1 {-# OPTIONS_GHC -fno-implicit-prelude #-}
3 -----------------------------------------------------------------------------
5 -- Module : Data.HashTable
6 -- Copyright : (c) The University of Glasgow 2003
7 -- License : BSD-style (see the file libraries/base/LICENSE)
9 -- Maintainer : libraries@haskell.org
10 -- Stability : provisional
11 -- Portability : portable
13 -- An implementation of extensible hash tables, as described in
14 -- Per-Ake Larson, /Dynamic Hash Tables/, CACM 31(4), April 1988,
15 -- pp. 446--457. The implementation is also derived from the one
16 -- in GHC's runtime system (@ghc\/rts\/Hash.{c,h}@).
18 -----------------------------------------------------------------------------
20 module Data.HashTable (
21 -- * Basic hash table operations
22 HashTable, new, insert, delete, lookup, update,
23 -- * Converting to and from lists
33 -- This module is imported by Data.Dynamic, which is pretty low down in the
34 -- module hierarchy, so don't import "high-level" modules
36 #ifdef __GLASGOW_HASKELL__
39 import Prelude hiding ( lookup )
41 import Data.Tuple ( fst )
44 import Data.List ( maximumBy, length, concat, foldl', partition )
45 import Data.Int ( Int32 )
47 #if defined(__GLASGOW_HASKELL__)
49 import GHC.Real ( fromIntegral )
50 import GHC.Show ( Show(..) )
51 import GHC.Int ( Int64 )
53 import GHC.IOBase ( IO, IOArray, newIOArray,
54 unsafeReadIOArray, unsafeWriteIOArray, unsafePerformIO,
55 IORef, newIORef, readIORef, writeIORef )
57 import Data.Char ( ord )
58 import Data.IORef ( IORef, newIORef, readIORef, writeIORef )
59 import System.IO.Unsafe ( unsafePerformIO )
60 import Data.Int ( Int64 )
61 # if defined(__HUGS__)
62 import Hugs.IOArray ( IOArray, newIOArray,
63 unsafeReadIOArray, unsafeWriteIOArray )
64 # elif defined(__NHC__)
65 import NHC.IOExtras ( IOArray, newIOArray, readIOArray, writeIOArray )
68 import Control.Monad ( mapM, mapM_, sequence_ )
71 -----------------------------------------------------------------------
76 -----------------------------------------------------------------------
78 readHTArray :: HTArray a -> Int32 -> IO a
79 writeMutArray :: MutArray a -> Int32 -> a -> IO ()
80 freezeArray :: MutArray a -> IO (HTArray a)
81 thawArray :: HTArray a -> IO (MutArray a)
82 newMutArray :: (Int32, Int32) -> a -> IO (MutArray a)
83 #if defined(DEBUG) || defined(__NHC__)
84 type MutArray a = IOArray Int32 a
85 type HTArray a = MutArray a
86 newMutArray = newIOArray
87 readHTArray = readIOArray
88 writeMutArray = writeIOArray
92 type MutArray a = IOArray Int32 a
93 type HTArray a = MutArray a -- Array Int32 a
94 newMutArray = newIOArray
95 readHTArray arr i = readMutArray arr i -- return $! (unsafeAt arr (fromIntegral i))
96 readMutArray :: MutArray a -> Int32 -> IO a
97 readMutArray arr i = unsafeReadIOArray arr (fromIntegral i)
98 writeMutArray arr i x = unsafeWriteIOArray arr (fromIntegral i) x
99 freezeArray = return -- unsafeFreeze
100 thawArray = return -- unsafeThaw
103 data HashTable key val = HashTable {
104 cmp :: !(key -> key -> Bool),
105 hash_fn :: !(key -> Int32),
106 tab :: !(IORef (HT key val))
108 -- TODO: the IORef should really be an MVar.
112 kcount :: !Int32, -- Total number of keys.
114 buckets :: !(HTArray [(key,val)])
117 -- ------------------------------------------------------------
118 -- Instrumentation for performance tuning
120 -- This ought to be roundly ignored after optimization when
121 -- iNSTRUMENTED=False.
123 -- STRICT version of modifyIORef!
