P2418R2: Add support for <code>std::generator</code>-like types to <code>std::format</code>

1. Proposal

[P2286] raised an issue of formatting std::generator from [P2168] and similar views with C++20 std::format. The issue is illustrated in the following example:

auto ints_coro(int n) -> std::generator<int> {
  for (int i = 0; i < n; ++i) {
    co_yield i;
  }
}
std::format("{}", ints_coro(10)); // error

Unfortunately we cannot make std::generator formattable because it is neither const-iterable nor copyable and std::format takes arguments by const&. This hasn’t been a problem in C++20 because range adapters which can also be not const-iterable are usually copyable. However, it will likely become a problem in the future once coroutines are more widely adopted.

This paper proposes solving the issue by making std::format and other formatting functions take arguments by forwarding references.

Other benefits of using forwarding references:

Formatting of non-const-iterable views can be more efficient by avoiding a copy.

It becomes possible to detect common lifetime errors, for example:

auto joined = std::format_join(std::vector{10, 20, 30, 40, 50, 60}, ":");
std::format("{:02x}", joined); // UB but can be made ill-formed with this proposal

2. Changes since R1

Removed previously added "All types that have formatter specializations satisfy the Formatter requirements unless specified otherwise." from [formatter.requirements].
Removed "[Note 1 : Enabled specializations meet the Formatter requirements, and disabled specializations do not. — end note]" in [format.formatter.spec], paragraph 2.
Added "and meets the Formatter requirements" to [format.formatter.spec], paragraph 4.
Replaced Formatter with BasicFormatter in [format.formatter.spec], paragraph 6, [format.arg], paragraph 4, [format.functions], paragraphs 20 and 25 and [format.arg.store], paragraph 2.

3. Changes since R0

Added LEWG poll results.
Improved the wording.
Added the Acknowledgements section.

4. LEWG polls

Poll: Send P2418R0 (Adding support for std::generator to std::format) to LWG for C++23 and as a DR for C++20, treated as an urgent matter.

SF	F	N	A	SA
11	5	1	0	0

5. Impact on existing code

This change will break formatting of bit fields:

struct S {
  int bit: 1;
};

auto s = S();
std::format("{}", s.bit); // will become ill-formed

Supporting bit fields was one of the reasons std::format passed arguments by const& in the first place. However, there are simple workarounds for this:

std::format("{}", +s.bit); // use + or cast to int

6. Implementation experience

The proposal has been implemented in the {fmt} library. Arguments have been passed by forwarding references since {fmt} 6.0 released about two years ago and non-const& argument support in formatter specializations was added recently.

7. Wording

All wording is relative to the C++ working draft [N4892].

Update the value of the feature-testing macro __cpp_lib_format to the date of adoption in [version.syn]:

Change in [format.syn]:

namespace std {
  // [format.functions], formatting functions
  template<class... Args>
    string format(format-string<Args...> fmt, const Args&... args);
  template<class... Args>
    wstring format(wformat-string<Args...> fmt, const Args&... args);
  template<class... Args>
    string format(const locale& loc, format-string<Args...> fmt,
                  const Args&... args);
  template<class... Args>
    wstring format(const locale& loc, wformat-string<Args...> fmt,
                   const Args&... args);

  ...

  template<class Out, class... Args>
    Out format_to(Out out, format-string<Args...> fmt, const Args&... args);
  template<class Out, class... Args>
    Out format_to(Out out, wformat-string<Args...> fmt, const Args&... args);
  template<class Out, class... Args>
    Out format_to(Out out, const locale& loc, format-string<Args...> fmt,
                  const Args&... args);
  template<class Out, class... Args>
    Out format_to(Out out, const locale& loc, wformat-string<Args...> fmt,
                  const Args&... args);

  ...

