However, I suggest avoiding trivially reducible problems like computing constants (e, pi, primes, fib) and redundant operations (binary trees). Make sure programs accept a non-trivial input (even if it is just an int over a wide range). Avoid unnecessary repeats (e.g. atom.hs). This will mean that transformations that improve performance on the benchmark suite will be more likely to improve the performance of real programs.
May I suggest using pureMD5 as one of the benchmarks? It makes sense in my mind that a tool with a real use and consistent run times be used. Not sure if I would consider the rolled (concise) or unrolled (ugly) version a better fit (or both). The rolled version showed excellent speed increase due to pointer tagging (17sec down from 27sec for hashing 200MB). On the other hand, the unrolled version is the obvious 'make as ugly a Haskell program as you can to complete with C'.