P2093R13
Formatted output

"Привет, κόσμος!"
― anonymous user

1. Introduction

A new I/O-agnostic text formatting library was introduced in C++20 ([FORMAT]). This paper proposes integrating it with standard I/O facilities via a simple and intuitive API achieving the following goals:

• Usability

• Unicode support

• Good performance

• Small binary footprint

2. Revision history

Changes since R12:

• Changed section numbers to permalinks in the wording.

• Replaced "string literal encoding" with "the ordinary literal encoding ([lex.charset]) in [print.fun].2.

• Added missing template parameter to forward in [print.fun].3.

• Replaced Let out = vformat(fmt, args) with "The function initializes an automatic variable via string out = vformat(fmt, args); in [print.fun].6 and [ostream.formatted.print].3.

• Replaced "As specified in [format.err.report] or" with "Any exception thrown by the call to vformat ([format.err.report])." in [print.fun].7 and [print.fun].11.

• Removed "�" from [print.fun].8 and [ostream.formatted.print].5.

• Clarified that writing to an output stream is done via an unformatted output function in [ostream.formatted.print].3 and [ostream.formatted.print].6.

• Added a missing definition of the ostream overload of println.

• Moved "The Unicode® Standard Version 14.0 – Core Specification" from Normative references to Bibliography per LWG request and verified that ISO 10646:2020 doesn’t include an equivalent of "U+FFFD Substitution of Maximal Subparts" section from the Unicode standard and therefore can’t be referenced instead.

• Rephrase Throws clauses per LWG feedback and clarify that exceptions from the underlying unformatted functions ([ostream.unformatted]) are propagated in [ostream.formatted.print].4 and [ostream.formatted.print].7.

Changes since R11:

• Corrected header names in the feature test macro definition replacing "also in <format>" with "also in <print> and <ostream>".

• Replaced stream with os in [ostream.formatted.print].2.

Changes since R10:

• Moved "If invoking the (native Unicode) API requires transcoding, implementations should substitute invalid code units with U+FFFD � REPLACEMENT CHARACTER per The Unicode Standard Version 13.0 – Core Specification, Chapter 3.9." to the Recommended practice clause in the definition of vprint_unicode per LEWG feedback.

• Made the definition of the ostream overload of vprint_unicode consistent with the FILE* overload.

• Namespace-qualified the calls to forward.

• Removed a call to make_format_args in [print.fun].1.

• Removed extra ) in the Effects clauses of vprint_unicode and vprint_nonunicode.

• Rendered str in italics in fmt.str.

• Defined the println overload that doesn’t take a stream in terms of the one that does.

• Add missing .str to fmt in the ostream overload of print.

• Updated The Unicode® Standard Version to 14.

Changes since R9:

• Rebased the wording onto the latest draft, most importantly adding support for types that are neither const-iterable nor copyable introduced by [P2418].

Changes since R8:

• Added new SG16 poll results.

• Improved wording for [print.syn].6 (previously [print.syn].31) to remove ambiguities per SG16 feedback.

• Fixed paragraph numbering in the wording.

• Clarified the difference from the Python’s print in § 10 API and naming.

Changes since R7:

• Added a reference to LLVM’s raw_ostream that implements similar mojibake prevention mechanism to § 15 Implementation.

Changes since R6:

• Added new SG16 poll results.

• Rebased the wording onto the latest draft, most importantly adding compile-time checks introduced by [P2216].

• Added "If out contains invalid code units, the behavior is undefined and implementations are encouraged to diagnose it." to the definition of vprint_unicode per SG16 feedback.

• Replaced "invalid code points are substituted with U+FFFD � REPLACEMENT CHARACTER" with " implementations should substitute invalid code units with U+FFFD � REPLACEMENT CHARACTER per The Unicode® Standard Version 13.0 – Core Specification, Chapter 3.9" in § 16 Wording per SG16 feedback.

• Added "The Unicode® Standard Version 13.0 – Core Specification" to Normative references in § 16 Wording.

• Clarified the behavior when mixing encodings in § 11 Unicode.

Changes since R5:

• Added new LEWG poll results.

• Added new SG16 poll results.

• Replaced <io> with <print> per LEWG feedback.

• Clarified the choice of U+FFFD � REPLACEMENT CHARACTER for transcoding errors in § 11 Unicode.

• Clarified the choice of literal encoding in § 11 Unicode.

• Clarified that ANSI escape codes for specifying coding systems are not considered a native system API that supports Unicode in § 11 Unicode.

• Added a reference to Rust’s standard output facility that implements the same mojibake prevention mechanism to § 15 Implementation

Changes since R4:

• Added SG16 Unicode poll results.

• Added a list of candidate headers formatted output functions can be added to.

• Moved the non-ostream overloads of formatted output functions to a separate header for cleaner separation of formatting and I/O facilities.

Changes since R3:

• Replaced _isatty(_fileno(stream)) with GetConsoleMode(_get_osfhandle(_fileno(stream)), ...) in a note in [format.functions] because the former may return 1 for streams not referring to a terminal.

