Skip to content

moehriegitt/vastringify

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

58 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Type-Safe Printf For C

This uses macro magic, compound literals, and _Generic to take printf() to the next level: type-safe printing, printing into compound literal char arrays, easy UTF-8, -16, and -32, with good error handling.

The goal is to be safe by removing the need for function varargs.

The usual C printf formatting syntax is used, with some restrictions and quite a few extensions.

This let's you mix UTF-8, -16, -32 strings seamlessly in input and output strings, without manual string format conversions, and without using different format specifiers or print function names.

This liberates you from thinking about %u vs. %lu vs. %llu vs. %zu, even in portable code with different integer types. . With this library, the compiler chooses the right function for your parameter and they all print fine with ~s, like also strings and pointers do.

'Type-safe' in this context does not mean that you get more compile errors or warnings, but that you cannot make a mistake, and you do not need the format string to specify the argument type. Format strings with this library do not need compile-time checking to be safe, because the compiler chooses the right formatting function for each parameter. You cannot pass the wrong size parameter and crash, because ... is avoided.

Examples

  • va_printf("~s ~s ~s", 65, (long long)65, "65") prints 65 65 65
  • va_printf("~c ~c ~c", 65, (long long)65, "65") prints A A 65
  • va_printf("~x ~x ~x", 65, (long long)65, "65") prints 41 41 65
  • va_printf("~p ~p ~p", 65, (long long)65, "65") prints 0x41 0x41 0x8838abc3932 (the pointer value of the string)
  • va_lprintf("~p", 65) returns 4, the length of 0x41
  • va_nprintf(10, "~p", 65) returns "0x41", the pointer to a compound literal (char[10]){} that was printed into
  • va_printf("~t x = ~=qzs", u"fo\020o") prints char16_t* x = u"fo\no"

Compatibility

This library requires at least a C11 compiler (for _Generic, char16_t, char32_t), and it uses a few gcc extensions that are also understood by Clang and a few other compilers (({...}), ,##__VA_ARGS__, __typeof__, __attribute__).

Synopsis

In the following, Char may be char, char16_t, or char32_t:

#include <va_print/file.h>

void
va_fprintf(FILE *f, Char const *format, ...);

void
va_ufprintf(FILE *f, Char const *format, ...);

void
va_Ufprintf(FILE *f, Char const *format, ...);

void
va_printf(Char const *format, ...);

va_stream_file_t
VA_STREAM_FILE(FILE *f);


#include <va_print/char.h>

Char *
va_snprintf(Char *s, size_t n, Char const *format, ...);

Char *
va_sprintf(Char s[], Char const *format, ...);

char *
va_nprintf(size_t n, Char const *format, ...);

char16_t *
va_unprintf(size_t n, Char const *format, ...);

char32_t *
va_Unprintf(size_t n, Char const *format, ...);

CharType *
va_gnprintf(CharType, size_t n, Char const *format, ...);

size_t
va_zprintf(Char const *format, ...);

size_t
va_uzprintf(Char const *format, ...);

size_t
va_Uzprintf(Char const *format, ...);

size_t
va_gzprintf(CharType, Char const *format, ...);

va_stream_charp_t
VA_STREAM_CHAR_P(Char *s, size_t n);


#include <va_print/alloc.h>

char *
va_axprintf(void *(*alloc)(void *, size_t, size_t), Char const *, ...);

char16_t *
va_uaxprintf(void *(*alloc)(void *, size_t, size_t), Char const *, ...);

char32_t *
va_Uaxprintf(void *(*alloc)(void *, size_t, size_t), Char const *, ...);

char *
va_asprintf(Char const *, ...);

char16_t *
va_uasprintf(Char const *, ...);

char32_t *
va_Uasprintf(Char const *, ...);

va_stream_vec_t
VA_STREAM_VEC(void *(*alloc)(void *, size_t, size_t));

va_stream_vec16_t
VA_STREAM_VEC16(void *(*alloc)(void *, size_t, size_t));

va_stream_vec32_t
VA_STREAM_VEC32(void *(*alloc)(void *, size_t, size_t));

void *
va_alloc(void *data, size_t nmemb, size_t size);


#include <va_print/fd.h>

void
va_dprintf(int fd, Char const *format, ...);

void
va_udprintf(int fd, Char const *format, ...);

void
va_Udprintf(int fd, Char const *format, ...);

va_stream_file_t
VA_STREAM_FD(int fd);


#include <va_print/len.h>

size_t
va_lprintf(Char const *format, ...);

va_stream_len_t
VA_STREAM_LEN();


#include <va_print/core.h>

va_stream_...t *
va_xprintf(va_stream_...t *s, Char const *format, ...);

void
va_iprintf(va_stream_...t *s, Char const *format, ...);

void
va_pprintf(va_stream_vtab_t *v, Char const *format, ...);

unsigned
va_stream_get_error(va_stream_...t const *s);

extern
char const *va_strerror(unsigned error_code);


#include <va_print/base.h>

typedef struct { ... } va_stream_t;

typedef struct { ... } va_stream_vtab_t;

typedef struct { unsigned code; } va_error_t;
#define VA_E_OK     ...
#define VA_E_NULL   ...
#define VA_E_DECODE ...
#define VA_E_ENCODE ...
#define VA_E_TRUNC  ...
#define VA_E_FORMAT ...
#define VA_E_ARGC   ...

va_stream_t
VA_STREAM(va_stream_vtab_t const *vtab);

#define VA_U_REPLACEMENT 0xfffd
#define VA_U_BOM         0xfeff
#define VA_U_SURR_MIN    0xd800
#define VA_U_SURR_MAX    0xdfff
#define VA_U_MAX         0x0010ffff
#define VA_U_MAXMAX      0x00ffffff

#define va_countof(A)    (sizeof(A)/sizeof((A)[0]))

Description

This library provides a type-safe printing mechanism to print any kind of string of base type char, char16_t, or char32_t, or any integer or pointer into a new string, an array, or a file.

The library also provides functions for user-defined output streams that can print into any other kind of stream.

The arguments to the formatted print are passed into a _Generic() macro instead of '...' and the resulting function call is thus type-safe based on the actual argument type, and cannot crash due to a wrong format specifier.

The format specifiers in this printing mechanism serve to define which output format should be used, as they are not needed for type information. The format specifier "~s" can be used as a generic 'default' output format.

Format Specifiers

The format is string is decoded and then output as is into the output stream where it is encoded, except for sequences of ~, which are interpreted as a format specifier.

For each format specifier in the format string, 0, 1, or more arguments are read (how many is specified below). If there are more arguments than what is needed for the format specifiers, the rest of the argumnets are ignored and the VA_E_ARGC stream error is set.

