Cuda ( 4.1 or future), LLVM and linking

Hello everybody,

Since I have Heard about the use of LLVM as compiler in cuda, one question arise in my mind.
Will we be able to ‘link’ multiples compilation units in a near or far future?

I explain the context where it will be useful for me:
I am working on an application that use an algorithm that is partially configurable by the end user.
( he might choose, which ‘noise’ he uses, some function in a pre-implemented pool to use, parameters, …)
The way we use cuda now is to pre-compile the algorithm and the pool of function in a ptx, edit the ptx at runtime according to the end-user wishes, and gives everything to the driver to compile. It works very well except that the final ptx can grow a lot and be very long to compile ( up to 10 minutes ).
We would be delighted to have a way to compile the ptxs of the algorithm and the pool of function during the initialization of the application and be able to link with external functions generated at runtime.
It would be wonderful

Thanks a lot