_for(100ms) != std::future_status::ready);
    

• Does not work well with thread_local variables
  • thread_local: global variables that are only “local” to each thread
  • As std::async does not guarantee a separate thread will be spawned, f could be in the same thread as the thread that calls .get or wait which will result in the thread_local variable to be accessible by it too.

Mitigations for default policy:

• Spin lock:
  • Use fut.wait_for(0s) == std::future_status::deferred to check what type of policy the library choose to use
    • If true -> deferred policy is used and get or wait needs to be executed
    • else -> async policy is used and spin lock with while(fut.wait_for(100ms) != std::future_status::ready) can be used
• Only use when the task need not run concurrently with the thread calling get or wait
• Thread’s thread_local variable does not matter
• Either it is ok if f is not called at all or it is guaranteed that get or wait is called later

Item 37: Make std::threads unjoinable on all paths

Active recall:

• What type of threads are joinable and unjoinable?
• What happens when a joinable thread is desctructed?
• What are the draw backs of implicit join on desctruction?
• What are the draw backs of implicit detach on desctruction?
• How do make threads unjoinable on all paths?

States of std::thread

• joinable:
  • When a thread object corresponds to an underlying asynchrnous thread. The thread is still executing.
  • Blocked or waiting threads are joinable
• unjoinable:
  • Threads that do not corresponds to underlying asynchrnous thread.
  • Includes:
    • std::thread that are default constructed
      • No function to execute
    • std::thread objects that have been move away
    • std::thread that have been joined. When a join is called on a thread
    • std::thread that have been detached

std::thread desctructor behaviour:

• If a joinable std::thread is desctructed, the program and all threads will be terminated
• Justification:
  • Draw backs of implicit join:
    • Could lead to performance anomalies
    • The thread could block the function from returning
  • Draw backs of implicit detach:
    • If a thread has a reference to a memory on the stack, a detached thread would allow the function to return
    • The stack frame will be popped off but the thread will be still be running
    • This could lead to buggy behaviour (UB?)

Solution: RAII-fy std::thread

class ThreadRAII {
public:
  enum class DtorAction { join, detach };

  ThreadRAII(std::thread&& t, DtorAction a) : action(a), t(std::move(t)) {}

  ~ThreadRAII()
  {
    if (t.joinable()) {
      if (action == DtorAction::join) {
        t.join();
      } else {
        t.detach();
      }
    }
  }

  std::thread& get() {return t;}

  private:
    DtorAction action;
    std::thread t;
}

• t should be supplied as a rvalue ref (must be moved)
• Good habit: Should declare t last as a thread should be constructed only when all variables are initialized
• Should not have race condition as when the desctructor is called, there will be no access to it already

Item 38: Be aware of varying thread handle desctructor behaviour

Active Recall

• What is a future?
• What is a promise?
• What happens when a future get destructed? Explain all the different scenarios
• How is the return value of a asynchrnous task communicated to caller?

Future: Act as an underlying handler to a software thread

Definitions:

• Caller: the function that calls the asynchrnous task
• Callee: the asynchrnous task that is being called

Future and Promise:

• There is a communication between the caller and callee for the callee to pass the return value to the caller
• The callee will wrap the return value in a Promise and pass it through the communication channel to the caller
• The caller will receive the return value from the communication channel as a Future

Shared State:

• The result of the task is stored as a Shared State on the heap
• Calle -> std::promise -> Shared State -> future -> caller
• Why not on callee?
  • The callee (asynchrnous task) might be completed before get is called.
  • This will result to be destroyed before the caller can get the result
• Why not caller?
  • A future can be used to create shared_future
  • If the result is stored on the caller’s stack, the future will be destroyed eventhough other shared_future will be pointing to it

Future destructor

• For most cases a future destructor just destroy all data members of a future object
  • Act as an implicit detach (the software thread is allowed to run free)
  • Unlike std::thread, it will not terminate the program
• Exception:
  • When a certain set of conditions are met, the future will block until the asynchrnous running taks is complete
    • Performs implicit join
  • Conditions:
    • The future refers to a Shared State created due to a call to std::async
      • std::future can be created using std::package_task which makes std::thread behave like a std::async. The destruction behaviour should not matter as std::thread destruction will take place.
    • When the task policy is std::launch::async can be either due to runtime system (thread manager) or it was specified in the code
    • The future is the last future referring to the shared state
      • std::futures are like shared pointers that points to an underlying resource (Shared State)
      • Determined using a reference count
  • Find a middle ground between program termination and drawbacks of implicit join or detach. Implicitly joining only for asynchrnous task and last reference is destroyed can avoid problems of dangling pointers with implicit detach
• This means that for detached task, they will never be executed if get is not called

Item 39: Consider void futures for one-shot even communication

Active Recall:

• How do condition variable work in c++?
• Why do condition variable need mutex?
• Ways a thread can notify another thread on an event?

