What shapes the loss landscape of self-supervised learning?