A prototypical concept you would expect to find in a dependently typed language is "the ability to have a function that concatenates to arrays of sizes X and Y and returns an array of size X+Y that can be safely verified against further functions that also require arrays given as arguments to have specific lengths".
C++ is a bit indirect in its semantics here, but you will note that most other languages with generic types cannot express this thought (using either real integers or simulating them with types, either of which can be done in C++).
One could easily imagine "filling out" C++ to allow compile time type argument manipulation of more value types without even stressing its model, and in fact that is exactly the direction the language has been going in with constexpr. Here is an article on how to build a compile-time printf using C++14.
http://blogs.microsoft.co.il/sasha/2015/01/29/compile-time-r...
The only real limitation in C++ is that these values have to be relatively "concrete": at some level there has to be a number that can be calculated, as you can only deal with values as you don't have the amazing ability to do unification of program fragments like you have access to in Idris (like, you'd like to be able to say "the length of this array is the same as the length of this runtime string", and then after a bunch of array manipulation claim that you return a string five characters longer than the input string). Again, though: this could be added to C++ without breaking its underlying model. And it definitely is not required by the concept of "dependent types" [edit: matt-noonan points out it is, in fact, required for a language to have "dependent types" :(.]