Changes since R2:

• Added better compatibility with other formatted I/O facilities as another advantage of using stdout in § 10 API and naming per SG16 feedback.

• Clarified that [P1885] can be used for literal encoding detection per SG16 feedback.

• Added comparison of Unicode handling in various languages in Appendix A: Unicode tests and a summary in § 11 Unicode per SG16 request.

• Removed incorrect "exposition-only" in § 11 Unicode.

• Replaced "both source and literal encodings are UTF-8, which is enabled by the /utf-8 compiler flag" with "the literal (execution) encoding is UTF-8, which is enabled by the /execution-charset:utf-8 compiler flag" since the source encoding is irrelevant there.

• Rephrased Effects of vprint_unicode for clarity.

Changes since R1:

• Added missing println overloads that take FILE* and ostream&.

• Moved the print functions that take ostream& to <ostream> since <format> shouldn’t depend on <ostream>.

• Clarified why it is useful to provide vprint* functions in § 13 Binary code.

• Rebased the wording onto the latest working draft, N4861, in particular updating the Throws clauses to match existing wording.

• Replaced std::system_error with system_error in the wording.

• Added paragraph numbers to the wording.

Changes since R0:

• Clarified that adding wchar_t overloads will be addressed in a separate paper per the UK C++ panel feedback.

3. SG16 polls (R8)

Poll: Use of UTF-8 as the literal encoding is sufficient for <print> facilities to establish encoding expectations.

Consensus: Consensus in favour.

Poll: Correct the P2093R8 wording for [print.syn].31 to remove ambiguities, and forward P2093 as revised to LEWG with a recommended ship vehicle of C++23.

Consensus: Consensus in favour.

4. SG16 polls (R6)

Poll: When <print> facilities must transcode their formatting results for display on a device, and during that process invalidly-encoded text is encountered, print() should replace the erroneously-encoded code units with U+FFFD REPLACEMENT CHARACTER.

Outcome: Consensus for the position

Poll: When <print> facilities need not transcode their formatting results for display on a device, and during that process invalidly-encoded text is encountered, print() should nevertheless replace the erroneously-encoded code units with U+FFFD REPLACEMENT CHARACTER.

Outcome: Consensus against the direction.

Poll: <format> and <print> facilities should have consistent behavior with respect to encoding expectations for the format string.

No objection to unanimous consent.

Poll: formatters should not be sensitive to whether they are being used with a <format> or <print> facility.

No objection to unanimous consent.

Poll: Regardless of format string encoding assumptions, <format> facilities may be used to format binary data.

Consensus: Strong consensus in favor.

Poll: Regardless of format string encoding assumptions, <print> facilities may be used to format binary data.

Consensus: Weak consensus in favor.

Poll: <print> facilities exhibit undefined behavior when an encoding expectation is present and a format string or formatter output does not match those expectations.

Consensus: Strong consensus in favor.

Poll: <print> facilities exhibit undefined behavior when an encoding expectation is present and a format string or formatter output does not match those expectations and output is directed to a device that has encoding expectations.

Consensus: Stronger consensus in favor relative to previous poll.

Poll: <print> facility implementors are encouraged to provide a run-time means for diagnosing format strings and formatter output that is not well-formed according to the expected encoding.

Consensus: Consensus in favor.

Poll: <format> and <print> facilities should have consistent behavior with respect to encoding expectations for the output of formatters.

Consensus: Strong consensus against.

Poll: <print> facility implementors are encouraged to substitute U+FFFD replacement characters following Unicode guidance when output is directed to a device and transcoding is necessary.

Consensus: Consensus in favor.

Poll: <print> facilities must provide an explicit program-controlled error handling mechanism for violations of encoding expectations.

Poll: Use of UTF-8 as the literal encoding is sufficient for <print> facilities to establish encoding expectations.

Consensus: Very weak consensus.

5. LEWG polls (R5)

Poll: Block P2093 until we have a proposal for a lower level facility that can query tty/console and perform direct output to the console

Outcome: Consensus Against

Poll: We want std::print et al. non ostream overloads to go to:

Outcome: both <print> and <format> have consensus, author can choose one for the next revision.

6. SG16 polls (R3)

Poll: Forward P2093R3 to LEWG.

Consensus? Yes

7. LEWG polls (R2)

Poll: We want P2093R2 to revert moving the print(ostream) overloads to the <ostream> header.

Outcome: Keep the paper as is

8. LEWG polls (R1)

Poll: We prefer std::cout as the default output target.

Poll: Add a member function on ostream instead of a std::print(ostream&, ...) free function overload.

Consensus against.

Poll: Remove std::println from the paper?

No consensus for change.

Poll: We are happy with the design with regards to UTF-8 output.

Unanimous consent.

