A proposal to add a utility class to represent optional objects (Revision 4)

ISO/IEC JTC1 SC22 WG21 N3672 2013-04-19

Introduction

Class template optional<T> proposed here is a type that may or may not store a value of type T in its storage space. Its interface allows to query if a value of type T is currently stored, and if so, to access it. The interface is based on Fernando Cacciola's Boost.Optional library^[2], shipping since March, 2003, and widely used. It requires no changes to core language, and breaks no existing code.

Table of contents

Revision history

Changes since R4C

• Removed all relational operators except operator== and operator<.

Changes since N3527:

• Renamed tag emplace to in_place based on feedback from LEWG.
• Removed the Motivation section.
• Added description of op = {} syntax and explained why we had to apply some tricks to support it. See The op = {} syntax.
• Clarified why we do not require in function template swap that T be MoveAssignamle. See Requirements for swap.
• Added short discussion on the choice of the header. See Header <utility> or <optional>?.
• Optional's operator< now uses less<T> rather than T::operator<. See Relational operators for rationale.
• Provided a comparison between optional in this proposal and boost::optional. See Comparison with Boost.Optional.
• Added "open questions" section.
• Reduced the discussion in Rationale section by omitting some alternatives we considered but rejected.
• Removed namespace experimental, because we are not targetting a TS anymore.
• Applied fixes in the standardese as per LWG's feedback.
• Function emplace now returns void — LWG recommendation.
• All mixed ordering relations now use only less<T> in implementation — LWG recommendation.
• Clarified (in section Overview of optional) that Optional cannot be implemented with aligned_storage .
• Reworded complicated Effects elements for copy/move assignment and swap, so that they are easier to read — LWG recommendation.

Changes since N3406=12-0096:

• Wording changes are now relative to N3485.
• optional<T> is hashable for hashable T’s.
• Optional refereces are now an auxiliary proposal — this gives the possibility to accept optional values without references.
• Optional refereces are now assignamble and swappable.
• get_value_or is now optional-specific member function, renamed to value_or.
• Added member function value — an alternative to operator* that checks if the object is engaged.
• optional<T> is a literal type.
• Mixed relational operations between optional<T> and T are now allowed.
• Removed reference implementation. We now only provide the implementation of the parts that we consider non-trivial.

Changes since N1878=05-0138:

• Revised wording; the changes are now relative to N3290.
• Removed any form of assignment for optional references.
• Removed duplicate interface for accessing the value stored by the optional.
• Added in-place construction and assignment.
• Now using different tag nullopt instead of nullptr to indicate the 'disengaged' (uninitialized) optional.
• Included C++11 features: move semantics, noexcept, variadic templates, perfect forwarding, static initialization.
• Changed the motivation section.
• Changed the design rationale section.
• Added reference implementation

Impact on the Standard

This proposal depends on library proposal N3471: it requires that Standard Library components move, forward and member functions of initializer_list are constexpr. The paper has already been incorporated into the Working Draft of the Standard N3485. This proposal also depends on language proposal N2439 (Rvalue references for *this). While this feature proposal has been incorporated into C++11, we are aware of only two compilers that implemented it: Clang and GCC 4.8.1. There is a risk that if compiler vendors do not implement it, they will also not be able to fully implement this proposal. In that case, the signature of member function optional<T>::value_or from this proposal will need to be modified.

N3507 (A URI Library for C++) depends on this library.

Overview of optional

The primary purpose of optional<T>'s interface is to be able to answer the quesiotn "do you contain a value of type T?", and iff the answer is "yes", to provide access to the contained value. Conceptually, optional can be illustrated by the following structure.

template <typename T>
struct optional
{
  bool is_initialized_;
  typename aligned_storage<sizeof(T), alignof(T)>::type storage_;
};

Flag is_initialized_ stores the information if optional has been assigned a value of type T. storage_ is a raw fragment of memory (allocated within optional object) capable of storing an object of type T. Objects in this storage are created and destroyed using placement new and pseudo destructor call as member functions of optional are executed and based on the value of flag is_initialized_.

Note: the above representation is only conceptual. In fact, our aggressive requirements for constexpr constructors prevent the usage of aligned_storage in the implementation of optional. See section Implementability for an overview of a possible reference implementation.

Alternatively, one can think of optional<T> as pair<bool, T>, with one important difference: if first is false, member second has never been even initialized, even with default constructor or value-initialization.

The basic usage of optional<T> can be illustrated with the following example.

optional<int> str2int(string);    // converts int to string if possible

int get_int_form_user()
{
  string s;

  for (;;) {
    cin >> s;
    optional<int> o = str2int(s); // 'o' may or may not contain an int
    if (o) {                      // does optional contain a value?
      return *o;                  // use the value
    }
  }
}

Interface of optional

Default construction of optional<T> creates an an object that stores no value of type T. No default constructor of T is called. T doesn't even need to be DefaultConstructible. We say that thus created optional object is disengaged. We use a pair of somewhat arbitrary names, 'engaged' and 'disengaged'. A default-constructed optional object is initialized (the lifetime of optional<T> has started), but it is disengaged (the lifetime of the contained object has not yet started). Trying to access the value of T in this state causes undefined behavior. The only thing we can do with a disengaged optional object is to query whether it is engaged, copy it, compare it with another optional<T>, or engage it.

optional<int> oi;                 // create disengaged object
optional<int> oj = nullopt;       // alternative syntax
oi = oj;                          // assign disengaged object
optional<int> ok = oj;            // ok is disengaged

if (oi)  assert(false);           // 'if oi is engaged...'
if (!oi) assert(true);            // 'if oi is disengaged...'

if (oi != nullopt) assert(false); // 'if oi is engaged...'
if (oi == nullopt) assert(true);  // 'if oi is disengaged...'

assert(oi == ok);                 // two disengaged optionals compare equal

Tag nullopt represents the disengaged state of an optional object. This reflects our conceptual model of optional objects: optional<T> can be thought of as a T with one additional value nullopt. Being disengaged is just another value of T. Thus, optional<unsigned> can be thought of as a type with possible range of values {nullopt, 0, 1, ...}. Value nullopt is always picked by the default constructor.

We can create engaged optional objects or engage existing optional objects by using converting constructor (from type T) or by assigning a value of type T.

optional<int> ol{1};              // ol is engaged; its contained value is 1
ok = 2;                           // ok becomes engaged; its contained value is 2
oj = ol;                          // oj becomes engaged; its contained value is 1

assert(oi != ol);                 // disengaged != engaged
assert(ok != ol);                 // different contained values
assert(oj == ol);                 // same contained value
assert(oi < ol);                  // disengaged < engaged
assert(ol < ok);                  // less by contained value

Copy constructor and copy assignment of optional<T> copies both the engaged/disengaged flag and the contained value, if it exists.

optional<int> om{1};              // om is engaged; its contained value is 1
optional<int> on = om;            // on is engaged; its contained value is 1
om = 2;                           // om is engaged; its contained value is 2
assert (on != om);                // on still contains 1. They are not pointers

We access the contained value using the indirection operator.

int i = *ol;                      // i obtains the value contained in ol
assert(i == 1);
*ol = 9;                          // the valuesagety  contained in ol becomes 9
assert(*ol == 9);

We also provide consistent operator->. Even though optional provides operators -> and * it is not a pointer. Unlike pointers, it has full value semantics: deep copy construction, deep copy assignmet, deep equality and less-than comparison, and constness propagation (from optional object to the contained value).

int p = 1;
optional<int> op = p;
assert(*op == 1);
p = 2;                         
assert(*op == 1);                 // value contained in op is separated from p

The typical usage of optional requires an if-statement.

if (ol)                      
  process(*ol);                   // use contained value if present
else
  processNil();                   // proceed without contained value
  
if (!om)   
  processNil();
else  
  process(*om);     

If the action to be taken for disengaged optional is to proceed with the default value, we provide a convenient idiom:

process(ol.value_or(0));      // use 0 if ol is disengaged

Sometimes the initialization from T may not do. If we want to skip the copy/move construction for T because it is too expensive or simply not available, we can call an 'emplacing' constructor, or function emplace.

optional<Guard> oga;                     // Guard is non-copyable (and non-moveable)     
optional<Guard> ogb(in_place, "res1");    // initializes the contained value with "res1"            
optional<Guard> ogc(in_place);            // default-constructs the contained value

oga.emplace("res1");                     // initializes the contained value with "res1"  
oga.emplace();                           // destroys the contained value and 
                                         // default-constructs the new one

There are two ways to disengage a perhaps engaged optional object:

ok = nullopt;                      // if ok was engaged calls T's dtor
oj = {};                           // assigns a temporary disengaged optional
oga = nullopt;                     // OK: disengage the optional Guard
ogb = {};                          // ERROR: Guard is not Moveable

Optional propagates constness to its contained value:

const optional<int> c = 4; 
int i = *c;                        // i becomes 4
*c = i;                            // ERROR: cannot assign to const int&

Exception safety

Type optional<T> is a wrapper over T, thus its exception safety guarantees depend on exception safety guarantees of T. We expect (as is the case for the entire C++ Standard Library) that destructor of T does not throw exceptions. Copy assignment of optional<T> provides the same exception guarantee as copy assignment of T and copy constructor of T (i.e., the weakest of the two). Move assignment of optional<T> provides the same exception guarantee as move assignment of T and move constructor of T. Member function emplace provides basic guarantee: if exception is thrown, optional<T> becomes disengaged, regardless of its prior state.

