Do the same optimization as x86-64: do __ffs() on the intermediate value
that found whether there is a zero byte, before we've actually computed
the final byte mask.

The logic is:

 has_zero():
	Check if the word has a zero byte in it, which indicates the end
	of the loop, and prepare a value to be used for the rest of the
	sequence.

	The standard LE implementation just creates a word that has the
	high bit set in each byte of the word that was zero.

        Example: 0xaa00bbccdd00eeff -> 0x0080000000800000

 prep_zero_mask():
	Possibly do more prep to then clean up the initial fast result
	from has_zero, so that it can be combined with another zero mask
	with a simple logical "or" to create a final mask.

	This is only used on big-endian machines that use a different
	algorithm, and is a no-op here.

 create_zero_mask():
	This is "step 1" of creating the count and the mask, and is
	meant for any common operations between the two.

	In the old implementation, this actually created the zero mask,
	that was then used for masking and for counting the number of
	bits in the mask.

	In the new implementation, this is a no-op.

 count_zero():
	This takes the mask bits, and counts the number of bytes before
	the first zero byte.

	In the old implementation, it counted the number of bits in the
	final byte mask (which was the same as the C standard "find last
	set bit" that uses the silly "starts at one" counting) and shifted
	the value down by three.

	In the new implementation, we know the intermediate mask isn't
	zero, and it just does "find first set" with the sane semantics
	without any off-by-one issues, and again shifts by three (which
	also masks off the bit offset in the zero byte itself).

	Example: 0x0080000000800000 -> 2

 zero_bytemask():
	This takes the mask bits, and turns it into an actual byte mask
	of the bytes preceding the first zero byte.

	In the old implementation, this was a no-op, because the work
	had already been done by create_zero_mask().

	In the new implementation, this does what create_zero_mask()
	used to do.

	Example: 0x0080000000800000 -> 0x000000000000ffff

The difference between the old and the new implementation is that
"count_zero()" ends up scheduling better because it is being done on a
value that is available earlier (before the final mask).

But more importantly, it can be implemented without the insane semantics
of the standard bit finding helpers that have the off-by-one issue and
have to special-case the zero mask situation.

On arm64, the new "count_zero()" ends up just "rbit + clz" plus the
shift right that then ends up being subsumed by the "add to final
length".
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>