Data.ByteString.Lazy.dropEnd: Use two-pointers technique (#629)

* Data.ByteString.Lazy.dropEnd: Use two-pointers technique

This can be seen as using the input itself as an implicit
queue; we formerly copied its chunks into an explicit queue.

By writing the key logic as a polymorphic `splitAtEndFold`,
it was easy to re-use it for `takeEnd`; the latter function
should now operate in constant stack space and can release
initial chunks of a very long input string sooner.

* Fix a very silly bug, and strengthen tests

...so that a plain 'cabal test' finds the bug almost every try
instead of finding it only every few dozen tries

* Actually work around the poison instance

(Some re-compilation check somewhere set a trap for me.)

This also replaces fromIntegral with intToIndexTy in a few places.

* Rewrite the poison instance using TypeError

* Rename "bsL" -> "toSplit" and "bsR" -> "toScan"

* Add basic benchmarks for lazy takeEnd/splitEnd

According to these benchmarks, the new implementation for takeEnd
is somewhat faster and the new implementation for dropEnd is roughly
3.5x to 4x as quick as its predecessor.

(cherry picked from commit 2bbc97e)

Data/ByteString/Lazy.hs

@@ -1,8 +1,11 @@
-{-# LANGUAGE BangPatterns #-}
 {-# OPTIONS_GHC -fno-warn-incomplete-patterns #-}
 {-# OPTIONS_HADDOCK prune #-}
 {-# LANGUAGE Trustworthy #-}
 
+{-# LANGUAGE BangPatterns #-}
+{-# LANGUAGE TypeApplications #-}
+{-# LANGUAGE ScopedTypeVariables #-}
+
 -- |
 -- Module : Data.ByteString.Lazy
 -- Copyright : (c) Don Stewart 2006
@@ -237,9 +240,9 @@ import qualified Data.ByteString as P (ByteString) -- type name only
 import qualified Data.ByteString as S -- S for strict (hmm...)
 import qualified Data.ByteString.Internal.Type as S
 import qualified Data.ByteString.Unsafe as S
-import qualified Data.ByteString.Lazy.Internal.Deque as D
 import Data.ByteString.Lazy.Internal
 
+import Control.Exception (assert)
 import Control.Monad (mplus)
 import Data.Word (Word8)
 import Data.Int (Int64)
@@ -790,15 +793,75 @@ take i cs0 = take' i cs0
 --
 -- @since 0.11.2.0
 takeEnd :: Int64 -> ByteString -> ByteString
-takeEnd i _ | i <= 0 = Empty
-takeEnd i cs0 = takeEnd' i cs0
- where takeEnd' 0 _ = Empty
- takeEnd' n cs =
- snd $ foldrChunks takeTuple (n,Empty) cs
- takeTuple _ (0, cs) = (0, cs)
- takeTuple c (n, cs)
- | n > fromIntegral (S.length c) = (n - fromIntegral (S.length c), Chunk c cs)
- | otherwise = (0, Chunk (S.takeEnd (fromIntegral n) c) cs)
+takeEnd i bs
+ | i <= 0 = Empty
+ | otherwise = splitAtEndFold (\_ res -> res) id i bs
+
+-- | Helper function for implementing 'takeEnd' and 'dropEnd'
+splitAtEndFold
+ :: forall result
+ . (S.StrictByteString -> result -> result)
+ -- ^ What to do when one chunk of output is ready
+ -- (The StrictByteString will not be empty.)
+ -> (ByteString -> result)
+ -- ^ What to do when the split-point is reached
+ -> Int64
+ -- ^ Number of bytes to leave at the end (must be strictly positive)
+ -> ByteString -- ^ Input ByteString
+ -> result
+{-# INLINE splitAtEndFold #-}
+splitAtEndFold step end len bs0 = assert (len > 0) $ case bs0 of
+ Empty -> end Empty
+ Chunk c t -> goR len c t t
+ where
+ -- Idea: Keep two references into the input ByteString:
+ -- "toSplit" tracks the current split point,
+ -- "toScan" tracks the yet-unprocessed tail.
+ -- When they are closer than "len" bytes apart, process more input. ("goR")
+ -- When they are at least "len" bytes apart, produce more output. ("goL")
+ -- We always have that "toScan" is a suffix of "toSplit",
+ -- and "toSplit" is a suffix of the original input (bs0).
+ goR :: Int64 -> S.StrictByteString -> ByteString -> ByteString -> result
+ goR !undershoot nextOutput@(S.BS noFp noLen) toSplit toScan =
+ assert (undershoot > 0) $
+ -- INVARIANT: length toSplit == length toScan + len - undershoot
+ -- (not 'assert'ed because that would break our laziness properties)
+ case toScan of
+ Empty
+ | undershoot >= intToInt64 noLen
+ -> end (Chunk nextOutput toSplit)
+ | undershootW <- fromIntegral @Int64 @Int undershoot
+ -- conversion Int64->Int is OK because 0 < undershoot < noLen
+ , splitIndex <- noLen - undershootW
+ , beforeSplit <- S.BS noFp splitIndex
+ , afterSplit <- S.BS (noFp `S.plusForeignPtr` splitIndex) undershootW
+ -> step beforeSplit $ end (Chunk afterSplit toSplit)
+
+ Chunk (S.BS _ cLen) newBsR
+ | cLen64 <- intToInt64 cLen
+ , undershoot > cLen64
+ -> goR (undershoot - cLen64) nextOutput toSplit newBsR
+ | undershootW <- fromIntegral @Int64 @Int undershoot
+ -> step nextOutput $ goL (cLen - undershootW) toSplit newBsR
+
+ goL :: Int -> ByteString -> ByteString -> result
+ goL !overshoot toSplit toScan =
+ assert (overshoot >= 0) $
+ -- INVARIANT: length toSplit == length toScan + len + intToInt64 overshoot
+ -- (not 'assert'ed because that would break our laziness properties)
+ case toSplit of
+ Empty -> splitAtEndFoldInvariantFailed
+ Chunk c@(S.BS _ cLen) newBsL
+ | overshoot >= cLen
+ -> step c $ goL (overshoot - cLen) newBsL toScan
+ | otherwise
+ -> goR (intToInt64 $ cLen - overshoot) c newBsL toScan
+
+splitAtEndFoldInvariantFailed :: a
+-- See Note [Float error calls out of INLINABLE things] in D.B.Internal.Type
+splitAtEndFoldInvariantFailed =
+ moduleError "splitAtEndFold"
+ "internal error: toSplit not longer than toScan"
 
