No matter how complex database is. The most important question is how complex domain model of application is.
Is query select * from anytable where anycol = @anyvalue optimized? I have no idea. Nobody has. Because there is only one true criteria of optimization - this is performance of such queries. You can save a lot of time with hibernate or other ORM, then use this time to find actually slow queries. As far as I know Hibernate has some ways to use optimized query.
Third your question is good. But also there is no one answer to the question 'Is dirty data good every time everywhere?'. Strictly saying, until locked, any data read from database are dirty, no matter how its were read and where its were stored. Data blocking is not good thing for performance, so usually you should find compromisse between actual data and performance.
There is no silver bullet. ORM has a lot of advantages, but there is only one serious case when it is not suitable: it is dynamic resultsets depends of parameters (when different parameters returns data with different column sets). Because object structure are static at compile time (in static typed languages) ORM can't help in this case.
Every other case can be solved. Entity sevices (changes tracking etc.) can be off, second-level cache can be disabled, and optimized query can be used instead of generated. I have no idea how to do all that things in Hibernate, but I'm sure it is possible.
ORM has a great advantage it concentrate all data access logic in manageable form, and put it in specific place. Also it supports few things are not so easy and direct to implement in your own data access library, like transaction management (including nested transactions, etc), identity mapping (one row - one object), complex hierarchy persisting (if you use objects and object hierarchies), optimistic locking etc, and ORM can greatly helps you with it.