Advanced use cases

With optional<T> problems described in Motivation section can be solved as follows. For lazy initialization:

class Car 
{
  mutable mutex m_;
  mutable optional<const Engine> engine_;
  mutable optional<const int> mileage_
  
public:
  const Engine& engine() const
  {
    lock_guard<mutex> _(m_);
    if (engine_ == nullopt) engine_.emplace( engineParams() );
    return *engine_;
  }
  
  const int& mileage() const
  {
    lock_guard<mutex> _(m_);
    if (!mileage_) mileage_ = initMileage();
    return *mileage_;
  }
};

The algorithm for finding the greatest element in vector can be written as:

optional<int> find_biggest( const vector<int>& vec )
{
  optional<int> biggest;  // initialized to not-an-int
  for (int val : vec) {
    if (!biggest || *biggest < val) {
      biggest = val;
      // or: biggest.emplace(val);
    }
  }
  return biggest;
} 

Missing return values are naturally modelled by optional values:

optional<char> c = stream.getNextChar();
optional<int> x = DB.execute("select ...");

storeChar( c.value_or('\0') );
storeCount( x.value_or(-1) );

Optional arguments can be implemented as follows:

template <typename T>
T getValue( optional<T> newVal = nullopt )
{
  if (newVal) {
    cached = *newVal;     
  }
  return cached;      
}

Manually controlling the life-time of guard-like objects can be achieved by emplacement operations and nullopt assignment:

{
  optional<Guard> grd1{in_place, "res1", 1};   // guard 1 initialized
  optional<Guard> grd2;

  grd2.emplace("res2", 2);                     // guard 2 initialized
  grd1 = nullopt;                              // guard 1 released

}                                              // guard 2 released (in dtor)

It is possible to use tuple and optional to emulate multiple return valuse for types without default constructor:

tuple<Date, Date, Date> getStartMidEnd();
void run(Date const&, Date const&, Date const&);
// ...

optional<Date> start, mid, end;           // Date doesn't have default ctor (no good default date)

tie(start, mid, end) = getStartMidEnd();
run(*start, *mid, *end); 

Comparison with value_ptr

N3339^[7] proposes a smart pointer template value_ptr. In short, it is a smart pointer with deep copy semantics. It has a couple of features in common with optional: both contain the notion of optionality, both are deep-copyable. Below we list the most important differences.

value_ptr requires that the pointed-to object is allocated in the free store. This means that the sizeof(value_ptr<T>) is fixed irrespective of T. value_ptr is 'polymorphic': object of type value_ptr<T> can point to an object of type DT, derived from T. The deep copy preserves the dynamic type. optional requires no free store allocation: its creation is more efficient; it is not "polymorphic".

Relational operations on value_ptr are shallow: only addresses are compared. Relational operations on optional are deep, based on object's value. In general, optional has a well defined value: being disengaged and the value of contained value (if it exists); this value is expressed in the semantics of equality operator. This makes optional a full value-semantic type. Comparison of value_ptr does not have this property: copy semantics are incompatible with equality comparison semantics: a copy-constructed value_ptr does not compare equal to the original. value_ptr is not a value semantic type.

Comparison with Boost.Optional

This proposal basically tries to follow Boost.Optional's interface. Here we list the significant differences.

Design rationale

The very minimum, ascetic interface for optional objects — apart from copy/move — could consist of two constructors and two functions:

// not proposed
int i = 9;
optional<int> oi{engaged, i};          // create engaged
optional<int> oj{disengaged};          // create disengaged
if (oi.is_engaged()) {                 // contains value?
  oi.get_value();                      // access the value
}

The reason we provide different and richer interface is motivated by users' convenience, performance improvements, secondary goals we want to acheive, and the attempt to standardize the existing practice.

Nearly every function in the interface of optional has risen some controversies. Different people have different expectations and different concerns and it is not possible to satisfy all conflicting requirements. Yet, we believe that optional is so universally useful, that it is worth standardizing it even at the expense of introducing a controversial interface. The current proposal reflects our arbitrary choice of balance between unambiguousity, genericity and flexibility of the interface. The interface is based on Fernanndo Cacciola's Boost.Optional library,^[2] and the users' feedback. The library has been widely accepted, used and even occasionally recommended ever since.

Conceptual model for optional<T>

Optional objects serve a number of purposes and a couple of conceptual models can be provided to answer the question what optional<T> really is and what interface it should provide. The three most common models are:

1. Just a T with deferred initialization (and additional interface to check if the object has already been initialized).
2. A discriminated union of types nullopt_t and T.
3. A container of T's with the maximum size of 1.

While (1) was the first motivation for optional, we do not choose to apply this model, because type optional<T> would not be a value semantic type: it would not model concept Regular (if C++ had concepts). In particular, it would be not clear whether being engaged or disengaged is part of the object's state. Programmers who wish to adopt this view, and don't mind the mentioned difficulties, can still use optional this way:

optional<int> oi;
initializeSomehow(oi);

int i = (*oi);
use(*oi);
(*oi) = 2;
cout << (*oi);

Note that this usage does not even require to check for engaged state, if one is sure that the object is engaged. One just needs to use indirection operator consistently anywhere one means to use the initialized value.

Model (2) treats optional<T> as either a value of type T or value nullopt, allocated in the same storage, along with the way of determining which of the two it is. The interface in this model requires operations such as comparison to T, comparison to nullopt, assignment and creation from either. It is easy to determine what the value of the optional object is in this model: the type it stores (T or nullopt_t) and possibly the value of T. This is the model that we propose.

Model (3) treats optional<T> as a special case container. This begs for a container-like interface: empty to check if the object is disengaged, emplace to engage the object, and clear to disengage it. Also, the value of optional object in this model is well defined: the size of the container (0 or 1) and the value of the element if the size is 1. This model would serve our pupose equally well. The choice between models (2) and (3) is to a certain degree arbitrary. One argument in favour of (2) is that it has been used in practice for a while in Boost.Optional.

Additionally, within the affordable limits, we propose the view that optional<T> just extends the set of the values of T by one additional value nullopt. This is reflected in initialization, assignment, ordering, and equality comparison with both T and nullopt.

optional<int> oi = 0;
optional<int> oj = 1;
optional<int> ok = nullopt;

oi = 1;
oj = nullopt;
ok = 0;

oi == nullopt;
oj == 0;
ok == 1;

Initialization of optional<T>

In cases T is a value semantic type capable of storing n distinct values, optional<T> can be seen as an extended T capable of storing n + 1 values: these that T stores and nullopt. Any valid initialization scheme must provide a way to put an optional object to any of these states. In addition, some Ts (like scope guards) are not MoveConstructible and their optional variants still should constructible with any set of arguments that work for T. Two models have been identified as feasible.

The first requires that you initialize either by providing either an already constructed T or the tag nullopt.

string s{"STR"};

optional<string> os{s};                    // requires Copyable<T>
optional<string> ot = s;                   // requires Copyable<T>
optional<string> ou{"STR"};                // requires Movable<T>
optional<string> ov = string{"STR"};       // requires Movable<T>

optional<string> ow;                       // disengaged
optional<string> ox{};                     // disengaged
optional<string> oy = {};                  // disengaged
optional<string> oz = optional<string>{};  // disengaged
optional<string> op{nullopt};              // disengaged
optional<string> oq = {nullopt};           // disengaged

In order to avoid calling move/copy constructor of T, we use a 'tagged' placement constructor:

optional<Guard> og;                        // disengaged
optional<Guard> oh{};                      // disengaged
optional<Guard> oi{in_place};               // calls Guard{} in place
optional<Guard> oj{in_place, "arg"};        // calls Guard{"arg"} in place

The in-place constructor is not strictly necessary. It could be dropped because one can always achieve the same effect with a two-liner:

optional<Guard> oj;                        // start disengaged
oj.emplace("arg");                         // now engage

Notably, there are two ways to create a disengaged optional object: either by using the default constructor or by calling the 'tagged constructor' that takes nullopt. One of these could be safely removed and optional<T> could still be initialized to any state.

The default constructor

This proposal provides a default constructor for optional<T> that creates a disengaged optional. We find this feature convenient for a couple of reasons. First, it is because this behaviour is intuitive as shown in the above example of function readChar. It avoids a certain kind of bugs. Also, it satisfies other expectations. If I declare optional<T> as a non-static member, without any initializer, I may expect it is already initialized to the most natural, disengaged, state regardless of whether T is DefaultConstructible or not. Also when declaring a global object, one could expect that default constructor would be initialized during static-initialization (this proposal guarantees that). One could argue that the tagged constructor could be used for that puropse:

optional<int> global = nullopt;

struct X
{
  optional<M> m = nullopt;
};

However, sometimes not providing the tag may be the result of an inadvertent omission rather than conscious decision. Because of our default constructor semantics we have to reject the initialization scheme that uses a perfect forwarding constructor. Even if this is fine, one could argue that we do not need a default constructor if we have a tagged constructor. We find this redundancy convenient. For instance, how do you resize a vector<optional<T>> if you do not have the default constructor? You could type:

vec.resize(size, nullopt);

However, that causes first the creation of disengaged optional, and then copying it multiple times. The use of copy constructor may incur run-time overhead and not be available for non-copyable Ts. Also, it would be not possible to use subscript operator in maps that hold optional objects.

Also, owing to this constructor, optional has a nice side-effect feature: it can make "almost Regular" types fully Regular if the lack of default constructor is the only thing they are missing. For instance consider type Date for representing calendar days: it is copyable movable, comparable, but is not DefaultConstructible because there is no meaningful default date. However, optional<Date> is Regular with a meaningful not-a-date state created by default.

Converting constructor (from T)

An object of type T is convertible to an engaged object of type optional<T>:

optional<int> oi = 1; // works

This convenience feature is not strictly necessary because you can achieve the same effect by using tagged forwarding constructor:

optional<int> oi{in_place, 1};

If the latter appears too inconvenient, one can always use function make_optional described below:

optional<int> oi = make_optional(1); 
auto oj = make_optional(1); 

The implicit converting constructor comes in handy in case of optional function arguments:

void fun(std::string s, optional<int> oi = nullopt);

fun("dog", 2);
fun("dog");
fun("dog", nullopt); // just to be explicit 

It has been argued that the constructor from T should be made explicit. It is not trivial to decide whether T should be convertiblle to optional<T>. This is not a clear situation where value of one type is stored in another type with greater resolution, or the situation where the same abstract value is stored in a type with different internal representation. On the other hand, given our conceptual model, optional<T> can store all values of T, so it is possible to apply a "lossless conversion". We decided to provide the conversion, in order to (1) adhere to our conceptual model, and (2) to enable the above convenience for function argument passing. The implicit conversion naturally implies that optional<T>'s can be compared with T's. This is discussed further down.

At some point we considered the possibility to make this constructor conditionally explicit: make it explicit if T has an explicit copy/move constructor, and make it non-explicit if T has a normal, non-explicit constructor. In the end, we find explicit copy constructor so unusual that we do not find it worthwile to addressing it at the expense of complicating the design.

Contextual conversion to bool for checking engaged state

Objections have been risen to this decision. When using optional<bool>, contextual conversion to bool (used for checking the engaged state) might be confused with accessing the stored value. While such mistake is possible, it is not the first such case in the standard: types bool*, unique_ptr<bool>, shared_ptr<bool> suffer from the same potential problem, and it was never considered a show-stopper. Some have suggested that a special case in the interface should be made for optional<bool> specialization. This was however rejected because it would break the generic use of optional.

