INLINE hash to retain fusion with low-arity foldl'

Originally, hash was eta-reduced and not explicitly inlined, and as a result the foldl' call here was not saturated. Thus the definition of hash was simple enough to inline into the call site, where the provided string allowed foldl' to inline into foldr, which then enabled fusion, avoiding potentially non-trivial allocation overhead.

In ghc!5259 (merged), we're reducing the arity of foldl', so that it can inline with just two arguments. This yields performance improvements in common idiomatic code, see ghc!5259 (merged) for details.

That reduced arity makes foldl' inline into even the eta-reduced hash, which then (for lack of an INLINE here) also inlined foldr and the fusion opportunity was lost.

Quoting Simon

* let's add that INLINE to hash in NofibUtils. We don't usually mess with
  nofib, but we don't want to perpetuate reliance on a fluke.

In addition, eta-expanding hash. It will now saturate both the original and the post-ghc!5259 (merged) foldl'.

common/NofibUtils.hs

@@ -10,7 +10,31 @@ import System.Environment (getArgs)
 -- | A very simple hash function so that we don't have to store and compare
 -- huge output files.
 hash :: String -> Int
-hash = foldl' (\acc c -> ord c + acc*31) 0
+hash str = foldl' (\acc c -> ord c + acc*31) 0 str
+{-# INLINE hash #-}
+{-
+Note: Originally, `hash` was eta-reduced and not explicitly inlined, and as a
+result the `foldl'` call here was not saturated.  Thus the definition of `hash`
+was simple enough to inline into the call site, where the provided string
+allowed `foldl'` to inline into `foldr`, which then enabled fusion, avoiding
+potentially non-trivial allocation overhead.
+
+In !5259, we're reducing the arity of `foldl'`, so that it can inline with just
+two arguments.  This yields performance improvements in common idiomatic code,
+see !5259 for details.
+
+That reduced arity makes `foldl'` inline into even the eta-reduced `hash`,
+which then (for lack of an INLINE here) also inlined `foldr` and the fusion
+opportunity was lost.
+
+Quoting Simon, <https://gitlab.haskell.org/ghc/ghc/-/merge_requests/5259#note_341945>
+
+    * let's add that INLINE to hash in NofibUtils. We don't usually mess with
+      nofib, but we don't want to perpetuate reliance on a fluke.
+
+In addition, eta-expanding `hash`.  It will now saturate both the original
+and the post-!5259 `foldl'`.
+-}
 
 -- | Using @salt xs@ on an loop-invariant @xs@ inside a loop prevents the
 -- compiler from floating out the input parameter.
@@ -27,4 +51,4 @@ salt xs = do
   --            executable, otherwise this isn't really 'pure'
   --            anymore.
   pure (take (max (maxBound - 1) s) xs)
-#endif
\ No newline at end of file
+#endif