#define	natural_t		uint64_t
#define	natural_logbytes	3	/* = log_2(sizeof(natural_t)) */
#define	atomic_cas_natural	atomic_cas_64

#define	natural_t		uint32_t
#define	natural_logbytes	2	/* = log_2(sizeof(natural_t)) */
#define	atomic_cas_natural	atomic_cas_32

#define	small_t			uint16_t
#define	small_bits		16	/* = sizeof(small_t)*CHAR_BIT */
#define	small_logbytes		1	/* = log_2(sizeof(small_t)) */
#define	atomic_cas_small	atomic_cas_16

#define	small_t			uint8_t
#define	small_bits		8	/* = sizeof(small_t)*CHAR_BIT */
#define	small_logbytes		0	/* = log_2(sizeof(small_t)) */
#define	atomic_cas_small	atomic_cas_8

small_t
atomic_cas_small(volatile small_t *ptr, small_t old, small_t new)
{
	const uintptr_t addr = (uintptr_t)ptr;
	const uintptr_t highmask = ~(uintptr_t)0 << natural_logbytes;
	volatile natural_t *const nptr = (volatile natural_t *)
	    (addr & highmask);
	const unsigned shift = small_bits *
	    ((addr & ~highmask) >> small_logbytes);
	const natural_t mask = ~(~(natural_t)0 << small_bits) << shift;
	const natural_t nold_part = (natural_t)old << shift;
	const natural_t nnew_part = (natural_t)new << shift;
	natural_t nold, nnew;

	assert((addr & ~(~(uintptr_t)0 << small_logbytes)) == 0);

	do {
		nold = *nptr;
		if ((nold & mask) != nold_part)
			return (nold & mask) >> shift;
		nnew = ((nold & ~mask) | nnew_part);
	} while (atomic_cas_natural(nptr, nold, nnew) != nold);

	return nnew;
}