124 modifyIORef :: IORef a -> (a -> a) -> IO ()
127 let z = f v in z `seq` writeIORef r z
131 insertions :: !Integer,
133 totBuckets :: !Integer,
134 maxEntries :: !Int32,
137 } deriving (Eq, Show)
139 {-# NOINLINE hashData #-}
140 hashData :: IORef HashData
141 hashData = unsafePerformIO (newIORef (HD { tables=0, insertions=0, lookups=0,
142 totBuckets=0, maxEntries=0,
143 maxChain=0, maxBuckets=tABLE_MIN } ))
145 instrument :: (HashData -> HashData) -> IO ()
146 instrument i | iNSTRUMENTED = modifyIORef hashData i
147 | otherwise = return ()
150 recordNew = instrument rec
151 where rec hd@HD{ tables=t, totBuckets=b } =
152 hd{ tables=t+1, totBuckets=b+fromIntegral tABLE_MIN }
154 recordIns :: Int32 -> Int32 -> [a] -> IO ()
155 recordIns i sz bkt = instrument rec
156 where rec hd@HD{ insertions=ins, maxEntries=mx, maxChain=mc } =
157 hd{ insertions=ins+fromIntegral i, maxEntries=mx `max` sz,
158 maxChain=mc `max` length bkt }
160 recordResize :: Int32 -> Int32 -> IO ()
161 recordResize older newer = instrument rec
162 where rec hd@HD{ totBuckets=b, maxBuckets=mx } =
163 hd{ totBuckets=b+fromIntegral (newer-older),
164 maxBuckets=mx `max` newer }
166 recordLookup :: IO ()
167 recordLookup = instrument lkup
168 where lkup hd@HD{ lookups=l } = hd{ lookups=l+1 }
170 -- stats :: IO String
171 -- stats = fmap show $ readIORef hashData
173 -- ----------------------------------------------------------------------------
174 -- Sample hash functions
178 -- This implementation of hash tables uses the low-order /n/ bits of the hash
179 -- value for a key, where /n/ varies as the hash table grows. A good hash
180 -- function therefore will give an even distribution regardless of /n/.
182 -- If your keyspace is integrals such that the low-order bits between
183 -- keys are highly variable, then you could get away with using 'fromIntegral'
184 -- as the hash function.
186 -- We provide some sample hash functions for 'Int' and 'String' below.
189 golden = 1013904242 -- = round ((sqrt 5 - 1) * 2^32) :: Int32
190 -- was -1640531527 = round ((sqrt 5 - 1) * 2^31) :: Int32
191 -- but that has bad mulHi properties (even adding 2^32 to get its inverse)
192 -- Whereas the above works well and contains no hash duplications for
195 hashInt32 :: Int32 -> Int32
196 hashInt32 x = mulHi x golden + x
198 -- | A sample (and useful) hash function for Int and Int32,
199 -- implemented by extracting the uppermost 32 bits of the 64-bit
200 -- result of multiplying by a 33-bit constant. The constant is from
201 -- Knuth, derived from the golden ratio:
202 -- > golden = round ((sqrt 5 - 1) * 2^32)
203 -- We get good key uniqueness on small inputs
204 -- (a problem with previous versions):
205 -- (length $ group $ sort $ map hashInt [-32767..65536]) == 65536 + 32768
207 hashInt :: Int -> Int32
208 hashInt x = hashInt32 (fromIntegral x)
210 -- hi 32 bits of a x-bit * 32 bit -> 64-bit multiply
211 mulHi :: Int32 -> Int32 -> Int32
212 mulHi a b = fromIntegral (r `shiftR` 32)
214 r = fromIntegral a * fromIntegral b
216 -- | A sample hash function for Strings. We keep multiplying by the
217 -- golden ratio and adding. The implementation is:
219 -- > hashString = foldl' f golden
220 -- > where f m c = fromIntegral (fromEnum c) * magic + hashInt32 m
221 -- > magic = 0xdeadbeef
223 -- Where hashInt32 works just as hashInt shown above.
225 -- Knuth argues that repeated multiplication by the golden ratio
226 -- will minimize gaps in the hash space, and thus it's a good choice
227 -- for combining together multiple keys to form one.
229 -- Here we know that individual characters c are often small, and this
230 -- produces frequent collisions if we use fromEnum c alone. A
231 -- particular problem are the shorter low ASCII and ISO-8859-1
232 -- character strings. We pre-multiply by a magic twiddle factor to
233 -- obtain a good distribution. In fact, given the following test:
235 -- > testp :: Int32 -> Int
236 -- > testp k = (n - ) . length . group . sort . map hs . take n $ ls
237 -- > where ls = [] : [c : l | l <- ls, c <- ['\0'..'\xff']]
238 -- > hs = foldl' f golden
239 -- > f m c = fromIntegral (fromEnum c) * k + hashInt32 m
242 -- We discover that testp magic = 0.
