| Document Number: | P2071R0 |
|---|---|
| Date: | 2020-01-13 |
| Audience: | SG16, EWG |
| Reply-to: | Tom Honermann <tom@honermann.net> Peter Bindels <peterbindels@gmail.com> |
This proposal continues the effort R. Martinho Fernandes initiated that culminated in P1097R2[P1097R2]. This proposal does not deviate from the general design intent in Fernandes' work, but does deviate in the following specific details:
C++ programmers have been able to portably use characters outside of the basic source character set in character and string literals since the introduction of universal-character-names in C++11. For example:
U'\u0100' // UTF-32 character literal with U+0100 {LATIN CAPITAL LETTER A WITH MACRON}
u8"\u0100\u0300" // UTF-8 string literal with U+0100 {LATIN CAPITAL LETTER A WITH MACRON} U+0300 {COMBINING GRAVE ACCENT}This proposal enables the above literals to be written using Unicode assigned names instead of Unicode code point values.
U'\N{LATIN CAPITAL LETTER A WITH MACRON}' // Equivalent to U'\u0100'
u8"\N{LATIN CAPITAL LETTER A WITH MACRON}\N{COMBINING GRAVE ACCENT" // Equivalent to u8"\u0100\u0300"Prior presentations of P1097 to EWG-I and EWG received strong encouragement:
| SF | F | N | A | SA |
|---|---|---|---|---|
| 5 | 9 | 7 | 0 | 0 |
| SF | F | N | A | SA |
|---|---|---|---|---|
| 8 | 16 | 8 | 1 | 1 |
Two areas of concern were raised during discussion in EWG in Belfast, 2019:
The introduction of universal-character-names in C++11 benefitted programmers by allowing them to portably encode characters outside of the basic source character set without having to resort to use of octal or hexadecimal escape-sequences to explicitly encode code units. However, Unicode code points by themselves do not clearly communicate to readers of the code which character is to be encoded; hence the code comments included with the code examples in the introduction. Allowing programmers to directly use Unicode assigned character names avoids the need for side channel communications, like code comments, that might get out of sync over time.
Use of UTF-8 as the encoding for source files has increased over time, but impediments to adoption remain. For example, Microsoft Visual C++ still defaults to a locale dependent encoding and that encourages limiting source files to ASCII. If the C++ community were to migrate en masse to UTF-8, then one might question whether universal-character-names would become a legacy backward compatibility feature since programmers could reliably type the intended character in their source code directly. And if universal-character-names were to become an anachronism, then what use would be served by introducing a named character escape?
Unicode defines a number of characters that, even when they can be typed directly, can result in confusion. These include invisible characters such as U+200B {ZERO WIDTH SPACE}, combining characters such as U+0300 {COMBINING GRAVE ACCENT}, visually indistinct characters such as U+003B {SEMICOLON} and U+037E {GREEK QUESTION MARK}, and characters with RTL (right-to-left) directionality. Consider how the following string literals containing these characters are rendered. In cases like these, use of escape sequences improves clarity; thus motivation for use of Unicode escape sequences will remain.
|
"" "" "̀" ";" ";" "´" "́" "´" "Ω" "Ω" "A" "Α" "А" "Ꭺ" "ꓮ" "𐊠" "𖽀" |
// U+0000200B {ZERO WIDTH SPACE} // U+0000200F {RIGHT-TO-LEFT MARK} // U+00000300 {COMBINING GRAVE ACCENT} // U+0000003B {SEMICOLON} // U+0000037E {GREEK QUESTION MARK} // U+000000B4 {ACUTE ACCENT} // U+00000301 {COMBINING ACUTE ACCENT} // U+00001FFD {GREEK OXIA} // U+000003A9 {GREEK CAPITAL LETTER OMEGA} // U+00002126 {OHM SIGN} // U+00000041 {LATIN CAPITAL LETTER A} // U+00000391 {GREEK CAPITAL LETTER ALPHA} // U+00000410 {CYRILLIC CAPITAL LETTER A} // U+000013AA {CHEROKEE LETTER GO} // U+0000A4EE {LISU LETTER A} // U+000102A0 {CARIAN LETTER A} // U+00016F40 {MIAO LETTER ZZYA} |
Named character escapes are supported in various forms in other programming languages. The following is the result of a brief survey of various languages. For languages that include such support, more details can be found in the Design considerations section.
| Language | Named character escape support |
|---|---|
| C# | No |
| D | Yes; HTML 5 named character references |
| Go | No |
| Java | No |
| Javascript | No |
| Perl | Yes; Unicode names, aliases, and named sequences |
| PHP | No |
| Python | Yes; Unicode names and aliases |
| Raku | Yes; Unicode names, aliases, named sequences, and emoji sequences |
| Ruby | No |
| Rust | No |
| Swift | No |
| Visual Basic | No |
There are numerous choices for how support for named characters can be integrated into C++. Useful questions for making design choices include:
Named character escapes are proposed as a more readable alternative to universal-character-names. As such, it is desirable that they be similar in syntax to universal-character-names and other existing escape sequences.
