Oneshot eta-expanding of Sem

[isovector]

I was reading this https://www.joachim-breitner.de/blog/763-Faster_Winter_5__Eta-Expanding_ReaderT, and it mentions using oneShot to convince GHC to float lambdas out. I suspect this is why the common pattern in polysemy of foo (Sem m) = Sem $ \k -> m $ \u -> blah performs better than the naive foo = runSem $ \u -> blah.

We can probably clean up a lot of this roundabout code by doing the same thing, and putting oneShot into the common combinators.

[KingoftheHomeless]

Impredicative types complicates implementing an etaSem à la etaReaderT. We have to work through a newtype to use oneShot.
If we want etaSem, we have one of two options: using coerce, or eta-expanding the argument to oneShot:

newtype Trans f g = Trans { runTrans :: forall x. f x -> g x}

etaSem :: forall r a. Sem r a -> Sem r a
etaSem (Sem m) = Sem $ \(k :: forall x. Union r (Sem r) x -> m x) ->
  oneShot
    (coerce (m :: (forall x. Union r (Sem r) x -> m x) -> m a))
    (Trans k)

etaSem' :: Sem r a -> Sem r a
etaSem' (Sem m) = Sem $ \k -> oneShot (\(Trans k') -> m k') (Trans k)

I don't know enough about the simplifier or oneShot to tell if either of these work. etaSem uses coerce, which, if the simplifier treats it as it does unsafeCoerce, causes it not to touch m anyway. etaSem''s eta-expansion on the argument could prevent lambda-floating if oneShot only looks one lambda deep.