244 hashString :: String -> Int32
245 hashString = foldl' f golden
246 where f m c = fromIntegral (fromEnum c) * magic + hashInt32 m
249 -- | A prime larger than the maximum hash table size
253 -- -----------------------------------------------------------------------------
257 tABLE_MAX = 32 * 1024 * 1024 -- Maximum size of hash table
262 hLOAD = 7 -- Maximum average load of a single hash bucket
265 hYSTERESIS = 64 -- entries to ignore in load computation
267 {- Hysteresis favors long association-list-like behavior for small tables. -}
269 -- -----------------------------------------------------------------------------
270 -- Creating a new hash table
272 -- | Creates a new hash table. The following property should hold for the @eq@
273 -- and @hash@ functions passed to 'new':
275 -- > eq A B => hash A == hash B
278 :: (key -> key -> Bool) -- ^ @eq@: An equality comparison on keys
279 -> (key -> Int32) -- ^ @hash@: A hash function on keys
280 -> IO (HashTable key val) -- ^ Returns: an empty hash table
284 -- make a new hash table with a single, empty, segment
285 let mask = tABLE_MIN-1
286 bkts' <- newMutArray (0,mask) []
287 bkts <- freezeArray bkts'
291 ht = HT { buckets=bkts, kcount=kcnt, bmask=mask }
294 return (HashTable { tab=table, hash_fn=hash, cmp=cmpr })
296 -- -----------------------------------------------------------------------------
297 -- Inserting a key\/value pair into the hash table
299 -- | Inserts a key\/value mapping into the hash table.
301 -- Note that 'insert' doesn't remove the old entry from the table -
302 -- the behaviour is like an association list, where 'lookup' returns
303 -- the most-recently-inserted mapping for a key in the table. The
304 -- reason for this is to keep 'insert' as efficient as possible. If
305 -- you need to update a mapping, then we provide 'update'.
307 insert :: HashTable key val -> key -> val -> IO ()
310 updatingBucket CanInsert (\bucket -> ((key,val):bucket, 1, ())) ht key
313 -- ------------------------------------------------------------
314 -- The core of the implementation is lurking down here, in findBucket,
315 -- updatingBucket, and expandHashTable.
317 tooBig :: Int32 -> Int32 -> Bool
318 tooBig k b = k-hYSTERESIS > hLOAD * b
320 -- index of bucket within table.
321 bucketIndex :: Int32 -> Int32 -> Int32
322 bucketIndex mask h = h .&. mask
324 -- find the bucket in which the key belongs.
325 -- returns (key equality, bucket index, bucket)
327 -- This rather grab-bag approach gives enough power to do pretty much
328 -- any bucket-finding thing you might want to do. We rely on inlining
329 -- to throw away the stuff we don't want. I'm proud to say that this
330 -- plus updatingBucket below reduce most of the other definitions to a
331 -- few lines of code, while actually speeding up the hashtable
332 -- implementation when compared with a version which does everything
334 {-# INLINE findBucket #-}
335 findBucket :: HashTable key val -> key -> IO (HT key val, Int32, [(key,val)])
336 findBucket HashTable{ tab=ref, hash_fn=hash} key = do
337 table@HT{ buckets=bkts, bmask=b } <- readIORef ref
338 let indx = bucketIndex b (hash key)
339 bucket <- readHTArray bkts indx
340 return (table, indx, bucket)
342 data Inserts = CanInsert
346 -- updatingBucket is the real workhorse of all single-element table
347 -- updates. It takes a hashtable and a key, along with a function
348 -- describing what to do with the bucket in which that key belongs. A
349 -- flag indicates whether this function may perform table insertions.
350 -- The function returns the new contents of the bucket, the number of
351 -- bucket entries inserted (negative if entries were deleted), and a
352 -- value which becomes the return value for the function as a whole.
353 -- The table sizing is enforced here, calling out to expandSubTable as
356 -- This function is intended to be inlined and specialized for every
357 -- calling context (eg every provided bucketFn).