 -- | /O(n\/c)/ 'drop' @n xs@ returns the suffix of @xs@ after the first @n@
 -- elements, or 'empty' if @n > 'length' xs@.
@@ -824,44 +887,9 @@ drop i cs0 = drop' i cs0
 --
 -- @since 0.11.2.0
 dropEnd :: Int64 -> ByteString -> ByteString
-dropEnd i p | i <= 0 = p
-dropEnd i p = go D.empty p
- where go :: D.Deque -> ByteString -> ByteString
- go deque (Chunk c cs)
- | D.byteLength deque < i = go (D.snoc c deque) cs
- | otherwise =
- let (output, deque') = getOutput empty (D.snoc c deque)
- in foldrChunks Chunk (go deque' cs) output
- go deque Empty = fromDeque $ dropEndBytes deque i
-
- len c = fromIntegral (S.length c)
-
- -- get a `ByteString` from all the front chunks of the accumulating deque
- -- for which we know they won't be dropped
- getOutput :: ByteString -> D.Deque -> (ByteString, D.Deque)
- getOutput out deque = case D.popFront deque of
- Nothing -> (reverseChunks out, deque)
- Just (x, deque') | D.byteLength deque' >= i ->
- getOutput (Chunk x out) deque'
- _ -> (reverseChunks out, deque)
-
- -- reverse a `ByteString`s chunks, keeping all internal `S.StrictByteString`s
- -- unchanged
- reverseChunks = foldlChunks (flip Chunk) empty
-
- -- drop n elements from the rear of the accumulating `deque`
- dropEndBytes :: D.Deque -> Int64 -> D.Deque
- dropEndBytes deque n = case D.popRear deque of
- Nothing -> deque
- Just (deque', x) | len x <= n -> dropEndBytes deque' (n - len x)
- | otherwise ->
- D.snoc (S.dropEnd (fromIntegral n) x) deque'
-
- -- build a lazy ByteString from an accumulating `deque`
- fromDeque :: D.Deque -> ByteString
- fromDeque deque =
- List.foldr chunk Empty (D.front deque) `append`
- List.foldl' (flip chunk) Empty (D.rear deque)
+dropEnd i p
+ | i <= 0 = p
+ | otherwise = splitAtEndFold Chunk (const Empty) i p
 
 -- | /O(n\/c)/ 'splitAt' @n xs@ is equivalent to @('take' n xs, 'drop' n xs)@.
 splitAt :: Int64 -> ByteString -> (ByteString, ByteString)
@@ -1688,6 +1716,9 @@ revNonEmptyChunks = List.foldl' (flip Chunk) Empty
 revChunks :: [P.ByteString] -> ByteString
 revChunks = List.foldl' (flip chunk) Empty
 