Attendance: 35

9. Motivating examples

Consider a common task of printing formatted text to stdout:

std::cout << std::format("Hello, {}!", name);

std::print("Hello, {}!", name);

The proposed std::print function improves usability, avoids allocating a temporary std::string object and calling operator<< which performs formatted I/O on text that is already formatted. The number of function calls is reduced to one which, together with std::vformat-like type erasure, results in much smaller binary code (see § 13 Binary code).

Existing alternatives in C++20:

std::cout << "Hello, " << name << "!";

std::printf("Hello, %s!", name);

auto msg = std::format("Hello, {}!", name);
std::fputs(msg.c_str(), stdout);

Another problem is formatting of Unicode text:

std::cout << "Привет, κόσμος!";

Привет, κόσμος!

With the proposed paper

std::print("Привет, κόσμος!");

>>> print("Привет, κόσμος!")
Привет, κόσμος!

This problem is independent of formatting char8_t strings but the same solution applies there. Adding charN_t and wchar_t overloads will be explored in a separate paper in a more general context.

10. API and naming

Many programming languages provide functions for printing text to standard output, often combined with formatting:

Variations of print[f] appear to be the most popular naming choice for this functionality. It is either provided as a free function (most common) or a member function (less common) together with a global object representing standard output stream. Notable exceptions are COBOL, Fortran, and Python 2 which have dedicated language statements and Rust where print! is a function-like macro.

We propose adding a free function called print with overloads for writing to the standard output (the default) and an explicitly passed output stream object. The default output stream can be either stdout or std::cout. We propose using stdout for the following reasons:

• stdout is considerably faster on at least two major implementations (see § 12 Performance).

• Better compatibility with other formatted I/O facilities compared to std::cout and its associated std::streambuf that suffer from private buffering, localization and conversion services that must be synchronized at a lower level.

• print won’t use any formatted output functionality of ostream.

In some languages like Python print only provides the default formatting although some formatting control may be achieved by other means such as named arguments and/or doing formatting manually via str.format or string interpolation. The current paper doesn’t propose such a default formatting facility. A search in a large Python codebase revealed that over 70% of print calls either take a string literal or use interpolation and this doesn’t even account for string variables. These use cases are covered by the current proposal with better usability and potentially better performance because there are no separate formatting function calls.

Since stdout doesn’t have an associated locale we propose using the current global locale for locale-specific formatting which is consistent with format. With cout or another explicitly passed stream, the stream’s locale will be used. In all cases the default formatting is locale-independent.

Another option is to make print a member function of basic_ostream. This would make usage somewhat more awkward:

std::cout.print("Hello, {}!", name);

There are multiple approaches to appending a trailing newline:

• Don’t append a newline automatically: printf in C and other languages.

• Append a newline but don’t format arguments: puts in C (inconsistent with fputs).

• Have two formatting functions/macros, one that appends newline and another that doesn’t: print/println in Java, print!/println! in Rust, Printf/Println in Go, Write/WriteLine in C#/.NET.

• Let the user choose a terminating string defaulting to "\n" and do limited formatting: print in Python and Swift.

We propose not appending a newline automatically for consistency with printf and iostreams:

std::print("Hello, {}!", name);    // doesn’t print a newline
std::print("Hello, {}!\n", name);  // prints a newline

Additionally we can provide a function that appends a newline:

std::println("Hello, {}!", name);  // prints a newline

Although println doesn’t provide much usability improvement compared to print with explicit '\n', it has been an occasionally requested feature in the fmt library ([FMT]).

Another question is which header non-ostream overloads of formatted output functions should go to. Possible options:

• <io>

• <print>

• <format>

• <ostream>

• <utility>

Earlier versions of the paper proposed <io> analogous to <cstdio> so that the future I/O facilities that don’t depend on ostream could be added there. This was changed to a more narrow-focused <print> but <io> can be added in the future once a symmetric input facility becomes available. Using <ostream> is undesirable because this header and its transitive dependencies are very big (42 thousand lines preprocessed on libc++):

% echo '#include <ostream>' | clang++ -E -x c++ - | wc -l
   42491

It also pulls in a lot of unrelated symbols such as ostream insertion operators, global cout, cerr, clog variables and their wchar_t counterparts.

ostream overloads are added to the <ostream> header.

11. Unicode

We can prevent mojibake in the Unicode example by detecting if the string literal encoding is UTF-8 and dispatching to a different function that correctly handles Unicode, for example:

constexpr bool is_utf8() {
  const unsigned char micro[] = "\u00B5";
  return sizeof(micro) == 3 && micro[0] == 0xC2 && micro[1] == 0xB5;
}

template <typename... Args>
void print(string_view fmt, const Args&... args) {
  if (is_utf8())
    vprint_unicode(fmt, make_format_args(args...));
  else
    vprint_nonunicode(fmt, make_format_args(args...));
}

#include <stdio.h>

int main() {
  puts("\xc3\x28"); // Invalid 2 Octet Sequence
}

In Visual C++ is_utf8 will return true if the literal (execution) encoding is UTF-8, which is enabled by the /execution-charset:utf-8 compiler flags or other means, and false otherwise. Literal encoding detection can be implemented in a more elegant way using [P1885].

