#include <stddef.h>
#include <stdint.h>

static inline size_t
countnl8(uint64_t x)
{
	uint64_t y = x & 0x8080808080808080;
	uint64_t z = x ^ 0x0a0a0a0a0a0a0a0a;

	z |= 0x8080808080808080;
	z -= 0x0101010101010101;
	z &= 0x8080808080808080;
	z |= y;
	z >>= 7;
	z += z >> 8;
	z += z >> 16;
	z += z >> 32;
	z &= 0xf;

	return z;
}

size_t
countnl(const uint64_t *p, size_t n)
{
	size_t nl = 0;
	size_t i = 0;

#if 1
	for (; i < 4*(n/4); i += 4) {
		uint64_t z0 = countnl8(p[i + 0]);
		uint64_t z1 = countnl8(p[i + 1]);
		uint64_t z2 = countnl8(p[i + 2]);
		uint64_t z3 = countnl8(p[i + 3]);

		nl += z0 + z1 + z2 + z3;
	}
#endif

	for (; i < n; i++)
		nl += countnl8(p[i]);

	return nl;
}