1. Changes from r3
- 
     Wording improvements 
2. Changes from r2
- 
     Rename map transform 
- 
     Remove "Alternative Names" section 
- 
     Address feedback on disallowing function pointers 
- 
     Remove capability for mapping void 
- 
     Tighten wording 
- 
     Discuss SFINAE-friendliness 
3. Changes from r1
- 
     Add list of programming languages to proposed solution section 
- 
     Change map 
- 
     Update wording to use std :: monostate void 
4. Changes from r0
- 
     More notes on P0650 
- 
     Discussion about mapping of functions returning void 
5. Motivation
image get_cute_cat ( const image & img ) { return add_rainbow ( make_smaller ( make_eyes_sparkle ( add_bow_tie ( crop_to_cat ( img )))); } 
But there’s a problem. What if there’s not a cat in the picture? What if there’s no good place to add a bow tie? What if it has its back turned and we can’t make its eyes sparkle? Some of these operations could fail.
One option would be to throw exceptions on failure. However, there are many code bases which do not use exceptions for a variety of reasons. There’s also the possibility that we’re going to get lots of pictures without cats in them, in which case we’d be using exceptions for control flow. This is commonly seen as bad practice, and has an item warning against it in the C++ Core Guidelines.
Another option would be to make those operations which could fail return a 
std :: optional < image > get_cute_cat ( const image & img ) { auto cropped = crop_to_cat ( img ); if ( ! cropped ) { return std :: nullopt ; } auto with_tie = add_bow_tie ( * cropped ); if ( ! with_tie ) { return std :: nullopt ; } auto with_sparkles = make_eyes_sparkle ( * with_tie ); if ( ! with_sparkles ) { return std :: nullopt ; } return add_rainbow ( make_smaller ( * with_sparkles )); } 
Our code now has a lot of boilerplate to deal with the case where a step fails. Not only does this increase the noise and cognitive load of the function, but if we forget to put in a check, then suddenly we’re down the hole of undefined behaviour if we 
Another possibility would be to call 
std :: optional < image > get_cute_cat ( const image & img ) { try { auto cropped = crop_to_cat ( img ); auto with_tie = add_bow_tie ( cropped . value ()); auto with_sparkles = make_eyes_sparkle ( with_tie . value ()); return add_rainbow ( make_smaller ( with_sparkles . value ())); catch ( std :: bad_optional_access & e ) { return std :: nullopt ; } } 
Again, this is using exceptions for control flow. There must be a better way.
6. Proposed solution
This paper proposes adding additional member functions to 
std :: optional < image > get_cute_cat ( const image & img ) { return crop_to_cat ( img ) . and_then ( add_bow_tie ) . and_then ( make_eyes_sparkle ) . transform ( make_smaller ) . transform ( add_rainbow ); } 
We’ve successfully got rid of all the manual checking. We’ve even improved on the clarity of the non-optional example, which needed to either be read inside-out or split into multiple declarations.
This is common in other programming languages. Here is a list of programming languages which have a 
- 
     Java: Optional 
- 
     Swift: Optional 
- 
     Haskell: Maybe 
- 
     Rust: Option 
- 
     OCaml: option 
- 
     Scala: Option 
- 
     Agda: Maybe 
- 
     Idris: Maybe 
- 
     Kotlin: T ? 
- 
     StandardML: option 
- 
     C#: Nullable 
Here is a list of programming languages which have a 
- 
     C++ 
- 
     I couldn’t find any others 
All that we need now is an understanding of what 
6.1. transform 
   
For example, if you have a 
auto  s  =  opt_string . transform ([]( auto &&  s )  {  return  s . size ();  }); which is somewhat equivalent to:
if ( opt_string ) { std :: size_t s = opt_string -> size (); } 
template < class T > class optional { template < class Return > std :: optional < Return > transform ( function < Return ( T ) > func ); }; 
It takes any callable object (like a function). If the 
If you come from a functional programming or category theory background, you may recognise this as a functor map.
6.2. and_then 
   
For example, say you have 
std :: optional < int >  i  =  opt_string . and_then ( stoi ); Which is roughly equivalent to:
if ( opt_string ) { std :: optional < int > i = stoi ( * opt_string ); } 
template < class T > class optional { template < class Return > std :: optional < Return > and_then ( function < std :: optional < Return > ( T ) > func ); }; 
It takes any callable object which returns a 
Again, those from an FP background will recognise this as a monadic bind.