The syntax proposed by Fernandes in P1097R2[P1097R2] is modeled after the syntax adopted for Python and consists of a \N escape introducer followed by a name enclosed in curly brackets. For example:
'\N{LATIN CAPITAL LETTER A}'
"\N{LATIN CAPITAL LETTER A WITH MACRON}"
Other choices for the escape introducer are possible; the Backward compatibility section discusses some possible motivation for preferring \u and/or \U and the Proposal options section includes this alternate syntax as an option.
Options for recognized names and how to match them are discussed in subsequent sections.
As proposed, only one name is allowed per named character escape, but that is an artificial limitation. Raku allows a sequence of comma separated names to be specified in a single escape. This is a natural extension if names are permitted to identify sequences of characters instead of a single character. The following would all be equivalent. This proposal leaves this option to a future extension; see the Possible future extensions section.
"\N{LATIN CAPITAL LETTER A WITH MACRON, COMBINING GRAVE ACCENT}"
"\N{LATIN CAPITAL LETTER A WITH MACRON}\N{COMBINING GRAVE ACCENT}"
"\u0100\u0300"
Perl and Raku both allow Unicode code point numbers to be specified as character names and could enable a syntax that avoids the strict 4 or 8 number requirements of universal-character-names as well as the natural U+NNNN style frequently used to identify Unicode characters. The following could all be equivalent. This proposal also leaves this option for a future extension as discussed in the Possible future extensions section.
"\N{U+0100}"
"\N{U+100}"
"\N{U+00000100}"
"\N{0x0100}"
"\N{256}"
"\u0100"
A named character escape feature is not particularly useful unless accompanied by at least one source of character names. The following list contains sources of character names that are consulted by at least one implementation of named character escapes in another programming language.
The first three are defined by the Unicode Consortium, part of the Unicode standard, and synchronized with ISO/IEC 10646. The names specified in each are designed in concert, share a common namespace, are immutable once published, and Unicode guarantees no conflicts between them. See the Unicode character encoding stability policy[UCESP] for more details. These sources are consulted for named character escapes in Perl, Python, and Raku.
The next two sources specify emoji character sequences. Though produced by the Unicode Consortium, they are not part of the Unicode standard, and are not covered by the Unicode character encoding stability policy[UCESP]. These two sources don't technically provide names; they provide optional descriptions. The provided descriptions use characters, particularly : and ,, that are disallowed in the names provided by the first three sources. These sources are consulted for named character escapes in Raku.
The last source is the specification of names recognized for use as named character references in HTML documents. This source is used for the implementation of named character escapes in the D programming language.
The stability guarantees offered by the Unicode standard are a strong motivator for their use and, as such, this proposal adopts them as the name sources to use.
The list of Unicode assigned names associates at most one name with each character. There are some characters that are not assigned a name in this list, for example, U+0080 is simply listed as a <control> character with no name. In some of these cases, the Unicode aliases list provides one or more names. For example, U+0080 has assigned aliases of PADDING CHARACTER (a figment alias) and PAD (an abbreviation alias).
Unicode aliases provide another critical service. As mentioned above, once assigned, names are immutable. Corrections are only offered by providing an alias. Aliases come in five varieties:
It is conceivable that implementors could desire, or be requested to, support additional implementation-defined names; perhaps including from the additional sources listed above. Since new characters and names will continue to be added to the Unicode standard, caution is warranted to avoid the possibility of introducing conflicting names over time. The description of the UAX44-LM2 name matching algorithm describes a historical case of how such a conflict once occurred. Any support for additional names should ensure that they occupy a non-overlapping namespace with the Unicode assigned names. Out of caution, this proposal disallows additional implementation-defined names.
Names can be finicky things. Having to remember whether a name is, for example, ZERO WIDTH SPACE or ZERO-WIDTH SPACE is likely to frustrate programmers. Some programmers might prefer zero width space.
