func xorIn(d *state, buf []byte) {
	if cpu.IsBigEndian {
		for i := 0; len(buf) >= 8; i++ {
			a := binary.LittleEndian.Uint64(buf)
			d.a[i] ^= a
			buf = buf[8:]
		}
	} else {
		ab := (*[25 * 64 / 8]byte)(unsafe.Pointer(&d.a))
		subtle.XORBytes(ab[:], ab[:], buf)
	}
}

// copyOut copies uint64s to a byte buffer.
func copyOut(d *state, b []byte) {
	if cpu.IsBigEndian {
		for i := 0; len(b) >= 8; i++ {