Note that ANSI escape codes for specifying coding systems ([ISO2022]) are not considered a native system API that supports Unicode for the purposes of this proposal.

We propose using the literal encoding for the following reasons:

1. Consistency with the design of std::format which is locale-independent by default ([P0645]) and disallows implicitly mixing encodings e.g. passing a narrow string into a wide std::format is ill-formed.

2. Consistency with the encoding used for width estimation ([P1868]). The standard wording doesn’t mention the literal encoding explicitly but the fact that the format strings are either literals or other compile-time strings ([P2216]) makes it the only conformant option.

3. Safety: the result of formatted_size does not depend on the global locale by default and a buffer allocated with this size can be passed safely to format_to even if the locale has been changed in the meantime, possibly from another thread.

4. Implementation and usage experience.

5. In the vast majority of cases format strings are literals. For example, analyzing a sample of 100 printf calls from [CODESEARCH] showed that 98 of them are string literals and 2 are string literals wrapped in the _ gettext macro.

6. The active code page and the terminal encoding being unrelated on popular Windows localizations such as Russian where the former is CP1251 while the latter is CP866. Instead of assuming one encoding regardless of the string origin which would often result in mojibake, an explicit encoding indication can be done via the standard extension API, e.g. (exposition only)

   print("Привет, {}!", locale_enc(string_in_locale_encoding));
   

   This is already possible to implement by providing appropriate std::formatter specializations.

This approach has been implemented in the fmt library ([FMT]), successfully tested and used on a variety of platforms.

Users can sometimes restrict the set of used characters to the common subset among multiple encodings (often ASCII) in which case encoding becomes mostly irrelevant. Such "polyglot" strings are fully supported for legacy encodings and partially supported for UTF-8 by the current proposal even though mixing encodings in such a way is a clearly bad practice from a general software engineering point of view.

Here’s an example output on Windows:

At the same time interoperability with legacy code is preserved when literal encoding is not UTF-8. In particular, in case of EBCDIC, Shift JIS or a non-Unicode Windows code page, print will perform no transcoding and the text will be printed as is.

The following table summarizes the behavior of formatted output facilities in different programming languages:

^* - the output is transcoded from a different UTF representation.

Correct means that the test message "Привет, κόσμος!" was fully readable in the terminal output. None of the tested language facilities were able to produce readable output when piped through the standard findstr command on Windows. Java gave the worst results producing both mojibake and replacement characters in this case: "╧ЁштхЄ, ??????!". Most other languages produced valid UTF-8 when the output of findstr was redirected to a file.

The current paper proposes following C, Go, JavaScript and Rust and preserving the original encoding (modulo UTF conversion). The only difference compared to printf is that we fix the console output on Windows. Java’s approach is problematic for the following reasons:

• There is a silent data loss for valid Unicode code points when the output is redirected to a file.

• It is more expensive because of transcoding.

• It may give an unusable result when piped through standard Windows commands like findstr.

• It transcodes into legacy encodings that are rarely used in practice nowadays. For example, usage of CP1251 dropped from 4.3% to 0.6% in the last 12+ years ([ENCODING-TRENDS]), including a 0.4% drop while the current paper was in review.

The full listings of test programs are given in Appendix A: Unicode tests.

12. Performance

All the performance benefits of std::format ([FORMAT]) automatically carry over to this proposal. In particular, locale-independence by default reduces global state and makes formatting more efficient compared to stdio and iostreams. There are fewer function calls (see § 13 Binary code) and no shared formatting state compared to iostreams.

The following benchmark compares the reference implementation of print with printf and ostream. This benchmark formats a simple message and prints it to the output stream redirected to /dev/null. It uses the Google Benchmark library [GOOGLE-BENCH] to measure timings:

#include <cstdio>
#include <iostream>

#include <benchmark/benchmark.h>
#include <fmt/ostream.h>

void printf(benchmark::State& s) {
  while (s.KeepRunning())
    std::printf("The answer is %d.\n", 42);
}
BENCHMARK(printf);

void ostream(benchmark::State& s) {
  std::ios::sync_with_stdio(false);
  while (s.KeepRunning())
    std::cout << "The answer is " << 42 << ".\n";
}
BENCHMARK(ostream);

void print(benchmark::State& s) {
  while (s.KeepRunning())
    fmt::print("The answer is {}.\n", 42);
}
BENCHMARK(print);

void print_cout(benchmark::State& s) {
  std::ios::sync_with_stdio(false);
  while (s.KeepRunning())
    fmt::print(std::cout, "The answer is {}.\n", 42);
}
BENCHMARK(print_cout);

void print_cout_sync(benchmark::State& s) {
  std::ios::sync_with_stdio(true);
  while (s.KeepRunning())
    fmt::print(std::cout, "The answer is {}.\n", 42);
}
BENCHMARK(print_cout_sync);

BENCHMARK_MAIN();

