C23 implications for C libraries

Introduction

The upcomming standard C23 has a lot of additions to the C library clause. Most of them are small, some of them are big but optional.

This document only marginally covers the latter, namely changes to floating point types and their implications for the C library, simply because that is not my domain of expertise.

The last publicly available working draft is https://open-std.org/JTC1/SC22/WG14/www/docs/n3096.pdf and section numbers below refer to that document. But this will not be the final version of C23. Against that document 239 national body comments had been raised to which WG14 answered in their June 2023 session https://open-std.org/JTC1/SC22/WG14/www/docs/n3148.doc. Most of the changes have no normative impact on the C library, those that we are aware of are marked below.

Unfortunately, the versions thereafter and in particular the final document fall under the weird ISO laws and can’t be made available.

The document will now be updated by the editors to the latest changes and an editorial review will follow. After that the document will be submitted for final approval by the national bodies. If there are no serious changes the final document will be submitted to ISO for publication January or February of 2024. Nevertheless, WG14 decided to stick to C23 as a name for this version, as this reflects the new version number 202311L in __STDC_VERSION__ (and others) which they will not be able to change in the procedure.

Impacted headers

There are a lot of impacted headers by the changes,

For C23, all these headers have to provide feature test macros of the form __STDC_VERSION_NAME_H__ where NAME is the capitalized form of the header name without .h suffix. The value of these macros has to be 202311L and should only be set to this once the transition is complete and the header is conforming.

Nevertheless, even before, such a macro can already be used as an include guard, it just has to be set to be empty or to a smaller numerical value, e.g. 0.

Unicode support (mandatory change)

C23 improves Unicode support. char8_t, char16_t, char32_t and strings and characters prefixed with u8, u and U are encoded with the proper UTF encoding. See <uchar.h>, below.

Unfortunately, there are still some escape hatches for hereditary implementations that are controlled by the feature test macros __STDC_ISO_10646__ and __STDC_MB_MIGHT_NEQ_WC__.

Unicode Identifier support has been updated to UAX #31 (Unicode Identifier and Pattern Syntax). This means that validity of identifiers rule as if the encoding where Unicode.

1. The character that starts an identifier must correspond to a Unicode codepoint that has the _{XID_Start} property (including the characters of the Latin alphabet) or be the character _. The possible use of the character $ as identifier start character is implementation-defined.

2. The set of continuation characters extents this to all characters that correspond to the _{XID_Continue} property (including decimal digits).

This means in particular that identifiers may not contain characters that have no Unicode equivalent or where the codepoint does not have the required properties. So any identifier ID can be stringified to a multi-byte encoded string mbs and then be transformed by mbrtoc32 to a UTF-32 encoded string s32 where the characters have the above properties. Transforming with mbrtoc16 or mbrtoc8 results in a valid UTF-16 or UTF-8 encoded string s16 and s8, respectively.

#include <stdbit.h>
#include <stdlib.h>
#include <uchar.h>

#define STRINGIFY_(X) #X
#define STRINGIFY(X) STRINGIFY_(X)

// For any accepted identifier or pp-number PP this is a valid mb-string
char const mbs[] = STRINGIFY(PP);

// For any accepted identifier or pp-number PP this will be a valid UTF-32 string
// where s32[0] has the XID_Start property or is an underscore character (for identifiers) or
// a decimal digit or a period character (for pp-numbers), and all
// subsequent characters have the XID_Continue property.
char32_t  s32[sizeof mbs];
char32_t* p32 = s32;
size_t len = sizeof mbs;
mbstate_t state32 = { };
for (char const* p = mbs; *p;) {
    // No error return possible, all mb characters must be
    // complete and have UTF-32 codepoints, os is never 0
    register size_t const os = mbrtoc32(p32, p, len, &state32);
    if (os-1 > len-1) unreachable();
    if (p32[0] > 0x10'FF'FF) unreachable();
    p32 += 1;
    p   += os;
    len -= os;
}
*p32 = 0;

// For any accepted identifier or pp-number PP this will be a valid UTF-16 string
char16_t  s16[2 * sizeof mbs];
char16_t* p16 = s16;
size_t len16 = sizeof mbs;
mbstate_t state16 = { };
for (char const* p = mbs; *p;) {
    // No error return possible, all mb characters must be
    // complete and have UTF-32 codepoints, os is never 0
    register size_t const os = mbrtoc16(p16, p, len16, &state16);
    if (os-1 > len-1) unreachable();
    // A surrogate character has been stored
    if ((0xD8'00 <= p16[0]) && (p16[0] < 0xE0'00)) {
        p16 += 1;
        // No error return possible, this provides the second surrogate.
        if (mbrtoc16(p16, p, len16, &state16) != -3) unreachable();
        if ((p16[0] < 0xD8'00) || (0xE0'00 >= p16[0])) unreachable();
    }
    p16   += 1;
    p     += os;
    len16 -= os;
}
*p16 = 0;

/* currently untested, use with care! */
// For any accepted identifier or pp-number PP this will be a valid UTF-8 string
char8_t  s8[4 * sizeof mbs];
char8_t* p8 = s8;
size_t len8 = sizeof mbs;
mbstate_t state8 = { };
for (char const* p = mbs; *p;) {
    // No error return possible, all mb characters must be
    // complete and have UTF-32 codepoints, os is never 0
    register size_t const os = mbrtoc8(p8, p, len8, &state8);
    if (os-1 > len-1) unreachable();
    // Determine the number of continuation bytes, if any.
    // This is found by looking at the highest order bit that is 0.
    // Mask with 0xFF is needed because p8[0] is promoted to int
    // before doing the complement.
    size_t const plus = 8-stdc_first_leading_one((~p8[0])&0xFFu);
    for (size_t i = 0; i < plus; i++) {
        p8 += 1;
        // No error return possible, this provides the next continuation character.
        if (mbrtoc8(p8, p, len8, &state8) != -3) unreachable();
    }
    p8   += 1;
    p    += os;
    len8 -= os;
}
*p8 = 0;

Similarly, preprocessing numbers have to adhere to admit all the above continuation characters as possible continuation after initial digits or a decimal point.

These changes may marginally impact C library implementations where they use preprocessor concatenation or stringification if their multi-byte or wide character encodings are not UTF-8 or UTF-32, respectively.

Thread safety of the C library (mostly optional changes)

In several places C23 now explicitly grants permission to C library implementations to use thread local state for stateful functions. This implies permission to change behavior of setlocale, multibyte character conversion functions, strtok, strerror, and time conversion functions. Some of the changes can be done silently. For others, the choice needs to be documented as they are now implementation-defined; so even if an implementation sticks to static state (as it was previously) this has to be documented, now. Previously, the C standard was ambiguous about this; but for the parts that make this an implementation-defined property (and thus force implementations to document) the possible changes are not normative but for quality of the implementation, only.

This is now implementation-defined for all functions that have character conversion state held in an internal object of type mbstate_t. Here now a static or a thread local object may be used, if necessary. A change to thread local could be of service for threaded applications, although this has to be taken with a grain of salt. Querying for thread local state can be about as expensive as the whole conversion operation. Therefore applications that are performance critical should never use the variants of the interfaces that use an internal buffer, anyhow. As this has been a serious safety risk prior to C23, hopefully not many applications use this feature, anyhow.

The following functions each have their own internal state object that has either static or thread local storage duration:

c16rtomb c8rtomb mbrtoc16 mbrtoc32 mbrtoc8 mbrtowc

The functions mbsrtowcs and mbrlen are impacted implicitly because they use the state of mbrtowc.

The functions need to use this state because they are “restartable”:

• They can be called with partial encodings for the input and a whole wide character can be collected over several calls. Here the state variable collects the input state.
• They may produce several output characters. For UTF-8 (up to 4 output characters) and UTF-16 (up to two output characters) the state is also used for the output. The functions that write multi-byte encodings expect sufficient storage to store a whole sequence for any such encoding at once; so these functions don’t need the state variable to manage their output.