Unicode provides a straight forward algorithm for matching names with various allowances including case-insensitivity, omission of some hyphens (-), and substitution of underscore (_) for space characters. UAX44-LM2 is included in the Unicode standard via Unicode Standard Annex #44[UAX#44].
The UAX44-LM2 matching rule would accept any of the following names as a match for U+200B {ZERO WIDTH SPACE}
ZERO WIDTH SPACE
ZERO-WIDTH SPACE
zero-width space
ZERO width S P_A_C EPortably using named character escapes will require implementations to agree on a minimum version of the name sources.
Thanks to the adoption of P1025R1[P1025R1] in Rapperswil, 2019, the C++ standard has a normative floating reference to ISO/IEC 10646[ISO/IEC10646], the ISO/IEC standard that specifies a subset of what is specified in the Unicode standard and is kept synchronized with it. ISO/IEC 10646:2017 includes the Unicode assigned names (in section 33), name aliases (in section 33), and named character sequences (in section 27).
The floating reference to ISO/IEC 10646 indicates a dependence on the version that is current at the time of standardization. Thus, conformance with the C++ standard will require conformance with the latest available publication of ISO/IEC 10646.
Implementors must be allowed, and encouraged, to conform to more recent versions of ISO/IEC 10646 as they are published.
Support for named escape sequences exists in several programming languages. The following details of existing practice were obtained from these documentation sources. The author has not verified the accuracy of this information.
Capabilities vary across languages:
| Language | Name sources | Comma separated names | Name matching | Matches code point numbers |
|---|---|---|---|---|
| D | HTML 5 | No | Exact match? | No |
| Perl | Unicode names Unicode name aliases Unicode named sequences registered custom aliases |
No |
Optionally, script qualified short names Optionally, loose matching (case insensitive, ignore underscore, most spaces, and most non-medial hyphens) |
Yes |
| Python | Unicode names Unicode name aliases |
No | Case-insensitive | No |
| Raku | Unicode names Unicode name aliases Unicode named sequences emoji ZWJ sequences emoji sequences |
Yes | Exact match? | Yes |
Examples:
| Language | Code |
|---|---|
| D |
|
| Perl |
|
| Python |
|
| Raku |
|
Escape sequences beyond those required in the standard are conditionally-supported ([lex.ccon]p7). For implementations that currently define a meaning for \N in character or string literals, the use of \N in this proposal is technically a breaking change.
Gcc, Clang, and Microsoft Visual C++ all accept \N as an escape sequence with the semantic effect of substituting N such that "\N{xxx}" is equivalent to "N{xxx}". However, they each emit a warning regarding an unrecognized escape sequence, so reliance on this behavior is not likely to be common. Still, there are likely to be some uses in the wild (probably some percentage of that were intended to be \n).
Another option would be to reuse the \u and/or \U introducer used for universal-character-names. Gcc and Clang both reject code like "\u{xxx}" and "\U{xxx}" as containing ill-formed universal-character-names. However, Microsoft Visual C++ accepts such uses without a warning and treats them as equivalent to "u{xxx} and "U{xxx}" respectively.
The implementation divergence that occurs for the \u and \U cases above suggests that repurposing them may result in less backward compatibility. Use of \u and/or \U would potentially require more wording changes to distinguish named character escapes from universal-character-names, but would be unlikely to pose a significant additional impact to implementors.
For now, this proposal adheres to Fernandes' original design and retains use of \N as the introducer for named character escapes.
The sources of character names listed in the Name sources section do not constitute big data by today's standards, but that does not mean that the volume of data and potential for impact to compiler distributions and compiler performance is insignificant. As mentioned earlier, some organizations have valid technical reasons to be sensitive to the size of the compiler distributions they use; in a distributed build environment that distributes compilers, the size of the distribution impacts latency and can therefore negatively impact build times.
The combined size of the Unicode 12.0 text files containing the Unicode assigned names, aliases, and named character sequences is approximately 1.5 MiB. A naive implementation might contribute 2+ MiB of code/data to a compiler. Some EWG members indicated that amount of increase is a cause for concern.
Fortunately, naive implementations are not the only option. Corentin Jabot has done some excellent work to demonstrate that an implementation should be possible that increases the code/data size of a compiler by less than 300 KiB. See the Implementation experience section for details. Corentin's approach is promising, but the additional complexity caries additional implementation cost and maintenance.