Notification prolem: A detector thread has to detect a condition that is only available to it and notify a reactor thread on the condition being met. The reactor thread should be blocked until the detctor thread notifies it.

Condition Variable C++

• C++ provides a std::condition_variable that allows thread to wait on the condition and notify on it.
  • Similar to java monitor
• When a thread calls wait(), conceptually it will block and be added to the queue.
• Another thread can call notify_one()/notify_all() and wake up one/all the threads in the queue.
• Takes in a callable object that will always be called before unblocking the thread
  • The callback should return true if the condition is met else false
• std::condition_variable::wait Takes a unique_lock that allows for the queue and condtions check call back to be thread-safe
  • Without the lock a thread could return true (condition met) but another thread could change the condition
• Common design
  • Condition variable to be associated to an arbitrary condition. A thread will check the condition and if the condition is not met it will call wait()
  • If another thread is able to make the condition to true, it will call notify()
• Functions like non-hoarse monitor (susceptible to spurious wakeup)

Solving notification problem with condition variable

std::condition_variable cv;
std::mutex m;

// detector thread
... // detect event
cv.notify_one(); // tell reactor about event

// reactor thread
{
  std::unique_lock<std::mutex> lk(m); // lock 
  cv.wait(lk);
}

• Code smell:
  • Mutex is used for shared data but there no shared data in this situation
• Problems:
  • If the detector notifies before the reactor waits on the variable. The reactor it will never be woken up after calling wait
  • Spurious wakeups: the condition might change after the detector notifies and the reactor wakes up. Can be resolved using a lambda to test for condition whenever a thread is a candidate for being unblocked. This will not be feasible if the thread is unable to test for the condition.

    cv.wait(lk, []{ //return ture when the event has occured })
    

Using promise as a notification channel

• Promise can be used as communication channel without needing to execute any task
• Will not need to use mutex, notification before wait is allowed and no suprious wakeups

// shared variable
std::promise<void> p;

// detector
... // detect event
p.set_value();

... // prepare to react
p.get_future().wait();

Disadvantages:

• Promise uses Shared State which will use heap-based allocation
• std::promise can only be set once (one-shot)

Other use case: If you would like to spawn a thread (get the overhead over and done with) and block immediately, while the main thread waits for the other work to be done.

std::promise<void> p;

void detect()
{
  auto sf = p.get_future().share();

  std::vector<std::thread> vt;
  for (int i = 0; i < threadsToRun; ++i) {
    vt.emplace_back([sf]{ 
      sf.wait();  // spawn thread and immediately wait for notification
      react(); 
      })
  }

  ...
  p.set_value(); // notfy waiting threads that configs are set
  ...
  for (auto& t : vt) {
    t.join();
  }
}

Item 40: Use std::atomic for concurrency, volatile for special memory

Active Recall:

• What does std::atomic do?
• Are all lines with std::atomic atomic?
• What does volatile do?
• Does volatile help with concurrency?

std::atomic

Functionality:

• provides operations that are guaranteed to be seen as atomic by other thread.
• Operations on std::atomic behave as if they were inside a mutex
• read-modify-write (RMW) operations are atomic (x++, x--)
• Prevent reodering:
  • Without the presence of std::atomic, compilers are allowed to reorder independent statements for optimisation
  • With std::atomic, statements before a write std::atomic cannot be reordered to occur afterwards.
    • Enforce on the compiler and the underlying hardware
  • Motivation: atomic variables are usually shared and reordering them could result in undesired behaviour

Non-functionallity:

• Does not guarantee that all statements are atomic
  • ie std::cout << ai;: the statements involves reading (atomic) atomic variable and writing to stdout (non-atomic)
• Copying of std::atomic is not allowed: compilers are unable to read a variable and write to another variable (copy) atomically. There are no hardware instructions to do so.
  • Workaround: std::atomic<int> y(x.load()) split up reading variable and writing to another variable into two atomic operations

volatile

Problem: compiler will can remove any redundant reads (continuous reading different values into the same variable) or superflous writes (writing to the same variable twice). However, these reads and write might desired as they are mapped to a special memory (shared memory). ie writing twice to a variable but the

Functionality:

• Prevents the compiler from removing any redundant reads or superflous writes to the variable.

