AnthonyCalandra/modern-c-features

modern-c-features

Overview

A collection of descriptions along with examples for C language and library features.

C23 includes the following language features:

• auto
• constexpr
• decimal floating-point types
• bit-width integers
• binary literals
• char8_t
• utf-8 character literals
• unicode string literals
• empty initializer
• attributes
• new keywords (C23)
• nullptr
• #embed
• enums with underlying type
• typeof
• improved compatibility for tagged types

C23 includes the following library features:

• floating-point formatting functions
• fscanf and fprintf format specifiers
• memset_explicit
• unreachable macro
• memccpy
• strdup and strndup
• gmtime_r and localtime_r
• timespec_getres

C17 contains defect reports and deprecations.

C11 includes the following language features:

• generic selection
• alignof
• alignas
• static_assert
• noreturn
• wide string literals
• anonymous structs and unions

C11 includes the following library features:

• bounds checking
• timespec_get
• aligned_malloc
• char32_t
• char16_t
• <stdatomic.h>
  • Atomic types
  • Atomic flags
  • Atomic variables
• <threads.h>
  • Threads
  • Mutexes
  • Condition variables
• quick exiting
• exclusive mode file opening

C23 Language Features

auto

auto-typed variables are deduced by the compiler according to the type of their initializer.

auto f = 123.0f; // `f` deduced to `float`

#include <tgmath.h> // import the `cos` macro
auto c = cos(x); // `c` deduced to a type depending on `x`

#define div(X, Y)            \
  _Generic((X)+(Y),          \
           int: div,         \
           long: ldiv,       \
           long long: lldiv) \
           ((X), (Y))
auto z = div(x, y); // deduced to either `int`, `long`, `long long`

See: generic selection.

constexpr

Scalars specified with constexpr are constants: they cannot be modified after being initialized, and must be fully initialized to a value that can be stored in the given type.

The C constexpr specifier does not support functions, structures, unions, or arrays. Additionally, constexpr cannot be specified for pointer, atomic, or volatile-qualified types.

constexpr size_t cache_line_size_bytes = 64;

Decimal floating-point types

Supports IEEE-754 Decimal floating-point types, _Decimal32, _Decimal64, _Decimal128. These floating point types are designed for base-10 floating point semantics.

As of September 2023, compiler support is lacking for most features.

_Decimal32 decsum = 0.0df;
for (int i = 0; i < 10; i++)
    decsum += 0.1df;
printf("%.17Ha\n", decsum); // would print "1.0"

float fsum = 0.0f;
for (int i = 0; i < 10; i++)
    fsum += 0.1f;
printf("%.17f\n", fsum); // could print "1.00000011920928955"

Bit-width integers

The _BitInt(N) type allows specifying an N bit integer signed or unsigned type.

_BitInt(4) sbi; // 4-bit width signed integer
unsigned _BitInt(4) ubi; // 4-bit width unsigned integer

Binary literals

Binary literals provide a convenient way to represent a base-2 number. It is possible to separate digits with '.

0b110 // == 6
0b1111'1111 // == 255

char8_t

An unsigned char type for holding 8-bit wide characters.

See: char32_t, char16_t, unicode string literals.

UTF-8 character literals

A character literal that begins with u8 is a character literal of type char8_t. The value of a UTF-8 character literal is equal to its ISO 10646 code point value.

char8_t x = u8'x';

See: char8_t, unicode string literals.

Unicode string literals

String literals prefixed with u8, u, and U represent UTF-8, UTF-16, and UTF-32 strings respectively.

See: char32_t, char16_t, char8_t, wide string literals.

Empty initializer

An object is empty-initialized if it is explicitly initialized using only a pair of braces. Arrays of unknown size cannot be initialized by an empty initializer. If an object is initialized using an empty initializer, default initialization occurs:

• Pointers are initialized to NULL;
• Decimal floating point types are initialized to positive zero;
• Arithmetic types are initialized to zero;
• Array types are initialized such that each slot is initialized according to these rules;
• Aggregate types are initialized such that each member is initialized according to these rules;
• Unions are initialized such that the first named member is initialized according to these rules.

char c = {}; // == 0
float f = {}; // == 0

struct
{
    int x;
    int y;
} f = {}; // == { x: 0, y: 0}

union
{
        char x;
        double y;
} u = {}; // u.x == 0

int ia[5] = {}; // == [0, 0, 0, 0, 0]

Attributes

Attributes provide a universal syntax over __attribute__(...), __declspec, etc. C++-styled attributes.

