Test

volatile uint8_t *port;
uint8_t High, Low, Mask;
uint8_t PinNumber = 2;
uint8_t Loop = 100;


void setup() {

  DDRD  = 0x04;
  PORTD = 0x00;

  port = portOutputRegister(digitalPinToPort(PinNumber));
  Mask = digitalPinToBitMask(PinNumber);
  High = *port |  Mask;
  Low  = *port | ~Mask;

  pinMode(16, OUTPUT);
}

void loop() {
  digitalWrite(16, HIGH);
  delay(500);
  digitalWrite(16, LOW);
  delay(500);
  Loop = 100;
  asm volatile(
    "Tone: \n\t"
      "rcall S3520HZ"        "\n\t"
      "dec %[Loop]"       "\n\t"
    "rcall Tone"         "\n\t"

    "S3520HZ:"                 "\n\t"  //3520 Hz A7 3564 Hz
      "st %a[port], %[High]"  "\n\t"
      "rcall delay7040"      "\n\t"
      "st %a[port], %[Low]"   "\n\t"
      "rcall delay7040"      "\n\t"
    "ret"                     "\n\t"

    "delay7040:" "\n\t"
      "ldi r21, 16" "\n\t"
      "delay7040_1:" "\n\t"
        "rcall delay_1 \n\t"
        "dec r21 \n\t"
      "brne delay7040_1 \n\t"
    "ret \n\t"

    "delay_1:"            "\n\t"  //0.0625us * (2 + 32 * 2) = 4.125 us
      "ldi r20,32"      "\n\t"
      "delay_11:"        "\n\t"
        "nop"           "\n\t"
        "dec r20"       "\n\t"
      "brne delay_11 \n\t"
    "ret"
    :[port]"+e"(port)
    :[High]"r"(High),
     [Low]"r"(Low),
     [Loop]"r"(Loop)
  );

  digitalWrite(16, HIGH);
  delay(500);
  digitalWrite(16, LOW);
  delay(500);
}