The generic Linux version of sem_post (nptl/sysdeps/unix/sysv/linux/sem_post.c) uses atomic_compare_and_exchange_bool_acq to increment the semaphore.  This is an unlock operation and should actually be using a release barrier.  The following test fails on some MIPS systems (MIPS uses the generic version of sem_post where various architectures use their own .S versions).  Testing a patch.

#include <pthread.h>
#include <semaphore.h>
#include <stdio.h>

#define NTHREADS 10
#define NITER 100000

sem_t sem;
int c;
volatile int thread_fail;

static void *
tf (void *arg)
{
  for (int i = 0; i < NITER; i++)
    {
      if (sem_wait (&sem) != 0)
        {
          perror ("sem_wait");
          thread_fail = 1;
        }
      ++c;
      if (sem_post (&sem) != 0)
        {
          perror ("sem_post");
          thread_fail = 1;
        }
    }
  return NULL;
}

int
main (void)
{
  if (sem_init (&sem, 0, 0) != 0)
    {
      perror ("sem_init");
      return 1;
    }

  pthread_t th[NTHREADS];
  for (int i = 0; i < NTHREADS; i++)
    {
      if (pthread_create (&th[i], NULL, tf, NULL) != 0)
        {
          puts ("pthread_create failed");
          return 1;
        }
    }

  if (sem_post (&sem) != 0)
    {
      perror ("sem_post");
      return 1;
    }

  for (int i = 0; i < NTHREADS; i++)
    if (pthread_join (th[i], NULL) != 0)
      {
        puts ("pthread_join failed");
        return 1;
      }

  if (c != NTHREADS * NITER)
    {
      printf ("c = %d, should be %d\n", c, NTHREADS * NITER);
      return 1;
    }
  return thread_fail;
}