_GNU_SOURCE is set inconsistently between configure and build

[alyssais]

Since b40c560 ("configure_func.sh: Add _GNU_SOURCE define and function signatures"), _GNU_SOURCE will be defined when checking whether a function is available, but then won't be set during the actual build,resulting in the build trying to use a function that isn't visible to it.

For example, musl only defines memrchr with _GNU_SOURCE. The configure check for this function sets _GNU_SOURCE, and will pass, but then during compileration calls to memrchr will result in errors because _GNU_SOURCE is not passed to the compiler then.

To fix this, _GNU_SOURCE should always be part of cflags to match the environment used by the configure checks. On platforms where _GNU_SOURCE isn't used, this should be a no-op.

[peterhurley-cisco]

I was actually about to report the same issue (with musl libc on yocto/oe-core).

This configure failure with musl libc is caused by a mismatch in the perl-cross feature test macros which correctly identify memrchr, memmem, etc.

Checking for memmem ... found
Checking for memmove ... found
Checking for memrchr ... found

but fail the GNU libc test

Checking if we're using GNU libc ... no

which results in d_memrchr = define

Checking for memrchr
| #define _GNU_SOURCE
| #include <stdlib.h>
| #include <string.h>
| int main(void) { memrchr(NULL, 0, 0); return 0; }
> aarch64-poky-linux-musl-gcc -mbranch-protection=standard -fstack-protector-strong -Wformat -Wformat-security -Werror=format-security --sysroot=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0/recipe-sysroot -Os -pipe -g -feliminate-unused-debug-types -fmacro-prefix-map=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0=/usr/src/debug/perl/5.34.3-r0 -fdebug-prefix-map=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0=/usr/src/debug/perl/5.34.3-r0 -fdebug-prefix-map=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0/recipe-sysroot= -fdebug-prefix-map=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0/recipe-sysroot-native= -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 -o try try.c -lm -lcrypt -lpthread -lgdbm -ldl -lgdbm_compat -O0 -fno-builtin
Setting d_memrchr=define
Result: found

but d_gnulibc = undef

 Checking if we're using GNU libc
| #include <stdio.h>
| #ifndef __GLIBC__
| #error here
| #endif
> aarch64-poky-linux-musl-gcc -mbranch-protection=standard -fstack-protector-strong -Wformat -Wformat-security -Werror=format-security --sysroot=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0/recipe-sysroot -Os -pipe -g -feliminate-unused-debug-types -fmacro-prefix-map=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0=/usr/src/debug/perl/5.34.3-r0 -fdebug-prefix-map=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0=/usr/src/debug/perl/5.34.3-r0 -fdebug-prefix-map=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0/recipe-sysroot= -fdebug-prefix-map=/mnt/buildspace/peter/yocto/router.kfom2/build/tmp/work/aarch64-poky-linux-musl/perl/5.34.3-r0/recipe-sysroot-native= -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 -c -o try.o try.c
try.c:3:2: error: #error here
    3 | #error here
      |  ^~~~~
Setting d_gnulibc=undef
Result: no

The resulting config.h emits this brokenness which means that _GNU_SOURCE is not defined in the perl build but HAS_MEMRCHR is :

/* HAS_MEMMEM:
 *	This symbol, if defined, indicates that the memmem routine is
 *	available to return a pointer to the start of the first occurrence
 *	of a substring in a memory area (or NULL if not found).
 *	In glibc, memmem is a GNU extension.  The function is visible in
 *	libc, but the prototype is only visible if _GNU_SOURCE is #defined.
 *	Thus we only define this if both the prototype and symbol are found.
 */
#define HAS_MEMMEM		/**/

/* HAS_MEMRCHR:
 *	This symbol, if defined, indicates that the memrchr routine is
 *	available to return a pointer to the last occurrence of a byte in
 *	a memory area (or NULL if not found).
 */
#define HAS_MEMRCHR		/**/

....

/* HAS_GNULIBC:
 *	This symbol, if defined, indicates to the C program that
 *	the GNU C library is being used.  A better check is to use
 *	the __GLIBC__ and __GLIBC_MINOR__ symbols supplied with glibc.
 */
/*#define HAS_GNULIBC	/ **/
#if defined(HAS_GNULIBC) && !defined(_GNU_SOURCE)
#   define _GNU_SOURCE
#endif

oe-core/yocto is carrying a workaround since yocto/scarthgap
openembedded/openembedded-core@6422e62

and I just asked for a cherry-pick of that workaround for yocto/kirkstone https://lists.openembedded.org/g/openembedded-core/message/216388

[peterhurley-cisco]

I'm happy to submit a fix but not sure which approach is acceptable:

1. emit need_gnusource from feature test macros that try_add '#define _GNU_SOURCE'
   and test that in the d_gnulibc feature test; or
2. check the various d_memmem, d_memrchr, in the d_gnulibc feature test; or
3. do like upstream perl and perform the d_gnulibc feature test first, and then all the dependent _GNU_SOURCE feature tests