barrier.h

barrier.h

#define atomic_load(address) __atomic_load_n(address, __ATOMIC_SEQ_CST)
#define atomic_store(address, value) __atomic_store_n(address, value, __ATOMIC_SEQ_CST)
#define atomic_fetch_add(address, value) __atomic_fetch_add(address, value, __ATOMIC_SEQ_CST)

extern void wait_on_address(u32* address, u32 while_value);
extern void wake_all_on_address(u32* address);

typedef struct {
  u32 barrier;
  u32 barrier_counter;
} SharedData;
void barrier(SharedData* shared_data, u32 thread_count) {
  u32 barrier = atomic_load(&shared_data->barrier);
  u32 barrier_stop = barrier + thread_count;
  u32 barrier_counter = atomic_add_fetch(&shared_data->barrier_counter, 1);
  if (barrier_counter != barrier_stop) {
    /* NOTE: On windows, WaitOnAddress() "is allowed to return for other reasons", same thing with futex() on linux */
    while (atomic_load(&shared_data->barrier) == barrier) {
      wait_on_address(&shared_data->barrier, barrier);
    }
  } else {
    atomic_store(&shared_data->barrier, barrier_stop);
    wake_all_on_address(&shared_data->barrier);
  }
}