If fewer arguments are given than needed for the format string, the rest of the format specifiers print empty and the VA_E_ARGC stream error is set.

A format specifier begins with ~, and what follows is similar to C:

  • a list of flag characters
  • a width specifier
  • a precision specifier
  • a list of integer mask and quotation specifiers
  • a conversion letter

~ is used instead of % to avoid confusion in source code that uses both this library and the standard C printf.

Flags

Generally, specifying multiple identical flags like ~008s is reserved for future use and should be avoided. It is unspecified how the current library handles such format strings.

  • # print in alternative form. For numeric format, a prefix to designate the base is prefixed to the value except to 0:

    • for o and base 8, 0 is prefixed
    • for b and base 2, 0b is prefixed,
    • for B and base 2, 0B is prefixed,
    • for x and base 16, 0x is prefixed,
    • for X and base 16, 0X is prefixed,
    • for e and base 32, 0e is prefixed,
    • for E and base 32, 0E is prefixed.

    In ~#p, the # flag switches off the implicit # that is contained in p, e.g., does not print the base prefix.

    For quoted strings, # inhibits printing of delimiting quotes.

  • 0 pads numerics with zero 0 on the left rather than with a space character . If a precision is given, this is ignored.

    For C and JSON quotation, this selects to quote non-US-ASCII characters using \u and \U instead of printing them in output encoding.

  • - selects to left flush instead of the default right flush.

  • (a space character U+0020) selects that a space is printed in front of positive signed integers. Nothing is printed if the precision is 0 and the value is 0 (this is different compared to the behaviour of C's printf).

  • + selects that a + is printed in front of positive signed integers and zero. Nothing is printed if the precision is 0 and the value is 0 (this is different compared to the behaviour of C's printf).

  • = specifies that the last value is printed again using this new format specifier. This is meager replacement for the $ position specifiers that are not implemented in this library.

A width is either a decimal integer, or a *. The * selects that the width is taken from the next function parameter. If fewer code points result from the conversion, the output is padded with white space up the width. A negative width is intepreted as a - flag followed by a positive width.

A precision is specified by a . (period) followed by either a decimal integer or a *. The * selects that the width is taken from the next function parameter. If the precision is just ., it is interpreted as zero. The precision defines the minimum number of digits in numeric conversions. For strings, this is the maximum number of raw code units read from the input string (not the number of converted code points, but the low-level number of elements in the string, so that non-NUL terminated arrays can be printed with their size passed as precision, even with multi-byte/multi-word encodings stored inside. Alternatively, there is va_span_t for a string prefix parameter type.

The input decoder will not read incomplete encodings at the end of limited strings, but will stop before. If a pointer to a string pointer is passed, then the pointer will be updated so that it points to the next character, i.e., the one after the last one that was read.

Integer Mask and Quotation Specifiers

Generally, specifying multiple identical mask and quotation specifiers or more than listed in the following list, like ~zzu or ~hhhx, is reserved for future use and should be avoided. It is unspecified how the current library handles such format strings.

  • h applies the mask 0xffff to an integer, then zero extends unsigned values, or sign extends signed values. E.g., va_printf("~#hx",0xabcdef) prints -0x3211.

  • hh applies the mask 0xff to an integer, then zero extends unsigned values, or sign extends signed values. E.g., va_printf("~hhX",0xabcdU) prints CD.

  • z reinterprets a signed integer as unsigned (mnemonic: zero extension). z is implicit in formats u (and U). E.g., va_printf("~hhu", -1) prints 255.

  • q selects C quotation for strings and char format. There is a separate section below to explain this.

  • Q selects JSON quotation for strings and char format. There is a separate section below to explain this.

  • k selects Bourne or Korn shell quotation. There is a separate section below to explain this.

  • K additional custom quotation

  • qq, QQ, kk more additional custom quotations

Note that most of the usual length specifiers (l, ll, etc.) known from C make no sense and are not recognised (nor ignored), because type casting control in varargs is not needed here due to the type-safety.

Conversions

The format specifiers is terminated by a single conversion character from the following list.

  • s prints anything in default notation (mnemonic: 'standard'). s is used by standard C printf for strings, and CommonLisp uses ~s for 'standard' format.

  • o selects octal integer notation for numeric printing (including pointers).

  • d or i selects decimal integer notation for numeric printing (including pointers).

  • u is equivalent to zd, i.e., prints a signed integer as unsigned in decimal notation. This implicitly sets the z option, which also affects quoted string printing, so ~qu prints strings like ~qzs.

  • x or X selects hexadecimal integer notation for numeric printing (including pointers). x uses lower case digits, X upper case. Note that this also prints signed numbers with a - if appropriate: va_printf("~#x", -5) prints -0x5. There's the z flag to print signed integers as unsigned.

  • b or B selects binary integer notation for numeric printing (including pointers). b uses lower case prefix, B uses upper case. The difference is only visible with the # flag.

  • e or E selects Base32 notation using the digits 'a'..'z','2'..'7'. e uses lower case digits and prefix, E uses upper case.

  • p prints like x, toggles the # flag, and for any strings, prints the pointer value instead of the contents. Note that it also prints signed numbers: va_print("~p",-5) prints -0x5.

    Note that ~#p prints pointers like ~x and ~p prints like ~#x, i.e., the # flag is toggled.

  • c prints integers (but not pointers) as characters, like a one-element string. Note that the NUL character is not printed, but behaves like an empty string, unless quotation is used. For string quotation where hexadecimals are printed, this uses lower case characters.

  • a, f, and g print like s, but will print differently when floating point support is added.

  • t prints the argument type in C syntax: int8_t..int64_t, uint8_t..uint64_t, char*, char16_t*, char32_t*, void*. Note that va_error_t* arguments never print, and never consume a ~ format, but always just return the stream error.

  • m prints the (error) status of the referenced item. This is for custom printers, and it is encoded as VA_MODE_STAT. The pre-defined data types have no error values and thus no error is printed. Retrieving the stream error using va_error_t is another topic, as it does not print anything. The letter m is inspired by the GNU extension %m, which prints strerror(errno), which this library does not support natively (to avoid depending on <errno.h>).

  • ~ prints ~ characters. By default, one is printed. The width gives the number of tildes, e.g. ~5~ prints ~~~~~, and ~0~ prints nothing. ~*~ reads the width from an argument. The use of precision and justification flags is reserved for future use, and it is unspecified how the library handles them.

  • any letter mentioned above in lowercase only also exists in uppercase, and then prints whatever is usually printed in lowercase in uppercase, like like hexadecimal digits or numeric base prefixes like 0B or 0X.