358 {-# INLINE updatingBucket #-}
360 updatingBucket :: Inserts -> ([(key,val)] -> ([(key,val)], Int32, a)) ->
361 HashTable key val -> key ->
363 updatingBucket canEnlarge bucketFn
364 ht@HashTable{ tab=ref, hash_fn=hash } key = do
365 (table@HT{ kcount=k, buckets=bkts, bmask=b },
366 indx, bckt) <- findBucket ht key
367 (bckt', inserts, result) <- return $ bucketFn bckt
369 table1 = table { kcount=k' }
370 bkts' <- thawArray bkts
371 writeMutArray bkts' indx bckt'
373 table2 <- if canEnlarge == CanInsert && inserts > 0 then do
374 recordIns inserts k' bckt'
376 then expandHashTable hash table1
379 writeIORef ref table2
382 expandHashTable :: (key -> Int32) -> HT key val -> IO (HT key val)
383 expandHashTable hash table@HT{ buckets=bkts, bmask=mask } = do
386 newmask = mask + mask + 1
387 recordResize oldsize (newmask+1)
389 if newmask > tABLE_MAX-1
393 newbkts' <- newMutArray (0,newmask) []
396 splitBucket oldindex = do
397 bucket <- readHTArray bkts oldindex
399 partition ((oldindex==). bucketIndex newmask . hash . fst) bucket
400 writeMutArray newbkts' oldindex oldb
401 writeMutArray newbkts' (oldindex + oldsize) newb
402 mapM_ splitBucket [0..mask]
404 newbkts <- freezeArray newbkts'
406 return ( table{ buckets=newbkts, bmask=newmask } )
408 -- -----------------------------------------------------------------------------
409 -- Deleting a mapping from the hash table
411 -- Remove a key from a bucket
412 deleteBucket :: (key -> Bool) -> [(key,val)] -> ([(key, val)], Int32, ())
413 deleteBucket _ [] = ([],0,())
414 deleteBucket del (pair@(k,_):bucket) =
415 case deleteBucket del bucket of
416 (bucket', dels, _) | del k -> dels' `seq` (bucket', dels', ())
417 | otherwise -> (pair:bucket', dels, ())
418 where dels' = dels - 1
420 -- | Remove an entry from the hash table.
421 delete :: HashTable key val -> key -> IO ()
423 delete ht@HashTable{ cmp=eq } key =
424 updatingBucket Can'tInsert (deleteBucket (eq key)) ht key
426 -- -----------------------------------------------------------------------------
427 -- Updating a mapping in the hash table
429 -- | Updates an entry in the hash table, returning 'True' if there was
430 -- already an entry for this key, or 'False' otherwise. After 'update'
431 -- there will always be exactly one entry for the given key in the table.
433 -- 'insert' is more efficient than 'update' if you don't care about
434 -- multiple entries, or you know for sure that multiple entries can't
435 -- occur. However, 'update' is more efficient than 'delete' followed
437 update :: HashTable key val -> key -> val -> IO Bool
439 update ht@HashTable{ cmp=eq } key val =
440 updatingBucket CanInsert
441 (\bucket -> let (bucket', dels, _) = deleteBucket (eq key) bucket
442 in ((key,val):bucket', 1+dels, dels/=0))
445 -- -----------------------------------------------------------------------------
446 -- Looking up an entry in the hash table
448 -- | Looks up the value of a key in the hash table.
449 lookup :: HashTable key val -> key -> IO (Maybe val)
451 lookup ht@HashTable{ cmp=eq } key = do
453 (_, _, bucket) <- findBucket ht key
454 let firstHit (k,v) r | eq key k = Just v
456 return (foldr firstHit Nothing bucket)
458 -- -----------------------------------------------------------------------------
459 -- Converting to/from lists
461 -- | Convert a list of key\/value pairs into a hash table. Equality on keys
462 -- is taken from the Eq instance for the key type.
464 fromList :: (Eq key) => (key -> Int32) -> [(key,val)] -> IO (HashTable key val)
465 fromList hash list = do
466 table <- new (==) hash
467 sequence_ [ insert table k v | (k,v) <- list ]
470 -- | Converts a hash table to a list of key\/value pairs.
472 toList :: HashTable key val -> IO [(key,val)]
473 toList = mapReduce id concat
475 {-# INLINE mapReduce #-}
476 mapReduce :: ([(key,val)] -> r) -> ([r] -> r) -> HashTable key val -> IO r
477 mapReduce m r HashTable{ tab=ref } = do
478 HT{ buckets=bckts, bmask=b } <- readIORef ref
479 fmap r (mapM (fmap m . readHTArray bckts) [0..b])
481 -- -----------------------------------------------------------------------------
484 -- | This function is useful for determining whether your hash
485 -- function is working well for your data set. It returns the longest
486 -- chain of key\/value pairs in the hash table for which all the keys
487 -- hash to the same bucket. If this chain is particularly long (say,
488 -- longer than 14 elements or so), then it might be a good idea to try
489 -- a different hash function.
491 longestChain :: HashTable key val -> IO [(key,val)]
492 longestChain = mapReduce id (maximumBy lengthCmp)
493 where lengthCmp (_:x)(_:y) = lengthCmp x y