Link to draft paper describing the project: https://github.com/nicolasstucki/specialized/blob/master/docs/2013-03-scala-workshop-draft.pdf
Generic code increases programmer productivity as it increases code reuse. For example, the LinkedList abstraction can be used in many contexts, from storing a list of numbers to storing representations of files on the disk. Unfortunately this comes at the expense of performance, as the generic code needs a common representation for all types. The common representation is usually a pointer to heap data. But for value types, such as integers, bytes and even value classes (see SIP-15) this leads to significant overheads, as they need to be allocated as objects on the heap and then pointed to, breaking data locality and adding an extra indirection.
To overcome this performance loss, many programming languages and virtual machines perform code specialization, which creates a copy of the generic class for each value type and adapts the data representation.
Scala does not require type parameters to be known at compile time, thus allowing truly-generic code to be generated. In this case, type information is not necessary and will not available in the generated bytecode. But this prohibits programmers from using specialized code from generic code, which might be desirable.
Scala can overcome the loss of type information in generic classes by attaching types in so-called ClassTags (formerly known as Manifests). This information can be used to dispatch to the correct specialized class implementation, and can be made either as a compiler plugin or as a macro. Either implementation is fine, as long as the syntactic overhead is reasonable. For this project one should research what is the best implementation and prepare a set of benchmarks that clearly show the performance hit of dispatching based on the ClassTag.