Original bug was a failed attempt to build llvm-11 with LDFLAGS="-flto -Wl,--as-needed": https://bugs.gentoo.org/749162.

There std::call_once is implemented via weak pthread_once() symbol: https://gcc.gnu.org/PR55394. Presence of -Wl,--as-needed removes libpthread.so and breaks the program at runtime as pthread_once() has to have a real implementation. gcc plans to fix std::call_once by avoiding pthread_once() for unrelated reason (solving it will take a while).

But the problem seems to be more general when local weak symbols are used to probe presence of a real implementation dynamically.

What do you think of changing 'ld' to prefer strong symbols from shared libraries and keeping DT_NEEDED entries even if local object files define weak symbols?

Toy example:

$ cat a.c
// gcc a.c -o a -Wl,--as-needed -pthread && ./a
#include <stdio.h>
#include <pthread.h>

static pthread_once_t once = PTHREAD_ONCE_INIT;

static __typeof(pthread_once) __gthrw_pthread_once __attribute__ ((__weakref__("pthread_once"), __copy__ (pthread_once)));

static void ir(void)
{
    printf("foo\n");
}

int main(void)
{
    if (__gthrw_pthread_once)
        __gthrw_pthread_once(&once, &ir);
}
$ gcc a.c -o a -Wl,--as-needed -pthread -fuse-ld=bfd && ./a
$ gcc a.c -o a -Wl,--as-needed -pthread -fuse-ld=gold && ./a
$ gcc a.c -o a -Wl,--as-needed -pthread -fuse-ld=lld && ./a

$ gcc a.c -o a -Wl,--no-as-needed -pthread -fuse-ld=bfd && ./a
foo

Ideally all 4 tests should call pthread_once from libpthread.so.0 instead of calling local weak reference.

Thanks!