Thus, if they are called with a null pointer for the state, all of them have to use a per-function internal state.

The following functions only need to use state if the C library implementation supports encodings that have so-called shift states, that is, where different types of encodings can be switched during parsing. Otherwise they should just ignore a possible state parameter (if any) and should not use any internal state.

c32rtomb mbtowc wcrtomb wcsrtombs wctomb

A call mbtowc(nullptr, nullptr, 0) can be used to determine if the current locale setting is such that multi-byte encodings have or don’t have shift states.

Note that C library implementations that only use UTF-8 as multi-byte encoding will never need shift state.

Rationale and wording for this change can be found in:

https://open-std.org/JTC1/SC22/WG14/www/docs/n2444.htm

Const-contract of the C library (mandatory change)

In C17 there are still interfaces that violate a const-contract: in some cases, pointer values to const-qualified objects that are passed as parameters to a function call can be returned as pointers to unqualified objects. C23 closes a lot of these loopholes per its default interfaces. This concerns the identifiers

bsearch_s bsearch memchr strchr strpbrk strrchr strstr wcschr wmemchr wcspbrk wcsrchr wcsstr

which now interface type-generic macros, see below. The function interfaces that may violate the const-contract still remain, they are kept for backwards compatibility, but unless they force the use of functions, user code sees the macros.

Code compiled with the C23 library interfaces is possibly rejected or diagnosed by compilers if they do not respect the const-contract. This is intentional.

Changes to integer types

One major change in the language is that two’s complement is now mandatory as a representation for signed types.

New feature test macros (mandatory change)

This makes it easier to characterize integer types. For all of them there are now ..._WIDTH macros, that, together with the sizeof operator, completely describe the types. Annex E gives a good overview of the macros that are required and inform about minimal values for these.

Note that in the future the ..._MAX values are not necessarily always suitable for comparison in #if conditionals. Usage of these should be changed to comparing the ..._WIDTH values, which are much smaller and usually more comprehensive numbers.

New valid token sequences for integer constants

All macros that deal with integer constants have to be capable to deal with new formats for these literals.

• Binary constants with a 0b or 0B prefix.
• Digit separators. The ' becomes a digit separator that can be placed anywhere between consecutive digits such as 1'000 or 0.333'333.

This concerns most importantly ..._C macros in <stdint.h>.

Exact-width integer types (mandatory change)

This holds for all width for which the platform has integer types without padding.

Extended integer types may be wider than intmax_t (optional change)

This concerns types that could be used as exact-width types, in particular [u]int128_t and [u]int256_t where many implementations already have extensions, but which cannot be “extended integer types” in the sense of C17.

On platforms with for example gcc (64 bit) that already have an extension __int128 (with predefined macro __SIZEOF_INT128__) the necessary addition to <stdint.h> could look as follows:

// No language version macro needed
#ifdef __SIZEOF_INT128__
typedef signed   __int128 int128_t;
typedef unsigned __int128 uint128_t;
typedef signed   __int128 int_fast128_t;
typedef unsigned __int128 uint_fast128_t;
typedef signed   __int128 int_least128_t;
typedef unsigned __int128 uint_least128_t;
# define UINT128_MAX         ((uint128_t)-1)
# define INT128_MAX          ((int128_t)+(UINT128_MAX/2))
# define INT128_MIN          (-INT128_MAX-1)
# define UINT_LEAST128_MAX   UINT128_MAX
# define INT_LEAST128_MAX    INT128_MAX
# define INT_LEAST128_MIN    INT128_MIN
# define UINT_FAST128_MAX    UINT128_MAX
# define INT_FAST128_MAX     INT128_MAX
# define INT_FAST128_MIN     INT128_MIN
# define INT128_WIDTH        128
# define UINT128_WIDTH       128
# define INT_LEAST128_WIDTH  128
# define UINT_LEAST128_WIDTH 128
# define INT_FAST128_WIDTH   128
# define UINT_FAST128_WIDTH  128
# if UINT128_WIDTH > ULLONG_WIDTH
#  define INT128_C(N)         ((int_least128_t)+N ## WB)
#  define UINT128_C(N)        ((uint_least128_t)+N ## WBU)
# else
#  define INT128_C(N)         ((int_least128_t)+N ## LL)
#  define UINT128_C(N)        ((uint_least128_t)+N ## LLU)
# endif
#endif

This might use the new mandatory bitprecise integer constants with suffix WB and WBU respectively by supposing that these will be implemented on these platforms to support at least 128 bit.

Note that the names for these types have been reserved since several C versions, so they are immediately available to the implementation and need not to be #ifdefed for the C version of the compilation. Nevertheless, integer literals and format specifiers for these types might then not be available.

#if UINT128_WIDTH > UINTMAX_WIDTH
# if __STDC_VERSION__ < 202311L
#  warning "extended integer type for 128 type is wider than intmax_t"
# endif
#endif

After implementing the mandatory changes to printf and scanf with length specifiers %w and %wf the corresponding macros should also be added to <inttypes.h>, see %w and %wf below.

Bit-precise integer types (mandatory change, compiler dependent)

There is a whole new series of integer types called bit-precise integer types. They are coded by the token sequence _BitInt(N) where N is an integer constant expression, and which can be combined with signed (default) and unsigned. The minimum value for N is 1 for unsigned types and 2 for signed. The maximum value is provided by a new macro BITINT_MAX which has to be added to <limits.h>.

Clang already implements these types in their recent compilers and exports a predefined macro __BITINT_MAX__. Hopefully gcc will choose the same. The name for the new macro had not previously been reserved, so its definition should be protected by a version macro.

#if __STDC_VERSION__ > 202300L
# ifdef __BITINT_MAX__
#  define BITINT_MAX __BITINT_MAX__
# endif
#endif

The constants for these types have the suffixes WB, wb, WBU, wbU, , WBu, or wbu, such that a suffixed constant has the type _BitInt(N) where N is minimal for the constant to fit. This can be useful (see above) to express integer constants in headers of a perhaps wider range than otherwise would be supported.

Otherwise these types have no direct library interfaces by themselves.

Nevertheless, these type may occur as arguments to generic interfaces

• all N for atomic_ functions in <stdatomic.h>, see below;
• certain N for the new <stdbit.h> interfaces, see below.

String to integer conversion

See strtol

ptrdiff_t

May now be 16 bit wide, instead of previously 17 bit. Only implementations that previously were not conforming because of this (some compilers for 16 bit hardware) should be concerned about this. If that was their only miss, they become conforming to C23 where they weren’t for C17.

char8_t added (mandatory change)

See <uchar.h>, below.

New header <stdbit.h> (mandatory change, independent)

This adds 14 × 5 functions and 14 tg macros for bit-manipulation; so in total 84 new interfaces.

On many architectures the functions themselves probably have builtins that just have to be interfaced.

The type-generic interfaces are a bit more tricky, since that have to work for

• standard unsigned integer types (bool excluded)
• extended unsigned integer types (don’t forget additions such as uint128_t)
• bit-precise types that have the same width as a standard or extended unsigned integer type (bool excluded).

On architectures where these integer types don’t have padding bits all of these should probably just switched by the size of the argument and the function arguments should just be cast to these.

ABI choice: There is one ABI choice to be made for this, the type described as generic_return_type for the type-generic functions. The question that has to be answered here is if an ABI community estimates that one day they will have extended integer types that have more than UINT_MAX bits. I personally don’t think that this is likely.

If we assume that unsigned int as a return type is just good enough, something along the lines of the following would probably work:

inline unsigned int stdc_leading_zeros_u8(uint8_t __x) [[__unsequenced__]] { /* do your thing */ }
...
inline int stdc_leading_zeros_u128(uint128_t __x) [[__unsequenced__]] { /* do your thing */ }
#define stdc_leading_zero(X)                              \
    _Generic((char(*)[sizeof(X)]){ 0 },                   \
        char(*)[8]: stdc_leading_zero_u8((uint8_t)X),       \
        ...                                               \
        char(*)[128]: stdc_leading_zeros_u128((uint128_t)X))

Here, the cast in the chosen branch would never lose bits; the cast in the other branches is well-defined and should not issue diagnostics. Also, the inline functions can be static since these are implementation details that do not have to be exported as linker interfaces on which user code may rely upon.