+intToInt64 :: Int -> Int64
+intToInt64 = fromIntegral @Int @Int64
+
 -- $IOChunk
 --
 -- ⚠ Using lazy I\/O functions like 'readFile' or 'hGetContents'

bench/BenchAll.hs

@@ -20,6 +20,8 @@ import Data.Semigroup
 import Data.String
 import Test.Tasty.Bench
 import Prelude hiding (words)
+import qualified Data.List as List
+import Control.DeepSeq
 
 import qualified Data.ByteString as S
 import qualified Data.ByteString.Char8 as S8
@@ -99,9 +101,12 @@ lazyByteStringData = case S.splitAt (nRepl `div` 2) byteStringData of
 
 {-# NOINLINE smallChunksData #-}
 smallChunksData :: L.ByteString
-smallChunksData
- = L.fromChunks [S.take sz (S.drop n byteStringData)
- | let sz = 48, n <- [0, sz .. S.length byteStringData]]
+smallChunksData = L.fromChunks $ List.unfoldr step (byteStringData, 1)
+ where
+ step (!s, !i)
+ | S.null s = Nothing
+ | otherwise = case S.splitAt i s of
+ (!s1, !s2) -> Just (s1, (s2, i * 71 `mod` 97))
 
 {-# NOINLINE byteStringChunksData #-}
 byteStringChunksData :: [S.ByteString]
@@ -419,6 +424,19 @@ main = do
  [ bench "strict" $ nf S.tails byteStringData
  , bench "lazy" $ nf L.tails lazyByteStringData
  ]
+ , bgroup "splitAtEnd (lazy)" $ let
+ testSAE op = \bs -> [op i bs | i <- [0,5..L.length bs]] `deepseq` ()
+ {-# INLINE testSAE #-}
+ in
+ [ bench "takeEnd" $
+ nf (testSAE L.takeEnd) lazyByteStringData
+ , bench "takeEnd (small chunks)" $
+ nf (testSAE L.takeEnd) smallChunksData
+ , bench "dropEnd" $
+ nf (testSAE L.dropEnd) lazyByteStringData
+ , bench "dropEnd (small chunks)" $
+ nf (testSAE L.dropEnd) smallChunksData
+ ]
  , bgroup "sort" $ map (\s -> bench (S8.unpack s) $ nf S.sort s) sortInputs
  , bgroup "stimes" $ let st = stimes :: Int -> S.ByteString -> S.ByteString
  in

bytestring.cabal

@@ -106,7 +106,6 @@ library
  Data.ByteString.Builder.RealFloat.Internal
  Data.ByteString.Builder.RealFloat.TableGenerator
  Data.ByteString.Internal.Type
- Data.ByteString.Lazy.Internal.Deque
  Data.ByteString.Lazy.ReadInt
  Data.ByteString.Lazy.ReadNat
  Data.ByteString.ReadInt

tests/Properties/ByteString.hs

@@ -80,6 +80,10 @@ import Test.Tasty
 import Test.Tasty.QuickCheck
 import QuickCheckUtils
 
+#ifdef BYTESTRING_LAZY
+import Data.Int
+#endif
+
 #ifndef BYTESTRING_CHAR8
 toElem :: Word8 -> Word8
 toElem = id
@@ -114,16 +118,25 @@ instance Arbitrary Natural where
 
 testRdInt :: forall a. (Arbitrary a, RdInt a) => String -> TestTree
 testRdInt s = testGroup s $
- [ testProperty "from string" $ \ prefix value suffix ->
+ [ testProperty "from string" $ int64OK $ \value prefix suffix ->
  let si = show @a value
  b = prefix <> B.pack si <> suffix
  in fmap (second B.unpack) (bread @a b)
  === sread @a (B.unpack prefix ++ si ++ B.unpack suffix)
- , testProperty "from number" $ \n ->
+ , testProperty "from number" $ int64OK $ \n ->
  bread @a (B.pack (show n)) === Just (n, B.empty)
  ]
 #endif
 
+intToIndexTy :: Int -> IndexTy
+#ifdef BYTESTRING_LAZY
+type IndexTy = Int64
+intToIndexTy = fromIntegral @Int @Int64
+#else
+type IndexTy = Int
+intToIndexTy = id
+#endif
+
 tests :: [TestTree]
 tests =
  [ testProperty "pack . unpack" $
@@ -308,7 +321,7 @@ tests =
 #endif
 