Some have also suggested that a member function like is_initialized would more clearly indicate the intent than explicit conversion to bool. However, we believe that the latter is a well established idiom in C++ community as well as in the C++ Standard Library, and optional appears so fundamental a type that a short and familiar notation appears more appropriate. It also allows us to combine the construction and checking for being engaged in a condition:

if (optional<char> ch = readNextChar()) {
  // ...
}

Using tag nullopt for indicating disengaged state

The proposed interface uses special tag nullopt to indicate disengaged optional state. It is used for construction, assignment and relational operations. This might rise a couple of objections. First, it introduces redundancy into the interface:

optional<int> opt1 = nullopt; 
optional<int> opt2 = {}; 

opt1 = nullopt;
opt2 = {};

if (opt1 == nullopt) ...
if (!opt2) ...
if (opt2 == optional<int>{}) ...

On the other hand, there are usages where the usage of nullopt cannot be replaced with any other convenient notation:

void run(complex<double> v);
void run(optional<string> v);

run(nullopt);              // pick the second overload
run({});                   // ambiguous

if (opt1 == nullopt) ...   // fine
if (opt2 == {}) ...        // illegal

bool is_engaged( optional<int> o)
{
  return bool(o);          // ok, but unclear
  return o != nullopt;     // familiar
}

While some situations would work with {} syntax, using nullopt makes the programmer's intention more clear. Compare these:

optional<vector<int>> get1() {
  return {};
}

optional<vector<int>> get2() {
  return nullopt;
}

optional<vector<int>> get3() {
  return optional<vector<int>>{};
}

The usage of nullopt is also a consequence of the adapted model for optional: a discriminated union of T and nullopt_t. Also, a similar redundancy in the interface already exists in a number of components in the standard library: unique_ptr, shared_ptr, function (which use literal nullptr for the same purpose); in fact, type requirements NullablePointer require of types this redundancy.

Name "nullopt" has been chosen because it clearly indicates that we are interested in creating a null (disengaged) optional<T> (of unspecified type T). Other short names like "null", "naught", "nothing" or "none" (used in Boost.Optional library) were rejected because they were too generic: they did not indicate unambiguously that it was optional<T> that we intend to create. Such a generic tag nothing could be useful in many places (e.g., in types like variant<nothing_t, T, U>), but is outside the scope of this proposal.

Note also that the definition of tag struct nullopt is more complicated than that of other, similar, tags: it has explicitly deleted default constructor. This is in order to enable the reset idiom (opt2 = {};), which would otherwise not work because of ambiguuity when deducing the right-hand side argument.

Why not nullptr

One could argue that since we have keyword nullptr, which already indicates a 'null-state' for a number of Standard Library types, not necessarily pointers (class template function), it could be equally well used for optional. In fact, the previous revision of this proposal did propose nullptr, however there are certain difficulties that arise when the null-pointer literal is used.

First, the interface of optional is already criticized for resembling too much the interface of a (raw or smart) pointer, which incorrectly suggests external heap storage and shallow copy and comparison semantics. The "ptr" in "nullptr" would only increase this confusion. While std::function is not a pointer either, it also does not provide a confusing operator->, or equality comparison, and in case it stores a function pointer it does shallow copying.

Second, using literal nullptr in optional would make it impossible to provide some of the natural and expected initialization and assignment semantics for types that themselves are nullable:

• optional<int*>,
• optional<const char*>,
• optional<M C::*>,
• optional<function<void(int)>>,
• optional<NullableInteger>,
• optional<nullptr_t>.

Should the following initialization render an engaged or a disengaged optional?

optional<int*> op = nullptr;

One could argue that if we want to initialize an engaged optional we should indicate that explicitly:

optional<int*> op{in_place, nullptr};

But this argument would not work in general. One of the goals of the design of optional is to allow a seamless "optionalization" of function arguments. That is, given the folowing function signature:

void fun(T v) {
  process(v);
}

It should be possible to change the signature and the implementation to:

void fun(optional<T> v) {
  if (v) process(*v);
  else   doSthElse();
}

and expect that all the places that call function fun are not affected. But if T happens to be int* and we occasionally pass value nullptr to it, we will silently change the intended behavior of the refactoring: because it will not be the pointer that we null-initialize anymore but a disengaged optional.

Note that this still does not save us from the above problem with refactoring function fun in case where T happens to be optional<U>, but we definately limit the amount of surprises.

In order to avoid similar problems with tag nullopt, instantiating template optional with types nullopt_t and in_place_t is prohibited.

There exist, on the other hand, downsides of introducing a special token in place of nullptr. The number of ways to indicate the 'null-state' for different library components will grow: you will have NULL, nullptr, nullopt. New C++ programmers will ask "which of these should I use now?" What guidelines should be provided? Use only nullptr for pointers? But does it mean that we should use nullopt for std::function? Having only one way of denoting null-state, would make the things easier, even if "ptr" suggests a pointer.

Why not a tag dependent on T?

It has been suggested that instead of 'typeless' nullopt a tag nested in class optional be used instead:

optional<int> oi = optional<int>::nullopt;

This has several advantages. Namespace std is not polluted with an additional optional-specific name. Also, it resolves certain ambiguities when types like optional<optional<T>> are involved:

optional<optional<int>> ooi = optional<int>::nullopt;           // engaged
optional<optional<int>> ooj = optional<optional<int>>::nullopt; // disengaged

void fun(optional<string>);
void fun(optional<int>);

fun(optional<string>::nullopt); // unambiguous: a typeless nullopt would not do 

Yet, we choose to propose a typeless tag because we consider the above problems rare and a typeless tag offers a very short notation in other cases:

optional<string> fun()
{
  optional<int> oi = nullopt;  // no ambiguity
  oi = nullopt;                // no ambiguity
  // ...
  return nullopt;              // no ambiguity
}

If the typeless tag does not work for you, you can always use the following construct, although at the expense of invoking a (possibly elided) move constructor:

optional<optional<int>> ooi = optional<int>{};           // engaged
optional<optional<int>> ooj = optional<optional<int>>{}; // disengaged

void fun(optional<string>);
void fun(optional<int>);

fun(optional<string>{}); // unambiguous 

Accessing the contained value

It was chosen to use indirection operator because, along with explicit conversion to bool, it is a very common pattern for accessing a value that might not be there:

if (p) use(*p);

This pattern is used for all sort of pointers (smart or dumb), and it clearly indicates the fact that the value may be missing and that we return a reference rather than a value. The indirection operator has risen some objections because it may incorrectly imply that optional is a (possibly smart) pointer, and thus provides shallow copy and comparison semantics. All library components so far use indirection operator to return an object that is not part of the pointer's/iterator's value. In contrast, optional indirects to the part of its own state. We do not consider it a problem in the design; it is more like an unprecedented usage of indirection operator. We believe that the cost of potential confusion is overweighed by the benefit of an easy to grasp and intuitive interface for accessing the contained value.

We do not think that providing an implicit conversion to T would be a good choice. First, it would require different way of checking for the empty state; and second, such implicit conversion is not perfect and still requires other means of accessing the contained value if we want to call a member function on it.

Using the indirection operator for a disengaged object is an undefined behavior. This behavior offers maximum runtime performance. In addition to indirection operator, we provide member function value that returns a reference to to the contained value if one exists or throws an exception (derived from logic_error) otherwise:

void interact()
{
  std::string s;
  cout << "enter number ";
  cin >> s;
  optional<int> oi = str2int(s);
  
  try {
    process_int(oi.value());
  }
  catch(bad_optional_access const&) {
    cout << "this was not a number";
  }
}

Relational operators

One of the design goals of optional is that objects of type optional<T> should be valid elements in STL containers and usable with STL algorithms (at least if objects of type T are). Equality comparison is essential for optional<T> to model concept Regular. C++ does not have concepts, but being regular is still essential for the type to be effectively used with STL. Ordering is essential if we want to store optional values in ordered associative containers. A number of ways of including the disengaged state in comparisons have been suggested. The ones proposed, have been crafted such that the axioms of equivalence and strict weak ordering are preserved: disengaged optional<T> is simply treated as an additional and unique value of T equal only to itself; this value is always compared as less than any value of T:

optional<unsigned> o0{0};
optional<unsigned> o1{1};
optional<unsigned> oN{nullopt};

assert (oN < o0);
assert (o0 < o1);
assert (!(oN  < oN));
assert (!(o1 < o1));

assert (oN != o0);
assert (o0 != o1);
assert (oN == oN);
assert (o0 == o0);

Value nullopt could have been as well considered greater than any value of T. The choice is to a great degree arbitrary. We choose to stick to what boost::optional does.

Behind the scenes, optional is using utility std::less<T> rather than T::operator<. This is almost the same thing except for one detail. Quoting 20.8.5 [comparisons] paragraph 8, "For templates greater, less, greater_equal, and less_equal, the specializations for any pointer type yield a total order, even if the built-in operators <, >, <=, >= do not."

Given that both nullopt_t and T are implicitly convertible to optional<T>, this implies the existence and semantics of mixed comparison between optional<T> and T, as well as between optional<T> and nullopt_t:

assert (oN == nullopt);
assert (o0 != nullopt);
assert (oN != 1);
assert (o1 == 1);

assert (oN < 1);
assert (o0 > nullopt);

Although it is difficult to imagine any practical use case of ordering relation between optional<T> and nullopt_t, we still provide it for completness's sake

The mixed relational operators, especially these representing order, between optional<T> and T have been accused of being dangerous. In code examples like the following, it may be unclear if the author did not really intend to compare two T's.

auto count = get_optional_count();
if (count < 20) {}                        // or did you mean: *count < 20 ?
if (count == nullopt || *count < 20) {}   // verbose, but unambiguous

Given that optional<T> is comparable and implicitly constructible from T, the mixed comparison is there already. We would have to artificially create the mixed overloads only for them to cause controlled compilation errors. A consistent approach to prohibiting mixed relational operators would be to also prohibit the convesion from T or to also prohibit homogenous relational operators for optional<T>; we do not want to do either, for other reasons discussed in this proposal. Also, mixed relational operations are available in Boost.Optional and were found useful by the users. Mixed operators come as something natural when we consider the model "T with one additional value".