As an extension, this would also work for

• bool because the conversion will just have a 0 or 1 of the appropriate size;
• signed integer types because conversion to the unsigned type is always well-defined.

New header <stdckdint.h> (mandatory change, independent)

The specification of the interfaces in C23 is type-generic:

#include <stdckdint.h>
bool ckd_add(type1 *result, type2 a, type3 b);
bool ckd_sub(type1 *result, type2 a, type3 b);
bool ckd_mul(type1 *result, type2 a, type3 b);

This adds overflow and wrap-around safe interfaces for integer addition, subtraction and multiplication. The result target is filled with the truncated result of the mathematical operation, the return value holds false if that result is correct, true otherwise.

Admissible types for these operations are all standard integer types with the exception of bool and char.

The intent of these interfaces is that they use the underlying hardware as efficient as possible. Most modern processors have overflow flags in hardware and these functions should in general just query these flags.

On most architectures there are probably builtins that just have to be interfaced, but since they are type-generic there might some more to do than just functions. On the other hand, compilers such as gcc and clang already have similar type-generic interfaces:

bool __builtin_add_overflow (type1 a, type2 b, type3 *res);
bool __builtin_sub_overflow (type1 a, type2 b, type3 *res);
bool __builtin_mul_overflow (type1 a, type2 b, type3 *res);

When relying on such an extension the contents of the header could just read

#ifndef __STDC_VERSION_STDCKDINT_H__
#define __STDC_VERSION_STDCKDINT_H__ 202311L

#ifdef __GNUC__
# define ckd_add(R, A, B) __builtin_add_overflow ((A), (B), (R))
# define ckd_sub(R, A, B) __builtin_sub_overflow ((A), (B), (R))
# define ckd_mul(R, A, B) __builtin_mul_overflow ((A), (B), (R))
#else
# error "we need a compiler extension for this"
#endif

#endif

Because of the type-generic nature, this is a header-only addition to the standard. No support for linker symbols for any variant of these functions is required. Effectively, this is an extension to the language and not a library feature.

The addition that has been integrated into C23 is the core proposal and some amendments. The history of this looks a bit confusing because later editions of the paper have removed parts that had already been adopted.

Alignment to IEC 60559 for standard floating point

Unfortunately not my field of expertise. There might be subtle changes to the FP model.

The following functions are added to _<math.h>, with the usual three versions for double (without suffix), float (with f suffix) and (with l suffix), and as type-generic macros to _<tgmath.h>:

acospi, asinpi, atan2pi, atanpi, compoundn, cospi, fmaximum_mag_num, fmaximum_mag, fmaximum_num, fmaximum, fminimum_mag_num, fminimum_mag, fminimum_num, fminimum, nextup, pown, powr, rootn, roundeven, sinpi, tanpi, fadd, dadd, fsub, dsub, fmul, dmul, fdiv, ddiv,

Similarly, the following functions and their variants are added to _<math.h>, but there are no type-generic macros in _<tgmath.h>:

canonicalize, fromfp, fromfpx, ufromfp, ufromfpx,

Optional IEC 60559 support for decimal floating point

Unfortunately not my field of expertise, but see some discussion for <stdarg.h> and <stdio.h>, below.

The main chunk of work here goes into to the huge amount of function that this requires for <math.h>, <stdlib.h> and maybe some other headers. This effort should probably not be repeated for every C library out there, but should be concentrated into sublementary libraries that can be added on demand, independently of the base C library.

The things that a genuin C library probably should support to make this possible is basic language support for these types

• Implementation of the feature macros in <float.h> such as described in section 5.2.4.2.3 p2 to p6 of C23.
• Implementation of the classification macros in <math.h>.
• Implementation of the functions frexpd32, frexpd64, frexpd128 functions in <math.h>.
• Implementation of the formatted IO in <stdio.h>.

This can be based on feature macros and builtins that e.g gcc already has since long, namely for N one of 32, 64 or 128:

• __DECN_EPSILON__ __DECN_MIN__ __DECN_MAX__ __DECN_MIN_EXP__ __DECN_SUBNORMAL_MIN__ __DECN_MAX_EXP__ __DECN_MANT_DIG__

• __builtin_signbitdN, __builtin_infdN, __builtin_nandN

Attributes

The introduction of attributes in C23 only implies a few changes in the library.

_Noreturn becomes [[noreturn]] (mandatory change)

Adapting to that will be a bit tedious because, _Noreturn lives on as a keyword for some time. The easiest would probably be to change occurrences of _Noreturn in the C library headers <setjmp.h>, <threads.h> and <stdlib.h> by something like __noreturn and then add the following in a general preamble or to the preambles of these headers:

#ifdef __has_c_attribute
# if __has_c_attribute(__noreturn__)
#  ifndef __noreturn
#   define __noreturn [[__noreturn__]]
#  endif
# endif
#endif
#ifndef __noreturn
# define __noreturn _Noreturn
#endif

The macros __STDC_VERSION_SETJMP_H__, __STDC_VERSION_STDLIB_H__ and __STDC_VERSION_THREADS_H__ should only be set to 202311L if this change is implemented.

Changes to implementation-defined occurences of _Noreturn are not necessary, but it would be a good occasion to do so.

The header <stdnoreturn.h> becomes obsolescent, but does not change.

Deprecating certain functions with [[deprecated]] (mandatory change)

Two functions in <time.h> receive such an attribute, see below.

Annotating C library functions with [[unsequenced]]

This new function type attribute models a generalization of what usually is called a pure function in CS. Because of the rules how function attributes accumulate in conditional operators, it is not possible to add this kind of attribute to C library functions; otherwise in rare cases this could change semantics of programs on architectures that implement the attribute.

The only C library functions with this attribute are the ones in <stdbit.h> because these are new and do not present problems with backwards compatibility.

In contrast to that user code that checks that the planets are aligned could augment C library interfaces with this attribute to help optimizers. This can even be done locally restricted to the body of a function, for example, where it is known that all calls to the function in question have the right properties:

typeof(sqrt) [[unsequenced]] sqrt; // We guarantee that all arguments are positive

There are a lot of functions that are unsequenced per definition of the standard and may thus be annotated in user code or in the internal definitions of functions of a C library implementation. All interfaces in 7.12.3, 7.18, 7.20 and the following families of functions do not access any global state. In particular they may never encounter exceptional conditions that would let them have read the rounding mode or write to the floating point flags or to errno.

abs atomic_init canonialize ceil copysign decodecdN difftime div encodecdN fabs floor fmax fmaximum[_mag][_num] fmin fminimum[_mag][_num] labs ldiv llabs lldiv memalignment memccpy memchr memcmp memcpy memset modf nan quantumdN round roundeven samequantumdN strchr strcmp strcpy strcspn strlen strncmp strncpy strpbrk strrchr strspn strstr trunc

This also holds for most functions in <math.h> and <complex.h> if it can be guaranteed that the function is never called with arguments that are invalid or result in out-of-range values, and if the rounding mode is previously fixed to a specific value for all uses.

Predefined macro names (mandatory change)

__STDC_UTF_16__ and __STDC_UTF_32__ are mandatory

They are now forced to the value 1

Remove trigraphs

Trigraphs are removed from the language. Any library headers that use them must be updated.

Generally, they are probably never used within strings or character literals in C library headers.

A more common use may be the trigraph ??= in preprocessing as a replacement for #. This can be easily be replaced by the digraphs %: and %:%: which have the following advantages

• They are proper tokens and will not be substituted in early compilation phases. So there is no messing with the contents of user strings anymore.
• The characters % and : are present in all source file encodings that are currently still in use.

Changes to <assert.h>

__STDC_VERSION_ASSERT_H__ (mandatory change)

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

assert (mandatory change)

Before this change, the assert macro commonly suffered from the one-macro-parameter problem, namely that expressions that contained a highlevel comma such as (mystruct){ a = 0, b = x}.b would not be seen as one argument but several.