6.3. or_else 
   
get_opt (). or_else ([]{ std :: cout << "get_opt failed" ;}); get_opt (). or_else ([]{ throw std :: runtime_error ( "get_opt_failed" )}); 
Users can easily abstract these patterns if they are common:
void opt_log ( std :: string_view msg ) { return [ = ] { std :: cout << msg ; }; } void opt_throw ( std :: string_view msg ) { return [ = ] { throw std :: runtime_error ( msg ); }; } get_opt (). or_else ( opt_log ( "get_opt failed" )); get_opt (). or_else ( opt_throw ( "get_opt failed" )); 
It has one overload (expositional):
template < class T > class optional { template < class Return > std :: optional < T > or_else ( function < Return () > func ); }; 
6.4. Chaining
With these two functions, doing a lot of chaining of functions which could fail becomes very clean:
std :: optional < int > foo () { return a (). and_then ( b ) . and_then ( c ) . and_then ( d ) . and_then ( e ); } 
Taking the example of 
std :: optional < int > i = opt_string . and_then ( stoi ) . transform ([]( auto i ) { return i * 2 ; }); 
We can also intersperse our chain with error checking code:
std :: optional < int > i = opt_string . and_then ( stoi ) . or_else ( opt_throw ( "stoi failed" )) . transform ([]( auto i ) { return i * 2 ; }); 
7. How other languages handle this
Rust has an 
8. Considerations
8.1. Disallowing function pointers
In San Diego, a straw poll on disallowing passing function pointers to the functions proposed was made with the following results:
| SF | F | N | A | SA | | 3 | 6 | 1 | 0 | 1 | 
This poll was carried out with the understanding that some parts of Ranges (in particular 
template < InputRange R , CopyConstructible F > requires View < R > && is_object_v < F &&> Invocable < F & , iter_reference_t < iterator_t < R >>> class transform_view ; 
I imagine the confusion comes from 
An object type is a (possibly cv-qualified) type that is not a function type, not a reference type, and not cv void.
As far as I am aware, if we disallowed passing function pointers here then this would be the only place in the standard library which does so. Alongside the issue of consistency, ease-of-use is damaged. For example:
int this_will_never_be_overloaded ( double ); my_optional_double . map ( this_will_never_be_overloaded ); //not allowed //have to write this instead my_optional_double . map ([]( double d ){ return this_will_never_be_overloaded ( d ); }); //or use some sort of macro my_optional_double . map ( LIFT ( this_will_never_be_overloaded )); 
For library extensions which are supposed to aid ergonomics and simplify code, this goes in the opposite direction.
The benefit of disallowing function pointers is avoidance of build breakages when a previously non-overloaded function is overloaded. This is not a new issue: it applies to practically every single higher order function in the C++ standard library, including functions in 
8.2. transform void 
   There are three main options for handling 
get_optional () // returns std::optional<T> . transform ( print_value ) // returns std::optional<std::monostate> . transform ( notify_success ); // Is only called when get_optional returned a value 
This proposal disallows mapping 
8.3. More functions
Rust’s Option class provides a lot more than 
8.4. transform 
   It would be possible to merge all of these into a single function which handles all use cases. However, I think this would make the function more difficult to teach and understand.
8.5. Operator overloading
We could provide operator overloads with the same semantics as the functions. For example, 
// original crop_to_cat ( img ) . and_then ( add_bow_tie ) . and_then ( make_eyes_sparkle ) . transform ( make_smaller ) . transform ( add_rainbow ); // rewritten crop_to_cat ( img ) >= add_bow_tie >= make_eyes_sparkle | make_smaller | add_rainbow ; 
Another option would be 
8.6. Applicative Functors
template < class Return > std :: optional < Return > transform ( std :: optional < function < Return ( T ) >> func ); 
This would give functional programmers the set of operations which they may expect from a monadic-style interface. However, I couldn’t think of many use-cases of this in C++. If some are found then we could add the extra overload.