  • any format character not mentioned above is reserved for future use. If used, the argument is skipped, and the VA_E_FORMAT error is set in the stream.

  • any combination of format character and type not mentioned above prints in default notation.

Parameter Types

The following function parameter types are recognised. Note that enums are not listed here, because of the weak type system of C, where enum constants have type 'int' and enum types match type 'int' in _Generic.

  • int, unsigned, char, signed char, unsigned char, short, unsigned short, long, unsigned long, long long, unsigned long long: these are integer and are printed in unsigned or signed decimal integer notation by default.

    This means that char, char16_t, and char32_t all print in numeric format by default, not in character format, as they are not distinct types. For interpreting them as a 1-element Unicode codepoint string, c format should be used.

    Also note that character constants like 'a' have type int in C and print numerically by default.

  • _Bool (or bool with <stdbool.h>): prints a boolean type. This is the only enum in C that does not match an int type in _Generic, so it is supported. Note that true and false still have type int and not bool, so only variables of type bool will print in boolean mode. This prints true or false by default or 0 and 1 if a numeric format is used: d, u, x, o, b.

  • char *, char const *: 8-bit character strings or arrays. They print as is by default.

    The default string encoding is UTF-8, It can be reset to a user encoding by #defining va_char_p_decode. Also see the section on encoding below.

    Unquoted, NULL prints empty and sets the VA_E_NULL error. Also see the section on quotation below.

    If the input decoder encounters an incomplete UTF-8 sequence right in front of the terminating NUL character, it will return the bytes of the incomplete sequence as decoding errors. However, if a precision, i.e., maximum string size is specified, it will stop decoding before the incomplete sequence without a decoding error. This way, strings can be printed in chunks without errors. Using a pointer to a string, the final string position, i.e., the first byte of the incomplete sequence at the end, can be queried in order to start a new string chunk with the incomplete sequence at the beginning, followed by, hopefully, the missing bytes of the UTF-8 sequence from the next chunk.

  • char16_t *, char16_t const *: 16-bit character strings or arrays. The default encoding is UTF-16, which can be switched using va_char16_p_decode.

    Unquoted, NULL prints empty and sets the VA_E_NULL error.

    If the input decoder encounters an high UTF-16 surrogate right in front of the terminating NUL character, it will return the high surrogate as a decoding error. However, if a precision, i.e., maximum string size is specified, it will stop decoding before the high surrogate. This can be used for chunked printing like with UTF-8.

  • char32_t *, char32_t const *: 32-bit character strings or arrays. The default encoding is UTF-32, which can be switched using va_char32_p_decode.

    Unquoted, NULL prints empty and sets the VA_E_NULL error.

  • Char **, Char const **: pointers to pointers to characters, i.e., pointers to string, will print the string and then update the pointer to point to the code unit just behind the last one that was read from the string. With no precision given in the format, they will point to the terminating NUL character. When these parameters are printed multiple times using the = flag, the string will be reset each time and the updated value will correspond to the end position during the last print of the string.

  • va_error_t*: this retrieves the error code from the stream and writes it into the passed struct. This can be used to check for encoding or decoding errors, out of memory conditions, or hitting the end of the output array. NULL must not be passed as a pointer.

  • va_read_iter_t*: this is an internal type to read from strings. There are quite a few constraints on how to define a proper va_read_iter_t, which are not all documented here.

  • va_span_t*: this is a length delimited string for printing non-NUL terminated strings or prefixes of strings. It is an alternative way to specify the string size in the argument directly instead of using the precision in the format specifier. When strings are specified this way, embedded U+0000 (NUL) characters are passed down, so quotation may print them as \u0000 or \000 or similar. NUL characters are never printed verbatim into the output stream, however, because the output stream is assumed to be text.

  • va_span16_t*: the same as va_span_t, but for char16_t strings.

  • va_span32_t*: the same as va_span_t, but for char32_t strings.

  • va_print_t*: user-defined printer for a value of an arbitrary type (there is a separate chapter on this, below).

  • anything else: is tried to be converted to a pointer and printed in hexadecimal encoding by default, i.e., in ~x format.

Library Modules

Printing Into Fixed Size Arrays

#include <va_print/char.h>

To print into an string of characters up to a given number of elements in the array, the following function can be used, and it returns the pointer to the string. The resulting string is always NUL terminated, i.e., the maximum string length is one less than the passed element count.

char s[20];
char *t = va_snprintf(s, sizeof(s), "foo~s", 5);
assert(s == t);

The target buffer's type may be 8, 16, or 32 bit characters -- no need to use a different function. For anything but char buffers, use va_countof() for the array size, so that the number of elements, not the number of bytes, is used as the string size.

char16_t s1[20];
char16_t *t1 = va_snprintf(s1, va_countof(s1), "foo~s", 5);

char32_t s2[20];
char32_t *t2 = va_snprintf(s2, va_countof(s2), "foo~s", 5);

The va_countof() values is inferred when using va_sprintf. This is different from the standard C sprintf function, which unsafely assumes a sufficiently large string -- you cannot express that with this library, but you'd have to use snprintf with a large size instead.

char s[20];
char *t = va_sprintf(s, "foo~s", 5);

It is possible to print into a compound literal of a given size and return the pointer to that string. In this case, no character array can be used to infer the string type, so it is encoded in the function name. There are functions for char, char16_t, and char32_t strings, as well as a generic version that takes as first argument the string character type.

char     *t1a = va_nprintf (20, "foo~s", 5);
char16_t *t2a = va_unprintf(20, "foo~s", 5);
char32_t *t3a = va_Unprintf(20, "foo~s", 5);
char     *t1b = va_gnprintf(char, 20, "foo~s", 5);
char16_t *t2b = va_gnprintf(char16_t, 20, "foo~s", 5);
char32_t *t3b = va_gnprintf(char32_t, 20, "foo~s", 5);

The stream for printing can also be generated separately and then used for iterative printing using va_iprintf. The stream makes sure not to print past the end of the char array.

char buff[20];
va_stream_char_p_t stream = VA_STREAM_CHAR_P(buff, va_countof(buff));
va_iprintf(&stream, "foo");
va_iprintf(&stream, "bar ~u", 55);
va_iprintf(&stream, "longer than the string, will be cropped");
...

There is a stream.pos counter for the current write index in the array, i.e., the string length of the encoded byte sequence. pos increments up to stream.size-1, but no further (note that size==0 is an illegal configuration, because then there is no space for NUL terminating the string).

Creating a va_stream_char_p_t with the buffer equal to NULL is explicitly allowed. Putting the bytes into a char array will then be inhibited. The printer still increments stream.pos up to stream.size-1, so by this, strnlen() functionality can be implemented on the resulting string. This is exactly how va_zprintf (mnemonic: siZe) works.

