Ahmad,
One of the key features (and limitations) of DMVPN Phase 2 is that each spoke can learn routes to every other spoke directly. It learns about these routes from the Hub, so it is ultimately up to the spoke to make its own determination via BGP or its own routing protocol, the best path to take to get to another spoke.
So think of it this way: In Phase 2, the Hub is still the “hub” for the control plane–all routes are learned through the hub. Spokes cannot exchange routes with each other directly. However, the Hub is no longer necessarily in the data plane–meaning spoke to spoke traffic can bypass the hub directly.
The limitation of this, by the way, (and why DMVPN Phase 2 is considered obsolete), is that in very large networks, it becomes a burden for every single spoke router to learn about all possible routes to every other spoke router (even if they are never needed). Phase 3 fixes this by allowing the use of summarization and default routes in combination with a “redirect” from the Hub. This allows spokes not become so burdened with unnecessary routes.