In server-side Java, system routines of a JVM are invoked frequently from various locations in Java byte-code. Thus in order to achieve high performance, it is necessary to do system-level optimization that is adaptive to each invocation site in bytecode. In this paper, we propose a method called SSJIT. SSJIT detects bottlenecks of a system using runtime statistics. Then it JIT-compiles the methods containing system-routine invocations related to the bottlenecks and optimizes each invocation. In this paper, we target a contention problem of mutex locks, which degrades throughput or fairness of a server. SSJIT detects contended locks by runtime statistics, and then invokes a JIT compiler. The compiler inline-expands the lock code that is modified to delay pre-emption during the critical section, which obviates contention. We evaluated SSJIT by measuring throughput, fairness, and reduced overhead compared with existing lock protocols.