So, purely functional languages have their own class of potentials due to the clear separation between pure and impure code. I have seen several features that are somewhat simpler to implement in Haskell like Nested Data Parallelism or Stream Fusion.
My question is, what are other improvements/optimizations that are more or less unique to Haskell in terms of feasibility/simplicity but not yet implemented? (I mostly care about GHC, but also love to hear about others)
One optimization I'd love to see in GHC is supercompilation. That seems unlikely in the near-future of GHC, though, because it's whole-program optimization, and GHC is very focused on module-at-a-time compilation.
Basically, supercompilation is executing as much of a program as possible at compile time. It naturally subsumes inlining, deforestation, specialization, and any number of other techniques. Early experimental results have been promising, but it's a very expensive process. It's hard to see it being a practical optimization, but the concept is ridiculously awesome.