The benchmark was compiled with Apple clang version 11.0.0 (clang-1100.0.33.17) with -O3 -DNDEBUG and run on macOS 10.15.4. Below are the results:

Run on (8 X 2800 MHz CPU s)
CPU Caches:
  L1 Data 32K (x4)
  L1 Instruction 32K (x4)
  L2 Unified 262K (x4)
  L3 Unified 8388K (x1)
Load Average: 1.83, 1.88, 1.82
---------------------------------------------------------​-
Benchmark                Time             CPU   Iterations
---------------------------------------------------------​-
printf                87.0 ns         86.9 ns      7834009
ostream                255 ns          255 ns      2746434
print                 78.4 ns         78.3 ns      9095989
print_cout            89.4 ns         89.4 ns      7702973
print_cout_sync       91.5 ns         91.4 ns      7903889

Both print and printf are ~3 times faster than cout even with synchronization to the standard C streams turned off. print is 14% faster when printing to stdout than to cout. For this reason and because print doesn’t use formatting facilities of ostream we propose using stdout as the default output stream and providing an overload for writing to ostream.

On Windows 10 with Visual C++ 2019 the results are similar although the difference between print writing to stdout and cout is smaller with stdout being 7% faster:

Run on (1 X 2808 MHz CPU )
CPU Caches:
  L1 Data 32K (x1)
  L1 Instruction 32K (x1)
  L2 Unified 262K (x1)
  L3 Unified 8388K (x1)
---------------------------------------------------------​-
Benchmark                Time             CPU   Iterations
---------------------------------------------------------​-
printf                 835 ns          816 ns       746667
ostream               2410 ns         2400 ns       280000
print                  580 ns          572 ns      1120000
print_cout             623 ns          614 ns      1120000
print_cout_sync        615 ns          614 ns      1120000

13. Binary code

We propose minimizing per-call binary code size by applying the type erasure mechanism from [P0645]. In this approach all the formatting and printing logic is implemented in a non-variadic function vprint. Inline variadic print function only constructs a format_args object, representing an array of type-erased argument references, and passes it to vprint*. Here is a simplified example:

void vprint(string_view fmt, format_args args);

template<class... Args>
  inline void print(string_view fmt, const Args&... args) {
    return vprint(fmt, make_format_args(args...));
  }

We provide vprint* overloads so that users can apply the same technique to their own code. For example:

void vlog(log_level level, string_view fmt, format_args args) {
  // Print the log level and use vprint* overloads to format and print the
  // message.
}

template<class... Args>
  inline void log(log_level level, string_view fmt, const Args&... args) {
    return vlog(level, fmt, make_format_args(args...));
  }

Here vlog that implements the logging logic is not parameterized on formatting argument types resulting in less code bloat compared to a naive templated version. As a real-world example, this technique has been applied in the Folly Logger ([FOLLY]) bringing ~5x binary size reduction per logging function call.

Below we compare the reference implementation of print to standard formatting facilities. All the code snippets are compiled with clang (Apple clang version 11.0.0 clang-1100.0.33.17) with -O3 -DNDEBUG -c -std=c++17 and the resulting binaries are disassembled with objdump -S:

void printf_test(const char* name) {
  printf("Hello, %s!", name);
}

__Z11printf_testPKc:
       0:       55      pushq   %rbp
       1:       48 89 e5        movq    %rsp, %rbp
       4:       48 89 fe        movq    %rdi, %rsi
       7:       48 8d 3d 08 00 00 00    leaq    8(%rip), %rdi
       e:       31 c0   xorl    %eax, %eax
      10:       5d      popq    %rbp
      11:       e9 00 00 00 00  jmp     0 <__Z11printf_testPKc+0x16>

void ostream_test(const char* name) {
  std::cout << "Hello, " << name << "!";
}

__Z12ostream_testPKc:
       0:       55      pushq   %rbp
       1:       48 89 e5        movq    %rsp, %rbp
       4:       41 56   pushq   %r14
       6:       53      pushq   %rbx
       7:       48 89 fb        movq    %rdi, %rbx
       a:       48 8b 3d 00 00 00 00    movq    (%rip), %rdi
      11:       48 8d 35 6c 03 00 00    leaq    876(%rip), %rsi
      18:       ba 07 00 00 00  movl    $7, %edx
      1d:       e8 00 00 00 00  callq   0 <__Z12ostream_testPKc+0x22>
      22:       49 89 c6        movq    %rax, %r14
      25:       48 89 df        movq    %rbx, %rdi
      28:       e8 00 00 00 00  callq   0 <__Z12ostream_testPKc+0x2d>
      2d:       4c 89 f7        movq    %r14, %rdi
      30:       48 89 de        movq    %rbx, %rsi
      33:       48 89 c2        movq    %rax, %rdx
      36:       e8 00 00 00 00  callq   0 <__Z12ostream_testPKc+0x3b>
      3b:       48 8d 35 4a 03 00 00    leaq    842(%rip), %rsi
      42:       ba 01 00 00 00  movl    $1, %edx
      47:       48 89 c7        movq    %rax, %rdi
      4a:       5b      popq    %rbx
      4b:       41 5e   popq    %r14
      4d:       5d      popq    %rbp
      4e:       e9 00 00 00 00  jmp     0 <__Z12ostream_testPKc+0x53>
      53:       66 2e 0f 1f 84 00 00 00 00 00   nopw    %cs:(%rax,%rax)
      5d:       0f 1f 00        nopl    (%rax)

