Ok cool. Sounds like we agree. Luite and I talked about something similar a few months ago, I'll try to dig up my notes and maybe we can turn this into a ghc feature request or patch
Great, I would definitely help with that!
are you sure you've evaluated how it interacts with this sort of optimization? I think it actually gets you pretty far!
Most of the magic occurs in reification. The trouble is that it expects a polymorphic function of rank 2 with a nontrivial context (Reifies), a function that by construction cannot be specialised, unless you specialise the receiving function (reify) for every application case. The beauty of reflection is that it's free by virtue of sharing. Unfortunately sharing is the exact opposite of inlining.
According to a benchmark I've done a few months ago it behaves exactly as if the reflected value was just a shared argument with no inlining performed; an expected and reasonable result.
What I'm really after is a sort of controlled inlining. That's pretty much what instance-based specialisation currently does for dictionaries. Technically a dictionary is just another argument, so there is no fundamental reason why we shouldn't have a more general specialiser.