Foldl' doesn't inline when partially applied

Summary

It is common to write functions that iterate over a structure using foldl' in an eta-reduced way, e.g. from a recent reddit thread:

-- SumLen.hs:
{-# LANGUAGE BangPatterns #-}
module SumLen where

import Data.List

sumlen :: [Float] -> (Float, Int)
sumlen = foldl' f (0, 0)
  where
    f (!s, !n) !x = (s + x, n + 1)

This, however, does not produce performant Core, the cleaned up Core looks like this:

sumlen4
  = \ ds x ->
      case ds of { (s, n) ->
      case s of { F# ipv ->
      case n of { I# ipv1 ->
      case x of { F# ipv2 ->
      (F# (plusFloat# ipv ipv2), I# (+# ipv1 1#))
      }
      }
      }
      }

sumlen = foldl' sumlen4 (F# 0.0#, I# 0#)

If instead we eta-expand the sumlen function:

sumlen :: [Float] -> (Float, Int)
sumlen xs = foldl' f (0, 0) xs
  where
    f (!s, !n) !x = (s + x, n + 1)

We do get efficient Core that does not allocate during the loop:

sumlen_$s$wgo
  = \ sc sc1 sc2 ->
      case sc2 of {
        [] -> (# F# sc1, I# sc #);
        : y ys ->
          case y of { F# ipv ->
          sumlen_$s$wgo (+# sc 1#) (plusFloat# sc1 ipv) ys
          }
      }

sumlen
  = \ w ->
      case sumlen_$s$wgo 0# 0.0# w of { (# ww1, ww2 #) -> (ww1, ww2) }

Steps to reproduce

Put the SumLen example in a file SumLen.hs and compile it with optimizations and simplifier dumping:

$ ghc -O2 -ddump-simpl -dsuppress-all -dsuppress-uniques SumLen.hs

This produces Core like the first example shown above.

Expected behavior

I expected it to produce more efficient Core like the second example shown above.

Environment

• GHC version used: 8.10.4 and 9.0.1

Optional:

• Operating System: Debian buster
• System Architecture: x86_64