To determine whether the stream was truncated, i.e., whether the buffer was too small for the print result, the stream's error code can be checked for the VA_E_TRUNC error code value after printing is done.

...
va_error_t e;
va_iprintf(&stream, "", &e);
if (e.code != VA_E_OK) {
    /* ... some stream error occurred ... */
}
...

Alternatively, if you have a stream anyway, there is va_stream_get_error() that returns the stream's error code. There is also va_strerror() to get the enum value as a string.

...
unsigned ec = va_stream_get_error(&stream);
if (ec != VA_E_OK) {
    va_fprintf(stderr, "ERROR: found ec=~s\n", va_strerror(ec));
    /* ... more error handling */
}

Note that there is no %n equivalent format specifier for reading the printed length; use stream.pos instead. Or use va_zprintf to compute the needed array size.

There is also va_lprintf to count the length of the string, i.e., the number of codepoints written, instead of the number of encoded bytes.

Printing Into Growing Vectors

#include <va_print/alloc.h>

It is possible to print into a string that is allocated grows using malloc():

char *c = va_asprintf("foo~s", msg);
...
free(c);

Here, the va_alloc() function is implicitly used to allocate, possibly reallocate while printing, and possibly freeing the string in case of an out-of-memory error.

For char16_t* and char32_t* target strings, there are va_uasprintf and va_Uasprintf, resp.

va_alloc is a wrapper around realloc and free. Any compatible function with the same prototype can be used instead.

A user defined allocation function can be supplied by using the va_axprintf function, which is just like va_asprintf, but takes the allocator function of type void *(void *, size_t nmemb, size_t size) as parameter, which is used for allocation, reallocation, and freeing (with nmemb==0). For char strings, the function is invoked with size==1.

char *c = va_axprintf(va_alloc, "foo~s", msg);
...
free(c);

For char16_t and char32_t output strings, there is va_uaxprintf and va_Uaxprintf, resp. The allocator function will then be invoked with a size==2 for char16_t and size==4 for char32_t.

It is also possible to create a stream for iterative printing.

va_stream_vec_t stream = VA_STREAM_VEC(va_alloc);
va_iprintf(&stream, "foo");
va_iprintf(&stream, "bar ~u", 55);

For 16 and 32 bit chars, there is va_stream_vec16_t plus VA_STREAM_VEC16 and va_stream_vec32_t plus VA_STREAM_VEC32.

Printing Into Files

#include <va_print/file.h>

To print into FILE* files, there is va_fprintf, which returns nothing.

va_fprintf(stderr, "foo~s", msg);

There is also va_printf that prints into stdout, and it also returns nothing.

va_printf("foo~s", msg);

To write char16_t or char32_t streams into files, the encodings UTF-16BE and UTF-32BE are used by default. Functions for this are called va_ufprintf and va_Ufprintf, resp.

For files, there is a stream type va_stream_file_t that can be constructed using VA_STREAM_FILE, e.g., to iteratively print.

va_stream_file_t stream = VA_STREAM_FILE(stderr);
va_iprintf(&stream, "foo");
va_iprintf(&stream, "bar ~u", 55);
va_iprintf(&stream, "longer than the string, will be cropped");
...

The 16-bit and 32-bit versions use the same stream type, and the constructors are called VA_STREAM_FILE16 and VA_STREAM_FILE32, resp. `

Printing Into Raw File Descriptors

#include <va_print/fd.h>

To print into int typed file descriptors, there is va_dprintf, which returns nothing.

va_dprintf(2, "foo~s", msg);

To write char16_t or char32_t streams into files, the encodings UTF-16BE and UTF-32BE are used by default. Functions for this are called va_udprintf and va_Udprintf, resp.

For file descriptors, there is a stream type va_stream_fd_t that can be constructed using VA_STREAM_FD, e.g., to iteratively print.

va_stream_fd_t stream = VA_STREAM_FD(2);
va_iprintf(&stream, "foo");
va_iprintf(&stream, "bar ~u", 55);
va_iprintf(&stream, "longer than the string, will be cropped");
...

The 16-bit and 32-bit versions use the same stream type, and the constructors are called VA_STREAM_FD16 and VA_STREAM_FD32, resp. `

Printing non-NUL Terminated Strings

One way to print non-NUL terminated strings or prefixes of strings is by specifying the 'precision' in the format. It is the length in char, char16_t, or char32_t elements and not the number of extracted codepoints, exactly for this purpose.

char const *data = "abcdef";
size_t size = 3;
va_fprintf(stderr, "token=~.*qs", size, data);

This prints token="abc".

An alternative way of controlling this is to pass a pointer to va_span_t to the printer, which contains the data and size:

char const *data = "abcdef";
size_t size = 3;
va_fprintf(stderr, "token=~qs", (&(va_span_t){ size, data }));

This also prints token="abc".

Note that strings specified by their size may contain U+0000 (NUL) characters, and they are quoted accordingly, if requested:

va_fprintf(stderr, "token=~qs", (&(va_span_t){ 1, "" }));

This also prints token="\000".

There are similar types va_span16_t and va_span32_t for wide character strings.

Computing String Lengths

#include <va_print/len.h>

The function va_lprintf returns the number of codepoints printed into an output stream. This is the string length regardless of output encoding.

size_t cp_count = va_lprintf("foo~s", msg);

This is not a good function for computing array sizes -- use the va_zprint family instead.

Computing String Array Sizes

#include <va_print/char.h>

To compute the size of the array needed to store a given printed string, there is va_zprintf.

size_t n = va_zprintf("foo~s", msg);
char *s = malloc(n);
va_error_t e;
va_snprintf(s, n, "foo~s", msg, &e);
assert(e.code == VA_E_OK);

This function counts the encoded size of the needed array, i.e., it also includes the NUL character in the count, and it counts for each codepoint, how many UTF-8 (or whatever encoding is used) bytes are used for each codepoint. This function is, therefore, useful for computing array sizes that fit the printed string exactly.

For char16_t and char32_t based strings, the function is called va_uzprintf and va_Uzprintf, resp.

There is a generic version that can be passed the array element type as the first parameter.

typedef SomeCharacterType MyChar;
...
size_t n = va_gzprintf(MyChar, "foo~s", msg);
MyChar *s = malloc(sizeof(MyChar) * n);
va_snprintf(s, n, "foo~s", msg);

Unicode

Internally, this library uses 32-bit codepoints with 24-bit payload and 8-bit tags for processing strings, and by default, the payload representation is Unicode. The library tries not to interpret the payload data unless necessary, so that other encodings could in principle be used and passed through the library.

