A click that counts once

Contacts bounce

When you press the stick, two pieces of metal inside the switch meet. They do not meet once: they touch, spring apart and touch again several times in the first few milliseconds before they settle, and the same happens when they part. The pull-down and the SW light see every one of those, and so does a loop fast enough to look.

Press it once and count. Counting every change from LOW to HIGH finds a click for each chatter, on the way down and on the way up. Counting the first change and then not looking for 20 ms finds one. The trace is a picture of the behaviour: this switch's bounce is not published.

Count the change, then wait

Two rules, and each fixes a different mistake. Count the change, not the level, or holding the stick down counts every time round the loop. And after a change, leave the pin alone for a short time, longer than the chatter and shorter than a person can click again:

const unsigned long DEBOUNCE_MS = 20;
int lastSw = LOW;
unsigned long quietUntil = 0;

void checkClick() {
  unsigned long now = millis();
  if (now < quietUntil) return;       // still settling
  int sw = digitalRead(SW_PIN);
  if (sw != lastSw) {
    if (sw == HIGH) Serial.println("click");
    lastSw = sw;
    quietUntil = now + DEBOUNCE_MS;
  }
}

Call checkClick() from loop() as often as you like. It prints once per press, and the release is caught by the same wait without counting. Eight directions uses exactly these lines.

Why not a delay

A delay(20) after each change does the same for the click, but it stops everything else too, and a joystick sketch is usually reading the stick at the same time. Comparing against millis() lets the loop keep reading X and Y while SW settles.