The test simplCore/should_compile/simpl015 was leading to massive compile times. This is weird: it looks like:

main = do { return ()
          ; return ()
          ... hundreds more times ...
          ; return () }

It turns out that the cause was over-eager eta-expansion. Each return () gives a top-level PAP thus

lvl17 :: IO ()
lvl17 = returnIO ()

But these definitions are then eta-expanded, thus:

lvl17 = ((\eta. returnIO () |> sym g) eta) |> g

    where   g :: State# RealWorld -> (# State# RealWorld, () #)
                 ~
                 IO ()

Now in general it makes sense to eta-expand through newtypes (as is done here), because it exposes more lambadas. But it really doesn't make sense to eta-expand a PAP like this.

Fortunately the fix is easy: use exprArity rather than manifestArity in SimplUtils.tryEtaExpandRhs.

The effect on simpl015 is dramatic: compiler allocation drops from 6.6G to 812M; and residency drops from 1.8G to 45M.