The only place the core library uses Unicode interpretation is when quoting C or JSON strings for codepoints >0x80 (e.g., when formatting with ~0qs), and if a decoding error is encountered or if the value is not valid Unicode, then it uses \ufffd to show this, because the quotation using \u or \U would otherwise be a lie.

The internal representation allows any value within 24 bits to be used for codepoints. 0 is interpreted as 'end of string' and is never printed into the output stream.

UTF-8, -16, and -32 encoders and decoders check that the Unicode constraints are met, like excluding anything above 0x10FFFF and high and low UTF-16 surrogates, and detecting decoding errors according to the Unicode recommendations and best practices. The encoder/decoder pairs usually try to pass through faulty sequences as is, if possible, e.g., reading ISO-8859-1 data from an UTF-8 ~s and printing it into an UTF-8 output stream preserves the original ISO-8859-1 byte sequence, although the intermediate steps do raise 'illegal sequence' errors.

Integers print without Unicode checks, i.e., if an integer is printed as a character using ~c, then the lower 24 bits is passed down to the output stream encoder as is. If integers larger than 0xffffff are tried to be printed with ~c, this results in a decoding error, and only the lower 24 bits are used.

Encodings

The library supports different string encodings for the format string, for input strings, and for output streams. The defaults are UTF-8, UTF-16, or UTF-32. This can be switched by setting the following #defines before including headers of this library, i.e., it cannot be switched dynamically out of the box, because this would mean that all the encoding modules would always be linked. Dynamic switching can be added by defining a new encoding that internally switches dynamically.

The following #defines switch function names:

Format String Encoding

The default is UTF-8, -16, or -32 encoding, and it can be changed by #defining before #include <va_print/...>:

#define va_char_p_format utf8
#define va_char16_p_format utf16
#define va_char32_p_format utf32

These macros are appended to an identifier to find the appropriate reader for the format string as follows:

va_char_p_read_vtab ## va_char_p_format
va_char16_p_read_vtab ## va_char16_p_format
va_char32_p_read_vtab ## va_char32_p_format

When using a different encoding than the default, it must be ensured that the corresponding vtab declarations are visible.

String Value Encoding

The default for reading string values is UTF-8, -16, or -32 encoding, for "...", u"...",and U"..." strings,resp. The default can be changed by defining one of the following macros before #include <va_print/...>:

#define va_char_p_decode utf8
#define va_char16_p_decode utf16
#define va_char32_p_decode utf32

These macros are appended to an identifier to find the appropriate reader for the string value as follows:

va_xprintf_char_p_ ## va_char_p_decode
va_xprintf_char_pp_ ## va_char_p_decode
va_xprintf_char_const_pp_ ## va_char_p_decode
va_xprintf_char16_p_ ## va_char16_p_decode
va_xprintf_char16_pp_ ## va_char16_p_decode
va_xprintf_char16_const_pp_ ## va_char16_p_decode
va_xprintf_char32_p_ ## va_char32_p_decode
va_xprintf_char32_pp_ ## va_char32_p_decode
va_xprintf_char32_const_pp_ ## va_char32_p_decode

Note that for each parameter type, a different printer function is used, so for a different encoding, three functions need to be provided. A typical such function implementation looks as follows:

va_stream_t *va_xprintf_char_p_utf8(
    va_stream_t *s,
    char const *x)
{
    va_read_iter_t iter = VA_READ_ITER(&va_char_p_read_vtab_utf8, x);
    return va_xprintf_iter(s, &iter);
}

Output Stream Encoding

For encoding strings into character arrays, the default encoding is UTF-8, UTF-16, or UTF-32, depending on the string type. To override the default, the following #defines can be set before #include <va_print/...>.

#define va_char_p_encode utf8
#define va_char16_p_encode utf16
#define va_char32_p_encode utf32

These are suffixed to find the vtab object for writing:

va_char_p_vtab_ ## va_char_p_encode
va_char16_p_vtab_ ## va_char16_p_encode
va_char32_p_vtab_ ## va_char32_p_encode

For dynamically allocated arrays, there are separate #definitions:

#define va_vec8_encode utf8
#define va_vec16_encode utf16
#define va_vec32_encode utf32

These are suffixed to find the vtab object for writing:

va_vec_vtab_ ## va_vec_encode
va_vec16_vtab_ ## va_vec16_encode
va_vec32_vtab_ ## va_vec32_encode

For FILE* output, the default encoding is UTF-8, UTF-16BE, and UTF-32BE, depending on output character width. The following #defines correspond to the encoding:

#define va_file8_encode utf8
#define va_file16_encode utf16be
#define va_file32_encode utf32be

These are suffixed to find the vtab object for writing:

va_file_vtab_ ## va_file_encode
va_file16_vtab_ ## va_file16_encode
va_file32_vtab_ ## va_file32_encode

For int file descriptor output, the default encoding is UTF-8, UTF-16BE, and UTF-32BE, depending on output character width. The following #defines correspond to the encoding:

#define va_fd8_encode utf8
#define va_fd16_encode utf16be
#define va_fd32_encode utf32be

These are suffixed to find the vtab object for writing:

va_fd_vtab_ ## va_fd_encode
va_fd16_vtab_ ## va_fd16_encode
va_fd32_vtab_ ## va_fd32_encode

Quotation

C/C++ quotation

  • q quotation option in format specifier
  • when printing integers, this is ignored
  • when printing pointers, this adds the # flag, i.e., the 0x prefix is printed
  • when printing strings, this selects C format quoted output
  • NULL strings print as NULL, and do not set the VA_E_NULL error, in contrast to unquoted printing.
  • without #, prints quotation marks, single for c and C, conversion, otherwise double.
  • with z prints the string size indicator based on the input string: empty for char, u for char16_t, and U for char32_t (and also U for 64-bit ints).
  • quotation of unprintable characters <U+0080 is done using octal quotation.
  • quotation of some characters in special notation: \t, \r, \n, \', \", \\.
  • 0 flag quotes all non-ASCII using \u or \U. Note that \x is not used, because it may not terminate, so quoting \x1 plus 1 is more complicated.
  • with 0 flag, chars that are marked as decoding errors are quoted as \ufffd, the replacement character, to avoid printing encoding errors with \u quotation, which would make the resulting string more wrong than with only the encoding errors. Without 0 flag, encoding errors are passed through if the input encoding equals output encoding, otherwise U+FFFD is encoded.
  • upper case formats use upper case letters in hexadecimals

Examples:

  • va_printf("~qs", "foo'bar") prints "foo\'bar".
  • va_printf("~qs", "foo'bar") prints "foo\'bar".
  • va_printf("~qzs", u"foo'bar") prints u"foo\'bar".
  • va_printf("~qc", 10) prints '\n'.
  • va_printf("~qzc", 10) prints U'\n'
  • va_printf("~#qc", 16) prints \020.
  • va_printf("~#0qc", 0x201c) prints \u201c.
  • va_printf("~#0qC", 0x201c) prints \u201C.
  • va_printf("~qa", (void*)18) prints 0x12 (on normal machines)
  • va_printf("~qa", 18) prints 18
  • va_printf("~0qa", u"\xd801") prints "\xfffd"

Java/JSON quotation

  • Q quotation option in format specifier
  • Like C, but always uses \u or \U and never octal
  • NULL strings print as null, and do not set the VA_E_NULL error, in contrast to unquoted printing.
  • the z flag is ignored (no u or U prefixes are printed).