There is a simple cure to this namely to define assert with .... If we suppose that there is a builtin function

[[noreturn]] void __builtin_assert(
    const char * __expr,            // textual version of the expression
    const char* __file,             // the name of the source
    unsigned long __line,           // the current line number, may be larger than INT_MAX
    const char* __func);            // the current function

Then the macro can be defined using a static function __assert_expression as

// Ensure that this version only accepts arguments that resolve to one
// C expression after preprocessing.
#ifndef __ASSERT_EXPRESSION
#define __ASSERT_EXPRESSION(...) __assert_expression(__VA_ARGS__)
static inline _Bool __assert_expression(_Bool __x) { return __x; }
#endif

#ifdef NDEBUG
#define assert(...) ((void)0)
#else
#if __STDC_VERSION >= 199901L
# define assert(...) ((__VA_ARGS__) ? (void)0 : __builtin_assert(#__VA_ARGS__, __FILE__, __LINE__, __func__))
#else
# define assert(...) ((void)(__ASSERT_EXPRESSION(__VA_ARGS__) || (__builtin_assert(#__VA_ARGS__, __FILE__, __LINE__, __func__),0)))
#endif

This ensures that an expression that is given as an argument and that happens to be several macro arguments, is still accepted and works as expected. On the other hand, an empty parameter or one that expands to more than one expression is diagnosed.

This change is conforming to all C versions after C99 where variable argument macros were introduced.

Changes to <float.h>

__STDC_VERSION_FLOAT_H__

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

Add INFINITY and NAN to <float.h>.

Missing macros

Not my field of expertise.

Changes to <inttypes.h>

__STDC_VERSION_INTTYPES_H__ (mandatory change)

This macro is mandatory and should be set to 202311L once the header is compliant to C23. (This macro is still missing from the latest draft, but should be added during CD2 ballot.)

PRI macros for narrow types

In a last minute change (GB-108 in CD2 ballot) it is now imposed that the macros for printf and similar functions convert values for narrow types (such as uint16_t) back to the narrow type before printing. This has an effect on implementations and on user code that used format specifiers for narrow types to print arbitrary signed or unsigned values.

PRI and SCN macros for binary numbers

Macros for the the new binary printf and scanf formats become mandatory, analogously as for x there is now b.

The printf specifier %B is optional. The macros with B (such as PRIB128) act as feature test macros.

Changes to <limits.h>

__STDC_VERSION_LIMITS_H__ (mandatory change)

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

Missing macros (mandatory change)

• BITINT_MAX for the maximal supported width of a bit precise integer type
• _WIDTH macros for all standard integer types

Changes to <stdarg.h>

This header provides a liaison between C language and C library. For implementations that have no good compiler support for va_... macros, support for C23 appears to be relatively challenging.

For implementations that just forward these macros to compiler builtins, there is no particular difficulty to be expected.

__STDC_VERSION_STDARG_H__ (mandatory change)

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

Additional arguments to va_start can be omitted (mandatory change)

void va_start(va_list ap, ...);

• Only the first argument shall be used and evaluated.
• Functions specified with ... parameter may have only that and va_start must be able to deal with that situation.

The formulation as given above is only valid if the preprocessor is conforming to C23, namely that it works well without a trailing comma. A good way to take care of that is by using the a feature test macro for the __VA_OPT__ feature:

#define __has_VA_OPT(X, ...) X ## __VA_OPT__(_OK)
#define __C23__VA_OPT__(...) 0
#define __C23_OK 1

#if __has_VA_OPT(__C23, __C23)
#define va_start(v, ...) __builtin_va_start(v __VA_OPT__(,) __VA_ARGS__)
#else
#define va_start(v,l)   __builtin_va_start(v,l)
#endif

Implementations that set __STDC_VERSION_STDARG_H__ have to verify that they conform to that reformulation.

Support for new types, standard or extended (mandatory change)

Users will expect va_arg to work for all types that a C23 compiler provides.

Compiler dependency: Implementations that set __STDC_VERSION_STDARG_H__ have to verify that they at least support the new standard types. Users would be quite surprised if a compiler implements optional types or certain common extensions and va_arg fails with them. This would make it in particular impossible for third party libraries to add support for these types.

va_arg can handle function arguments with type nullptr_t (mandatory change)

Functions that are called with a nullptr argument in the variable argument list can deal with such an argument by calls

va_arg(ap, nullptr_t)
va_arg(ap, char*)

or any other pointer type that has the same representation as char*, such as void*. In particular, on implementations that have exactly one pointer model (such as POSIX) any pointer type can be used as type name argument.

The result of the macro call is then a null pointer of the indicated type.

Compiler dependency: To support the variant that uses nullptr_t compiler support is needed.

va_arg may be called with _BitInt(N) types (mandatory change)

This is a bit tricky for narrow bit-precise types, because they do not promote to int as would other narrow types.

va_arg and decimal floating types (optional change)

Full support for decimal floating point is indicated by the feature test macro __STDC_IEC_60559_DFP__. Setting this macro implies not only compiler support (arithmetic, fuction calls) but also the implementation of about 600 C library interfaces (5.5 pages in the library summary Annex B). So implementations might be hesitant to support this. Hopefully the open source implementations will join forces to supply an add-on external library that completes support for these types.

Compiler dependency: Decimal floating point types are optional, but they are already implemented on a major compiler (gcc) and expectations will be high that any C library supports these types on such platforms in <stdarg.h>, even if they do not intend to set __STDC_IEC_60559_DFP__ themselves.

va_arg and extended integer types (optional change)

With relaxation for wide integer types that exceed the width of intmax_t, compiler platforms may start to provide support for such types. This may for example be the case of __int128 types on x86_64, powerpc64 or aarch64, and which are already supported by compilers.

For any width N such there is an extended integer type that has no padding there has to be support for types intN_t and uintN_t. So in particular ..._C macros and wN length modifiers for printf are mandatory. So C library implementations also have to support these extended integer types for va_arg.

Changes to <stdatomic.h>

__STDC_VERSION_STDATOMIC_H__ (mandatory change)

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

Remove ATOMIC_VAR_INIT ? (optional change)

This macro is removed from the standard because it is not deemed necessary to implement atomics, normal initialization syntax should be sufficient.

Implementations may keep this macro, though, because the ATOMIC_... prefix is still reserved and so they do not impede on the users name space for identifiers.

New _BitInt(N) and decimal floating point types must be supported by type-generic interfaces (mandatory change)

For implementations that mostly rely on _Generic or similar features to provide operations such as atomic_fetch_add, for example, this addition to the interface might be quite challenging.

ABI choice: Also ABI decisions have to be taken which of these new types, if any, are to be lock-free. There are no particular feature test macros for these types concerning this property (so no preprocessor conditionals can be used) but the generic function atomic_is_lock_free has to cope with them. This function could, for example, return true for all types where there is a supporting basic type with the same size that is lock-free.

New macro ATOMIC_CHAR8_T_LOCK_FREE and type atomic_char8_t (mandatory change)

This is necessary because of the addition of char8_t to <uchar.h>. This can be done by inserting the following two lines at the appropriate places of <stdatomic.h>.

#define ATOMIC_CHAR8_T_LOCK_FREE ATOMIC_CHAR_LOCK_FREE
typedef _Atomic(unsigned char) atomic_char8_t;

Because the ATOMIC_ and atomic_ prefixes are reserved, no preprocessor conditional is needed.

Changes to <stdbool.h>

__STDC_VERSION_BOOL_H__ ?

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

Remove/protect bool, false and true

The following could be the complete contents of this header:

#ifndef __STDC_VERSION_BOOL_H__
#define __STDC_VERSION_BOOL_H__ 202311L

#if (__STDC_VERSION__ < 202300L) && !defined(__cplusplus)
#define true  1
#define false 0
#define bool  _Bool
#else
#define true  true
#define false false
#define bool  bool
#endif

#define __bool_true_false_are_defined 1

#endif

Note that this unconditionally defines the macros, because some application code might do preprocessor conditionals on these.

