#define	word_t			uint64_t
#define	word_logbytes		3	/* = log_2(sizeof(word_t)) */
#define	atomic_cas_word		atomic_cas_64

#define	word_t			uint32_t
#define	word_logbytes		2	/* = log_2(sizeof(word_t)) */
#define	atomic_cas_word		atomic_cas_32

#define	subword_t		uint16_t
#define	subword_logbytes	1	/* = log_2(sizeof(subword_t)) */
#define	atomic_cas_subword	atomic_cas_16

#define	subword_t		uint8_t
#define	subword_logbytes	0	/* = log_2(sizeof(subword_t)) */
#define	atomic_cas_subword	atomic_cas_8

subword_t
atomic_cas_subword(volatile subword_t *ptr, subword_t old, subword_t new)
{
	const uintptr_t addr = (uintptr_t)ptr;
	const uintptr_t addrmask = ~(uintptr_t)0 << word_logbytes;
	volatile word_t *const wptr = (volatile word_t *)
	    (addr & addrmask);
	const unsigned shift = sizeof(subword_t)*CHAR_BIT *
	    ((addr & ~addrmask) >> subword_logbytes);
	const word_t mask = ~(~(word_t)0 << sizeof(subword_t)*CHAR_BIT)
	    << shift;
	const word_t wold_part = (word_t)old << shift;
	const word_t wnew_part = (word_t)new << shift;
	word_t wold, wnew;

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

	do {
		wold = *wptr;
		if ((wold & mask) != wold_part)
			return (subword_t)((wold & mask) >> shift);
		wnew = ((wold & ~mask) | wnew_part);
	} while (atomic_cas_word(wptr, wold, wnew) != wold);

	return old;
}