Examples:

  • va_printf("~Qs", "foo'bar") prints "foo\'bar".
  • va_printf("~Qc", 10) prints '\n'.
  • va_printf("~#Qc", 16) prints \u0010.
  • va_printf("~#0Qc", 0x201c) prints \u201c.
  • va_printf("~#0QC", 0x201c) prints \u201C.
  • va_printf("~Qa", (void*)18) prints 0x12 (on normal machines)
  • va_printf("~Qa", 18) prints 18

Bourne Shell quotation

  • k quotation option in format specifier (mnemonic: Korn Shell quotation)
  • when printing integers, this is ignored
  • when printing pointers, this adds the # flag, i.e., the 0x prefix is printed
  • when printing strings, this selects Shell quoted format
  • NULL strings print as empty string, and set the VA_E_NULL error, just like unquoted printing.
  • the empty string is printed as ''
  • uses single quotes if necessary
  • without #, prints quotation marks if necessary
  • others print no quotation marks for in-string printing
  • this actually quotes nothing except the single quotation mark.
  • chars marked as decoding errors are not quoted, but passed through.

Examples:

  • va_printf("~ks", "ab") prints ab.
  • va_printf("~ks", "") prints ''.
  • va_printf("~#ks", "") prints ``.
  • va_printf("~ks", "a b") prints 'a b'.
  • va_printf("~ks", "a'b") prints 'a'\''b'.
  • va_printf("~#ks", "a'b") prints a'\''b.
  • va_printf("~ka", (void*)18) prints 0x12 (on normal machines)
  • va_printf("~ka", 18) prints 18

User-Defined Quotation

The quotation mechanism of the library can be extended by own quotation techniques. The API for this is currently preliminary and may change.

For implementing custom quotation, the impl.h header file needs to be included to get access to the internal programming API:

#include <va_print/impl.h>

There is a definition of a struct va_quotation_t which has three entries to be defined for a quotation mechanism:

  • unsigned delim[2]: the delimiter with which to quote. array index [0] is used for characters and index [1] is used for string quotation. The VA_DELIM(prefix, frontquote, backquote) macro constructs an entry. The quotation characters frontquote and backquot must be in the BMP, i.e., smaller than or equal to U+FFFF. The prefix character must be smaller than or equal to U+00FF. If it is 0xff, then the prefix is selected if the z modifier is used based on the string character type: empty for char, u for char16_t and U for char32_t.

  • bool (*check_quote)(va_stream_t *s, unsigned c): if NULL, quotation is always used. If non-NULL, this function is used on each character of the string to check whether quotation is needed. If the function returns non-false for any of the characters, then quotation is needed.

  • bool (*check_flush)(va_stream_t *s): if non-NULL, will be invoked at the end of the string quotation check (only if check_quote is non-NULL) to check again whether quotation is needed. check_quote and check_flush may use s->qctxt for storing some state, e.g., for detecting and empty string (which needs quotation in Shell quotation).

  • void (*render_quote)(va_stream_t *s, unsigned ch): for the actual quotation of a character. This must invoke one of the va_stream_render*() functions for writing the quoted representation of the character into the output stream. The renderer can store context in s->qctxt, an unsigned, during quotation rendering. It is initialised to 0 when rendering starts.

  • void (*render_flush)(va_stream_t *s): callback for the quotation to signal the end of the quoted string. Maybe NULL if not needed.

There are the following rendering functions for the render_quote method:

  • va_stream_render(s,c): prints the character verbatim into the output stream. Note that this cannot be done manually by a different function, because this function also contains the logic for counting string widths, etc.
  • va_stream_render_quote_u(s,c): prints as \u0123 or \U01234567 in hexadecimal notation
  • va_stream_render_quote_oct(s,c): prints as \012 in octal notation.

For setting a quotation technique, a va_quotation_t needs to be initialised and set using va_quotation_set().

static va_quotation_t const my_quotation = {
    .delim = { VA_DELIM(0, '<', '>'), VA_DELIM(0, '|', '|') },
    .render_quote = my_render_quote,
};
void my_init(void)
{
    va_quotation_t const *old = va_quotation_set(VA_QUOTE_qq, &my_quotation);
}

This sets the qq prefix to use my_quotation as a quotation method. There are currently 8 different quotation method slots:

  • 0: the default if no q, Q, k, or K modifier is specified
  • VA_QUOTE_q: used if the modifier q is given
  • VA_QUOTE_Q: used if the modifier Q is given
  • VA_QUOTE_k: used if the modifier k is given
  • VA_QUOTE_K: used if the modifier K is given
  • VA_QUOTE_qq: used if the modifier qq is given
  • VA_QUOTE_QQ: used if the modifier QQ is given
  • VA_QUOTE_kk: used if the modifier kk is given

The prefixes q, k, and Q are predefined as described in the previous sections. The others do not quote by default and are free for adding user quotation methods. Note that is possible to set quotation 0 so that it is used when no quotation modifier is given.

User Defined Printers

For user types, printers can be defined so that you do not need to print into an intermediate string buffer, but you can directly print into the output stream. This saves space for the temporary string buffer and avoids thinking about buffer sizes.

The library provides the type va_print_t* that can be passed as an argument to the printing functions instead of the value itself. This provides the library with a user-defined callback for printing, and also encapsulates the value to be printed.

va_print_t has a callback function void (*print)(va_stream_t *, va_print_t*) that the user can fill in to print the user value. The function can make use of va_iprintf to stringify the value. The original stream's format is passed via width, prec, and opt value in the va_print_t struct so that the printer can query them.