Changes to <stddef.h>

__STDC_VERSION_STDDEF_H__

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

unreachable macro (mandatory change)

Defined in 7.21.1.

• https://open-std.org/JTC1/SC22/WG14/www/docs/n2826.pdf, the macro option has been chosen by WG14.

Compiler dependency: Quality implementations need compiler support for this.

gcc and clang already have __builtin_unreachable so in general here the “implementation” for the interface is only

#ifdef unreachable
# error unreachable() is a standard macro for C23
#endif
#if __STDC_VERSION__ >= 202311L
# define unreachable() __builtin_unreachable()
#else
# define unreachable static_assert(0, "unreachable becomes a standard macro for C23")
#endif

This also detects usage of the now-reserved identifier unreachable in code that is not yet ready for C23.

Because it is just undefined in the primary sense of the word if a call to that feature is reached, theoretically low quality implementations could do nonsense as the following:

#define unreachable() ((void)0)
#define unreachable() ((void)(1/0))
#define unreachable() ((void)puts("reached the ureachable"))
#define unreachable() abort()
#define unreachable() give_me_your_credit_card_number()

All of this would result in suboptimal code for their users, because this feature is meant such that whole branches of the control flow graph can be pruned from the executable.

nullptr_t

C23 has a new keyword and constant nullptr and provides access to the underlying type via <stddef.h> as

#if __STDC_VERSION__ > 202311L
typedef typeof(nullptr) nullptr_t;
#endif

On general implementations, a preprocessor conditional is needed, because the identifiers nullptr and nullptr_t had not been reserved before C23. For the type alone, such a conditional is not needed on POSIX systems, since their types with suffix _t are already reserved.

NULL

Because of the ambiguity of its definition, this macro has some portability and safety problems and so C23 has integrated nullptr (introduced to C++ in 2011) to replace it on the long run.

There is no requirement, yet, to have this macro point to nullptr, though, but on non-POSIX implementations it might be a good idea to move to something like the following:

#if (__cplusplus >= 201103L) || (__STDC_VERSION__ >= 202311L)
#define NULL nullptr
#elif defined(__cplusplus)
#define NULL 0L              /* Any of 0, 0L, 0LL as wide as a void* */
#else
#define NULL ((void*)0)
#endif

POSIX implementations should for the moment stay with ((void*)0) since this is a requirement, there, until this constraint is lifted. Since nullptr values have the same representation as void* this should not result in much difficulties, anyhow.

A preprocessor conditional is needed, because nullptr is new for C23.

Changes to <stdio.h>

__STDC_VERSION_STDIO_H__

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

Changes to formatted IO

This implies changes for printf, scanf and friends.

Printing of narrow types

See <inttypes.h>

"%lc" conversion of L'\0'

In a last minute change (GB-141 in CD2 ballot) the strategy for printing wide characters has changed. The only effective change is for a nul wide character, which previously had no output and now results in NUL. This is consistent with other printing of nul characters, but is a normative change nevertheless.

printing NaN

There is a new _PRINTF_NAN_LEN_MAX macro that holds the maximum length of output corresponding to a NaN value that would be issued by printf and friends.

• https://open-std.org/JTC1/SC22/WG14/www/docs/n2446.htm

the b conversion specifier

This specifies binary input and output analogous to hexadecimal, only that the character b plays the role of x.

For scanf, this is a semantic change, because input may be accepted or rejected according the version of the C library that is linked to the excutable, see strtol below. Depending on the solutions that WG14 might still find for that problems, it might be possible that a whole second set of scanf interfaces is needed.

Introduction of a similar B conversion specifier (comparable to X) for printf is optional and not imposed because implementations could already have occupied this space. So if an implementations does currently nothing particular for B it is expected that they also implement B analogous to X

The macros in <inttypes.h> with B (such as PRIB128) act as feature test macros.

• https://open-std.org/JTC1/SC22/WG14/www/docs/n2630.pdf
• https://open-std.org/JTC1/SC22/WG14/www/docs/n3072.htm

wN length modifiers

Where N is any of the values (decimal without leading 0) for all supported minimum-width integer types provided by 7.22.1.2. The exact width integer types (7.22.1.1) now are a subset of these, so in particular these must all be supported. To implement these, it must be known to which size a given width corresponds, so in particular which widths are natively supported by the architecture. For 8, 16, 32 and 64 the minimum-width integer types are mandatory, so at least these values must be supported.

• https://open-std.org/JTC1/SC22/WG14/www/docs/n2680.pdf

Implementations should try to support N and exact-width types for values where there might not even be compiler support, yet. Now that we know that signed integers are two’s complement, once the byte interpretation (endianess) of int128_t for example is known, I/O on these represented values should not be difficult. Adding such values for which there might be no full support for the type otherwise is allowed, it just has to be documented. In any case, applications may easily check with #ifdef.

Note that it is not expected that these macros work for the new _BitInt(N) types even if N is one of the standard value. This is because the argument-passing convention for these types may be different from the standard integer types.

For versions of the C library that support these formats (and thus set __STDC_VERSION_STDIO_H__) the macros PRIxN etc should use these new length modifiers. This could look something like

#define _PRI(F, N) "w" #N #F
#define PRIX128 _PRI(X, 128)
#define PRIX16  _PRI(X, 16)
#define PRIX256 _PRI(X, 256)
#define PRIX32  _PRI(X, 32)
#define PRIX64  _PRI(X, 64)
#define PRIX8   _PRI(X, 8)
#define PRId128 _PRI(d, 128)
#define PRId16  _PRI(d, 16)
#define PRId256 _PRI(d, 256)
#define PRId32  _PRI(d, 32)
#define PRId64  _PRI(d, 64)
#define PRId8   _PRI(d, 8)
#define PRIi128 _PRI(i, 128)
#define PRIi16  _PRI(i, 16)
#define PRIi256 _PRI(i, 256)
#define PRIi32  _PRI(i, 32)
#define PRIi64  _PRI(i, 64)
#define PRIi8   _PRI(i, 8)
#define PRIo128 _PRI(o, 128)
#define PRIo16  _PRI(o, 16)
#define PRIo256 _PRI(o, 256)
#define PRIo32  _PRI(o, 32)
#define PRIo64  _PRI(o, 64)
#define PRIo8   _PRI(o, 8)
#define PRIu128 _PRI(u, 128)
#define PRIu16  _PRI(u, 16)
#define PRIu256 _PRI(u, 256)
#define PRIu32  _PRI(u, 32)
#define PRIu64  _PRI(u, 64)
#define PRIu8   _PRI(u, 8)
#define PRIx128 _PRI(x, 128)
#define PRIx16  _PRI(x, 16)
#define PRIx256 _PRI(x, 256)
#define PRIx32  _PRI(x, 32)
#define PRIx64  _PRI(x, 64)
#define PRIx8   _PRI(x, 8)

#define PRIXLEAST128 PRIX128
#define PRIXLEAST16 PRIX16
#define PRIXLEAST256 PRIX256
#define PRIXLEAST32 PRIX32
#define PRIXLEAST64 PRIX64
#define PRIXLEAST8 PRIX8
#define PRIdLEAST128 PRId128
#define PRIdLEAST16 PRId16
#define PRIdLEAST256 PRId256
#define PRIdLEAST32 PRId32
#define PRIdLEAST64 PRId64
#define PRIdLEAST8 PRId8
#define PRIiLEAST128 PRIi128
#define PRIiLEAST16 PRIi16
#define PRIiLEAST256 PRIi256
#define PRIiLEAST32 PRIi32
#define PRIiLEAST64 PRIi64
#define PRIiLEAST8 PRIi8
#define PRIoLEAST128 PRIo128
#define PRIoLEAST16 PRIo16
#define PRIoLEAST256 PRIo256
#define PRIoLEAST32 PRIo32
#define PRIoLEAST64 PRIo64
#define PRIoLEAST8 PRIo8
#define PRIuLEAST128 PRIu128
#define PRIuLEAST16 PRIu16
#define PRIuLEAST256 PRIu256
#define PRIuLEAST32 PRIu32
#define PRIuLEAST64 PRIu64
#define PRIuLEAST8 PRIu8
#define PRIxLEAST128 PRIx128
#define PRIxLEAST16 PRIx16
#define PRIxLEAST256 PRIx256
#define PRIxLEAST32 PRIx32
#define PRIxLEAST64 PRIx64
#define PRIxLEAST8 PRIx8