Staying up to date with new Unicode releases will also, of course, pose an additional cost on implementors.
As indicated previously, at least one EWG member in Belfast was strongly interested in a more general core language feature, presumably a string interpolation facility, that would allow named character escapes to be implemented as a library feature. Such a feature could take many forms, but might look something like the following where \{ is an escape sequence followed by a call to a constexpr function named nce with arguments passed in some form.
"\{nce(LATIN CAPITAL LETTER A WITH GRAVE)}"Such a feature could certainly be implemented, but would seem to necessarily be more verbose and would necessitate inclusion of appropriate headers; headers that would be quite large in the case of a named character database or that would make use of a compiler intrinsic; which would put the complexity back in the compiler (though in implementation-defined territory rather than in standard core language). The verbosity concern could potentially be reduced by introducing core language sugar for lowering the proposed syntax to the example string interpolation syntax above.
The wording included in this proposal is for the following design:
The following options are not currently proposed, but could be adopted as modifications of the current proposal.
The following options are not currently proposed but could be considered for future extension.
This proposal has not yet been implemented in an existing compiler. However, the implementation concerns raised in Belfast prompted Corentin Jabot to conduct an experiement to determine how small the implementation overhead, in terms of data and code within the compiler, could be reduced to. His blog post[CJ_BLOG] on the experiment reported that he was able to implement a function (cp_from_name) that accepts a Unicode 12.0 name or name alias and returns a code point value in under 300 KiB. His implementation is available in the cp_to_name branch of his ext-unicode-db GitHub repository at https://github.com/cor3ntin/ext-unicode-db/tree/name_to_cp[CJ_IMPL].
Thank you to R. Martinho Fernandes for taking the initiative to research and first propose support for named character escapes and for contributing his considerable expertise in general to SG16.
Thank you to Corentin Jabot for the excellent work he did experimenting with and analyzing implementation impact. Without his work, the data necessary to respond to the implementation concerns raised in Belfast would not have been available at this time, thereby delaying further progress on this proposal.
Thank you to Peter Bindels and Corentin Jabot for providing feedback on an initial draft that I delivered to then less than two hours before the Prague pre-meeting mailing deadline!
| [CJ_BLOG] |
Corentin Jabot,
"Storing Unicode: Character Name to Codepoint Mapping", 2019. https://cor3ntin.github.io/posts/cp_to_name |
| [CJ_IMPL] |
Corentin Jabot,
"ext-unicode-db", 2019. https://github.com/cor3ntin/ext-unicode-db/tree/name_to_cp |
| [ISO/IEC10646] |
"Information technology — Universal Coded Character Set (UCS)", ISO/IEC 10646:2017, 2017. https://www.iso.org/standard/69119.html |
| [N4835] |
"Working Draft, Standard for Programming Language C++", N4835, 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/n4835.pdf |
| [P1025R1] |
Steve Downey, et al.
"Update The Reference To The Unicode Standard", P1025R1, 2018. https://wg21.link/p1025r1 |
| [P1097R1] |
R. Martinho Fernandes,
"Named character escapes", P1097R1, 2018. https://wg21.link/p1097r1 |
| [P1097R2] |
R. Martinho Fernandes,
"Named character escapes", P1097R2, 2019. https://wg21.link/p1097r2 |
| [P2029R0] |
Tom Honermann,
"Proposed resolution for core issues 411, 1656, and 2333;
numeric and universal character escapes in character and string literals",
P2029R0, 2020. https://wg21.link/p2029r0 |
| [UCESP] |
"Unicode Character Encoding Stability Policies", 2017. https://www.unicode.org/policies/stability_policy.html |
| [UAX#44] |
Ken Whistler and Laurențiu Iancu,
"Unicode Standard Annex #44 - Unicode Character Database", Revision 24, Unicode 12.0.0, 2019. https://www.unicode.org/reports/tr44/tr44-24.html |
These changes are relative to N4835[N4835].
If P2029R0[P2029R0] were to be adopted, substantial wording updates will be required.
