d31ecd5 · text · git.iko.soy

git.iko.soy/text

on master help sign in

Speed up partialCopyI when data is aligned

d31ecd5Bryan O'Sullivan16 years ago

Data/Text/Array.hs+46−9

compare with master

unifiedsplit

@@ -57,9 +57,10 @@ if (_k_) < 0 || (_k_) >= (_len_) then error ("Data.Text.Array." ++ (_func_) ++ "

57#if defined(ASSERTS)

58import Control.Exception (assert)

59#endif

60+import Data.Bits ((.&.))

60import Data.Text.UnsafeShift (shiftL, shiftR)

61import Data.Text.UnsafeShift (shiftL, shiftR)

61import GHC.Base (ByteArray#, MutableByteArray#, Int(..),

62import GHC.Base (ByteArray#, MutableByteArray#, Int(..),

62− indexWord16Array#, newByteArray#,

63+ indexWord16Array#, indexWordArray#, newByteArray#,

63 readWord16Array#, readWordArray#, unsafeCoerce#,

64 readWord16Array#, readWordArray#, unsafeCoerce#,

64 writeWord16Array#, writeWordArray#)

65 writeWord16Array#, writeWordArray#)

65import GHC.ST (ST(..), runST)

66import GHC.ST (ST(..), runST)

@@ -125,6 +126,14 @@ unsafeIndex (Array len ba#) i@(I# i#) =

125 case indexWord16Array# ba# i# of r# -> (W16# r#)

126 case indexWord16Array# ba# i# of r# -> (W16# r#)

126{-# INLINE unsafeIndex #-}

127{-# INLINE unsafeIndex #-}

127

128

129+-- | Unchecked read of an immutable array. May return garbage or

130+-- crash on an out-of-bounds access.

131+unsafeIndexWord :: Array -> Int -> Word

132+unsafeIndexWord (Array len ba#) i@(I# i#) =

133+ CHECK_BOUNDS("unsafeIndexWord",len,i)

134+ case indexWordArray# ba# i# of r# -> (W# r#)

135+{-# INLINE unsafeIndexWord #-}

136+

128-- | Unchecked read of a mutable array. May return garbage or

137-- | Unchecked read of a mutable array. May return garbage or

129-- crash on an out-of-bounds access.

138-- crash on an out-of-bounds access.

130unsafeRead :: MArray s -> Int -> ST s Word16

139unsafeRead :: MArray s -> Int -> ST s Word16

@@ -185,6 +194,15 @@ run2 k = runST (do

185 arr <- unsafeFreeze marr

194 arr <- unsafeFreeze marr

186 return (arr,b))

195 return (arr,b))

187

196

197+-- | The amount to divide or multiply by to switch between units of

198+-- 'Word16' and units of 'Word'.

199+wordFactor :: Int

200+wordFactor = SIZEOF_HSWORD `shiftR` 1

201+

202+-- | Indicate whether an offset is word-aligned.

203+wordAligned :: Int -> Bool

204+wordAligned i = i .&. (wordFactor - 1) == 0

205+

188-- | Copy an array in its entirety. The destination array must be at

206-- | Copy an array in its entirety. The destination array must be at

189-- least as big as the source.

207-- least as big as the source.

190copy :: MArray s -- ^ source array

208copy :: MArray s -- ^ source array

@@ -194,9 +212,9 @@ copy dest@(MArray dlen _) src@(MArray slen _)

194 | dlen >= slen = fast_loop 0

212 | dlen >= slen = fast_loop 0

195 | otherwise = fail "Data.Text.Array.copy: array too small"

213 | otherwise = fail "Data.Text.Array.copy: array too small"

196 where

214 where

197− nwds = slen `div` (SIZEOF_HSWORD `shiftR` 1)

215+ nwds = slen `div` wordFactor

198 fast_loop !i

216 fast_loop !i

199− | i >= nwds = copy_loop (i * (SIZEOF_HSWORD `shiftR` 1))

217+ | i >= nwds = copy_loop (i * wordFactor)

200 | otherwise = do unsafeReadWord src i >>= unsafeWriteWord dest i

218 | otherwise = do unsafeReadWord src i >>= unsafeWriteWord dest i

201 fast_loop (i+1)

219 fast_loop (i+1)

202 copy_loop !i

220 copy_loop !i

@@ -205,7 +223,12 @@ copy dest@(MArray dlen _) src@(MArray slen _)

205 copy_loop (i+1)

223 copy_loop (i+1)

206

224

207-- | Copy some elements of a mutable array.

225-- | Copy some elements of a mutable array.

208−partialCopyM :: MArray s -> Int -> MArray s -> Int -> Int -> ST s ()

226+partialCopyM :: MArray s -- ^ Destination

227+ -> Int -- ^ Destination offset

228+ -> MArray s -- ^ Source

229+ -> Int -- ^ Source offset

230+ -> Int -- ^ Count

231+ -> ST s ()

209partialCopyM dest didx src sidx count =

232partialCopyM dest didx src sidx count =

210#if defined(ASSERTS)

233#if defined(ASSERTS)

211 assert (sidx + count <= length src) .

234 assert (sidx + count <= length src) .

@@ -219,9 +242,23 @@ partialCopyM dest didx src sidx count =

219 copy_loop (i+1) (j+1) (c+1)

242 copy_loop (i+1) (j+1) (c+1)

220

243

221-- | Copy some elements of an immutable array.

244-- | Copy some elements of an immutable array.

222−partialCopyI :: MArray s -> Int -> Array -> Int -> Int -> ST s ()

245+partialCopyI :: MArray s -- ^ Destination

246+ -> Int -- ^ Destination offset

247+ -> Array -- ^ Source

248+ -> Int -- ^ Source offset

249+ -> Int -- ^ First offset in source /not/ to

250+ -- copy (i.e. /not/ length)

251+ -> ST s ()

223−partialCopyI dest i0 src j0 top = go i0 j0

252+partialCopyI dest i0 src j0 top

253+ | wordAligned i0 && wordAligned j0 = fast (i0 `div` wordFactor) (j0 `div` wordFactor)

254+ | otherwise = slow i0 j0

224 where

255 where

256+ topwds = top `div` wordFactor

257+ fast !i !j

258+ | i >= topwds = slow (i * wordFactor) (j * wordFactor)

259+ | otherwise = do unsafeWriteWord dest i (src `unsafeIndexWord` j)

260+ fast (i+1) (j+1)

261+ slow !i !j

225− go !i !j | i >= top = return ()

262+ | i >= top = return ()

226− | otherwise = do unsafeWrite dest i (src `unsafeIndex` j)

263+ | otherwise = do unsafeWrite dest i (src `unsafeIndex` j)

227− go (i+1) (j+1)

264+ slow (i+1) (j+1)