8.7. Overloaded and_then 
   Instead of an additional 
o . and_then ( a ,  []{ throw  std :: runtime_error ( "oh no" );}); The above will throw if 
8.8. SFINAE-friendliness
struct foo { void non_const () {} }; std :: optional < foo > f = foo {}; auto l = []( auto && x ) { x . non_const (); }; //error: passing 'const foo' as 'this' argument discards qualifiers f . transform ( l ); 
If p0847 is accepted and this proposal is rebased on top of it, then this would no longer be an issue, and 
9. Pitfalls
Users may want to write code like this:
std :: optional < int > foo ( int i ) { return a (). and_then ( b ) . and_then ( get_func ( i )); } 
The problem with this is 
One possible solution to this would be to add an additional function, 
std :: optional < int > foo ( int i ) { return a (). and_then ( b ) . bind_with ([ i ](){ return get_func ( i )}); } 
9.1. Other solutions
There is a proposal for adding a general monadic interface to C++. Unfortunately doing the kind of composition described above would be very verbose with the current proposal without some kind of Haskell-style 
std :: optional < int > get_cute_cat ( const image & img ) { return functor :: map ( functor :: map ( monad :: bind ( monad :: bind ( crop_to_cat ( img ), add_bow_tie ), make_eyes_sparkle ), make_smaller ), add_rainbow ); } 
My proposal is not necessarily an alternative to this proposal; compatibility between the two could be ensured and the generic proposal could use my proposal as part of its implementation. This would allow users to use both the generic syntax for flexibility and extensibility, and the member-function syntax for brevity and clarity.
If 
Another option would be to use a ranges-style interface for the general monadic interface:
std :: optional < int > get_cute_cat ( const image & img ) { return crop_to_cat ( img ) | monad :: bind ( add_bow_tie ) | monad :: bind ( make_eyes_sparkle ) | functor :: map ( make_smaller ) | functor :: map ( add_rainbow ); } 
9.2. Interaction with other proposals
Deducing this would ease implementation and solve the issue of SFINAE-friendliness for 
There is a proposal for std::expected which would benefit from many of these same ideas. If the idea to add monadic interfaces to standard library classes on a case-by-case basis is chosen rather than a unified non-member function interface, then compatibility between this proposal and the 
Mapping functions which return 
Any proposals which make lambdas or overload sets easier to write and pass around will greatly improve this proposal. In particular, proposals for lift operators and abbreviated lambdas would ensure that the clean style is preserved in the face of many anonymous functions and overloads.
If an operator overloading approach is taken and the workflow operator (
10. Implementation experience
This proposal has been implemented here.
11. Proposed Wording
11.1. Add declarations for the monadic operations to the synopsis of class template optional in [optional.optional]:
//... template < class U > constexpr T value_or ( U && ) const & ; template < class U > constexpr T value_or ( U && ) && ; // [optional.monadic], monadic operations template < class F > constexpr auto and_then ( F && f ) & ; template < class F > constexpr auto and_then ( F && f ) && ; template < class F > constexpr auto and_then ( F && f ) const & ; template < class F > constexpr auto and_then ( F && f ) const && ; template < class F > constexpr auto transform ( F && f ) & ; template < class F > constexpr auto transform ( F && f ) && ; template < class F > constexpr auto transform ( F && f ) const & ; template < class F > constexpr auto transform ( F && f ) const && ; template < class F > constexpr optional or_else ( F && f ) & ; template < class F > constexpr optional or_else ( F && f ) && ; template < class F > constexpr optional or_else ( F && f ) const & ; template < class F > constexpr optional or_else ( F && f ) const && ; // [optional.mod], modifiers void reset () noexcept ; //... 
11.2. Add new subclause "Monadic operations [optional.monadic]" between [optional.observe] and [optional.mod]:
template < class F > constexpr auto and_then ( F && f ) & ; template < class F > constexpr auto and_then ( F && f ) const & ; 
Let 
Mandates: 
Returns:  The return type is 
template < class F > constexpr auto and_then ( F && f ) && ; template < class F > constexpr auto and_then ( F && f ) const && ; 
Let 
Mandates: 
Returns: The return type is 
template < class F > constexpr auto transform ( F && f ) & ; template < class F > constexpr auto transform ( F && f ) const & ; 
Let 
Mandates: 
Returns: The return type is 
template < class F > constexpr auto transform ( F && f ) && ; template < class F > constexpr auto transform ( F && f ) const && ; 
Let 
Mandates: 
Returns: The return type is 
template < class F > constexpr optional or_else ( F && f ) & ; template < class F > constexpr optional or_else ( F && f ) const & ; 
Mandates: 
Expects: 
Effects: If 
template < class F > constexpr optional or_else ( F && f ) && ; template < class F > constexpr optional or_else ( F && f ) const && ; 
Mandates: 
Expects: 
Effects: If 
11.3. Acknowledgements
Thank you to Michael Wong and Chris Di Bella for representing this paper to the committee. Thanks to Kenneth Benzie, Vittorio Romeo, Jonathan Müller, Adi Shavit, Nicol Bolas, Vicente Escribá, Barry Revzin, and especially Casey Carter for review and suggestions.