gcc-15 switched to C23

Tl;DR

In November gcc merged the switch from C17 (-std=gnu17) to C23 (-std=c23) language standard used by default for C code. This will cause quite a few build failures in projects written in C. A few example fixes:

• libffi: optional va_start parameter.
• ncompress: void foo() changed the meaning to void foo(void).
• ell bool, true and false are new keywords. And specifically false is not equals to 0 or NULL.

more words

C23 has a few high-visibility breaking changes compared to C17.

bool, true, and false are unconditionally defined now

true and false are now predefined constants (instead of being a part of <stdbool.h> macros and typedefs). Thus, code like below does not compile any more:

enum { false = 0; }
typedef int bool;

Error messages:

$ printf 'enum { false = 0 };' | gcc -std=c17 -c -x c -
$ printf 'enum { false = 0 };' | gcc -c -x c -
<stdin>:1:8: error: expected identifier before 'false'

$ printf 'typedef int bool;' | gcc -std=c17 -c -x c -
$ printf 'typedef int bool;' | gcc -c -x c -
<stdin>:1:13: error: two or more data types in declaration specifiers
<stdin>:1:1: warning: useless type name in empty declaration

The fix is usually to use <stdbool.h> or avoid name collisions. Example affected project is linux.

partially defined int (*)() function prototypes are just int (*)(void) now

This one is trickier to fix when intentionally used. C happened to allow the following code:

// $ cat a.c
typedef void (*PF)();

static int f0(void)  { return 42; }
static int f1(int a) { return 42 + a; }

int main() {
    PF pf;

    // 0-argument function pointer
    pf = f0;
    pf();

    // 1-argument function pointer
    pf = f1;
    pf(42);

    // 3-argument function pointer: an odd one, but happens to work
    pf(42,42,42);
}

But not any more:

$ gcc -std=c17 -c a.c
$ gcc -c a.c
a.c: In function 'main':
a.c:15:8: error: assignment to 'PF' {aka 'int (*)(void)'} from incompatible pointer type 'int (*)(int)' [-Wincompatible-pointer-types]
   15 |     pf = f1;
      |        ^
a.c:16:5: error: too many arguments to function 'pf'
   16 |     pf(42);
      |     ^~
a.c:19:5: error: too many arguments to function 'pf'
   19 |     pf(42,42,42);
      |     ^~

This hack is used at least in ski, ghc and ncompress. But more frequently it’s use is an accident (ell, iwd, bash and a few others).

parting words

Quick quiz: the above changes look like they tickle some very obscure case. How many packages are affected on a typical desktop system? What would be your guess? 1? 5? 100? 1000?

So far on my system (~2000 installed packages) I observed the failures of the following projects:

• linux
• speechd
• vde2
• sane-backends
• timidity
• neovim
• bluez
• samba
• weechat
• iwd
• protobuf
• netpbm
• mariadb-connector-c
• liblqr1
• sqlite-odbc-driver
• python:typed-ast
• python2
• perl:XS-Parse-Keyword
• pgpdump
• ell
• SDL-1
• ruby-3.1
• dnsmasq
• ghc
• gnupg
• ghostscript
• procmail
• jq
• libsndfile
• ppp
• time
• postfix
• mcpp
• xmlrpc-c
• unifdef
• hotdoc
• mypy
• rustc
• xorg:libXt
• rsync
• oniguruma
• ltrace
• sudo
• lsof
• lv
• dbus-glib
• argyllcms
• valgrind
• postgresql-14
• gdb
• git
• ncompress
• w3m
• freeglut
• xcur2png
• vifm
• p11-kit
• cyrus-sasl
• xvidcore
• guile
• editline
• e2fsprogs
• gsm
• libconfig
• db
• libtirpc
• nghttp2
• libkrb5
• libgpg-error
• cpio
• sharutils
• gpm
• expect
• ncurses
• yasm
• texinfo-6.7
• gettext
• unzip
• gdbm
• m4
• binutils
• ed
• gmp
• bash

That’s more than 80 packages, or about 4% of all the packages I have installed.

Looks like gcc-15 will be a disruptive release (just like gcc-14) that will require quite a few projects to adapt to new requirements (either by fixing code or by slapping -std=gnu17 as a requirement).

Most of the build failures above are not yet fixed upstream. These can be good first contribution fixes if you are thinking of making one.

Have fun!