  template<class Out, class... Args>
    format_to_n_result<Out> format_to_n(Out out, iter_difference_t<Out> n,
                                        format-string<Args...> fmt,
                                        const Args&... args);
  template<class Out, class... Args>
    format_to_n_result<Out> format_to_n(Out out, iter_difference_t<Out> n,
                                        wformat-string<Args...> fmt,
                                        const Args&... args);
  template<class Out, class... Args>
    format_to_n_result<Out> format_to_n(Out out, iter_difference_t<Out> n,
                                        const locale& loc,
                                        format-string<Args...> fmt,
                                        const Args&... args);
  template<class Out, class... Args>
    format_to_n_result<Out> format_to_n(Out out, iter_difference_t<Out> n,
                                        const locale& loc,
                                        wformat-string<Args...> fmt,
                                        const Args&... args);

  template<class... Args>
    size_t formatted_size(format-string<Args...> fmt, const Args&... args);
  template<class... Args>
    size_t formatted_size(wformat-string<Args...> fmt, const Args&... args);
  template<class... Args>
    size_t formatted_size(const locale& loc, format-string<Args...> fmt,
                          const Args&... args);
  template<class... Args>
    size_t formatted_size(const locale& loc, wformat-string<Args...> fmt,
                          const Args&... args);
      
  ...

  template<class Context = format_context, class... Args>
    format-arg-store<Context, Args...>
      make_format_args(const Args&... fmt_args);
  template<class... Args>
    format-arg-store<wformat_context, Args...>
      make_wformat_args(const Args&... args);

  ...
}

Change in [format.functions]:

template<class... Args>
  string format(format-string<Args...> fmt, const Args&... args);

...

template<class... Args>
  wstring format(wformat-string<Args...> fmt, const Args&... args);

...

template<class... Args>
  string format(const locale& loc, format-string<Args...> fmt,
                const Args&... args);

...

template<class... Args>
  wstring format(const locale& loc, wformat-string<Args...> fmt,
                 const Args&... args);

...

template<class Out, class... Args>
  Out format_to(Out out, format-string<Args...> fmt, const Args&... args);

...

template<class Out, class... Args>
  Out format_to(Out out, wformat-string<Args...> fmt, const Args&... args);

...

template<class Out, class... Args>
  Out format_to(Out out, const locale& loc, format-string<Args...> fmt,
                const Args&... args);

...

template<class Out, class... Args>
  Out format_to(Out out, const locale& loc, wformat-string<Args...> fmt,
                const Args&... args);

...

template<class Out, class... Args>
  format_to_n_result<Out> format_to_n(Out out, iter_difference_t<Out> n,
                                      format-string<Args...> fmt,
                                      const Args&... args);
template<class Out, class... Args>
  format_to_n_result<Out> format_to_n(Out out, iter_difference_t<Out> n,
                                      wformat-string<Args...> fmt,
                                      const Args&... args);
template<class Out, class... Args>
  format_to_n_result<Out> format_to_n(Out out, iter_difference_t<Out> n,
                                      const locale& loc,
                                      format-string<Args...> fmt,
                                      const Args&... args);
template<class Out, class... Args>
  format_to_n_result<Out> format_to_n(Out out, iter_difference_t<Out> n,
                                      const locale& loc,
                                      wformat-string<Args...> fmt,
                                      const Args&... args);

...

20 Preconditions: Out models output_iterator<const charT&>, and formatter<T_i, charT> meets the Formatter requirements ([formatter.requirements]) for each T_i in Args.

...

template<class... Args>
  size_t formatted_size(format-string<Args...> fmt, const Args&... args);
template<class... Args>
  size_t formatted_size(wformat-string<Args...> fmt, const Args&... args);
template<class... Args>
  size_t formatted_size(const locale& loc, format-string<Args...> fmt,
                        const Args&... args);
template<class... Args>
  size_t formatted_size(const locale& loc, wformat-string<Args...> fmt,
                        const Args&... args);

...

25 Preconditions: formatter<T_i, charT> meets the Formatter requirements ([formatter.requirements]) for each T_i in Args.

...

Change in [formatter.requirements]:

A type F meets the Formatter requirements if:

it meets the
- Cpp17DefaultConstructible (Table 27),
- Cpp17CopyConstructible (Table 29),
- Cpp17CopyAssignable (Table 31), and
- Cpp17Destructible (Table 32)
requirements,
it is swappable ([swappable.requirements]) for lvalues, and
the expressions shown in ~~Table 67~~ are valid and have the indicated semantics.

...

Given character type charT, output iterator type Out, and formatting argument type T, in ~~Table~~ 67:

f is a value of type F,
u is an lvalue of type T,
t is a value of a type convertible to (possibly const) T,
PC is basic_format_parse_context<charT>,
FC is basic_format_context<Out, charT>,
pc is an lvalue of type PC, and
fc is an lvalue of type FC.

pc.begin() points to the beginning of the format-spec ([format.string]) of the replacement field being formatted in the format string. If format-spec is empty then either pc.begin() == pc.end() or *pc.begin() == '}'.