C23 provides the following attributes:

• [[deprecated("reason")]] - the entity declared with the attribute (such as a function), is deprecated. Specifying a reason is optional and can be ommitted.
• [[fallthrough]] - indicates falling-through in a switch statement case is intentional.
• [[nodiscard("reason")]] - encourages a compiler warning if the return value for the function declared with this attribute is discarded. Specifying a reason is optional and can be omitted.
• [[maybe_unused]] - suppresses a compiler warning on the entity where this attribute is declared if it's unused (such as an unused function parameter), if applicable.
• [[noreturn]] - indicates that the function this attribute is declared on does not return.
• Compiler-specific directives - such as [[clang::no_sanitize]].

// `noreturn` attribute indicates `f` doesn't return.
[[noreturn]] void f()
{
    exit(0);
}

See: noreturn.

New keywords (C23)

New keywords replacing traditional macros definitions:

• true and false
• thread_local
• static_assert

See: static_assert.

nullptr

Introduces a new null pointer type designed to replace NULL. nullptr itself is of type nullptr_t which can be implicitly converted into pointer types and bool, and -- unlike NULL -- is not convertible to integral types.

void foo(int);
foo(NULL); // valid
foo(nullptr); // error

#embed

#embed is a preprocessor directive to include text and binary resources directly into source code. This pattern is common in applications such as games that create custom fonts, load images into memory, etc. This allows the C preprocessor to replace external tools that convert resources to byte representations as C arrays.

The #embed preprocessor directive also includes optional parameters including suffix, prefix, and if_empty.

const uint8_t image_bytes[] = {
#embed "image.bmp"
};
 
const char message_text[] = {
#embed "message.txt" if_empty('M', 'i', 's', 's', 'i', 'n', 'g', '\n')
,'\0'
};

Enums with underlying type

C23 enums provide the ability to optionally specify the underlying type.

enum e : unsigned short
{
    x // `x` is an `unsigned short`
};

typeof

Gets the type of an expression, similar to decltype in C++. Also includes typeof_unqual which removes cv-qualifiers from the type.

int a;
const volatile int b;
typeof(a) c; // has type of `int`
typeof_unqual(b) d; // has type of `int

Improved compatibility for tagged types

Tagged types that have the same tag name and content become compatible not only across translation units but also inside the same translation unit. Also, redeclaration of the same tagged types is now allowed.

// header
#define PRODUCT(A ,B) struct prod { A a; B b; }
#define SUM(A, B) struct sum { bool flag; union { A a; B b; }; }

// source
void foo(PRODUCT(int, SUM(float, double)) x)
{
    // ...
}

void bar(PRODUCT(int, SUM(float, double)) y)
{
    foo(y); // compatible type -- compiles
}

C23 Library Features

Floating-point formatting functions

Converts floating-point values to byte strings. Convert floats, doubles, and long doubles using strfromf, strfromd, and strfroml respectively.

char buf[BUFFER_SIZE] = {};
strfromf(&buf, BUFFER_SIZE, "%f", 123.0f);

fscanf and fprintf format specifiers

Introduces format specifiers and modifiers for fscanf() and fprintf() function families:

• %wN, %wfN modifiers for uintN_t, intN_t, uint_fastN_t, and int_fastN_t types.
• H, D, DD modifiers for _Decimal32, _Decimal64, and _Decimal128 types.
• b, B specifiers for converting unsigned int to unsigned binary.

uint64_t num = 1234;
printf("A uint64_t number: %w64u\n");
printf("A uint64_t number in binary: %w64b\n");

memset_explicit

Copies the given value into each of the first count characters of the object pointed to by dest. Unlike memset, this function cannot be optimized away by the compiler; therefore, this function is guaranteed to perform the memory write.

char str[] = "foo";
memset_explicit(str, 0, sizeof(str));

unreachable macro

Provides a standard macro for (historically) compiler-specific macros for denoting an unreachable area of the code.

if (1 > 0) ...
else unreachable();

memccpy

Copies bytes from source to destination stopping after either the terminating byte is found (and also copied) or the given number of bytes is copied.