Similar for scanf:

#define _SCN(F, N) "w" #N #F
#define SCNX128 _SCN(X, 128)
#define SCNX16  _SCN(X, 16)
#define SCNX256 _SCN(X, 256)
#define SCNX32  _SCN(X, 32)
#define SCNX64  _SCN(X, 64)
#define SCNX8   _SCN(X, 8)
#define SCNd128 _SCN(d, 128)
#define SCNd16  _SCN(d, 16)
#define SCNd256 _SCN(d, 256)
#define SCNd32  _SCN(d, 32)
#define SCNd64  _SCN(d, 64)
#define SCNd8   _SCN(d, 8)
#define SCNi128 _SCN(i, 128)
#define SCNi16  _SCN(i, 16)
#define SCNi256 _SCN(i, 256)
#define SCNi32  _SCN(i, 32)
#define SCNi64  _SCN(i, 64)
#define SCNi8   _SCN(i, 8)
#define SCNo128 _SCN(o, 128)
#define SCNo16  _SCN(o, 16)
#define SCNo256 _SCN(o, 256)
#define SCNo32  _SCN(o, 32)
#define SCNo64  _SCN(o, 64)
#define SCNo8   _SCN(o, 8)
#define SCNu128 _SCN(u, 128)
#define SCNu16  _SCN(u, 16)
#define SCNu256 _SCN(u, 256)
#define SCNu32  _SCN(u, 32)
#define SCNu64  _SCN(u, 64)
#define SCNu8   _SCN(u, 8)
#define SCNx128 _SCN(x, 128)
#define SCNx16  _SCN(x, 16)
#define SCNx256 _SCN(x, 256)
#define SCNx32  _SCN(x, 32)
#define SCNx64  _SCN(x, 64)
#define SCNx8   _SCN(x, 8)

#define SCNXLEAST128 SCNX128
#define SCNXLEAST16 SCNX16
#define SCNXLEAST256 SCNX256
#define SCNXLEAST32 SCNX32
#define SCNXLEAST64 SCNX64
#define SCNXLEAST8 SCNX8
#define SCNdLEAST128 SCNd128
#define SCNdLEAST16 SCNd16
#define SCNdLEAST256 SCNd256
#define SCNdLEAST32 SCNd32
#define SCNdLEAST64 SCNd64
#define SCNdLEAST8 SCNd8
#define SCNiLEAST128 SCNi128
#define SCNiLEAST16 SCNi16
#define SCNiLEAST256 SCNi256
#define SCNiLEAST32 SCNi32
#define SCNiLEAST64 SCNi64
#define SCNiLEAST8 SCNi8
#define SCNoLEAST128 SCNo128
#define SCNoLEAST16 SCNo16
#define SCNoLEAST256 SCNo256
#define SCNoLEAST32 SCNo32
#define SCNoLEAST64 SCNo64
#define SCNoLEAST8 SCNo8
#define SCNuLEAST128 SCNu128
#define SCNuLEAST16 SCNu16
#define SCNuLEAST256 SCNu256
#define SCNuLEAST32 SCNu32
#define SCNuLEAST64 SCNu64
#define SCNuLEAST8 SCNu8
#define SCNxLEAST128 SCNx128
#define SCNxLEAST16 SCNx16
#define SCNxLEAST256 SCNx256
#define SCNxLEAST32 SCNx32
#define SCNxLEAST64 SCNx64
#define SCNxLEAST8 SCNx8

wfN length modifiers

Where N is any of the values (decimal without leading 0) for all supported fastest minimum-width integer types provided by 7.22.1.3.

ABI choice: To implement these, it must be known to which size a given width corresponds, this is an ABI decision. For 8, 16, 32 and 64 the fastest minimum-width integer types are mandatory, so at least these values must be supported.

• https://open-std.org/JTC1/SC22/WG14/www/docs/n2680.pdf

For versions of the C library that support these formats (and thus set __STDC_VERSION_STDIO_H__) the macros PRIxN etc should use these new length modifiers. This could look something like

#define _PRIFAST(F, N) "wf" #N #F
#define PRIXFAST128 _PRIFAST(X, 128)
#define PRIXFAST16  _PRIFAST(X, 16)
#define PRIXFAST256 _PRIFAST(X, 256)
#define PRIXFAST32  _PRIFAST(X, 32)
#define PRIXFAST64  _PRIFAST(X, 64)
#define PRIXFAST8   _PRIFAST(X, 8)
#define PRIdFAST128 _PRIFAST(d, 128)
#define PRIdFAST16  _PRIFAST(d, 16)
#define PRIdFAST256 _PRIFAST(d, 256)
#define PRIdFAST32  _PRIFAST(d, 32)
#define PRIdFAST64  _PRIFAST(d, 64)
#define PRIdFAST8   _PRIFAST(d, 8)
#define PRIiFAST128 _PRIFAST(i, 128)
#define PRIiFAST16  _PRIFAST(i, 16)
#define PRIiFAST256 _PRIFAST(i, 256)
#define PRIiFAST32  _PRIFAST(i, 32)
#define PRIiFAST64  _PRIFAST(i, 64)
#define PRIiFAST8   _PRIFAST(i, 8)
#define PRIoFAST128 _PRIFAST(o, 128)
#define PRIoFAST16  _PRIFAST(o, 16)
#define PRIoFAST256 _PRIFAST(o, 256)
#define PRIoFAST32  _PRIFAST(o, 32)
#define PRIoFAST64  _PRIFAST(o, 64)
#define PRIoFAST8   _PRIFAST(o, 8)
#define PRIuFAST128 _PRIFAST(u, 128)
#define PRIuFAST16  _PRIFAST(u, 16)
#define PRIuFAST256 _PRIFAST(u, 256)
#define PRIuFAST32  _PRIFAST(u, 32)
#define PRIuFAST64  _PRIFAST(u, 64)
#define PRIuFAST8   _PRIFAST(u, 8)
#define PRIxFAST128 _PRIFAST(x, 128)
#define PRIxFAST16  _PRIFAST(x, 16)
#define PRIxFAST256 _PRIFAST(x, 256)
#define PRIxFAST32  _PRIFAST(x, 32)
#define PRIxFAST64  _PRIFAST(x, 64)
#define PRIxFAST8   _PRIFAST(x, 8)

Similar for scanf:

#define _SCNFAST(F, N) "wf" #N #F
#define SCNXFAST128 _SCNFAST(X, 128)
#define SCNXFAST16  _SCNFAST(X, 16)
#define SCNXFAST256 _SCNFAST(X, 256)
#define SCNXFAST32  _SCNFAST(X, 32)
#define SCNXFAST64  _SCNFAST(X, 64)
#define SCNXFAST8   _SCNFAST(X, 8)
#define SCNdFAST128 _SCNFAST(d, 128)
#define SCNdFAST16  _SCNFAST(d, 16)
#define SCNdFAST256 _SCNFAST(d, 256)
#define SCNdFAST32  _SCNFAST(d, 32)
#define SCNdFAST64  _SCNFAST(d, 64)
#define SCNdFAST8   _SCNFAST(d, 8)
#define SCNiFAST128 _SCNFAST(i, 128)
#define SCNiFAST16  _SCNFAST(i, 16)
#define SCNiFAST256 _SCNFAST(i, 256)
#define SCNiFAST32  _SCNFAST(i, 32)
#define SCNiFAST64  _SCNFAST(i, 64)
#define SCNiFAST8   _SCNFAST(i, 8)
#define SCNoFAST128 _SCNFAST(o, 128)
#define SCNoFAST16  _SCNFAST(o, 16)
#define SCNoFAST256 _SCNFAST(o, 256)
#define SCNoFAST32  _SCNFAST(o, 32)
#define SCNoFAST64  _SCNFAST(o, 64)
#define SCNoFAST8   _SCNFAST(o, 8)
#define SCNuFAST128 _SCNFAST(u, 128)
#define SCNuFAST16  _SCNFAST(u, 16)
#define SCNuFAST256 _SCNFAST(u, 256)
#define SCNuFAST32  _SCNFAST(u, 32)
#define SCNuFAST64  _SCNFAST(u, 64)
#define SCNuFAST8   _SCNFAST(u, 8)
#define SCNxFAST128 _SCNFAST(x, 128)
#define SCNxFAST16  _SCNFAST(x, 16)
#define SCNxFAST256 _SCNFAST(x, 256)
#define SCNxFAST32  _SCNFAST(x, 32)
#define SCNxFAST64  _SCNFAST(x, 64)
#define SCNxFAST8   _SCNFAST(x, 8)

H, D and DD length modifiers for decimal floating point

For _Decimal32, _Decimal64 and _Decimal128. This is optional depending on support for decimal floating point. Implementation should not be too difficult. In particular these new number types have prescribed representation formats (2 possible choices and endianess), so implementation should even be possible without complete support for these types on the compilation platform of the C library. Some support is needed from <math.h> namely classification for infinite or NaN values.

Compiler dependency: Support for this is important such that the rest of any library support for decimal floating point can be added by an independent library.

Changes to <stdlib.h>

__STDC_VERSION_STDLIB_H__

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

strtol and friends

With base 0 or 2, these functions now accept the new binary integer constants. When used with the optional the prefixes 0b or 0B, this is a semantic change: for example the following code

long res = strtol("0b1", 0, 0);

has res ≡ 0 for C17 and res ≡ 1 for C23. This is because for the first the interpretation stops before the b.

C library implementations that want to support previous versions of C, have to take care of that semantic change. Therefore they probably have to have two functions for each of the interfaces

// don't recognize 0xb or 0xB for base 0 or 2
long int strtol_c17(const char *restrict nptr, char **restrict endptr, int base);
long long int strtoll_c17(const char *restrict nptr, char **restrict endptr, int base);
unsigned long int strtoul_c17(const char *restrict nptr, char **restrict endptr, int base);
unsigned long long int strtoull_c17(const char *restrict nptr, char **restrict endptr, int base);

// recognize 0xb or 0xB for base 0 or 2
long int strtol_c23(const char *restrict nptr, char **restrict endptr, int base);
long long int strtoll_c23(const char *restrict nptr, char **restrict endptr, int base);
unsigned long int strtoul_c23(const char *restrict nptr, char **restrict endptr, int base);
unsigned long long int strtoull_c23(const char *restrict nptr, char **restrict endptr, int base);

#if __STDC_VERSION__ >= 202311L
#define strtol   strtol_c23
#define strtoll  strtoll_c23
#define strtoul  strtoul_c23
#define strtoull strtoull_c23
#else
#define strtol   strtol_c17
#define strtoll  strtoll_c17
#define strtoul  strtoul_c17
#define strtoull strtoull_c17
#endif

Not yet stable? It is possible that there will be an NB comment for this and that this interface might still change, e.g by adding [[deprecated]] to the existing interfaces and by creating new interfaces with prefix stdc_.

Alignment requirements for memory management functions

These were reformulated for C23. Implementations that set __STDC_VERSION_STDLIB_H__ have to verify that they conform to that reformulation, see

• https://open-std.org/JTC1/SC22/WG14/www/docs/n2293.htm

calloc overflow

Implementations that set __STDC_VERSION_STDLIB_H__ have to verify that they conform to 7.24.3.2 p3:

The calloc function returns either a pointer to the allocated space or a null pointer if the space cannot be allocated or if the product nmemb * size would wraparound size_t.

bsearch becomes a const-preserving tg macro

This is specified as

QVoid *bsearch(const void *, QVoid*, size_t nmemb, size_t size, int (*)(const void*, const void*));

to emphasize that the return is exactly the same void pointer type as the argument. volatile or restrict types are not accepted.

An implementation of this type-generic macro could look as follows.

// The function itself stays exactly the same.
void* (bsearch)(const void*, const void*, size_t, size_t, int(*)(const void*, const void*));
#if __STDC_VERSION__ > 202300L
# define bsearch(K, B, N, S, C)                                                  \
    _Generic(                                                                    \
        /* ensure conversion to a void pointer */                                \
        true ? (B) : (void*)1,                                                   \
        void const*: (void const*)bsearch((K), (void const*)(B), (N), (S), (C)), \
        /* volatile qualification of *B is an error for this call */             \
        default:     bsearch((K), (B), (N), (S), (C))                            \
)
#endif

A preprocessor conditional is needed, because the type of call expressions potentially changes with this.

User code that misused these calls and stored the result for a call with a const qualified array in a pointer with unqualified target type may see their code diagnosed or even rejected. This is intentional.

Functions free_sized and free_aligned_sized

Add once_flag and call_once also to <stdlib.h>

This type and function now become mandatory for C23.

Similar to size_t it can appear in several headers. Something along the lines of should be added to <stdlib.h>

#if (__STDC_VERSION__ >= 201311L) && !defined(ONCE_FLAG_INIT)
typedef int once_flag;
#define ONCE_FLAG_INIT 0
void call_once(once_flag*, void (*)(void));
#endif

A protection by a preprocessor conditional is not strictly necessary because these names are otherwise only potentially reserved, so adding them to a C library is always possible without impeding on the user identifier space.

A rationale for introducing this feature to the general C library and reference implementations for C libraries that might not have that feature, yet, because they don’t support threads can be found in

https://open-std.org/JTC1/SC22/WG14/www/docs/n2840.htm

Implementations that know how to avoid to link against the whole threads and atomic options could have an alternative version of this function

typedef int volatile once_flag;
#define ONCE_FLAG_INIT 0
void call_once(once_flag* flag, void (*func)(void)) {
    if (!*flag) {
        func();
        *flag = 1;
    }
}

In this case, the necessary synchronization guarantees are given, because the call to the function and the assignment to flag are sequenced and cannot be optimized away. Note that this version is not guaranteed to be asynchronous signal safe: not even the type sig_atomic_t does not give the guarantees that are needed for implementing this function. Only an implementation with lock-free require-release atomics (or similar) could be asynchronous signal safe.

New functions strfromd, strfromf and strfroml

New function memalignment

Changes to <time.h>

__STDC_VERSION_TIME_H__

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

strftime and wcsftime

Two new O modified specifications are added to strftime and wcsftime:

%Ob is replaced by the locale’s abbreviated alternative month name.

%OB is replaced by the locale’s alternative appropriate full month name.

These have also been accepted for POSIX.

The C standard says nothing about erroneous specifications or some that would be implementation-defined extensions, so using new specifiers is just undefined for pre-C23. Therefore these can be added to any implementation without making it non-conforming for any C version.

On the other hand, __STDC_VERSION_TIME_H__ should only be taken to 202311L once this addition is implemented.

Deprecation of asctime and ctime

C23 follows POSIX in deprecating these functions. Their interfaces now have the corresponding attribute:

#if __STDC_VERSION__ > 202300L
[[deprecated]] char *asctime(const struct tm *timeptr);
[[deprecated]] char *ctime(const time_t *timer);
#else
char *asctime(const struct tm *timeptr);
char *ctime(const time_t *timer);
#endif

A preprocessor conditional is needed, because the attribute syntax is new for C23.

gmtime_r and localtime_r

C23 inherits these two interfaces from POSIX. C libraries that implement POSIX can simply expose them to C23 aware code by something similar to

#if __STDC_VERSION__ > 202300L
struct tm *gmtime_r(const time_t*, struct tm*);
struct tm *localtime_r(const time_t*, struct tm*);
#endif

A preprocessor conditional is needed, because the names had not been reserved before C23.

mktime

In a literal last minute change (GB-159 in CD2 ballot and N3147 ) it the behavior for mktime has been restricted. If the time that is provided in the argument is not representable the tm_wday member has to remain untouched.

timegm

The timegm function from BSD is added. Specifications can be found here.

https://open-std.org/JTC1/SC22/WG14/www/docs/n2833.htm

with an amendment in

https://open-std.org/JTC1/SC22/WG14/www/docs/n3147.txt

A preprocessor conditional is needed, because the name had not been reserved before C23.