For completeness sake, we also provide ordering relations between optional<T> and nullopt_t, even though we see no practical use case for them:

bool test(optional<int> o)
{
  assert (o >= nullopt);    // regardless of o's state
  assert (!(o < nullopt));  // regardless of o's state
  assert (nullopt <= o);    // regardless of o's state 
  return (o > nullopt);     // o != nullopt is cleaner
}

This is similar to comparing values of type unsigned int with 0:

bool test(unsigned int i)
{
  assert (i >= 0);          // regardless of i's state
  assert (0 <= i);          // regardless of i's state
  return (i > 0);           // i != 0 is cleaner
}

There exist two ways of implementing operator> for optional objects:

bool operator>(const optional<T>& x, const optional<T>& y)
{
  return (!x) ? false : (!y) ? true : greater{}(*x, *y);  // use T::operator>
}

bool operator>(const optional<T>& x, const optional<T>& y)
{
  return y < x;                                 // use optional<T>::operator<
}

In case T::operator> and T::operator< are defined consistently, both above implementations are equivalent. If the two operators are not consistent, the choice of implementation makes a difference. For homogenous relational operations (between two optional<T>s), we chose the former specification. This is consistent with a similar choice for std::tuple. The only downside of this solution would be visible if some type defined the two predicates inconsistently. We have never seen such a case. The only examples of "clever" abusage of relational operators in domain-speciffic languages we are aware of are expression templates. But in this case operators < and > do not return type bool and we cannot imagine, you would use optional for them. On the other hand, there is one small benefit that comes with our proposal: T::operator> may not even be defined in order for optional<T>::operator< to work.

For heterogenous relational operations (between optional<T> and T), we also choose to implement all in terms of std::less. This way we are consistent with the rest of the standard library. This is a change compared to the previous revisions, where mixed relops were implemented in terms of corresponding relational operators of T.

Resetting the optional value

Assigning the value of type T to optional<T> object results in doing two different things based on whether the optional object is engaged or not. If optional object is engaged, the contained value is assigned a new value. If optional object is disengaged, it becomes engaged using T's copy/move constructor. This behavior is based on a silent assumption that T's copy/move constructor is copying a value in a similar way to copy/move assignment. A similar logic applies to optional<T>'s copy/move assignment, although the situation here is more complicated because we have two engaged/disengaged states to be considered. This means that optional<T>'s assignment does not work (does not compile) if T is not assignable:

optional<const int> oi = 1;  // ok
oi = 2;                      // error 
oi = oi;                     // error 
oi = nullopt;                // ok

There is an option to reset the value of optional object without resorting to T's assignment:

optional<const int> oj = 1;  // ok
oj.emplace(2);               // ok 

Function emplace disengages the optional object if it is engaged, and then just engages the object anew by copy-constructing the contained value. It is similar to assignment, except that it is guaranteed not to use T's assignment and provides only a basic exception safety guarantee. In contrast, assignment may provide a stronger guarantee if T's assignment does.

To sumarize, this proposal offers three ways of assigning a new contained value to an optional object:

optional<int> o;
o = make_optional(1);         // copy/move assignment
o = 1;                        // assignment from T
o.emplace(1);                 // emplacement 

The first form of assignment is required to make optional a regular object, useable in STL. We need the second form in order to reflect the fact that optional<T> is a wrapper for T and hence it should behave as T as much as possible. Also, when optional<T> is viewed as T with one additional value, we want the values of T to be directly assignable to optional<T>. In addition, we need the second form to allow the interoperability with function std::tie as shown above. The third option is required to be able to reset an optional non-assignable T.

Tag in_place

This proposal provides an 'in-place' constructor that forwards (perfectly) the arguments provided to optional's constructor into the constructor of T. In order to trigger this constructor one has to use the tag struct in_place. We need the extra tag to disambiguate certain situations, like calling optional's default constructor and requesting T's default construction:

optional<Big> ob{in_place, "1"}; // calls Big{"1"} in place (no moving)
optional<Big> oc{in_place};      // calls Big{} in place (no moving)
optional<Big> od{};             // creates a disengaged optional

We no longer propose name emplace for the tag. We follow the recomendation from LEWG. Name emplace should be reserved in namespace std in case an algorithm with that name needs to be added in the future. Also it might be confusing to see a member function and a tag with the same name?

Requirements on T

Class template optional imposes little requirements on T: it has to be either an lvalue reference type, or a complete object type satisfying the requirements of Destructible. It is the particular operations on optional<T> that impose requirements on T: optional<T>'s move constructor requires that T is MoveConstructible, optional<T>'s copy constructor requires that T is CopyConstructible, and so on. This is because optional<T> is a wrapper for T: it should resemble T as much as possible. If T is EqualityComparable then (and only then) we expect optional<T> to be EqualityComparable.

Optional references

In this revision, optional references are presented as an auxiliary proposal. The intention is that the Committee should have an option to accept optional values without optional references, if it finds the latter concept inacceptable. Users that in generic contexts require to also store optional lvalue references can achieve this effect, even without direct support for optional references, with a bit of meta-programming.

template <class T>
struct generic
{
  typedef T type;
};

template <class U>
struct generic<U&>
{
  typedef std::reference_wrapper<U> type;
};

template <class T>
using Generic = typename generic<T>::type;

template <class X>
void generic_fun()
{
  std::optional<Generic<X>> op;
  // ...
}

Although the behavior of such "emulated" optional references will be slightly different than that of "normal" optional references.

Exception specifications

First draft of this revision required an aggressive usage of conditional noexcept specifications for nearly every, member- or non-member-, function in the interface. For instance equality comparison was to be declared as:

template <class T>
  bool operator==(const optional<T>& lhs, const optional<T>& rhs)
  noexcept(noexcept(*lhs == *rhs));

This was based on one of our goals: that we want optional<T> to be applicable wherever T is applicable in as many situations as reasonably possible. One such situation occurs where no-throw operations of objects of type T are used to implement a strong exception safety guarantee of some operations. We would like objects of type optional<T> to be also useable in such cases. However, we do not propose this aggressive conditional no-throw guarantees at this time in order for the proposed library component to adhere to the current Library guidelines for conditional noexcept: it is currently only used in move constructor, move assignment and swap. One exception to this rule, we think could be made for optional's move constructor and assignment from type T&&, however we still do not propose this at this time in order to avoid controversy.

Constructors and mutating functions that disengage an optional object are required to be noexcept(true): they only call T's destructor and impose no precondition on optional object's or contained value's state. The same applies to the observers that check the disengaged/engaged state.

The observers that access the contained value — operator* and operator-> — are not declared as noexcept(true) even though they have no good reason to throw. This is because they impose a precondition that optional object shall be engaged, and as per observations from N3248^[6], library vendors may need to use exceptions to test if the implementation has all the necessary precondition-checking code inside. These observer functions are still required not to throw exceptions.

In general, operations on optional objects only throw, when operations delegated to the contained value throw.

Making optional a literal type

We propose that optional<T> be a literal type for trivially destructible T's.

constexpr optional<int> oi{5};
static_assert(oi, "");            // ok
static_assert(oi != nullopt, ""); // ok
static_assert(oi == oi, "");      // ok
int array[*oi];                   // ok: array of size 5 

Making optional<T> a literal-type in general is impossible: the destructor cannot be trivial because it has to execute an operation that can be conceptually described as:

~optional() {
  if (is_engaged()) destroy_contained_value();
}

It is still possible to make the destructor trivial for T's which provide a trivial destructor themselves, and we know an efficient implementation of such optional<T> with compile-time interface — except for copy constructor and move constructor — is possible. Therefore we propose that for trivially destructible T's all optional<T>'s constructors, except for move and copy constructors, as well as observer functions are constexpr. The sketch of reference implementation is provided in this proposal.

We need to make a similar exception for operator-> for types with overloaded operator&. The common pattern in the Library is to use function addressof to avoid the surprise of overloaded operator&. However, we know of no way to implement constexpr version of function template addressof. The best approach we can take is to require that for normal types the non-overloaded (and constexpr) operator& is used to take the address of the contained value, and for the tricky types, implementations can use the normal (non-constexpr) addressof. Similar reasoning applies to operations on optional references in the auxiliary proposal below.

Moved-from state

When a disengaged optional object is moved from (i.e., when it is the source object of move constructor or move assignment) its state does not change. When an engaged object is moved from, we move the contained value, but leave the optional object engaged. A moved-from contained value is still valid (although possibly not specified), so it is fine to consider such optional object engaged. An alternative approach would be to destroy the contained value and make the moved-from optional object disengaged. However, we do not propose this for performance reasons.

In contexts, like returning by value, where you need to call the destructor the second after the move, it does not matter, but in cases where you request the move explicitly and intend to assign a new value in the next step, and if T does not provide an efficient move, the chosen approach saves an unnecessary destructor and constructor call:

optional<array<Big, 1000>> oo = ... // array doesn't have efficient move
op = std::move(oo);
oo = std::move(tmp);

The following is an even more compelling reason. In this proposal std::optional<int> is allowed to be implemented as a TriviallyCopyable type. Therefore, the copy constructor of type std::array<std::optional<int>, 1000> can be implemented using memcpy. With the additional requirement that optional's move constructor should not be trivial, we would be preventing the described optimization.

The fact that the moved-from optional is not disengaged may look "uncomfortable" at first, but this is an invalid expectation. The requirements of library components expressed in 17.6.5.15 (moved-from state of library types) only require that moved-from objects are in a valid but unspecified state. We do not need to guarantee anything above this minimum.

The op = {} syntax

We put the extra requirements in the standardese to make sure that the following syntax works for resetting the optional:

We consider that this will become a common idiom for resetting (putting into default-constructed state) values in C++. While you get that syntax for free for POD types, in optional we have to take extra care to enable it; this is because optional provides three assignment operators: copy/move assignment, assignment from T and from nullopt_t. If we just provided the "intuitive" declaration of assignment from const T& and T&&, the expression above would become ambiguous. The expression above is processed as:

where DEDUCED needs to be deduced from all available overloads. We would have two candidates: move assignment and assignemnt from T&&, which would cause an ambiguity. Therefore, we require that the assignment from T&& is declared in a more convoluted way — as a template:

The additional requirement that decay<U> == T says that the only valid instantiations from this template are these for const T& and T&& (and some other less relevant variations of references to T). But it is still a template, and templates do not participate in the resolution of type DEDUCED.

For the same reason, we require that tag nullopt_t is not DefaultConstructible. Otherwise, because optional provides an assignment from nullopt_t, DEDUCED might also have been deduced as nullopt_t.

Note that it is not the only way to disengage an optional object. You can also use:

op = std::nullopt;

Requirements for swap

For function swap, we require that T is "swappable for lvalues and is_move_constructible<T>::value is true". Similarly, we require that it is declared with the following excption specification:

noexcept(is_nothrow_move_constructible<T>::value && noexcept(swap(declval<T&>(), declval<T&>())))

You can see that we require things of the move constructor but not of move assignment. This is the consequence of the behaviour of function swap, which does three different things based on the engagement states of the two optional objects in question. Conceptually, the behaviour could be described as:

void swap(optional<T>& lhs, optional<T>& rhs)
{
  if ( lhs &&  rhs) swap(*lhs, *rhs);
  if (!lhs &&  rhs) move-construct in lhs, destroy in rhs;
  if ( lhs && !rhs) move-construct in rhs, destroy in lhs;
  if (!lhs && !rhs) /* no-op */;
}

No move-assignment is involved in swap. This is part of the function's contract.