void print_test(const char* name) {
  print("Hello, {}!", name);
}

__Z10print_testPKc:
       0:	55 	pushq	%rbp
       1:	48 89 e5 	movq	%rsp, %rbp
       4:	48 83 ec 10 	subq	$16, %rsp
       8:	48 89 7d f0 	movq	%rdi, -16(%rbp)
       c:	48 8d 3d 19 00 00 00 	leaq	25(%rip), %rdi
      13:	48 8d 4d f0 	leaq	-16(%rbp), %rcx
      17:	be 0a 00 00 00 	movl	$10, %esi
      1c:	ba 0d 00 00 00 	movl	$13, %edx
      21:	e8 00 00 00 00 	callq	0 <__Z10print_testPKc+0x26>
      26:	48 83 c4 10 	addq	$16, %rsp
      2a:	5d 	popq	%rbp
      2b:	c3 	retq

The code generated for the print_test function that uses the reference implementation of print described in this proposal is more than 2x smaller than the ostream code and has one function call instead of three. The printf code is further 2x smaller but doesn’t have any error handling. Adding error handling would make its code size closer to that of print.

The following factors contribute to the difference in binary code size between print and printf:

• Passing format string as string_view instead of const char*.

• Capturing and passing argument type information.

• Preparing the array of formatting arguments.

14. Impact on existing code

The current proposal adds new functions to the headers <print> and <ostream> and should have no impact on existing code.

15. Implementation

The proposed print function has been implemented in the open-source fmt library [FMT] and has been in use for about 6 years.

Rust’s standard output facility uses essentially the same approach for preventing mojibake when printing to console on Windows ([RUST-STDIO]). The main difference is that invalid code units are reported as errors in Rust.

LLVM’s raw_ostream [LLVM-OSTREAM] also implements this approach when writing to console on Windows. The main difference is that in case of invalid UTF-8 it falls back on writing raw (not transcoded) data.

16. Wording

Add an entry for __cpp_lib_print to section "Header <version> synopsis [version.syn]", in a place that respects the table’s current alphabetic order:

#define __cpp_lib_print 202112L **placeholder**  // also in <print> and <ostream>

Add the header <print> to the "C++ library headers" table in [headers], in a place that respects the table’s current alphabetic order.

Add after subsection "Header synopsis [iomanip.syn]":

namespace std {
  template<class... Args>
    void print(​format-string<Args...> fmt, Args&&... args);
  template<class... Args>
    void print(FILE* stream, ​format-string<Args...> fmt, Args&&... args);

  template<class... Args>
    void println(​format-string<Args...> fmt, Args&&... args);
  template<class... Args>
    void println(FILE* stream, ​format-string<Args...> fmt, Args&&... args);

  void vprint_unicode(string_view fmt, format_args args);
  void vprint_unicode(FILE* stream, string_view fmt, format_args args);

  void vprint_nonunicode(string_view fmt, format_args args);
  void vprint_nonunicode(FILE* stream, string_view fmt, format_args args);
}

Modify subsection "Header <ostream> synopsis [ostream.syn]":

...
template<class charT, class traits, class T>
basic_ostream<charT, traits>& operator<<(basic_ostream<charT, traits>&& os, const T& x);

template<class... Args>
  void print(ostream& os, ​format-string<Args...> fmt, Args&&... args);
template<class... Args>
  void println(ostream& os, ​format-string<Args...> fmt, Args&&... args);

void vprint_unicode(ostream& os, string_view fmt, format_args args);
void vprint_nonunicode(ostream& os, string_view fmt, format_args args);

Add a new subsection to "Formatting and manipulators [iostream.format]":

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

1 Effects: Equivalent to:

  print(stdout, fmt, std::forward<Args>(args)...);

template<class... Args>
  void print(FILE* stream, ​format-string<Args...> fmt, Args&&... args);

2 Effects: If the ordinary literal encoding ([lex.charset]) is UTF-8, equivalent to:

  vprint_unicode(stream, fmt.​str, make_format_args(std::forward<Args>(args)...));

  vprint_nonunicode(stream, fmt.​str, make_format_args(std::forward<Args>(args)...));

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

3 Effects: Equivalent to:

  println(stdout, fmt, std::forward<Args>(args)...);

template<class... Args>
  void println(FILE* stream, ​format-string<Args...> fmt, Args&&... args);

4 Effects: Equivalent to:

  print(stream, "{}\n", format(fmt, std::forward<Args>(args)...));

void vprint_unicode(string_view fmt, format_args args);

5 Effects: Equivalent to:

  vprint_unicode(stdout, fmt, args);

void vprint_unicode(FILE* stream, string_view fmt, format_args args);

string out = vformat(fmt, args);

[ Note: On POSIX and Windows, stream referring to a terminal means that, respectively, isatty(fileno(stream)) and GetConsoleMode(_get_osfhandle(_fileno(stream)), ...) return nonzero. — end note ]

[ Note: On Windows, the native Unicode API is WriteConsoleW. — end note ]

7 Throws: Any exception thrown by the call to vformat ([format.err.report]). system_error if a call by the implementation to an operating system or other underlying API results in an error that prevents the function from meeting its specifications.