To define a printer for a custom type, a derived struct of va_print_t can be used to encapsulate the value, or the provided value can be used to store store a pointer you your value into va_print_t::value if that is sufficient information. E.g., for a simple pair of integers:

typedef struct {
    unsigned a,b;
} my_pair_t;

a printing function is defined for values of this user type:

static void my_print_pair(va_stream_t *s, va_print_t *p)
{
    my_pair_t const *v = p->value;
    va_iprintf(s, "(.a=~s .b=~s)", v->a, v->b);
}

When this is in place, to make the invocation easy, it is a good idea to encapsulate the creation of a temporary va_print_t into a macro. The following macro uses an ({...}) block to implement a type check, because va_print_t::value is a void pointer.

#define P_PAIR(_v) (&VA_PRINT(&my_print_pair,({my_pair_t const *v_=(v); v_;}))

With these definitions, values of the user type can be printed:

my_pair_t pair = { 1, 2 };
va_fprintf(stderr, "pair=~s\n", P_PAIR(&pair));

This prints pair=(.a=1 .b=2) with the above definitions.

The custom printer framework handles width and quotation just like with normal string types.

va_fprintf(stderr, "quote(pair)=~-10qs\n", P_PAIR(&pair));

In contrast to the width, alignment, and quotation, The print precision is not handled by the framework -- the print function needs to handle it, because the semantics of a precision depends on the type. Also, NULL values, are not handled specially, but such values are printed via the normal mechanism -- the framework does not examine the va_print_t::value at all.

Different print formats must be handled by the user print function, e.g., for printing the type or a pointer values -- nothing of this is done by the framework. E.g., with VA_BGET(p->opt, VA_OPT_MODE), the mode can be queries and one of the VA_MODE_* constants can be checked. See va_print/impl.h for the implementation API to access the format options.

In the P_PAIR macro above, one could apply some _Generic magic to select the right printer for a given object type, maybe even to improve on the ({...}) type checking.

Extensions

  • This is type-safe, i.e., printing an int using "~s" will not crash, but just print the integer.

  • ~b and ~B print binary, with optional 0b or 0B prefix.

  • ~e and ~E print integers in Base32, with optional 0e or 0E prefix. This could be handy for writing error codes: 0EINVAL, 0EAGAIN, 0EIO, ...

  • any meaningless format specifier (=letter) defaults to 'print in natural default form'. It is recommended to use ~s for default format printing of anything.

  • The = modifier prints the last value again, possibly with a different format. Note that the format containing = should not contain any *, because then the width/precision will be printed, not the last value, which is probably not what you want.

  • The q, Q, k, and K modifiers mark different kinds of quotation. q is for C, Q is for Java/JSON, and k for Bourne/Korn Shells.

  • The t format prints the input value type in C syntax.

  • The m format is a custom format to print the status of an object, usually for the error status.

Differences

  • The ~x specifier also prints negative signed numbers, again, due to type-safety. Reinterpreting them as unsigned can be done with the z flag.

  • The format specifiers are not needed to prevent the program from crashing, because the information about the type that is passed is not needed. The format really only specifies 'print like ...', so by default it is recommended to just print with ~s.

  • Due to the type-safety, most length modifiers are not supported nor needed. See h, hh, and z modifiers.

  • for strings, the precision counts the number of output bytes in the standard, but in this library, it is the number of input elements in the array, i.e., the precision specifies the array size the string points to. It is felt that input count is more useful, because it allows to print non-NUL terminated strings, while the output width can be controlled by a delimited printer. Also, different glyph widths in Unicode means that the visual width cannot really be controlled by the output count, either.

  • for strings, the width counts the number of characters that are printed, before encoding them in the output encoding. This includes all characters needed for quotation.

  • This library assumes that text is printed, not binary, so it will not output plain \0.

Restrictions

  • %n is not implemented, because pointers to integers are already used for strings, and the ambiguity between size_t* and char32_t* is common on many 32-bit systems, where both are unsigned* in C. Distinguishing whether to read or to write based on the format string alone is also the opposite of what this library tries to do, and accidentally writing the print size into an char32_t* string is a weird bug I'd rather not make possible.

  • m$ syntax for reordering format strings is not supported, because it would require storing the parameters in an array and would counteract all the magic of the recursive expressions. This would make the code much more complex and stack usage infeasible. In fact, it would probably make the whole point of this library infeasible. There is the extended = option for at least printing the same value multiple times, so ~d ~=#x prints the same value decimal and hexadecimal, and ~qs ~=p prints a string in C quotation and its pointer value.

  • no floats, because support would be too large for a small library. Maybe it is added later -- it could be in a separate .o file that is only used if float arguments are actually used (the magic of _Generic: you would not pay for floats unless you use them).

  • Of the size flags, hh, h, l, ll, L, q, j, z, Z, t, only h, hh, and z are implemented, with slightly different semantics to print unsigned integers: h applies a mask 0xffff, hh applies a mask 0xff and z reinterprets the given number as unsigned. Due to the type-safety, the other flags are not needed as the library just prints whatever is thrown at it.

  • '...' literals have type int in C, so values >0x7f, with its highest bit set, will be misinterpreted as illegal Unicode on compilers where char is signed. On my compiler, printing "~c",'\xfe' prints a replacement characters, because \xfe equals (int)0xfffffffe, which is not valid Unicode, and this library has no chance to find out that this is in fact (char)0xfe. So printing '...' literals is unfortunately broken, without a fix. Printing with ~hhc works as expected (but ~zc does not, because, \xfe is an int). Printing (char)'\xfe' also works, but is more ugly in my opinion (I do not like casts much).

  • Compiling printf with any modern compiler gives you compile time warnings about the argument type vs. format string consistency. If these warnings are gone, there are usually no typing problems left for the target architecture (but compiling for other architectures may still have warnings and produce crashing code).

    For this library, no such warnings are be issued, because the argument passing is type safe and you cannot crash it with a wrong format specifier. However, if the argument count and format specifier count do not match, then the output is likely wrong, and the function will have undesirable behaviour. There is no compile time warning for this.

  • gcc 6: The library itself uses relatively little stack. But gcc (and also clang 3.8) accumulates the temporary stack objects in each function without reusing the stack space, i.e., each call to some print function builds up more stack at the call site. The temporary objects are clearly dead, but gcc keeps them. It does not help to add ({...}) or do{...}while(0) to formally restrict the official lifetime of the object to a block -- the compilers keep the object around. This is highly undesirable here, but I have no idea how to prevent this. -fconserve-stack and any other optimisations I tried don't change anything.

    gcc 11 fixes this (or maybe some earlier version), but it requires a block to limit the lifetime, even if the object is clearly dead. I added ({...}) to the macros so that newer compilers produce much less stack usage at call sites.

Q&A

  • Q: Why formatted printing?

    A: Because it is nicer, and also it is feasible for Gnu gettext, which e.g. C++'s cout<< is not. A: Because I like the string template based approach and find it more concise and can read it with less effort.

