With the development of techniques, the uncertainty in software systems is continuously increasing. Data race is a typical bug in current programs, which is a classic type of uncertainty programs. Despite significant progress in recent years, the important problem of practical static race detection remains open. Previous static techniques either suffer from a high false positive rate due to the compromise of precision, or scalability issues caused by a highly precise analysis. This paper presents GUARD, a staged approach to resolve this paradox. First, it performs a lightweight context-sensitive data access analysis, based on the value flow of a program, to identify the candidate data race subpaths instead of the whole program paths. Second, may-happen-in-parallel (MHP) analysis is employedto identify whether two data accesses in a program may execute concurrently. This stage is scalable, due to the design of the thread flow graph (TFG), which encodes thread information to query MHP relationship of the subpaths. Finally, for each subpath whose two data accesses are MHP, the heavyweight path-sensitive analysis is appliedto verify the feasibility of the data races. The evaluation demonstrates that GUARD can finish checking industrial-sized projects, up to 1.3MLoC, in 1 870s with an average false positive rate of 16.0%. Moreover, GUARD is faster than the state-of-the-art techniques with the average speedup 6.08X and significantly fewer false positives. Besides, GUARD has found 12 new race bugs in real-world programs. All of them are reportedtothe developers and 8 of them have been confirmed.