Non-functionallity:

• Does not help with concurrency

Item 41: Consider pass by value for copyable parameters that are cheap to move and always copied.

Active Recall:

• What is the lifetime for a by-value parameters
• Alternatives for functions that needs to copy the arguments into a container

Problem: Implementing a function that needs to copy the provided argument into a container.

Overload rlvaue and lvalue:

class Widget {
public:
  void addName(const std::string& newName) { 
    names.push_back(newName);
  }
  void addName(const std::string&& newName) { 
    names.push_back(std::move(newName));
  }
}

• Takes the arguments by reference and copy construct for lvalue ref and move construct for rvalue ref.
• Drawbacks:
  • Code duplication: two identical functions
  • Two function in object code
• Cost:
  • one copy for lvalue and one move for rvalue.
  • Taking the args by reference are no-cost operations.

Forwarding reference:

class Widget {
public:
  template<typename T>
  void addName(T&& newName) {
    names.push_back(std::forwad<T>(newName));
  }
}

• Takes the argument by forwarding reference and forward the reference to push_back. push_back will move construct if the original argument is rvalue reference or copy construct if the original value is lvalue reference.
• Disadvantages:
  • Implementation must typically be in header files
  • Eventhough only one function in source code, could be instantiated to many functions in object code. Each type that can be convertible to std::string will instantiate a function.
  • Bad error messages

Solution: pass by value

class Widget {
public:
  void addName(std::string newName) {
    names.push_back(std::move(newName));
  }
}

• Copy the provided argument into the parameter. If the argument is lvalue, it will be copy constructed and if the argument is rvalue, it will be move constructed.
• Move the parameter into the container.
• When to use by value:
  • Cheap to move: it incurs an additional move operation.
  • Should be copyable parameter:
    • If the parameter is move-only, only rvalue arguments can be provided to move construct the parameter.
    • Will be better to take the argument by rvalue reference and forward it to the container. Incur one less move constructor.
  • When the parameters are always copied on all path:
    • If a path would not lead to the parameter being copied, passing by value will incur additional uneccassry copy right from the start.
  • Not move assigning:
    • It might be easier to copy assign to lvalue reference.
    • If the parameter has a pointer to a dynamically allocated memory, passing by value will result in memory allocation and moving the parameter afterwards will result in deallocating the memory.
    • If the parameter is passed by reference, we will not copy the arguments and copy assignment operation might not incurr additional cost if the dynamically allocated memory can be reused
    • Note: For string copy assignment, if the capacity of the lhs is more than the length of the string on the rhs, there might not be a need to allocate memory to the lhs.
• Advantage:
  • No code duplication
  • No bad error message
  • Only one function in object code
• Disadvantages:
  • Incur additional move construct for both lvalue and rvalue argument. This is due to additional move operation of the parameter (value) into the container.
  • Object slicing: if the parameter is a base class, passing a derived class into the parameter will result in object slicing (derived class characteristics will be sliced off)

Item 42: Consider emplacement instead of insertion

Problem with insertion:

std::vector<std::string> vs;
vs.push_back("xyzzy");

1. A temporary (tmp) std::string object is created from the string literal "xyzzy".
2. tmp is passed to push_back as a rvalue ref and a copy of tmp is constructed in the memory for std::vector. The copy is move constructed.
3. tmp is destroyed and call the destructor

With emplace:

1. Forwards the string literal to the constructor of the element in the memory of the vector
2. No temporary object created -> no extra construction and destruction of the temporary variable

Heuristic to use emplace

• The value being added is constructed into the container, not assigned:
  • If the value is inserted to a place that has been constructed, the compiler will most probably create a temporary object and move assign the object into place.
  • A temporary object will be created even though it uses emplace
  • Node-based containers always use construction to add new value and emplace_back/front
• The argument type(s) begin passed differ from the type held by the container:
  • push_back will result in a temporary object being created but emplace does not
• Container is unlikely to reject the new value as a duplicate:
  • If the container rejects duplicate, it will first create a node using the emplace arguments and compare the nodes. This will result in the node being destroyed.
  • Temp nodes are created more often for emplace than insertion

Caevet: resource management objects (smart pointers) should be constructed outside of emplace_back

• Using emplace will result the resource being constructed separately from it being wraped in the object. A leak could occur when an exception is thrown before the resource is wrapped.
• emplace uses direct initialization but push_back uses copy initialization. Copy initialization cannot use explicit constructor but direct initialization can.