// Assume `src` is populated elsewhere.
char dest[MAX_LEN] = {};
// Copy to `dest` until either the NUL terminator (zero) byte
// is found or we are at `MAX_LEN - 1` bytes copied.
memccpy(dest, src, 0, MAX_LEN - 1);

strdup and strndup

These functions duplicate a given string by allocating a buffer and copying. The returned string is always null-terminated. Be sure to free the memory returned by thhese functions after use.

const char* src = "foobarbaz";

char* src2 = strdup(src); // "foobarbaz"
free(src2);

char* src3 = strndup(src, 3); // "foo"
free(src3);

gmtime_r and localtime_r

Works as gmtime and localtime does but uses a given storage buffer for the result. Returns a NULL pointer on error.

time_t t = time(NULL);

struct tm buf;
struct* tm ret = gmtime_r(&t, &buf);
// ...
struct* tm ret2 = localtime_r(&t, &buf);

timespec_getres

Stores the resolution of time provided by the given base. Returns zero on failure.

struct timespec ts;
const int res = timespec_getres(&ts, TIME_UTC);
if (res == TIME_UTC)
{
    // ...
}

C11 Language Features

Generic selection

Using the _Generic keyword, select an expression based on the type of a given controlling expression. The format of _Generic is as follows:

_Generic( controlling-expression, T1: E1, ... )

where:

• controlling-expresion is an expression that yields a type present in the type list (T1: E1, ...).
• T1 is a type.
• E1 is an expression.

Optionally, specifying default in a type list, will match any controlling expression type.

#define abs(expr) _Generic((expr), \
    int: abs(expr), \
    long int: labs(expr), \
    float: fabs(expr), \
    /* ... */ \
    /* Don't call abs for unsigned types, etc. */ \
    default: expr \
)

printf("%d %li %f %d\n", abs(-123), abs(-123l), abs(-3.14f), abs(123u));
// prints: 123 123 3.14 99 123

alignof

Queries the alignment requirement of a given type using the _Alignof keyword or alignof convenience macro defined in <stdalign.h>.

// On a 64-bit x86 machine:
alignof(char); // == 1
alignof(int); // == 4
alignof(int*); // == 8

// Queries alignment of array members.
alignof(int[5]); // == 4

struct foo { char a; char b; };
alignof(struct foo); // == 1

// 3 bytes of padding between `a` and `b`.
struct bar { char a; int b; };
alignof(struct bar); // == 4

max_align_t is a type whos alignment is as large as that of every scalar type.

alignas

Sets the alignment of the given object using the _Alignas keyword, or alignof macro in <stdalign.h>.

struct sse_t
{
    // Aligns `sse_data` on a 16 byte boundary.
    alignas(16) float sse_data[4];
};

struct buffer
{
    // Align `buffer` to the same alignment boundary as an `int`.
    alignas(int) char buf[sizeof(int)];
};

max_align_t is a type whos alignment is as large as that of every scalar type.

static_assert

Compile-time assertion either using the _Static_assert keyword or the static_assert keyword macro defined in assert.h.

static_assert(sizeof(int) == sizeof(char), "`int` and `char` sizes do not match!");

See: new keywords (C23).

noreturn

Specifies the function does not return. If the function returns, either by returning via return or reaching the end of the function body, its behaviour is undefined. noreturn is a keyword macro defined in <stdnoreturn.h>.

noreturn void foo()
{
    exit(0);
}

See: attributes.

Wide string literals

Create 16-bit or 32-bit wide string literals and character constants.

char16_t c1 = u'貓';
char32_t c2 = U'🍌';

char16_t s1[] = u"a猫🍌"; // => [0x0061, 0x732B, 0xD83C, 0xDF4C, 0x0000]
char32_t s2[] = U"a猫🍌"; // => [0x00000061, 0x0000732B, 0x0001F34C, 0x00000000]

See: char32_t, char16_t, unicode string literals.

Anonymous structs and unions

Allows unnamed (i.e. anonymous) structs or unions. Every member of an anonymous union is considered to be a member of the enclosing struct or union, keeping the layout intact. This applies recursively if the enclosing struct or union is also anonymous.

struct v
{
   union // anonymous union
   {
       int a;
       long b;
   };
   int c;
} v;
 
v.a = 1;
v.b = 2;
v.c = 3;

printf("%d %ld %d", v.a, v.b, v.c); // prints "2 2 3"

union v
{
   struct // anonymous struct
   {
       int a;
       long b;
   };
   int c;
} v;
 
v.a = 1;
v.b = 2;
v.c = 3;

printf("%d %ld %d", v.a, v.b, v.c); // prints "3 2 3"

C11 Library Features

Bounds checking

Many standard library functions now have bounds-checking versions of existing functions. You can specify a callback function using set_constraint_handler_s as a constraint handler when a function fails its boundary check. The standard library includes two constraint handlers: abort_handler_s which writes to stderr and terminates the program; and ignore_handler_s ignores the violation and continue the program.