IO operations

The proposed interface for optional values does not contain IO operations: operator<< and operator>>. While we believe that they would be a useful addition to the interface of optional objects, we also observe that there are some technical obstacles in providing them, and we choose not to propose them at this time. Library components like containers, pairs, tuples face the same issue. At present IO operations are not provided for these types. Our preference for optional is to provide an IO solution compatible with this for containers, pairs and tuples, therefore at this point we refrain from proposing a solution for optional alone.

Function value_or

This function template returns a value stored by the optional object if it is engaged, and if not, it falls back to the default value specified in the second argument. It used to be called get_value_or in the previous revisions, but we decided to rename it, as a consequence of disscussions, so that it is similar to another new member function value. This method for specifying default values on the fly rather than tying the default values to the type is based on the observation that different contexts or usages require different default values for the same type. For instance the default value for int can be 0 or -1. The callee might not know what value the caller considers special, so it returns the lack of the requested value explicitly. The caller may be better suited to make the choice what special value to use.

optional<int> queryDB(std::string);
void setPieceCount(int);
void setMaxCount(int);

setPieceCount( queryDB("select piece_count from ...").value_or(0) );
setMaxCount( queryDB("select max_count from ...").value_or(numeric_limits<int>::max()) );

The decision to provide this function is controversial itself. As pointed out by Robert Ramey, the goal of the optional is to make the lack of the value explicit. Its syntax forces two control paths; therefore we will typically see an if-statement (or similar branching instruction) wherever optional is used. This is considered an improvement in correctness. On the other hand, using the default value appears to conflict with the above idea. One other argument against providing it is that in many cases you can use a ternary conditional operator instead:

auto&& cnt = queryDB("select piece_count from ...");
setPieceCount(cnt ? *cnt : 0);

auto&& max = queryDB("select max_count from ...");
setMaxCount(max ? std::move(*max) : numeric_limits<int>::max());

However, in case optional objects are returned by value and immediately consumed, the ternary operator syntax requires introducing an lvalue. This requires more typing and explicit move. This in turn makes the code less safe because a moved-from lvalue is still accessible and open for inadvertent misuse.

There are reasons to make it a free-standing function. (1) It can be implemented by using only the public interface of optional. (2) This function template could be equally well be applied to any type satisfying some requirements NullableProxy. In this proposal, function value_or is defined as a member function and only for optionals. Making a premature generalization would risk standardizing a function with suboptimal performance/utility. While we know what detailed semantics (e.g., the return type) value_or should have for optional, we cannot claim to know the ideal semantics for a generic function. Also, it is not clear to us if this convenience function is equally useful for pointers, as it is for optional objects. By making value_or a member function we leave the room for this name in namespace std for a possible future generalization.

The second argument in the function template's signature is not T but any type convertible to T:

template <class T, class V> 
  typename decay<T>::type optional<T>::value_or(V&& v) const&;
template <class T, class V> 
  typename decay<T>::type optional<T>::value_or(V&& v) &&;

This allows for a certain run-time optimization. In the following example:

optional<string> op{"cat"};
string ans = op.value_or("dog");

Because the optional object is engaged, we do not need the fallback value and therefore to convert the string literal "dog" into type string.

It has been argued that the function should return by constant reference rather than value, which would avoid copy overhead in certain situations:

void observe(const X& x);

optional<X> ox { /* ... */ };
observe( ox.value_or(X{args}) );    // unnecessary copy

However, the benefit of the function value_or is only visible when the optional object is provided as a temporary (without the name); otherwise, a ternary operator is equally useful:

optional<X> ox { /* ... */ };
observe(ox ? *ok : X{args});            // no copy

Also, returning by reference would be likely to render a dangling reference, in case the optional object is disengaged, because the second argument is typically a temporary:

optional<X> ox {nullopt};
auto&& x = ox.value_or(X{args});
cout << x;                              // x is dangling!

There is also one practical problem with returning a reference. The function takes two arguments by reference: the optional object and the default value. It can happen that one is deduced as lvalue reference and the other as rvalue reference. In such case we would not know what kind of reference to return. Returning lvalue reference might prevent move optimization; returning an rvalue reference might cause an unsafe move from lvalue. By returning by value we avoid these problems by requiring one unnecessary move in some cases.

We also do not want to return a constant lvalue reference because that would prevent a copy elision in cases where optional object is returned by value.

It has also been suggested (by Luc Danton) that function optional<T>::value_or<V> should return type decay<common_type<T, V>::type>::type rather than decay<T>::type. This would avoid certain problems, such as loss of accuracy on arithmetic types:

// not proposed
std::optional<int> op = /* ... */;
long gl = /* ... */;

auto lossless = op.value_or(gl);   // lossless deduced as long rather than int

However, we did not find many practical use cases for this extension, so we do not propose it at this time.

Together with function value, value_or makes a set of similarly called functions for accessing the contained value that do not cause an undefined behavior when invoked on a disengaged optional (at the expense of runtime overhead). They differ though, in the return type: one returns a value, the other a reference.

One other similar convenience function has been suggested. Sometimes the default value is not given, and computing it takes some time. We only want to compute it, when we know the optional object is disengaged:

optional<int> oi = /* ... */;

if (oi) {
  use(*oi);
}
else {
  int i = painfully_compute_default();
  use(i);
}

The solution to that situation would be another convenience function which rather than taking a default value takes a callable object that is capable of computing a default value if needed:

use( oi.value_or_call(&painfully_compute_default) );       
// or
use( oi.value_or_call([&]{return painfully_compute_default();} );

We do not propose this, as we prefer to standardize the existing practice. Also, it is not clear how often the above situations may occur, and the tool prove useful.

Function make_optional

We also propose a helper function make_optional. Its semantics is closer to that of make_pair or make_tuple than that of make_shared. You can use it in order for the type of the optional to be deduced:

int i = 1;
auto oi = make_optional(i);          // decltype(oi) == optional<int>

This may occasionally be useful when you need to pick the right overload and not type the type of the optional by hand:

void fun(optional<complex<double>>);
void fun(Current);                   // complex is convertible to Current

complex<double> c{0.0, 0.1};
fun(c);                              // ambiguous
fun({c});                            // ambiguous
fun(make_optional(c));               // picks first overload

This is not very useful in return statements, as long as the converting constructor from T is implicit, because you can always use the brace syntax:

optional<complex<double>> findC()
{
  complex<double> c{0.0, 0.1};
  return {c};
}

make_shared-like function does not appear to be useful at all: it is no different than manually creating a temporary optional object:

// not proposed
fun( make_optional<Rational>(1, 2) );
fun( optional<Rational>{1, 2} );     // same as above

It would also not be a good alternative for tagged placement constructor, because using it would require type T to be MoveConstructible:

// not proposed
auto og = make_optional<Guard>("arg1"); // ERROR: Guard is not MoveConstructible

Such solution works for shared_ptr only because its copy constructor is shallow. One useful variant of shared_ptr-like make_optional would be a function that either creates an engaged or a disengaged optional based on some boolean condition:

// not proposed
return make_optional_if<Rational>(good(i) && not_bad(j), i, j);

// same as:
if (good(i) && not_bad(j)) {
  return {i, j};
}
else {
  return nullopt;
}

// same as:
optional<Rational> or = nullopt;
if (good(i) && not_bad(j)) or.emplace(i, j);
return or; // move-construct on return

Since this use case is rare, and the function call not that elegant, and a two-liner alternative exists, we do not propose it.

Header <utility> or <optional>?

Should optional be packed into header file <utility> or does it deserve its own header? The answer to this question is arbitrary. To great extent (where we believed it made sense) we try to follow what tuple does. The latter has its own header. Optional is pretty much a "stand-alone" library with vast usage. Also, we do not want to overload <utility> too much. There is no problem, on the other hand, with putting optional to <utility>, if this is the feedback from LWG.

Handling initializer_list

Another feature worth considering is a "sequence constructor" (one that takes initializer_list as its argument). It would be enabled (in enable_if sense) only for these Ts that themself provide a sequence constructor. This would be useful to fully support two features we already mentioned above (but chose not to propose).

First, our goal of "copy initialization forwarding" for optional also needs to address the following usages of initializer_list:

vector<int> v = {1, 2, 4, 8};
optional<vector<int>> ov = {1, 2, 4, 8};

assert (v == *ov);

This is not only a syntactical convenience. It also avoids subtle bugs. When perfect forwarding constructor is implemented naively with one variadic constructor, optional vector initialization may render surprising result:

optional<vector<int>> ov = {3, 1};

assert (*ov == vector{3, 1});    // FAILS!
assert (*ov == vector{1, 1, 1}); // TRUE!

However this sequence constructor feature is incompatible with another one: default constructor creating a disengaged optional. This is because, as outlined in the former example, initializer {}, that looks like 0-element list, is in fact interpreted as the request for value-initialization (default constructor call). This may hit programmers that use initializer list in "generic" context:

template <class ...A> // enable_if: every A is int
void fwd(const A&&... a)
{
  optional<vector<int>> o = {a...};
  assert (bool(o)); // not true for empty a
}

If this feature were to be added, we would need to provide an assignment from initializer list and variadic 'emplacement' constructor with the first forwarded argument being initializer_list:

ov = {1, 2, 4, 8};

allocator<int> a;
optional<vector<int>> ou { in_place, {1, 2, 4, 8}, a };

assert (ou == ov);

Since we are not proposing neither perfect forwarding constructor, nor the "copy initialization forwarding", we are also not proposing the sequence constructor. However, in this proposal, the following constructs work:

optional<vector<int>> ov{in_place, {3, 1}};
assert (*ov == vector{3, 1});

ov.emplace({3, 1});
assert (*ov == vector{3, 1});

optional<optional<T>>

The necessity to create a "double" optional explicitly does not occur often. Such type may appear though in generic contexts where we create optional<V> and V only happens to be optional<T>. Some special behavior to be observed in this situation is the following. When copy-initializing with nullopt, the "outermost" optional is initialized to disengaged state. Thus, changing function argument from optional<T> to optional<optional<T>> will silently break the code in places where the argument passed to function happens to be of type nullopt_t:

// before change
void fun(optional<T> v) {
  process(v);
}

fun(nullopt); // process() called

// after change
void fun(optional<optional<T>> v) {
  if (v) process(*v);
  else   doSthElse();
}

fun(nullopt); // process() not called!