#if __STDC_VERSION__ > 202300L
time_t timegm(struct tm *);
#endif

timespec_getres

This function is meant to port POSIX’ clock_getres to C by translating base to the equivalent POSIX clock:

#if __STDC_VERSION__ > 202300L
int timespec_getres(struct timespec*, int);
#endif

A preprocessor conditional is needed, because the name had not been reserved before C23.

TIME_MONOTONIC, TIME_ACTIVE, TIME_THREAD_ACTIVE

C23 adds three optional time bases TIME_MONOTONIC, TIME_ACTIVE and TIME_THREAD_ACTIVE which are modeled after the POSIX clocks CLOCK_MONOTONIC, CLOCK_PROCESS_CPUTIME_ID and CLOCK_THREAD_CPUTIME_ID, respectively.

Having time bases for C other than TIME_UTC is at the liberty of the implementation, so any C library that runs on a POSIX system could easily provide the equivalent to all POSIX clocks that it interfaces. For example, for Linux currently these are

#define CLOCK_REALTIME           0
#define CLOCK_MONOTONIC          1
#define CLOCK_PROCESS_CPUTIME_ID 2
#define CLOCK_THREAD_CPUTIME_ID  3
#define CLOCK_MONOTONIC_RAW      4
#define CLOCK_REALTIME_COARSE    5
#define CLOCK_MONOTONIC_COARSE   6
#define CLOCK_BOOTTIME           7
#define CLOCK_REALTIME_ALARM     8
#define CLOCK_BOOTTIME_ALARM     9
#define CLOCK_SGI_CYCLE         10
#define CLOCK_TAI               11

This could easily be done by using something as

#define TIME_UTC              (CLOCK_REALTIME+1)
#define TIME_MONOTONIC        (CLOCK_MONOTONIC+1)
#define TIME_ACTIVE           (CLOCK_PROCESS_CPUTIME_ID+1)
#define TIME_THREAD_ACTIVE    (CLOCK_THREAD_CPUTIME_ID+1)
#define TIME_MONOTONIC_RAW    (CLOCK_MONOTONIC_RAW+1)
#define TIME_UTC_COARSE       (CLOCK_REALTIME_COARSE+1)
#define TIME_MONOTONIC_COARSE (CLOCK_MONOTONIC_COARSE+1)
#define TIME_BOOTTIME         (CLOCK_BOOTTIME+1)
#define TIME_UTC_ALARM        (CLOCK_REALTIME_ALARM+1)
#define TIME_BOOTTIME_ALARM   (CLOCK_BOOTTIME_ALARM+1)
#define TIME_SGI_CYCLE        (CLOCK_SGI_CYCLE+1)
#define TIME_TAI              (CLOCK_TAI+1)

and then adapting the corresponding implementation of timespec_get a bit. This would be conforming to current and future C, because the TIME_ prefix is already reserved for that purpose.

ABI choice: Unfortunately the choice of the values is an ABI choice, so before doing so it has to be ensured that other C libraries on the same platform use the same values.

Changes to <string.h>

__STDC_VERSION_STRING_H__

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

memccpy, strdup and strndup

C23 borrows these interfaces from POSIX

void* memccpy(void*, const void*, int, size_t);
char* strdup(const char*);
char* strndup(const char*, size_t);

A preprocessor conditional is not needed, because the names (with prefixes mem and str) had been reserved before C23.

memset_explicit

The following function is added with the same normative specification as memset:

void* memset_explicit(void*, int, size_t);

The difference to that is in expectation. In the description

The purpose of this function is to make sensitive information stored in the object inaccessible³⁷⁹⁾.

And with the following text in the footnote.

The intention is that the memory store is always performed (i.e., never elided), regardless of optimizations.

There has been long and heated debate in WG14 and WG21 about this, and this is the best that we came up with. The standard does not have the language to describe that memory that is e.g freed has to be zeroed out when given back to the system; such a feature is not observable from within the program.

So this puts the responsibility for the intended purpose (hiding information from one part of the execution to other parts and to the outer world) entirely a point of “quality of implementation”. Implementations should take care:

• A call to this function should never be optimized out. This can often be achieved by having it in a separate TU from all other functions and by disabling link-time optimization for this TU or at least for this function.
• No store to any byte of the function should be optimized out.
• The return of the function should synchronize with all read and write operations. This could for example be achieved by issuing a call atomic_signal_fence(memory_order_seq_cst) or equivalent, such that even a signal that kicks in right after the call could not not read the previous contents of the byte array.
• All caches for the byte array should have been invalidated on return.
• To avoid side-channel attacks, the implementation of the function should make no explicit or implicit reference to the contents of the byte array nor the value that has been chosen for the overwrite. Each write operation should use the same time (and other resources) per byte.
• Good performance is not expected, security first.

A preprocessor conditional is not needed, because the name (with prefix mem) had been reserved before C23.

memchr, strchr, strpbrk, strrchr and strstr become const-preserving tg macros

For example for memchr this is specified as

QVoid* memchr(QVoid*, int, size_t);

to emphasize that the return is exactly the same void pointer type as the argument. volatile or restrict types are not accepted.

An implementation of this type-generic macro could look as follows.

// The function itself stays exactly the same.
void* (memchr)(void const*, int, size_t);
#if __STDC_VERSION__ > 202300L
# define memchr(S, C, N)                                              \
    _Generic(                                                         \
        /* ensure conversion to a void pointer */                     \
        true ? (S) : (void*)1,                                        \
        void const*: (void const*)memchr((void const*)(S), (C), (N)), \
        /* volatile qualification is an error for this call */        \
        default:     memchr((S), (C), (N))                            \
)
#endif

A preprocessor conditional is needed, because the type of call expressions potentially changes with this.

User code that misused these calls and stored the result for a call with a const qualified array in a pointer with unqualified target type may see their code diagnosed or even rejected. This is intentional.

Changes to <uchar.h>

The required Unicode support has been straightened out and complemented. The types charN_t now are designated for UTF-N encoding without exception. For N 16 or 32 there should not be much changes to existing C libraries: WG14 found none for which the previously optional macros __STDC_UTF_16__ and __STDC_UTF_32__ had not been set.

Also, see the discussion about thread safety above.

__STDC_VERSION_UCHAR_H__

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

Conversion functions and type for UTF-8

The set of conversion functions from and to UTF encodings is completed by adding functions for UTF-8. For implementations that have UTF-8 as internal representation, anyhow, these functions are almost no-opts; they just have to iterate through the character sequence that composes the multi-byte character in UTF-8 encoding.

#if __STDC_VERSION__ >= 202300L
typedef unsigned char char8_t;
size_t mbrtoc8(char8_t * restrict pc8, const char * restrict s, size_t n, mbstate_t * restrict ps);
size t c8rtomb(char * restrict s, char8_t c8, mbstate_t * restrict ps);
#endif

The preprocessor conditional is needed, because the names had not been previously reserved.

Changes to <wchar.h>

__STDC_VERSION_WCHAR_H__

This macro is mandatory and should be set to 202311L once the header is compliant to C23.

wcschr, wcspbrk, wcsrchr, wcsstr and wmemchr become const-preserving tg macros

Similar to memchr and similar, above.

Change to fputwc

This function now also sets the error indicator of the stream if an encoding error occurs.

Changes in Annex K

bsearch_s

Similar to bsearch, above.