static inline int
atomic·cas(volatile int *p, int t, int s)
{
	__asm__ __volatile__ (
		"lock ; cmpxchg %3, %1"
		: "=a"(t), "=m"(*p) : "a"(t), "r"(s) : "memory" );
	return t;
}

static inline void
*atomic·casp(volatile void *p, void *t, void *s)
{
	__asm__( "lock ; cmpxchg %3, %1"
		: "=a"(t), "=m"(*(void *volatile *)p)
		: "a"(t), "r"(s) : "memory" );
	return t;
}

static inline int
atomic·swap(volatile int *p, int v)
{
	__asm__ __volatile__(
		"xchg %0, %1"
		: "=r"(v), "=m"(*p) : "0"(v) : "memory" );
	return v;
}

static inline int
atomic·fetchadd(volatile int *p, int v)
{
	__asm__ __volatile__(
		"lock ; xadd %0, %1"
		: "=r"(v), "=m"(*p) : "0"(v) : "memory" );
	return v;
}

static inline void
atomic·and(volatile int *p, int v)
{
	__asm__ __volatile__(
		"lock ; and %1, %0"
		: "=m"(*p) : "r"(v) : "memory" );
}

static inline void
atomic·or(volatile int *p, int v)
{
	__asm__ __volatile__(
		"lock ; or %1, %0"
		: "=m"(*p) : "r"(v) : "memory" );
}

static inline void
atomic·and64(volatile uint64 *p, uint64 v)
{
	__asm__ __volatile(
		"lock ; and %1, %0"
		 : "=m"(*p) : "r"(v) : "memory" );
}

static inline void
atomic·or64(volatile uint64 *p, uint64 v)
{
	__asm__ __volatile__(
		"lock ; or %1, %0"
		 : "=m"(*p) : "r"(v) : "memory" );
}

static inline void
atomic·inc(volatile int *p)
{
	__asm__ __volatile__(
		"lock ; incl %0"
		: "=m"(*p) : "m"(*p) : "memory" );
}

static inline void
atomic·dec(volatile int *p)
{
	__asm__ __volatile__(
		"lock ; decl %0"
		: "=m"(*p) : "m"(*p) : "memory" );
}

static inline void
atomic·store(volatile int *p, int x)
{
	__asm__ __volatile__(
		"mov %1, %0 ; lock ; orl $0,(%%rsp)"
		: "=m"(*p) : "r"(x) : "memory" );
}

static inline void
atomic·barrier()
{
	__asm__ __volatile__( "" : : : "memory" );
}

static inline void
atomic·spin()
{
	__asm__ __volatile__( "pause" : : : "memory" );
}

static inline void
atomic·crash()
{
	__asm__ __volatile__( "hlt" : : : "memory" );
}

static inline int
atomic·ctz64(uint64 x)
{
	__asm__( "bsf %1,%0" : "=r"(x) : "r"(x) );
	return x;
}

static inline int
atomic·clz64(uint64 x)
{
	__asm__( "bsr %1,%0 ; xor $63,%0" : "=r"(x) : "r"(x) );
	return x;
}