This issue would not arise if nullopt were T-specific:

fun(optional<T>::nullopt);            // process() called
fun(optional<optional<T>>::nullopt);  // process() not called

Since T-dependent nullopt is not proposed, in order to create an engaged optional containing a disengaged optional, one needs to use one of the following constructs:

optional<optional<T>> ot {in_place};
optional<optional<T>> ou {in_place, nullopt};
optional<optional<T>> ov {optional<T>{}};

Also note that make_optional will create a "double" optional when called with optional argument:

optional<int> oi;
auto ooi = make_optional(oi);
static_assert( is_same<optional<optional<int>>, decltype(ooi)>::value, "");

Open questions

Allocator support

Optional does not allocate memory. So it can do without allocators. However, it can be useful in compound types like:

typedef vector< optional<vector<int, MyAlloc>>, MyAlloc>; MyVec;
MyVec v{ v2, MyAlloc{} };

One could expect that the allocator argument is forwarded in this constructor call to the nested vectors that use the same allocator. Allocator support would enable this. std::tuple offers this functionality.

Proposed wording

After subclause 17.3.10 ([defns.default.behavior.func]), insert a new subclause. (Subclause [defns.handler] becomes 17.3.12.)

17.3.11 [defns.direct-non-list-init]

dierect-non-list-initialization

A direct-initialization that is not list-initialization.

Add new header in Table 14 (C++ library headers).

Table 14 — C++ library headers

<algorithm>	<fstream>	<list>	<ratio>	<tuple>
<array>	<functional>	<locale>	<regex>	<typeindex>
<atomic>	<future>	<map>	<set>	<typeinfo>
<bitset>	<initializer_list>	<memory>	<sstream>	<type_traits>
<chrono>	<iomanip>	<mutex>	<stack>	<unordered_map>
<codecvt>	<ios>	<new>	<stdexcept>	<unordered_set>
<complex>	<iosfwd>	<numeric>	<streambuf>	<utility>
<condition_variable>	<iostream>	<optional>	<string>	<valarray>
<dequeue>	<istream>	<ostream>	<strstream>	<vector>
<exception>	<iterator>	<queue>	<system_error>
<forward_list>	<limits>	<random>	<thread>

After chapter 20.4 Tuples [tuple], insert a new paragraph. (Chapter [template.bitset] (Class template bitset) becomes 20.6.)

20.5 Optional objects [optional]

20.5.1 In general [optional.general]

This subclause describes class template optional that represents optional objects. An optional object for object types is an object that contains the storage for another object and manages the lifetime of this contained object, if any. The contained object may be initialized after the optional object has been initialized, and may be destroyed before the optional object has been destroyed. The initialization state of the contained object is tracked by the optional object.

20.5.2 Header <optional> synopsis [optional.synop]

namespace std {
  // 20.5.4, optional for object types
  template <class T> class optional;

  // 20.5.5, In-place construction
  struct in_place_t{};
  constexpr in_place_t in_place{};

  // 20.5.6, Disengaged state indicator
  struct nullopt_t{see below};
  constexpr nullopt_t nullopt(unspecified);
  
  // 20.5.7, class bad_optional_access
  class bad_optional_access;

  // 20.5.8, Relational operators
  template <class T>
    constexpr bool operator==(const optional<T>&, const optional<T>&);
  template <class T>
    constexpr bool operator<(const optional<T>&, const optional<T>&);

  // 20.5.9, Comparison with nullopt
  template <class T> constexpr bool operator==(const optional<T>&, nullopt_t) noexcept;
  template <class T> constexpr bool operator==(nullopt_t, const optional<T>&) noexcept;
  template <class T> constexpr bool operator<(const optional<T>&, nullopt_t) noexcept;
  template <class T> constexpr bool operator<(nullopt_t, const optional<T>&) noexcept;

  // 20.5.10, Comparison with T
  template <class T> constexpr bool operator==(const optional<T>&, const T&);
  template <class T> constexpr bool operator==(const T&, const optional<T>&);
  template <class T> constexpr bool operator<(const optional<T>&, const T&);
  template <class T> constexpr bool operator<(const T&, const optional<T>&);

  // 20.5.11, Specialized algorithms
  template <class T> void swap(optional<T>&, optional<T>&) noexcept(see below);
  template <class T> constexpr optional<see below> make_optional(T&&);

  // 20.5.12, hash support
  template <class T> struct hash;
  template <class T> struct hash<optional<T>>;
} // namespace std

A program that necessitates the instantiation of template optional for a reference type, or for possibly cv-qualified types in_place_t or nullopt_t is ill-formed.

20.5.3 Definitions [optional.defs]

An instance of optional<T> is said to be disengaged if it has been default constructed, constructed with or assigned with a value of type nullopt_t, constructed with or assigned with a disengaged optional object of type optional<T>.

An instance of optional<T> is said to be engaged if it is not disengaged.

20.5.4 optional for object types [optional.object]

namespace std {

  template <class T>
  class optional
  {
  public:
    typedef T value_type;

    // 20.5.4.1, constructors
    constexpr optional() noexcept;
    constexpr optional(nullopt_t) noexcept;
    optional(const optional&);
    optional(optional&&) noexcept(see below);
    constexpr optional(const T&);
    constexpr optional(T&&);
    template <class... Args> constexpr explicit optional(in_place_t, Args&&...);
    template <class U, class... Args>
      constexpr explicit optional(in_place_t, initializer_list<U>, Args&&...);

    // 20.5.4.2, destructor
    ~optional();

    // 20.5.4.3, assignment
    optional& operator=(nullopt_t) noexcept;
    optional& operator=(const optional&);
    optional& operator=(optional&&) noexcept(see below);
    template <class U> optional& operator=(U&&);
    template <class... Args> void emplace(Args&&...);
    template <class U, class... Args>
      void emplace(initializer_list<U>, Args&&...);

    // 20.5.4.4, swap
    void swap(optional&) noexcept(see below);

    // 20.5.4.5, observers
    constexpr T const* operator ->() const;
    T* operator ->();
    constexpr T const& operator *() const;
    T& operator *();
    constexpr explicit operator bool() const noexcept;
    constexpr T const& value() const;
    T& value();
    template <class U> constexpr T value_or(U&&) const&;
    template <class U> T value_or(U&&) &&;

  private:
    bool init; // exposition only
    T*   val;  // exposition only
  };

} // namespace std

Engaged instances of optional<T> where T is of object type shall contain a value of type T within its own storage. This value is referred to as the contained value of the optional object. Implementations are not permitted to use additional storage, such as dynamic memory, to allocate its contained value. The contained value shall be allocated in a region of the optional<T> storage suitably aligned for the type T.

Members init and val are provided for exposition only. Implementations need not provide those members. init indicates whether the optional object's contained value has been initialized (and not yet destroyed); val points to (a possibly uninitialized) contained value.

T shall be an object type and shall satisfy the requirements of Destructible (Table 24).

20.5.4.1 Constructors [optional.object.ctor]

constexpr optional<T>::optional() noexcept;
constexpr optional<T>::optional(nullopt_t) noexcept;

Postconditions:

*this is disengaged.

Remarks:

No T object referenced is initialized. For every object type T these constructors shall be constexpr constructors (7.1.5).

optional<T>::optional(const optional<T>& rhs);

Requires:

is_copy_constructible<T>::value is true.

Effects:

If rhs is engaged initializes the contained value as if direct-non-list-initializing an object of type T with the expression *rhs.

Postconditions:

bool(rhs) == bool(*this).

Throws:

Any exception thrown by the selected constructor of T.

optional<T>::optional(optional<T> && rhs) noexcept(see below);

Requires:

is_move_constructible<T>::value is true.

Effects:

If rhs is engaged initializes the contained value as if direct-non-list-initializing an object of type T with the expression std::move(*rhs). bool(rhs) is unchanged.

Postconditions:

bool(rhs) == bool(*this).

Throws:

Any exception thrown by the selected constructor of T.

Remarks:

The expression inside noexcept is equivalent to:

is_nothrow_move_constructible<T>::value

optional<T>::optional(const T& v);

Requires:

is_copy_constructible<T>::value is true.

Effects:

Initializes the contained value as if direct-non-list-initializing an object of type T with the expression v.

Postconditions:

*this is engaged.

Throws:

Any exception thrown by the selected constructor of T.

Remarks:

If T's selected constructor is a constexpr constructor, this constructor shall be a constexpr constructor.

optional<T>::optional(T&& v);

Requires:

is_move_constructible<T>::value is true.

Effects:

Initializes the contained value as if direct-non-list-initializing an object of type T with the expression std::move(v).

Postconditions:

*this is engaged.

Throws:

Any exception thrown by the selected constructor of T.

Remarks:

If T's selected constructor is a constexpr constructor, this constructor shall be a constexpr constructor.

template <class... Args> constexpr explicit optional(in_place_t, Args&&... args);

Requires:

is_constructible<T, Args&&...>::value is true.

Effects:

Initializes the contained value as if constructing an object of type T with the arguments std::forward<Args>(args)....

Postconditions:

*this is engaged.

Throws:

Any exception thrown by the selected constructor of T.

Remarks:

If T's constructor selected for the initialization is a constexpr constructor, this constructor shall be a constexpr constructor.

template <class U, class... Args>
explicit optional(in_place_t, initializer_list<U> il, Args&&... args);

Requires:

is_constructible<T, initializer_list<U>&, Args&&...>::value is true.

Effects:

Initializes the contained value as if constructing an object of type T with the arguments il, std::forward<Args>(args)....

Postconditions:

*this is engaged.

Throws:

Any exception thrown by the selected constructor of T.

Remarks:

The function shall not participate in overload resolution unless is_constructible<T, initializer_list<U>&, Args&&...>::value is true.

Remarks:

If T's constructor selected for the initialization is a constexpr constructor, this constructor shall be a constexpr constructor.

20.5.4.2 Destructor [optional.object.dtor]

optional<T>::~optional();

Effects:

If is_trivially_destructible<T>::value != true and *this is engaged, calls val->T::~T().

Remarks:

If is_trivially_destructible<T>::value == true then this destructor shall be a trivial destructor.

20.5.4.3 Assignment [optional.object.assign]

optional<T>& optional<T>::operator=(nullopt_t) noexcept;

Effects:

If *this is engaged calls val->T::~T() to destroy the contained value; otherwise no effect.

Returns:

*this.

