Performance regression in 9.2.4 and 9.4.2 compared to 9.0.2

This is the code (adapted from this benchmark):

{-# LANGUAGE BangPatterns #-}

import Control.Parallel.Strategies
import Data.Word
import System.Environment

data Tree = Leaf Word32 | Node Tree Tree

unravel :: Int -> Tree -> [Tree]
unravel 0 x = [x]
unravel _ x@Leaf{} = [x]
unravel n (Node x y) = unravel (n - 1) x ++ unravel (n - 1) y

-- Creates a tree with 2^n elements
gen 0 = Leaf 1
gen n = Node (gen(n - 1)) (gen(n - 1))

-- Adds all elements of a tree
sun !s (Leaf x)   = x + s
sun !s (Node a b) = sun (sun s b) a

sun' :: Tree -> Word32
sun' = sum . parMap rseq (sun 0) . unravel 6

-- Performs 2^n additions
main = do
  n <- read.head <$> getArgs :: IO Word32
  print $ sun' (gen n)

Here are my timings for different ghc versions (compiled with -O2 -threaded and running with ./T5 30 +RTS -s -N8 on an 8 core machine):

GHC	Total	MUT	GC
9.0.2	3.552s	1.577s	1.965s
9.2.4	21.721s	13.308s	8.399s
9.4.2	6.858s	4.378s	2.468s