How guarantees compiler not to optimize within a guarded code in C++

Question

If I use a code locked by a mutex and guarded by e.g. a scoped_lock, the instruction order can be guaranteed. However, I do not really understand, why the varia