darling-libobjc2/spinlock.h

#ifdef __MINGW32__
#include <windows.h>
static unsigned sleep(unsigned seconds)
{
	Sleep(seconds*1000);
	return 0;
}
#else
#include <unistd.h>
#endif

/**
 * Number of spinlocks.  This allocates one page on 32-bit platforms.
 */
#define spinlock_count (1<<10)
static const int spinlock_mask = spinlock_count - 1;
/**
 * Integers used as spinlocks for atomic property access.
 */
extern int spinlocks[spinlock_count];
/**
 * Get a spin lock from a pointer.  We want to prevent lock contention between
 * properties in the same object - if someone is stupid enough to be using
 * atomic property access, they are probably stupid enough to do it for
 * multiple properties in the same object.  We also want to try to avoid
 * contention between the same property in different objects, so we can't just
 * use the ivar offset.
 */
static inline volatile int *lock_for_pointer(const void *ptr)
{
	intptr_t hash = (intptr_t)ptr;
	// Most properties will be pointers, so disregard the lowest few bits
	hash >>= sizeof(void*) == 4 ? 2 : 8;
	intptr_t low = hash & spinlock_mask;
	hash >>= 16;
	hash |= low;
	return spinlocks + (hash & spinlock_mask);
}

/**
 * Unlocks the spinlock.  This is not an atomic operation.  We are only ever
 * modifying the lowest bit of the spinlock word, so it doesn't matter if this
 * is two writes because there is no contention among the high bit.  There is
 * no possibility of contention among calls to this, because it may only be
 * called by the thread owning the spin lock.
 */
inline static void unlock_spinlock(volatile int *spinlock)
{
	__sync_synchronize();
	*spinlock = 0;
}
/**
 * Attempts to lock a spinlock.  This is heavily optimised for the uncontended
 * case, because property access should (generally) not be contended.  In the
 * uncontended case, this is a single atomic compare and swap instruction and a
 * branch.  Atomic CAS is relatively expensive (can be a pipeline flush, and
 * may require locking a cache line in a cache-coherent SMP system, but it's a
 * lot cheaper than a system call).
 *
 * If the lock is contended, then we just sleep and then try again after the
 * other threads have run.  Note that there is no upper bound on the potential
 * running time of this function, which is one of the great many reasons that
 * using atomic accessors is a terrible idea, but in the common case it should
 * be very fast.
 */
inline static void lock_spinlock(volatile int *spinlock)
{
	int count = 0;
	// Set the spin lock value to 1 if it is 0.
	while(!__sync_bool_compare_and_swap(spinlock, 0, 1))
	{
		count++;
		if (0 == count % 10)
		{
			// If it is already 1, let another thread play with the CPU for a
			// bit then try again.
			sleep(0);
		}
	}
}