8 Recommended practice: If invoking the native Unicode API requires transcoding, implementations should substitute invalid code units with U+FFFD REPLACEMENT CHARACTER per The Unicode Standard Version 14.0 – Core Specification, Chapter 3.9.

void vprint_nonunicode(string_view fmt, format_args args);

9 Effects: Equivalent to:

  vprint_nonunicode(stdout, fmt, args);

void vprint_nonunicode(FILE* stream, string_view fmt, format_args args);

11 Throws: Any exception thrown by the call to vformat ([format.err.report]). system_error if a call by the implementation to an operating system or other underlying API results in an error that prevents the function from meeting its specifications.

Add subsection "Print [ostream.formatted.print]" to "Formatted output functions [ostream.formatted]":

template<class... Args>
  void print(ostream& os, ​format-string<Args...> fmt, Args&&... args);

1 Effects: If the ordinary literal encoding ([lex.charset]) is UTF-8, equivalent to:

  vprint_unicode(os, fmt.​str, make_format_args(std::forward<Args>(args)...));

  vprint_nonunicode(os, fmt.​str, make_format_args(std::forward<Args>(args)...));

template<class... Args>
  void println(ostream& os, ​format-string<Args...> fmt, Args&&... args);

2 Effects: Equivalent to:

  print(os, "{}\n", format(fmt, std::forward<Args>(args)...));

void vprint_unicode(ostream& os, string_view fmt, format_args args);

string out = vformat(os.getloc(), fmt, args);

4 Throws: Any exception thrown by the call to vformat ([format.err.report]) or the unformatted output function ([ostream.unformatted]). system_error if a call by the native Unicode API results in an error that prevents the function from meeting its specifications.

5 Recommended practice: If invoking the native Unicode API requires transcoding, implementations should substitute invalid code units with U+FFFD REPLACEMENT CHARACTER per The Unicode Standard Version 14.0 – Core Specification, Chapter 3.9.

void vprint_nonunicode(ostream& os, string_view fmt, format_args args);

7 Throws: Any exception thrown by the call to vformat ([format.err.report]) or the unformatted output function ([ostream.unformatted]).

Add to Bibliography:

Appendix A: Unicode tests

This appendix gives full listings of programs for testing Unicode handling in various formatting facilities as well as test commands and their output on different platforms. The code contains additional sanity checks to ensure that the strings are encoded in some form of UTF as opposed to a legacy encoding.

C (test.c):

#include <stdio.h>
#include <stdlib.h>

int main() {
  const char* message = "Привет, κόσμος!\n";
  if ((unsigned char)message[0] != 0xD0 && (unsigned char)message[1] != 0x9F)
    abort();
  printf(message);
}

Go (test.go):

package main

import "fmt"
import "log"

func main() {
  var message = "Привет, κόσμος!"
  if message[0] != 0xD0 && message[1] != 0x9F {
    log.Fatal("wrong encoding")
  }
  fmt.Println(message)
}

Java (Test.java):

class Test {
  public static void main(String[] args) {
    String message = "Привет, κόσμος!\n";
    if (message.charAt(0) != 0x41F) throw new RuntimeException();
    System.out.print(message);
  }
}

JavaScript / Node.js (test.js):

message = "Привет, κόσμος!";
if (message.charCodeAt(0) != 0x41F) throw "wrong encoding";
console.log(message);

Python (test.py):

message = "Привет, κόσμος!"
if ord(message[0]) != 0x41F:
    raise Exception()
print(message)

Rust (test.rs):

fn main() {
  if "Привет, κόσμος!".chars().nth(0).unwrap() as u32 != 0x41F {
    panic!();
  }
  println!("Привет, κόσμος!");
}

Linux:

$ cc test.c -o c-test
$ ./c-test
Привет, κόσμος!
$ ./c-test > out-c-linux.txt

$ go build -o go-test test.go
$ ./go-test
Привет, κόσμος!
$ ./go-test > out-go-linux.txt

$ java Test
Привет, κόσμος!
$ java Test > out-java-linux.txt

$ node test.js
Привет, κόσμος!
$ node test.js > out-js-linux.txt

$ python3 test.py
Привет, κόσμος!
$ python3 test.py > out-py-linux.txt

$ rustc test.rs -o rust-test
$ ./rust-test
Привет, κόσμος!
$ ./rust-test > out-rust-linux.txt

