FloatOut reverts eta expansion of thunks

Call Arity works hard to identify thunks that can be eta expanded, because they are only called once, with arity > 0. As it does for lvl this example

{-# OPTIONS_GHC -O2 -fforce-recomp #-}
{-# LANGUAGE BangPatterns #-}

module Lib (main) where

expensive :: a -> a
expensive x = x
{-# NOINLINE expensive #-}

lvl :: Int -> Int
lvl = expensive (+ 42)
{-# NOINLINE lvl #-}

go :: Int -> Int -> Int
go 0 x = lvl x
go 1 x = lvl (x+1)
go n x = go (n-1) (x+1)

res :: Int
res = go 15 25

main :: IO ()
main = print res

Before SetLevels (after Call Arity), indeed there is

lvl
  = \ (eta_B1 :: Int) ->
      expensive
        @ (Int -> Int)
        (\ (ds_d2K7 :: Int) ->
           case ds_d2K7 of { GHC.Types.I# x_a2KB [Dmd=<S,U>] ->
           GHC.Types.I# (GHC.Prim.+# x_a2KB 42#)
           })
        eta_B1

lvl got eta expanded. But if you look at the output of -ddump-simpl, you'll see

lvl2_r3jd = expensive_rqN @ (Int -> Int) lvl1_r3dx

which is an un-eta-expanded form of lvl above. lvl2 is due to FloatOut, which identifies expensive (+ 42) as a MFE of lvl = \eta. expensive (+ 42) eta.

I'm proposing that lvl should stay eta expanded. Here's how to ensure that:

That \eta. should have a one-shot annotation. The only world in which it is OK to eta-expand a thunk is when the new manifest lambdas are one-shot (and that the thunk is not otherwise seqed, of course). And then we have to make sure that we preserve that one-shot-ness (#18355), so that FloatIn and the Simplifier will be able to float it back in.

Preserving that one-shot-ness is the hard part, but fortunately that seems to have been solved by Kinds are Calling Conventions.