  • Q: Is this perfect?

    A: Well, no. It is hard to extend for other types to print. The macro mechanisms used are near impossible to understand and causse weird error messages and wrong error positions (in my gcc). The _Generic mechanism causes a ton of C code to be emitted for each print call -- the compiler throws almost all of it away, based on argument type, but looking at the pre-processed code is interesting. The number of arguments may be inconsistent with the format string without compile time warning.

  • Q: I stack usage really low?

    A: Kind of, but not as much as I'd like. It's around 250 bytes worst case on my x86-64.

  • Q: What about code size?

    A: This generates more code at the call site, because each argument is translated to another function call. Also, the temporary objects cause more stack to be used at the call site. Interally, the library is OK wrt. code size, I think.

  • Q: What about speed?

    A: Really? This is about printing messages -- probably short ones (less than a few kB, I'd guess). So while I did try not to mess it up, this is not optimised for speed.

  • Q: Is this safer than printf?

    A: Definitely, I think. There is absolutely no chance to give a wrong format specifier and access the stack (like printf does via stdarg.h) in undefined ways. This is particularly true for multi-arch development where with printf you need to be careful about length specifiers, and you might not get a warning on your machine, but the next person will and it will crash there. I usually need to compile a few times on multiple architectures to get the integer length right, e.g., %u vs %lu vs. %llu vs. %zu.

    This library's mechanism is also more convenient, because you do not need to think much about what you're printing with what format specifier, and there are no PRId16 macros that obfuscate your portable code. And you can use UTF-8, -16, -32 strings seamlessly and mix them freely. You can print into a alloced or stack allocated compound literal safely, with error checking (end of string, out of memory, etc.) and guaranteed NUL termination.

  • Q: Why do you use ~s and not %s?

    A1: This did use %s at the beginning. But the format strings must not be be confused with the standard C printf. The format is not compatible with a printf call, and this is not a drop-in replacement. E.g., in a larger code base, both this and old printf might be mixed, maybe during a transition period, or just because. So programmers may see both styles. With the different sigil, it is immediately clear which format is used in a print call when editing code, and confusion can hopefully be avoided.

TODO

  • ISO-8859-1 (because why not)

Examples

Open a file with computed name, up to a fixed path length:

#include <va_print/char.h>

FILE *open_text_rd(char const *dir, char const *file, unsigned suffix)
{
    return fopen(va_nprintf(80, "~s/~s~.s", dir, file, suffix), "rt");
}

The same with error checking about truncated string or en- or decoding errors:

FILE *open_text_rd(
    char const *dir, char const *file, unsigned suffix)
{
    va_error_t e;
    char *fn = va_nprintf(80, "~s/~s~.s", dir, file, suffix, &e);
    if (e.code != VA_E_OK) {
        return NULL;
    }
    return fopen(fn, "rt");
}

You can use 8-bit, 16-bit, or 32-bit characters seamlessly. The following uses UTF-16 as a parameter, but calls fopen() with an UTF-8 string. The only change is the parameter type. Just for fun, let's use an UTF-32 format string:

FILE *open_text_rd(
    char16_t const *dir, char16_t const *file, unsigned suffix)
{
    va_error_t e;
    char *fn = va_nprintf(80, U"~s/~s~.s", dir, file, suffix, &e);
    if (e.code != VA_E_OK) {
        return NULL;
    }
    return fopen(fn, "rt");
}

This can also be done by creating a dynamically allocated string. The va_error_t mechanism then protects against out-of-memory situation (in which case the function deallocates what it allocated before and returns NULL), and also against Unicode decoding/encoding errors, and other traps.

#include <va_print/alloc.h>

FILE *open_text_rd(char const *dir, char const *file, unsigned suffix)
{
    FILE *f = NULL;
    va_error_t e;
    char *fn = va_asprintf("~s/~s~.s", dir, file, suffix, &e);
    if (e.code == VA_E_OK) {
        f = fopen(fn, "rt");
    }
    free(fn);
    return f;
}

Using VLA, do the same with arbitrary length by pre-computing the length using va_lprintf():

#include <va_print/len.h>

FILE *open_text_rd(char const *dir, char const *file, unsigned suffix)
{
    char s[va_zprintf("~s/~s~.s", dir, file, suffix)];
    return fopen(va_sprintf(s, "~s/~s~.s", dir, file, suffix), "rt");

}

How Does This Work?

The main idea is to use macro magic (both standard C99 and some extensions from gcc, like allowing __VA_ARGS__ to be empty etc.) to convert the printf calls:

x_printf(format);
x_printf(format, arg1);
x_printf(format, arg1, arg2);
...

Into a recursive call sequence:

init(&STREAM(...), format);
render(init(&STREAM(...), format), arg1);
render(render(init(&STREAM(...), format), arg1), arg2);
...

The STREAM() is a temporary stream object, a compound literal, that is used for state information when parsing the format string, and for storing the output printer. The pointer to this temporary object is returned by all of the functions to the next layer of recursion. The init() initialises the format parser and the output stream (e.g. for NUL termination and initial alloc()), and each render() consumes one argument by printing it (once or more times) or using it as a width or precision.

The macro magic is called VA_REC(). Additional to what is described above, it passes a first parameter to the init() and render() calls to show whether the call is the last one of the expression. This is done to be able to optimise the call site code generation, and to allow sane error handling if the number of format specifiers and arguments do not match. You can try it with gcc -E:

VA_REC(render, init, stream);
VA_REC(render, init, stream, a);
VA_REC(render, init, stream, a, b);
...

This becomes:

init(0,stream);
render(0, init(1,stream), a);
render(0, render(1, init(1,stream), a), b);

The init(0,...) macro call is an extern function call that initialises the stream, initialises the output stream (e.g., NUL terminates a char array and/or allocs initial memory), and parses the format string, so that even with no arguments, the expression behaves in a sane way.

The init(1,...) macro call instead resolves to a fast inline function that initialises the stream by just setting all the slots. The format and output stream initialisation is then done by the first render(...) invocation. This way, there is the minimal number of extern calls to keep the call site code small.

The render() resolves to a _Generic() call that selects the appropriate printer based on the type of the argument, and based on whether it's the last call of the expression, so that for each argument, a different C functions may be invoked. E.g.:

int i;
render(1,s,i)    -> print_int(s,i)
render(0,s,i)    -> print_last_int(s,i)

char const *x;
render(1,s,x)    -> print_string(s,x)
render(0,s,x)    -> print_last_string(s,x)

The _last variant finishes printing the format string even if no more argument is given -- this is used to make error handling sane and not just stop in the middle of the format string.