Standard library functions with bounds-checked equivalents will be appended with _s. For example, some bounds-checked versions include fopen_s, get_s, asctime_s, etc.

Example of a custom constraint handler with get_s:

void custom_handler_s(const char* restrict msg, void* restrict ptr, errno_t error)
{
    fprintf(stderr, "ERROR: %s\n", msg);
    abort();
}

set_constraint_handler_s(custom_handler_s);
char buffer[BUFFER_SIZE];
gets_s(buffer, BUFFER_SIZE);
printf("Entered: %s\n", buffer);

If a user enters a string greater than or equal to BUFFER_SIZE the custom constraint handler will be called.

timespec_get

Populates a struct timespec object with the current time given a time base.

struct timespec ts;
timespec_get(&ts, TIME_UTC);
char buff[BUFFER_SIZE];
strftime(buff, sizeof(buff), "%D %T", gmtime(&ts.tv_sec));
printf("Current time: %s.%09ld UTC\n", buff, ts.tv_nsec);

aligned_malloc

Allocates bytes of storage whose alignment is specified.

// Allocate at a 256-byte alignment.
int* p = aligned_alloc(256, sizeof(int));
// ...
free(p);

char32_t

An unsigned integer type for holding 32-bit wide characters.

See: wide string literals, char16_t, unicode string literals.

char16_t

An unsigned integer type for holding 16-bit wide characters.

See: wide string literals, char32_t, unicode string literals.

<stdatomic.h>

Atomic types

The _Atomic type specifier/qualifier is used to denote that a variable is an atomic type. An atomic type is treated differently than non-atomic types because access to atomic types provide freedom from data races. For example, the following code was compiled on x86 clang 4.0.0:

_Atomic int a;
a = 1; // mov     ecx, 1
       // xchg    dword ptr [rbp - 8], ecx

int b;
b = 1; // mov     dword ptr [rbp - 8], 1

Notice that the atomic type used a mov then an xchg instruction to assign the value 1 to a, while b uses a single mov. Other operations may include +=, ++, and so on... Ordinary read and write access to atomic types are sequentially-consistent.

See this StackOverflow post for more information.

A variety of typedefs exist which expand to _Atomic T. For example, atomic_int is typedef'd to _Atomic int. See this list for a more complete list.

Atomic types can also be tested for lock-freedom by using the atomic_is_lock_free function or the various ATOMIC_*_LOCK_FREE macro constants. Since atomic_flags are guaranteed to be lock-free, the following example will always return 1:

atomic_flag flag = ATOMIC_FLAG_INIT;
atomic_is_lock_free(&flag); // == 1

Atomic flags

An atomic_flag is a lock-free (guaranteed), atomic boolean type representing a flag. An example of an atomic operation which uses a flag is test-and-set. Initialize atomic_flag objects using the ATOMIC_FLAG_INIT macro.

A simple spinlock implementation using atomic_flag as the "spin value":

struct spinlock
{
    // false - lock is free, true - lock is taken
    atomic_flag flag;
};

void acquire_spinlock(struct spinlock* lock)
{
    // `atomic_flag_test_and_set` returns the value of the flag.
    // We keep spinning until the lock is free (value of the flag is `false`).
    while (atomic_flag_test_and_set(&lock->flag) == true);
}

void release_spinlock(struct spinlock* lock)
{
    atomic_flag_clear(&lock->flag);
}

void print_foo(void* lk)
{
    struct spinlock* lock = (struct spinlock*) lk;
    acquire_spinlock(lock);
    printf("foo\n");
    release_spinlock(lock);
}

void print_bar(void* lk)
{
    struct spinlock* lock = (struct spinlock*) lk;
    acquire_spinlock(lock);
    printf("bar\n");
    release_spinlock(lock);
}

struct spinlock lock = { .flag = ATOMIC_FLAG_INIT };

// In Thread A:
print_foo(&lock);
// ==============
// In Thread B:
print_bar(&lock);

Atomic variables

An atomic variable is a variable which is declared as an atomic type. Atomic variables are meant to be used with the atomic operations which operate on the values held by these variables (except for the flag operations, ie. atomic_flag_test_and_set). Unlike atomic_flag, these are not guaranteed to be lock-free. Initialize atomic variables using the ATOMIC_VAR_INIT macro or atomic_init if it has been default-constructed already.