Hide inserted textChange in 5.2 [lex.phases] paragraph 5:
Each basic source character set member in a character literal or a string literal, as well as each escape sequenceand, universal-character-name, and named-escape-sequence in a character literal or a non-raw string literal, is converted to the corresponding member of the execution character set ([lex.ccon], [lex.string]); if there is no corresponding member, it is converted to an implementation defined member other than the null (wide) character. 8
Change in 5.13.3 [lex.ccon]:
character-literal:encoding-prefixopt ' c-char-sequence '
encoding-prefix: one ofu8 u U L
c-char-sequence:
c-char:any member of the basic source character set except the single-quote ', backslash \, or new-line character
escape-sequence:
simple-escape-sequence: one of\' \" \? \\
\a \b \f \n \r \t \v
octal-escape-sequence:
hexadecimal-escape-sequence:
named-escape-sequence:\N{ n-char-sequence }
n-char-sequence:
n-char: one ofA B C D E F G H I J K L M N O P Q R S T U V W X Y Za b c d e f g h i j k l m n o p q r s t u v w x y z0 1 2 3 4 5 6 7 8 9_ - space
Change in 5.13.3 [lex.ccon] paragraph 7:
Certain non-graphic characters, the single quote ', the double quote ", the question mark ?,19 and the backslash \, can be represented according to Table 8. The double quote " and the question mark ?, can be represented as themselves or by the escape sequences \" and \? respectively, but the single quote ' and the backslash \ shall be represented by the escape sequences \' and \\ respectively. Escape sequences in which the character following the backslash is not listed in Table 8 are conditionally-supported, with implementation-defined semantics. An escape sequence specifies a single character.
Table 8: Escape sequences [tab:lex.ccon.esc]
new-line NL(LF) \n horizontal tab HT \t vertical tab VT \v backspace BS \b carriage return CR \r form feed FF \f alert BEL \a backslash \ \\ question mark ? \? single quote ' \' double quote " \" octal number ooo \ooo hex number hhh \xhhh named escape sequence named character \N{xxx}
Add a new paragraph (X) after
5.13.3 [lex.ccon] paragraph 9:
Drafting Note: Associated character names and character
name aliases are listed in section 33 of ISO/IEC 10646:2017. Named UCS
sequence identifiers are listed in section 27.
A named-escape-sequence is translated to the encoding, in the appropriate execution character set, of the character or character sequence associated with the ISO/IEC 10646 associated character name or character name alias that matches the name specified by the n-char-sequence. Matching of names is performed by:If no name is matched, then the program is ill-formed. If the matched name is HANGUL JUNGSEONG OE, then steps 2 and 3 are performed against the name HANGUL JUNGSEONG O-E and, if the names match, U+1180 {HANGUL JUNGSEONG O-E} is encoded, otherwise U+116C {HANGUL JUNGSEONG OE} is encoded. Otherwise, the character associated with the matched name is encoded. [ Note: The special handling of U+1180 {HANGUL JUNGSEONG O-E} resolves an ambiguity in the matching algorithm; this is the only case of ambiguity. — end note ]
(X.1) — removing all medial hyphens. (X.2) — removing all space and underscore characters. (X.3) — lowercasing all capital letters.
Change in 5.13.5 [lex.string] paragraph 14:
Escape sequences and universal-character-names in non-raw string literals have the same meaning as in character literals ([lex.ccon]), except that the single quote ' is representable either by itself or by the escape sequence \', and the double quote " shall be preceded by a \, and except that a universal-character-name or named-escape-sequence in a UTF-16 string literal may yield a surrogate pair. In a narrow string literal, a universal-character-name or named-escape-sequence may map to more than one char or char8_t element due to multibyte encoding. The size of a char32_t or wide string literal is the total number of escape sequences, universal-character-names, named-escape-sequences, and other characters, plus one for the terminating U'\0' or L'\0'. The size of a UTF-16 string literal is the total number of escape sequences, universal-character-names, named-escape-sequences, and other characters, plus one for each character requiring a surrogate pair, plus one for the terminating u'\0'. [ Note: The size of a char16_t string literal is the number of code units, not the number of characters. — end note ] Within char32_t and char16_t string literals, any universal-character-names shall be within the range 0x0 to 0x10FFFF. The size of a narrow string literal is the total number of escape sequences and other characters, plus at least one for the multibyte encoding of each universal-character-name, named-escape-sequences, plus one for the terminating '\0'.
Change in table 17 of
15.11 [cpp.predefined] paragraph 1.8:
Drafting note: the final value for the
__cpp_named_character_escapes feature test macro will be selected by
the project editor to reflect the date of approval.
Table 17 — Feature-test macros [tab:cpp.predefined.ft]
Macro name Value […] […] __cpp_modules 201907L __cpp_named_character_escapes XXXXXXL ** placeholder ** __cpp_namespace_attributes 201411L […] […]