#include "precompile.h"
#include "spinlock.h"


static void tkspause()
{
	#if defined __x86_64__ || defined __i386__

#ifdef __SSE2__
	__asm__ __volatile__ ("pause");
#else
	__asm__ __volatile__ ("rep; nop");
#endif

#elif defined __arm__

#if __ARM_ARCH == 7
	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");
#endif

#elif defined __OCTEON__

	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");
	__asm__ __volatile__ ("nop");

#elif defined __powerpc__
	__asm__ __volatile__ ("ori 27, 27, 27");

#endif
}

// please refer http://book.opensourceproject.org.cn/kernel/kernel3rd/opensource/0596005652/understandlk-chp-5-sect-2.html

static __inline void backoff(unsigned int nloop)
{
	//and d, s              d <- d & s
	//if ZF d, s 
#if defined _WIN32 // __x86_64__ || defined __i386__
	__asm {
		and ecx, nloop;
		cmovz ecx, nloop;
		rep nop;
	}
#elif defined(__linux__)

#if defined(__arm__) || defined(__aarch64__) 
	//TODO:
#else
	//TODO: need to confirm its correctness.
	__asm__ __volatile__ ("and %edi, %ecx;"
									"cmovz %edi, %ecx;"
									"rep;"
									"nop");
#endif 

#endif

}


TOOLKIT_API void spinlock_enter(volatile spinlock_t *sl, int spin_count)
{
	DWORD tid = GetCurrentThreadId();
	if (sl->_value == tid) {
		sl->_reentrance++;
	} else {
		int i = 0;
		if (spin_count <= 0)
			spin_count = DEFAULT_SPIN_COUNT;
		while (InterlockedCompareExchange((LONG*)&sl->_value, (LONG)tid, 0) != 0) {
			for (; i < spin_count; ++i)
				if (sl->_value == 0)
					break;
			if (i >= spin_count)
				Sleep(1); /* yield cpu */
		}
		sl->_reentrance++;
	}
}
TOOLKIT_API void spinlock_leave(volatile spinlock_t *sl)
{
	_ReadWriteBarrier();
	if (--sl->_reentrance == 0)
		InterlockedExchange((LONG*)&sl->_value, 0);	
}

TOOLKIT_API void rw_spinlock_read_lock(rw_spinlock_t *rwlock)
{
	union rw_spinlock_barrier bnew, b;
	for (;;) {
		bnew.nAtomic = b.nAtomic = rwlock->barrier.nAtomic    ;
		bnew.readerCount++      ;   
		bnew.writerActive = b.writerActive = 0  ;
		bnew.writerCount = b.writerCount = 0    ;
		if (InterlockedCompareExchange(&rwlock->barrier.nAtomic, bnew.nAtomic, b.nAtomic) == b.nAtomic)
			break;
		backoff(DEFAULT_SPIN_COUNT) ;
	}
}

TOOLKIT_API void rw_spinlock_read_unlock(rw_spinlock_t *rwlock)
{
	union rw_spinlock_barrier bnew, b ;
	_ReadWriteBarrier();
	for (;;) {
		bnew.nAtomic = b.nAtomic = rwlock->barrier.nAtomic    ;
		bnew.readerCount--      ;
		TOOLKIT_ASSERT( b.writerActive == 0 ) ;
		if (InterlockedCompareExchange(&rwlock->barrier.nAtomic, bnew.nAtomic, b.nAtomic) == b.nAtomic)
			break;
		backoff(DEFAULT_SPIN_COUNT) ;
	}
}

TOOLKIT_API void rw_spinlock_write_lock(rw_spinlock_t *rwlock)
{
	union rw_spinlock_barrier bnew, b ;

	for (;;) {
		bnew.nAtomic = b.nAtomic = rwlock->barrier.nAtomic    ;
		bnew.writerCount++      ;
		if (InterlockedCompareExchange(&rwlock->barrier.nAtomic, bnew.nAtomic, b.nAtomic) == b.nAtomic)
			break;
		backoff(DEFAULT_SPIN_COUNT)	;
	}

	for (;;) {
		bnew.nAtomic = b.nAtomic = rwlock->barrier.nAtomic    ;
		bnew.readerCount = 
			b.readerCount = 0   ;
		b.writerActive = 0      ;					
		bnew.writerActive = 1   ;
		if (InterlockedCompareExchange(&rwlock->barrier.nAtomic, bnew.nAtomic, b.nAtomic) == b.nAtomic)
			break;
		backoff(DEFAULT_SPIN_COUNT)	;
	}
}

TOOLKIT_API void rw_spinlock_write_unlock(rw_spinlock_t *rwlock)
{
	union rw_spinlock_barrier bnew, b ;

	_ReadWriteBarrier();

	for (;;) {
		bnew.nAtomic = b.nAtomic = rwlock->barrier.nAtomic    ;
		TOOLKIT_ASSERT( b.writerActive == 1 ) ;
		TOOLKIT_ASSERT( b.readerCount == 0 )  ;
		bnew.writerActive = 0   ;
		--bnew.writerCount      ;
		if (InterlockedCompareExchange(&rwlock->barrier.nAtomic, bnew.nAtomic, b.nAtomic) == b.nAtomic)
			break;
		backoff(DEFAULT_SPIN_COUNT)	;
	}
}

#define BACKOFF_LIMIT 1000

TOOLKIT_API void fastlock_enter( lock_t l )
{
	int i = 0;
	int spin_count = 0;
	int backoffs = 0;
	while (InterlockedCompareExchange(l, 1, 0) == 1) {
		for (spin_count = i + 10000; i < spin_count; ++i) {
			if (*l == 0)
				break;
#if defined(_MSC_VER)
#if _MSC_VER < 1400
			__asm { rep nop }
#else
			YieldProcessor();		//关闭sphost时，调试在这里停止，且无法通过taskkill关闭sphost
#endif
#else
			tkspause();
#endif //_MSC_VER
		}
		backoffs++;
		if (backoffs % BACKOFF_LIMIT == 0) {
			Sleep(500);
		}
		SwitchToThread();
	}
}

int fastlock_tryenter(lock_t l)
{
	int i = 0;
	int spin_count = 0;
	int backoffs = 0;
	while (InterlockedCompareExchange(l, 1, 0) == 1) {
		for (spin_count = i + 10000; i < spin_count; ++i) {
			if (*l == 0)
				break;
#if defined(_MSC_VER)
#if _MSC_VER < 1400
			__asm { rep nop }
#else
			YieldProcessor();		//关闭sphost时，调试在这里停止，且无法通过taskkill关闭sphost
#endif
#else
			tkspause();
#endif //_MSC_VER
		}
		backoffs++;
		if (backoffs % BACKOFF_LIMIT == 0) {
			Sleep(500);
			if (*l == 0)
				break;
			else
				return 0;
		}
		SwitchToThread();
	}
	return 1;
}