CmmCommonBlockElim.hs 11.7 KB
Newer Older
1
{-# LANGUAGE GADTs, BangPatterns #-}
2
module CmmCommonBlockElim
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
3 4 5 6 7
  ( elimCommonBlocks
  )
where


8
import BlockId
9
import Cmm
10
import CmmUtils
11
import CmmSwitch (eqSwitchTargetWith)
Simon Marlow's avatar
Simon Marlow committed
12
import CmmContFlowOpt
13
-- import PprCmm ()
14
import Prelude hiding (iterate, succ, unzip, zip)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
15

Simon Marlow's avatar
Simon Marlow committed
16
import Hoopl hiding (ChangeFlag)
17
import Data.Bits
Peter Wortmann's avatar
Peter Wortmann committed
18
import Data.Maybe (mapMaybe)
Ian Lynagh's avatar
Ian Lynagh committed
19
import qualified Data.List as List
20
import Data.Word
Peter Wortmann's avatar
Peter Wortmann committed
21
import qualified Data.Map as M
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
22 23
import Outputable
import UniqFM
niteria's avatar
niteria committed
24 25
import UniqDFM
import qualified TrieMap as TM
26 27
import Unique
import Control.Arrow (first, second)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
28

29 30 31
-- -----------------------------------------------------------------------------
-- Eliminate common blocks

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
32 33 34 35 36 37 38 39 40 41 42
-- If two blocks are identical except for the label on the first node,
-- then we can eliminate one of the blocks. To ensure that the semantics
-- of the program are preserved, we have to rewrite each predecessor of the
-- eliminated block to proceed with the block we keep.

-- The algorithm iterates over the blocks in the graph,
-- checking whether it has seen another block that is equal modulo labels.
-- If so, then it adds an entry in a map indicating that the new block
-- is made redundant by the old block.
-- Otherwise, it is added to the useful blocks.

43 44 45 46 47 48 49
-- To avoid comparing every block with every other block repeatedly, we group
-- them by
--   * a hash of the block, ignoring labels (explained below)
--   * the list of outgoing labels
-- The hash is invariant under relabeling, so we only ever compare within
-- the same group of blocks.
--
50 51
-- The list of outgoing labels is updated as we merge blocks (that is why they
-- are not included in the hash, which we want to calculate only once).
52 53 54 55 56
--
-- All in all, two blocks should never be compared if they have different
-- hashes, and at most once otherwise. Previously, we were slower, and people
-- rightfully complained: #10397

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
57 58
-- TODO: Use optimization fuel
elimCommonBlocks :: CmmGraph -> CmmGraph
Peter Wortmann's avatar
Peter Wortmann committed
59
elimCommonBlocks g = replaceLabels env $ copyTicks env g
60
  where
61
     env = iterate mapEmpty blocks_with_key
62
     groups = groupByInt hash_block (postorderDfs g)
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78
     blocks_with_key = [ [ (successors b, [b]) | b <- bs] | bs <- groups]

-- Invariant: The blocks in the list are pairwise distinct
-- (so avoid comparing them again)
type DistinctBlocks = [CmmBlock]
type Key = [Label]
type Subst = BlockEnv BlockId

-- The outer list groups by hash. We retain this grouping throughout.
iterate :: Subst -> [[(Key, DistinctBlocks)]] -> Subst
iterate subst blocks
    | mapNull new_substs = subst
    | otherwise = iterate subst' updated_blocks
  where
    grouped_blocks :: [[(Key, [DistinctBlocks])]]
    grouped_blocks = map groupByLabel blocks
79

80 81 82 83 84 85
    merged_blocks :: [[(Key, DistinctBlocks)]]
    (new_substs, merged_blocks) = List.mapAccumL (List.mapAccumL go) mapEmpty grouped_blocks
      where
        go !new_subst1 (k,dbs) = (new_subst1 `mapUnion` new_subst2, (k,db))
          where
            (new_subst2, db) = mergeBlockList subst dbs
86

87 88
    subst' = subst `mapUnion` new_substs
    updated_blocks = map (map (first (map (lookupBid subst')))) merged_blocks
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
89

90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108
mergeBlocks :: Subst -> DistinctBlocks -> DistinctBlocks -> (Subst, DistinctBlocks)
mergeBlocks subst existing new = go new
  where
    go [] = (mapEmpty, existing)
    go (b:bs) = case List.find (eqBlockBodyWith (eqBid subst) b) existing of
        -- This block is a duplicate. Drop it, and add it to the substitution
        Just b' -> first (mapInsert (entryLabel b) (entryLabel b')) $ go bs
        -- This block is not a duplicate, keep it.
        Nothing -> second (b:) $ go bs

mergeBlockList :: Subst -> [DistinctBlocks] -> (Subst, DistinctBlocks)
mergeBlockList _ [] = pprPanic "mergeBlockList" empty
mergeBlockList subst (b:bs) = go mapEmpty b bs
  where
    go !new_subst1 b [] = (new_subst1, b)
    go !new_subst1 b1 (b2:bs) = go new_subst b bs
      where
        (new_subst2, b) =  mergeBlocks subst b1 b2
        new_subst = new_subst1 `mapUnion` new_subst2
109 110 111 112 113 114 115