Postconditions:

*this is disengaged.

optional<T>& optional<T>::operator=(const optional<T>& rhs);

Requires:

is_copy_constructible<T>::value is true and is_copy_assignable<T>::value is true.

Effects:

• If *this is disengaged and rhs is disengaged, no effect, otherwise
• if *this is engaged and rhs is disengaged, destroys the contained value by calling val->T::~T(), otherwise
• if *this is disengaged and rhs is engaged, initializes the contained value as if direct-non-list-initializing an object of type T with *rhs, otherwise
• (if both *this and rhs are engaged) assigns *rhs to the contained value.

Returns:

*this.

Postconditions:

bool(rhs) == bool(*this).

Exception Safety:

If any exception is thrown, values of init and rhs.init remain unchanged. If an exception is thrown during the call to T's copy constructor, no effect. If an exception is thrown during the call to T's copy assignment, the state of its contained value is as defined by the exception safety guarantee of T's copy constructor.

optional<T>& optional<T>::operator=(optional<T>&& rhs) noexcept(see below);

Requires:

is_move_constructible<T>::value is true and is_move_assignable<T>::value is true.

Effects:

• If *this is disengaged and rhs is disengaged, no effect, otherwise
• if *this is engaged and rhs is disengaged, destroys the contained value by calling val->T::~T(), otherwise
• if *this is disengaged and rhs is engaged, initializes the contained value as if direct-non-list-initializing an object of type T with std::move(*rhs), otherwise
• (if both *this and rhs are engaged) assigns std::move(*rhs) to the contained value.

Returns:

*this.

Postconditions:

bool(rhs) == bool(*this).

Remarks:

The expression inside noexcept is equivalent to:

is_nothrow_move_assignable<T>::value && is_nothrow_move_constructible<T>::value

Exception Safety:

If any exception is thrown, values of init and rhs.init remain unchanged. If an exception is thrown during the call to T's move constructor, the state of *rhs.val is determined by exception safety guarantee of T's move constructor. If an exception is thrown during the call to T's move assignment, the state of *val and *rhs.val is determined by exception safety guarantee of T's move assignment.

template <class U> optional<T>& optional<T>::operator=(U&& v);

Requires:

is_constructible<T, U>::value is true and is_assignable<U, T>::value is true.

Effects:

If *this is engaged assigns std::forward<U>(v) to the contained value; otherwise initializes the contained value as if direct-non-list-initializing object of type T with std::forward<U>(v).

Returns:

*this.

Postconditions:

*this is engaged.

Exception Safety:

If any exception is thrown, value of init remains unchanged. If an exception is thrown during the call to T's constructor, the state of v is determined by exception safety guarantee of T's constructor. If an exception is thrown during the call to T's assignment, the state of *val and v is determined by exception safety guarantee of T's assignment.

Remarks:

The function shall not participate in overload resolution unless is_same<typename remove_reference<U>::type, T>::value is true.

[Note: The reason to provide such generic assignment and then constraining it so that effectively T == U is to guarantee that assignment of the form o = {} is unambiguous. —end note]

template <class... Args> void optional<T>::emplace(Args&&... args);

Requires:

is_constructible<T, Args&&...>::value is true.

Effects:

Calls *this = nullopt. Then initializes the contained value as if constructing an object of type T with the arguments std::forward<Args>(args)....

Postconditions:

*this is engaged.

Throws:

Any exception thrown by the selected constructor of T.

Exception Safety:

If an exception is thrown during the call to T's constructor, *this is disengaged, and the previous *val (if any) has been destroyed.

template <class U, class... Args> void optional<T>::emplace(initializer_list<U> il, Args&&... args);

Requires:

is_constructible<T, initializer_list<U>&, Args&&...>::value is true.

Effects:

Calls *this = nullopt. Then initializes the contained value as if constructing an object of type T with the arguments il, std::forward<Args>(args)....

Postconditions:

*this is engaged.

Throws:

Any exception thrown by the selected constructor of T.

Exception Safety:

If an exception is thrown during the call to T's constructor, *this is disengaged, and the previous *val (if any) has been destroyed.

Remarks:

The function shall not participate in overload resolution unless is_constructible<T, initializer_list<U>&, Args&&...>::value is true.

20.5.4.4 Swap [optional.object.swap]

void optional<T>::swap(optional<T>& rhs) noexcept(see below);

Requires:

LValues of type T shall be swappable and is_move_constructible<T>::value is true.

Effects:

• If *this is disengaged and rhs is disengaged, no effect, otherwise
• if *this is engaged and rhs is disengaged, initializes the contained value of rhs by direct-initialization with std::move(*(*this)), followed by val->T::~T(), swap(init, rhs.init), otherwise
• if *this is disengaged and rhs is engaged, initializes the contained value of *this by direct-initialization with std::move(*rhs), followed by rhs.val->T::~T(), swap(init, rhs.init), otherwise
• (if both *this and rhs are engaged) calls swap(*(*this), *rhs).

Throws:

Any exceptions that the expressions in the Effects clause throw.

Remarks:

The expression inside noexcept is equivalent to:

is_nothrow_move_constructible<T>::value && noexcept(swap(declval<T&>(), declval<T&>()))

Exception Safety:

If any exception is thrown, values of init and rhs.init remain unchanged. If an exception is thrown during the call to function swap the state of *val and *rhs.val is determined by the exception safety guarantee of swap for lvalues of T. If an exception is thrown during the call to T's move constructor, the state of *val and *rhs.val is determined by the exception safety guarantee of T's move constructor.

20.5.4.5 Observers [optional.object.observe]

constexpr T const* optional<T>::operator->() const;
T* optional<T>::operator->();

Requires:

*this is engaged.

Returns:

val.

Throws:

Nothing.

Remarks:

Unless T is a user-defined type with overloaded unary operator&, the first function shall be a constexpr function.

constexpr T const& optional<T>::operator*() const;
T& optional<T>::operator*();

Requires:

*this is engaged.

Returns:

*val.

Throws:

Nothing.

Remarks:

The first function shall be a constexpr function.

constexpr explicit optional<T>::operator bool() noexcept;

Returns:

init.

Remarks:

this function shall be a constexpr function.

constexpr T const& optional<T>::value() const;
T& optional<T>::value();

Returns:

*val, if bool(*this).

Throws:

bad_optional_access if !*this.

Remarks:

The first function shall be a constexpr function.

template <class U> constexpr T optional<T>::value_or(U&& v) const&;

Requires:

is_copy_constructible<T>::value is true and is_convertible<U&&, T>::value is true.

Returns:

bool(*this) ? **this : static_cast<T>(std::forward<U>(v)).

Throws:

Any exception thrown by the selected constructor of T.

Exception Safety:

If init == true and exception is thrown during the call to T's constructor, the value of init and v remains unchanged and the state of *val is determined by the exception safety guarantee of the selected constructor of T. Otherwise, when exception is thrown during the call to T's constructor, the value of *this remains unchanged and the state of v is determined by the exception safety guarantee of the selected constructor of T.

Remarks:

If the selected constructor of T is a constexpr constructor, this function shall be a constexpr function.

template <class U> T optional<T>::value_or(U&& v) &&;

Requires:

is_move_constructible<T>::value is true and is_convertible<U&&, T>::value is true.

Returns:

bool(*this) ? std::move(**this) : static_cast<T>(std::forward<U>(v)).

Throws:

Any exception thrown by the selected constructor of T.

Exception Safety:

If init == true and exception is thrown during the call to T's constructor, the value of init and v remains unchanged and the state of *val is determined by the exception safety guarantee of the T's constructor. Otherwise, when exception is thrown during the call to T's constructor, the value of *this remains unchanged and the state of v is determined by the exception safety guarantee of the selected constructor of T.

20.5.5 In-place construction [optional.inplace]

struct in_place_t{};
constexpr in_place_t in_place{};

The struct in_place_t is an empty structure type used as a unique type to disambiguate constructor and function overloading. Specifically, optional<T> has a constructor with in_place_t as the first argument followed by an argument pack; this indicates that T should be constructed in-place (as if by a call to placement new expression) with the forwarded argument pack as parameters.

20.5.6 Disengaged state indicator [optional.nullopt]

struct nullopt_t{see below};
constexpr nullopt_t nullopt(unspecified);

The struct nullopt_t is an empty structure type used as a unique type to indicate a disengaged state for optional objects. In particular, optional<T> has a constructor with nullopt_t as single argument; this indicates that a disengaged optional object shall be constructed.

Type nullopt_t shall not have a default constructor. It shall be a literal type. Constant nullopt shall be initialized with an argument of literal type.

20.5.7 Class bad_optional_access [optional.bad_optional_access]

namespace std {
  class bad_optional_access : public logic_error {
  public:
    explicit bad_optional_access(const string& what_arg);
    explicit bad_optional_access(const char* what_arg);
  };
}

The class bad_optional_access defines the type of objects thrown as exceptions to report the situation where an attempt is made to access the value of a disengaged optional object.

bad_optional_access(const string& what_arg);

Effects:

Constructs an object of class bad_optional_access.

Postcondition:

strcmp(what(), what_arg.c_str()) == 0.

bad_optional_access(const char* what_arg);

Effects:

Constructs an object of class bad_optional_access.

Postcondition:

strcmp(what(), what_arg) == 0.

20.5.8 Relational operators [optional.relops]

template <class T> constexpr bool operator==(const optional<T>& x, const optional<T>& y);

Requires:

T shall meet the requirements of EqualityComparable.

Returns:

If bool(x) != bool(y), false; otherwise if bool(x) == false, true; otherwise *x == *y.

Remarks:

Instantiations of this function template for which *x == *y is a core constant expression, shall be constexpr functions.

template <class T> constexpr bool operator<(const optional<T>& x, const optional<T>& y);

Requires:

Expression less<T>{}(*x, *y) shall be well-formed.

Returns:

If (!y), false; otherwise, if (!x), true; otherwise less<T>{}(*x, *y).

Remarks:

Instantiations of this function template for which less<T>{}(*x, *y) is a core constant expression, shall be constexpr functions.

20.5.9 Comparison with nullopt [optional.nullops]

template <class T> constexpr bool operator==(const optional<T>& x, nullopt_t) noexcept;
template <class T> constexpr bool operator==(nullopt_t, const optional<T>& x) noexcept;

Returns:

(!x).

template <class T> constexpr bool operator<(const optional<T>& x, nullopt_t) noexcept;

Returns:

false.

template <class T> constexpr bool operator<(nullopt_t, const optional<T>& x) noexcept;

Returns:

bool(x).

