In this paper, we study the problem of compression efficiency of SRH (Segment Routing Header) in G-SRv6 (Generalized SRv6). We propose the Function-decoupled Segment Routing Mechanism (FDSRM) to optimize the SID list in SRH while ensuring that the routing policy decision would not be affected. FDSRM logically decouples the functions of SID/G-SID according to the function in routing decision and instruction indication. Based on FDSRM, we propose a mathematical optimization framework that leverages the LSTM neural network to optimize the SID allocation to adapt to future traffic. Simulation results indicate that FDSRM can improve the probability of SRH compression by 50.86%, and compress more than 24.50 bytes when the hop count is greater than 20.