Conventional symbol-level precoding (SLP) designs assume fixed modulations and detection rules at the receivers for simplifying the transmit precoding optimizations, which greatly limits the flexibility of SLP and the communication quality-of-service (QoS). To overcome the performance bottleneck of these approaches, in this letter we propose an end-to-end learning based approach to jointly optimize the modulation orders, the transmit precoding and the receive detection for an SLP communication system. A neural network composed of the modulation order prediction (MOP-NN) module and the symbol-level precoding and detection (SLPD-NN) module is developed to solve this mathematically intractable problem. Simulations verify the notable performance improvement brought by the proposed end-to-end learning approach.