I'd focus on the shared memory locks first (the mutexes and semaphores) as they are most likely to cause issues. Look at which state is being protected by locks and then determine which state is under the protection of several locks. This will give you a sense of potential conflicts. Look at situations where code that holds a lock calls out to methods (don't forget virtual methods). Try to eliminate these calls where possible (by reducing the time the lock is held).
Given the list of mutexes that are held and a rough idea of the state that they protect, assign a locking order (i.e., mutex A should always be taken before mutex B). Try to enforce this in the code.
See if you can combine several locks into one if concurrency won't be adversely affected. For example, if mutex A and B seem like they might have deadlocks and an ordering scheme is not easily done, combine them to one lock initially.
It's not going to be easy but I'm for simplifying the code at the expense of concurrency to get a handle of the problem.