20.5.10 Comparison with T [optional.comp_with_t]

template <class T> constexpr bool operator==(const optional<T>& x, const T& v);

Returns:

bool(x) ? *x == v : false.

template <class T> constexpr bool operator==(const T& v, const optional<T>& x);

Returns:

bool(x) ? v == *x : false.

template <class T> constexpr bool operator<(const optional<T>& x, const T& v);


Returns:

bool(x) ? less<T>{}(*x, v) : true.

20.5.11 Specialized algorithms [optional.specalg]

template <class T> void swap(optional<T>& x, optional<T>& y) noexcept(noexcept(x.swap(y)));

Effects:

calls x.swap(y).

template <class T>
  constexpr optional<typename decay<T>::type> make_optional(T&& v);

Returns:

optional<typename decay<T>::type>(std::forward<T>(v)).

20.5.12 Hash support [optional.hash]

template <class T> struct hash<optional<T>>;

Requires:

the template specilaization hash<T> shall meet the requirements of class template hash (20.9.12). The template specilaization hash<optional<T>> shall meet the requirements of class template hash. For an object o of type optional<T>, if bool(o) == true, hash<optional<T>>()(o) shall evaluate to the same value as hash<T>()(*o).

Auxiliary proposal — optional references

We propose optional references as an auxiliary proposal. This is to give the Committee the freedom to make the decision to accept or not optional references independently of the decision to accept optional values.

Overview

Optional references are surprising to many people because they do not appear to add any more functionality than pointers do. There exist though a couple of arguments in favour optional references:

• optional<T> can be used in generic code, were T can be either a reference or an object.
• It is slightly easier to pass arguments to functions, beacause addressof operator is not required.
• Raw pointers historically add confusion in the sense that it is not clear whether we should delete the value they point to or not, as well as whether we should expect nullptr or not. With optional references the answer to these questions is obvious.

The interface for optional references is somewhat limited compared to that for optional values.

int i = 1;
int j = 2;
optional<int&> ora;                // disengaged optional reference to int
optional<int&> orb = i;            // contained reference refers to object i

*orb = 3;                          // i becomes 3
ora = j;                           // ERROR: optional refs do not have assignment from T
ora = optional<int&>{j};           // contained reference refers to object j
ora = {j};                         // same as line above
orb = ora;                         // rebinds orb to refers to the same object as ora
*orb = 7;                          // j becomes 7
ora.emplace(j);                    // OK: contained reference refers to object j
ora.emplace(i);                    // OK: contained reference now refers to object i
ora = nullopt;                     // OK: ora becomes disengaged

In some aspects optional lvalue references act like raw pointers: they are rebindable, may become disengaged, and do not have special powers (such as extending the life-time of a temporary). In other aspects they act like C++ references: they do not provide pointer arithmetic, operations like comparisons, hashing are performed on referenced objects.

hash<int> hi;
hash<optional<int&>> hoi;

int i = 0;
int j = 0;
optional<int&> ori = i;            // orj and ori refer to two different objects
optional<int&> orj = j;            // but with the same value

assert (hoi(ori) == hi(i));        
assert (hoi(ori) == hoi(orj));     // because we are hashing the values
assert (ori == orj);               // because we are comparing the values

Because optional references are easily implementable with a raw pointer, we require that almost no operation on optional references (except value and value_or) throws exceptions.

Optional references do not popagate constness to the object they indirect to:

int i = 9;
const optional<int&> mi = i;
int& r = *mi;                      // OK: decltype(*mi) == int&
 
optional<const int&> ci = i;
int& q = *ci;                      // ERROR: decltype(*ci) == const int&

Note also the peculiar way in which function make_optional can create optional references:

int i = 1;
auto oi = make_optional(i);          // decltype(oi) == optional<int>
auto ri = make_optional(ref(i));     // decltype(ri) == optional<int&>

Practical use cases

While used significantly less often, optional references are useful in certain cases. For instance, consider that you need to find a possibly missing element in some sort of a collection, and if it is found, change it. The following is a possible implementation of such function that will help with the find.

optional<int&> find_biggest( vector<int>& vec )
{
  optional<int&> biggest; 
  for (int & val : vec) {
    if (!biggest || *biggest < val) {
      biggest.emplace(val);
    }
  }
  return biggest;
} 

This could be alternatively implemented using a raw pointer; however, optional reference makes it clear that the caller will not be owing the object. For another example, consider that we want a function to modify the passed object (storeHere in the example) as one of its responsibilities, but we sometimes cannot provide the object, but we do want the function to perform other responsibilities. The following is the possible implementation

template <typename T>
T getValue( optional<T> newVal = nullopt, optional<T&> storeHere = nullopt )
{
  if (newVal) {
    cached = *newVal;
    
    if (storeHere) {
      *storeHere = *newVal; // LEGAL: assigning T to T
    }      
  }
  return cached;      
}

Again, this can also be implemented with pointers; however, with optional references the ownership of the memory is clear. Additionally, we do not have to use the indirection operator:

static int global = 0;
const int newVal = 2;
return getValue(newVal, global);

Assignment for optional references

The semantics of optional references' copy assignment turned out to be very controversial. This is because whatever semantics for such assignment is chosen, it is confusing to many programmers. An optional reference can be seen as a reference with postponed initialization. In this case, assignment (to engaged optional reference) is expected to have deep copy semantics: it should assign value to the referred object. On the other hand, an optional reference can be seen as a pointer with different syntax. In this case the assignment (to engaged optional reference) should change the reference, so that it refers to the new object. Neither of these models appears more valid than the other. On the other hand, the majority of people insist that optional should be copy-assignable. We choose somewhat arbitralrily to provide a rebinding semantics for std::optional. This is to follow the practice adapted by std::reference_wrapper and boost::optional. In consequence, optional references are not value semantic types: operator== compares something else than copy assignment and constructor are copying. This should not be surprising, though, for a component that is called a "reference".

The other semantics can be implemented by using the following idiom:

void assign_norebind(optional<T&>& optref, T& obj)
{
  if (optref) *optref = obj;
  else        optref.emplace(obj);
}

No rvalue binding for optional references

Optional referencs cannot provide an essential feature of native references: extending the life-time of temporaries (rvalues). Temporaries can be bound to (1) rvalue references and to (2) lvalue references to const. In order to avoid dangling reference problems we need to prevent either type of binfing to optional references. In order to prevent the former, we disallow optional rvalue references altogether. We are not aware of any practical use case for such entities. Since optional lvalue references to const appear useful, we avoid the rvalue binding problem by requiring implementations to "poison" rvalue reference constructors for optional lvalue references to const. This may appear surprising as it is inconsistent with normal (non-optional) reference behavior:

const int& nr = int{1};           // ok
optional<const int&> or = int{1}; // error: int&& ctor deleted

Implementability

This proposal can be implemented as pure library extension, without any compiler magic support, in C++11. An almost full rerefence implementation of this proposal can be found at https://github.com/akrzemi1/Optional/. Below we demonstrate how one can implement optional's constexpr constructors to engaged and disengaged state as well as constexpr operator* for TriviallyDestructible T's.

namespace std {

#if defined NDEBUG
# define ASSERTED_EXPRESSION(CHECK, EXPR) (EXPR)
#else
# define ASSERTED_EXPRESSION(CHECK, EXPR) ((CHECK) ? (EXPR) : (fail(#CHECK, __FILE__, __LINE__), (EXPR)))
  inline void fail(const char* expr, const char* file, unsigned line) { /*...*/ }
#endif

struct dummy_t{};

template <class T>
union optional_storage
{
  static_assert( is_trivially_destructible<T>::value, "" );

  dummy_t dummy_;
  T       value_;

  constexpr optional_storage()            // null-state ctor
    : dummy_{} {}

  constexpr optional_storage(T const& v)  // value ctor
    : value_{v} {}

  ~optional_storage() = default;          // trivial dtor
};


template <class T>
// requires: is_trivially_destructible<T>::value
class optional
{
  bool initialized_;
  optional_storage<T> storage_;

public:
  constexpr optional(nullopt_t) : initialized_{false}, storage_{} {}

  constexpr optional(T const& v) : initialized_{true}, storage_{v} {}

  constexpr T const& operator*() 
  {
    return ASSERTED_EXPRESSION(bool(*this), storage_.value_);
  }
  
  constexpr T const& value()
  {
    return *this ? storage_.value_ : (throw bad_optional_access(""), storage_.value_);
  }

  // ...
};

} // namespace std

Acknowledgements

Many people from the Boost community, participated in the developement of the Boost.Optional library. Sebastian Redl suggested the usage of function emplace.

Daniel Krügler provided numerous helpful suggestions, corrections and comments on this paper; in particular he suggested the addition of and reference implementation for "perfect initialization" operations.

Tony Van Eerd offered many useful suggestions and corrections to the proposal.

People in discussion group "ISO C++ Standard - Future Proposals" provided numerous insightful suggestions: Vladimir Batov (who described and supported the perfect forwarding constructor), Nevin Liber, Ville Voutilainen, Richard Smiths, Dave Abrahams, Chris Jefferson, Jeffrey Yasskin, Nikolay Ivchenkov, Matias Capeletto, Olaf van der Spek, Vincent Jacquet, Kazutoshi Satoda, Vicente J. Botet Escriba, Róbert Dávid, Vincent Jacquet, Luc Danton, Greg Marr, and many more.

Joe Gottman suggested the support for hashing some optional objects.

Nicol Bolas suggested to make operator-> conditionally constexpr based on whether T::operator& is overloaded.

References

1. John J. Barton, Lee R. Nackman, "Scientific and Engineering C++: An Introduction with Advanced Techniques and Examples".
2. Fernando Cacciola, Boost.Optional library (http://www.boost.org/doc/libs/1_49_0/libs/optional/doc/html/index.html)
3. MSDN Library, "Nullable Types (C# Programming Guide)", (http://msdn.microsoft.com/en-us/library/1t3y8s4s.aspx
4. Code Synthesis Tools, "C++ Object Persistence with ODB", (http://www.codesynthesis.com/products/odb/doc/manual.xhtml#7.3)
5. Jaakko Järvi, Boost Tuple Library (http://www.boost.org/doc/libs/1_49_0/libs/tuple/doc/tuple_users_guide.html)
6. Alisdair Meredith, John Lakos, "noexcept Prevents Library Validation" (N3248, http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3248.pdf)
7. Walter E. Brown, "A Preliminary Proposal for a Deep-Copying Smart Pointer" (N3339, http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2012/n3339.pdf)
8. Andrzej Krzemieński, Optional library implementation in C++11 (https://github.com/akrzemi1/Optional/)