The C11 Atomics library provides many additional atomic operations, memory fences, and allows specifying memory orderings for atomic operations.

struct spinlock
{
    // false - lock is free, true - lock is taken
    atomic_bool flag;
};

void acquire_spinlock(struct spinlock* lock)
{
    bool expected = false;
    // `atomic_compare_exchange_weak` returns `false` when the value of the
    // flag is not equal to `desired`. We keep spinning until the lock is
    // free (value of the flag is `false`).
    while (atomic_compare_exchange_weak(&lock->flag, &expected, true) == false)
    {
        // `expected` will get set to the value of the flag for every call.
        // Reset it since we always "expect" `false`.
        expected = false;
    }
}

void release_spinlock(struct spinlock* lock)
{
    atomic_store(&lock->flag, false);
}

void print_foo(void* lk)
{
    struct spinlock* lock = (struct spinlock*) lk;
    acquire_spinlock(lock);
    printf("foo\n");
    release_spinlock(lock);
}

void print_bar(void* lk)
{
    struct spinlock* lock = (struct spinlock*) lk;
    acquire_spinlock(lock);
    printf("bar\n");
    release_spinlock(lock);
}

struct spinlock lock;
atomic_init(&lock.flag, false);

// In Thread A:
print_foo(&lock);
// ==============
// In Thread B:
print_bar(&lock);

<threads.h>

C11 provides an OS-agnostic thread library supporting thread creation, mutexes, and condition variables. Threading library is located in <threads.h>. As of September 2023, it is poorly supported in most major compilers.

Threads

Creates a thread and executes print_n in the new thread.

void print_n(int* n)
{
    printf("%d\n", *n); // prints "123"
}

int n = 123;
thrd_t thr;
const int ret = thrd_create(
    &thr, (thrd_start_t) &print_n, (void*) &n);
thrd_join(thr, NULL);

Mutexes

C11 provides mutexes that support:

• Timed locking - Given a timespec object, blocks the current thread until the mutex is locked or until it times out.
• Recursive locking - Supports recursive locking (e.g. re-locking in recursive functions). Recursive timed locks can also be created, capturing both these operations.

Specify which type of mutex to be created by passing any of mtx_plain, mtx_timed, mtx_recursive, or any combination of these to mtx_init's type parameter.

mtx_t mutex;
const int ret = mtx_init(&mutex, mtx_plain);

// In each thread: =============
mtx_lock(&mutex);
// Critical section
mtx_unlock(&mutex);
// ==============================

mtx_destroy(&mutex);

Mutexes must be cleaned up by calling mtx_destroy.

Condition variables

C11 provides condition variables as part of its concurrency library. Condition variables will block when waiting, support timed-waiting, and can be signalled and broadcasted. Note that spurious wakeups may occur.

Condition variables must be cleaned up by calling cnd_destroy.

// Assume remove_from_queue, add_to_queue, and
// can_consume are defined.

mtx_t mutex;
// Initialize mutex...

cnd_t cond;
const int ret = cnd_init(&cond);

// In a consumer thread: =====
// Check in a loop due to spurious wakeups.
while (!can_consume && cnd_wait(&cond, &mutex));
remove_from_queue();
// =========================

// In a producer thread: =====
add_to_queue();
can_consume = true;
cnd_signal(&cond);
// ===========================

cnd_destroy(&cond);

Quick exiting

Causes program termination to occur where clean termination is not possible or impractical; for example if cooperative cancellation between different threads is not possible or cancellation order is unachievable. This would have resulted in some threads trying to access static-duration objects while they are/have been destructed.

Quick exiting allows applications to register handlers to be called on exit while being able to access static duration objects (unlike atexit). Quick exiting does not execute static-duration object destructors in order to achieve this.

Therefore, quick exiting can be thought of as being "in between" choosing to normally terminate a program, and abnormally terminating using abort.

void f()
{
    // called first
}
 
void g()
{
    // called second
}
 
int main()
{
    at_quick_exit(f);
    at_quick_exit(g);
    quick_exit(0);
}

Exclusive mode file opening

File access mode flag x can optionally be appended to w or w+ specifiers. This flag forces the function to fail if the file exists, instead of overwriting it.

FILE* fp = fopen(fname, "w+x");
if (!fp)
{
    // File either exists or there was an error
}
else
{
    // File opened successfully.
    fclose(fp);
}

Author

Anthony Calandra

License

MIT