Table 67: Formatter requirements [tab:formatter]

Expression	Return type	Requirement
~~`f.parse(pc)`~~	~~`PC::iterator`~~	Parses format-spec (20.20.2) for type T in the range `[pc.begin(), pc.end())` until the first unmatched character. Throws `format_error` unless the whole range is parsed or the unmatched character is `}`. [Note 1: This allows formatters to emit meaningful error messages. — end note] Stores the parsed format specifiers in `*this` and returns an iterator past the end of the parsed range.
`f.format(t, fc)`	`FC::iterator`	Formats `t` according to the specifiers stored in `*this`, writes the output to `fc.out()` and returns an iterator past the end of the output range. The output shall only depend on `t`, `fc.locale()`, `fc.arg(n)` for any value `n` of type `size_t`, and the range `[pc.begin(), pc.end())` from the last call to `f.parse(pc)`.
`f.format(u, fc)`	`FC::iterator`	As above, but does not modify `u`.

Change in [format.formatter.spec]:

2 Let charT be either char or wchar_t. Each specialization of formatter is either enabled or disabled, as described below.

~~[Note 1 : Enabled specializations meet the Formatter requirements, and disabled specializations do not. — end note]~~

...

4 If the library provides an explicit or partial specialization of formatter<T, charT>, that specialization is enabled except as noted otherwise.

...

6 An enabled specialization formatter<T, charT> meets the Formatter requirements (20.20.6.1).

Change in [format.arg]:

namespace std {
  template
  class basic_format_arg {
  private:
    ...
    template<class T> explicit basic_format_arg(const T& v) noexcept;  // exposition only
    ...
}

...

template<class T> explicit basic_format_arg(const T& v) noexcept;

4 Constraints: The template specialization

typename Context::template formatter_type<T>

meets the Formatter requirements ([formatter.requirements]). The extent to which an implementation determines that the specialization meets the Formatter requirements is unspecified, except that as a minimum the expression

typename Context::template formatter_type<T>()
  .format(declval<const T&>(), declval<Context&>())

shall be well-formed when treated as an unevaluated operand.

...

The class handle allows formatting an object of a user-defined type.

namespace std {
  template<class Context>
  class basic_format_arg<Context>::handle {
    const void* ptr_;                                             // exposition only
    void (*format_)(basic_format_parse_context<char_type>&,
                    Context&, const void*);                       // exposition only

    template<class T> explicit handle(const T& val) noexcept;  // exposition only
    ...
  };
}

template<class T> explicit handle(const T& val) noexcept;

Effects: Initializes ptr_ with addressof(val) and format_ with

[](basic_format_parse_context<char_type>& parse_ctx,
   Context& format_ctx, const void* ptr) {
  typename Context::template formatter_type<T> f;
  parse_ctx.advance_to(f.parse(parse_ctx));
  format_ctx.advance_to(f.format(
    *static_cast<const T*>(ptr), format_ctx));
}

Change in [format.arg.store]:

template<class Context = format_context, class... Args>
  format-arg-store<Context, Args...> make_format_args(const Args&... fmt_args);

2 Preconditions: The type typename Context::template formatter_type<T_i> meets the Formatter requirements ([formatter.requirements]) for each T_i in Args.

...

template<class... Args>
  format-arg-store<wformat_context, Args...> make_wformat_args(const Args&... args);

Add to [diff.cpp20.utilities]:

8. Acknowledgements

Thanks Barry Revzin for bringing up the issue of formatting std::generator in [P2286]. Thanks Tim Song and Tomasz Kamiński for wording improvement suggestions.

P2418R2
Add support for `std::generator`-like types to `std::format`

Published Proposal, 2021-09-24

1. Proposal

2. Changes since R1

3. Changes since R0

4. LEWG polls

5. Impact on existing code

6. Implementation experience

7. Wording

8. Acknowledgements

References

Informative References

P2418R2Add support for std::generator-like types to std::format

Published Proposal, 2021-09-24

1. Proposal

2. Changes since R1

3. Changes since R0

4. LEWG polls

5. Impact on existing code

6. Implementation experience

7. Wording

8. Acknowledgements

References

Informative References

P2418R2
Add support for `std::generator`-like types to `std::format`