All output files are in UTF-8:

• out-c-linux.txt

• out-go-linux.txt

• out-java-linux.txt

• out-js-linux.txt

• out-py-linux.txt

• out-rust-linux.txt

Linux configuration:

• Ubuntu Focal 20.04 with the ru_RU.UTF-8 locale

• cc: gcc 9.3.0

• go: go1.13.8

• java: openjdk 11.0.9.1

• node: v14.5.0

• python3: 3.7.5

• rustc: 1.47.0

macOS:

% cc test.c -o c-test
% ./c-test
Привет, κόσμος!
% ./c-test > out-c-macos.txt

% go build -o test-go test.go
% ./test-go
Привет, κόσμος!
% ./test-go > out-go-macos.txt

% java Test
Привет, κόσμος!
% java Test > out-java-macos.txt

% node test.js
Привет, κόσμος!
% node test.js > out-js-macos.txt

% python3 test.py
Привет, κόσμος!
% python3 test.py > out-py-macos.txt

% rustc test.rs -o rust-test
% ./rust-test
Привет, κόσμος!
% ./rust-test > out-rust-macos.txt

All output files are in UTF-8:

• out-c-macos.txt

• out-go-macos.txt

• out-java-macos.txt

• out-js-macos.txt

• out-py-macos.txt

• out-rust-macos.txt

macOS configuration:

• macOS Catalina 10.15.7 with the ru_RU.UTF-8 locale which is the default for Russian

• cc: Apple clang version 12.0.0 (clang-1200.0.32.27)

• go: go1.15.5

• java: openjdk 14.0.1

• node: v14.5.0

• python3: 3.7.5

• rustc: 1.47.0

Windows:

>cl /Fe:c-test.exe test.c
...
>c-test
╨Я╤А╨╕╨▓╨╡╤В, ╬║╧М╧Г╬╝╬┐╧В!
>c-test > out-c-windows.txt
>c-test | findstr ,
╨Я╤А╨╕╨▓╨╡╤В, ╬║╧М╧Г╬╝╬┐╧В!

>go build -o go-test.exe test.go
>go-test
Привет, κόσμος!
>go-test > out-go-windows.txt
>go-test | findstr ,
╨Я╤А╨╕╨▓╨╡╤В, ╬║╧М╧Г╬╝╬┐╧В!

>java Test
Привет, ??????!
>java Test > out-java-windows.txt
>java Test | findstr ,
╧ЁштхЄ, ??????!

>node test.js
Привет, κόσμος!
>node test.js > out-js-windows.txt
>node test.js | findstr ,
╨Я╤А╨╕╨▓╨╡╤В, ╬║╧М╧Г╬╝╬┐╧В!

>python test.py
Привет, κόσμος!
>python test.py > out-py-windows.txt
Traceback (most recent call last):
  File "...\test.py", line 4, in <module>
    print(message)
  File "...\Python39\lib\encodings\cp1251.py", line 19, in encode
    return codecs.charmap_encode(input,self.errors,encoding_table)[0]
UnicodeEncodeError: 'charmap' codec can’t encode characters in position 8-13: character maps to <undefined>
>python test.py | findstr ,
Traceback (most recent call last):
  File "...\test.py", line 4, in <module>
    print(message)
  File "...\Python39\lib\encodings\cp1251.py", line 19, in encode
    return codecs.charmap_encode(input,self.errors,encoding_table)[0]
UnicodeEncodeError: 'charmap' codec can’t encode characters in position 8-13: character maps to <undefined>

>rustc test.rs -o rust-test.exe
>rust-test
Привет, κόσμος!
>rust-test > out-rust-windows.txt
>rust-test | findstr ,
╨Я╤А╨╕╨▓╨╡╤В, ╬║╧М╧Г╬╝╬┐╧В!

C, JavaScript (node.js), Rust and Go produced valid UTF-8 when the output was redirected to files. Java produced a file in the legacy CP1251 encoding with ? for non-representable code points. Python failed on transcoding to CP1251. Output files:

• out-c-windows.txt

• out-go-windows.txt

• out-java-windows.txt

• out-js-windows.txt

• out-py-windows.txt

• out-rust-windows.txt

Windows configuration:

• Windows 10 10.0.19041 with Russian Region and Language settings

• cl: Microsoft (R) C/C++ Optimizing Compiler Version 19.28.29335

• go: go1.15.6

• java: Java HotSpot(TM) 64-Bit Server VM (build 15.0.1+9-18)

• node: v14.15.3

• python: 3.9.1

• rustc: v14.15.3

17. Acknowledgements

Thanks to Corentin Jabot for his work on text encodings in C++ and in particular [P1885] that will simplify implementation of the current proposal.

Thanks to Roger Orr, Peter Brett, Hubert Tong, the BSI C++ panel and Tom Honermann for their feedback, support, constructive criticism and contributions to the proposal.