  , testProperty "drop" $
- \n x -> B.unpack (B.drop n x) === List.genericDrop n (B.unpack x)
+ \(intToIndexTy -> n) x -> B.unpack (B.drop n x) === List.genericDrop n (B.unpack x)
  , testProperty "drop 10" $
  \x -> let n = 10 in B.unpack (B.drop n x) === List.genericDrop n (B.unpack x)
  , testProperty "drop 2^31" $
@@ -325,7 +338,7 @@ tests =
 #endif
 
  , testProperty "take" $
- \n x -> B.unpack (B.take n x) === List.genericTake n (B.unpack x)
+ \(intToIndexTy -> n) x -> B.unpack (B.take n x) === List.genericTake n (B.unpack x)
  , testProperty "take 10" $
  \x -> let n = 10 in B.unpack (B.take n x) === List.genericTake n (B.unpack x)
  , testProperty "take 2^31" $
@@ -342,11 +355,11 @@ tests =
 #endif
 
  , testProperty "dropEnd" $
- \n x -> B.dropEnd n x === B.take (B.length x - n) x
+ \(intToIndexTy -> n) x -> B.dropEnd n x === B.take (B.length x - n) x
  , testProperty "dropWhileEnd" $
  \f x -> B.dropWhileEnd f x === B.reverse (B.dropWhile f (B.reverse x))
  , testProperty "takeEnd" $
- \n x -> B.takeEnd n x === B.drop (B.length x - n) x
+ \(intToIndexTy -> n) x -> B.takeEnd n x === B.drop (B.length x - n) x
  , testProperty "takeWhileEnd" $
  \f x -> B.takeWhileEnd f x === B.reverse (B.takeWhile f (B.reverse x))
 
@@ -366,7 +379,7 @@ tests =
  , testProperty "compareLength 4" $
  \x (toElem -> c) -> B.compareLength (B.snoc x c <> undefined) (B.length x) === GT
  , testProperty "compareLength 5" $
- \x n -> B.compareLength x n === compare (B.length x) n
+ \x (intToIndexTy -> n) -> B.compareLength x n === compare (B.length x) n
  , testProperty "dropEnd lazy" $
  \(toElem -> c) -> B.take 1 (B.dropEnd 1 (B.singleton c <> B.singleton c <> B.singleton c <> undefined)) === B.singleton c
  , testProperty "dropWhileEnd lazy" $
@@ -470,7 +483,8 @@ tests =
  (l1 == l2 || l1 == l2 + 1) && sum (map B.length splits) + l2 == B.length x
 
  , testProperty "splitAt" $
- \n x -> (B.unpack *** B.unpack) (B.splitAt n x) === List.genericSplitAt n (B.unpack x)
+ \(intToIndexTy -> n) x -> (B.unpack *** B.unpack) (B.splitAt n x)
+ === List.genericSplitAt n (B.unpack x)
  , testProperty "splitAt 10" $
  \x -> let n = 10 in (B.unpack *** B.unpack) (B.splitAt n x) === List.genericSplitAt n (B.unpack x)
  , testProperty "splitAt (2^31)" $
@@ -594,13 +608,13 @@ tests =
 #endif
 
  , testProperty "index" $
- \(NonNegative n) x -> fromIntegral n < B.length x ==> B.index x (fromIntegral n) === B.unpack x !! n
+ \(NonNegative n) x -> intToIndexTy n < B.length x ==> B.index x (intToIndexTy n) === B.unpack x !! n
  , testProperty "indexMaybe" $
- \(NonNegative n) x -> fromIntegral n < B.length x ==> B.indexMaybe x (fromIntegral n) === Just (B.unpack x !! n)
+ \(NonNegative n) x -> intToIndexTy n < B.length x ==> B.indexMaybe x (intToIndexTy n) === Just (B.unpack x !! n)
  , testProperty "indexMaybe Nothing" $
- \n x -> (n :: Int) < 0 || fromIntegral n >= B.length x ==> B.indexMaybe x (fromIntegral n) === Nothing
+ \n x -> n < 0 || intToIndexTy n >= B.length x ==> B.indexMaybe x (intToIndexTy n) === Nothing
  , testProperty "!?" $
- \n x -> B.indexMaybe x (fromIntegral (n :: Int)) === x B.!? (fromIntegral n)
+ \(intToIndexTy -> n) x -> B.indexMaybe x n === x B.!? n
 
 #ifdef BYTESTRING_CHAR8
  , testProperty "isString" $