-- -----------------------------------------------------------------------------
-- Hashing and equality on blocks

-- Below here is mostly boilerplate: hashing blocks ignoring labels,
-- and comparing blocks modulo a label mapping.
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
116

117
-- To speed up comparisons, we hash each basic block modulo jump labels.
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
118 119
-- The hashing is a bit arbitrary (the numbers are completely arbitrary),
-- but it should be fast and good enough.
120

121 122 123 124
-- We want to get as many small buckets as possible, as comparing blocks is
-- expensive. So include as much as possible in the hash. Ideally everything
-- that is compared with (==) in eqBlockBodyWith.

125 126
type HashCode = Int

127
hash_block :: CmmBlock -> HashCode
128 129
hash_block block =
  fromIntegral (foldBlockNodesB3 (hash_fst, hash_mid, hash_lst) block (0 :: Word32) .&. (0x7fffffff :: Word32))
130
  -- UniqFM doesn't like negative Ints
131 132 133 134 135
  where hash_fst _ h = h
        hash_mid m h = hash_node m + h `shiftL` 1
        hash_lst m h = hash_node m + h `shiftL` 1

        hash_node :: CmmNode O x -> Word32
Peter Wortmann's avatar
Peter Wortmann committed
136
        hash_node n | dont_care n = 0 -- don't care
Peter Wortmann's avatar
Peter Wortmann committed
137
        hash_node (CmmUnwind _ e) = hash_e e
138 139 140
        hash_node (CmmAssign r e) = hash_reg r + hash_e e
        hash_node (CmmStore e e') = hash_e e + hash_e e'
        hash_node (CmmUnsafeForeignCall t _ as) = hash_tgt t + hash_list hash_e as
141
        hash_node (CmmBranch _) = 23 -- NB. ignore the label
142
        hash_node (CmmCondBranch p _ _ _) = hash_e p
143
        hash_node (CmmCall e _ _ _ _ _) = hash_e e
144
        hash_node (CmmForeignCall t _ _ _ _ _ _) = hash_tgt t
145
        hash_node (CmmSwitch e _) = hash_e e
Peter Wortmann's avatar
Peter Wortmann committed
146
        hash_node _ = error "hash_node: unknown Cmm node!"
147

148
        hash_reg :: CmmReg -> Word32
149
        hash_reg   (CmmLocal localReg) = hash_unique localReg -- important for performance, see #10397
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
150
        hash_reg   (CmmGlobal _)    = 19
151

152
        hash_e :: CmmExpr -> Word32
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
153 154 155
        hash_e (CmmLit l) = hash_lit l
        hash_e (CmmLoad e _) = 67 + hash_e e
        hash_e (CmmReg r) = hash_reg r
156
        hash_e (CmmMachOp _ es) = hash_list hash_e es -- pessimal - no operator check
157
        hash_e (CmmRegOff r i) = hash_reg r + cvt i
158
        hash_e (CmmStackSlot _ _) = 13
159

160
        hash_lit :: CmmLit -> Word32
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
161 162
        hash_lit (CmmInt i _) = fromInteger i
        hash_lit (CmmFloat r _) = truncate r
163
        hash_lit (CmmVec ls) = hash_list hash_lit ls
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
164
        hash_lit (CmmLabel _) = 119 -- ugh
165 166
        hash_lit (CmmLabelOff _ i) = cvt $ 199 + i
        hash_lit (CmmLabelDiffOff _ _ i) = cvt $ 299 + i
167
        hash_lit (CmmBlock _) = 191 -- ugh
168
        hash_lit (CmmHighStackMark) = cvt 313
169

170 171
        hash_tgt (ForeignTarget e _) = hash_e e
        hash_tgt (PrimTarget _) = 31 -- lots of these
172 173 174

        hash_list f = foldl (\z x -> f x + z) (0::Word32)

175
        cvt = fromInteger . toInteger
Peter Wortmann's avatar
Peter Wortmann committed
176

177 178 179
        hash_unique :: Uniquable a => a -> Word32
        hash_unique = cvt . getKey . getUnique

Peter Wortmann's avatar
Peter Wortmann committed
180 181 182 183 184 185
-- | Ignore these node types for equality
dont_care :: CmmNode O x -> Bool
dont_care CmmComment {}  = True
dont_care CmmTick {}     = True
dont_care _other         = False

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
186 187 188
-- Utilities: equality and substitution on the graph.

-- Given a map ``subst'' from BlockID -> BlockID, we define equality.
189
eqBid :: BlockEnv BlockId -> BlockId -> BlockId -> Bool
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
190
eqBid subst bid bid' = lookupBid subst bid == lookupBid subst bid'
191
lookupBid :: BlockEnv BlockId -> BlockId -> BlockId
192
lookupBid subst bid = case mapLookup bid subst of
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
193 194 195
                        Just bid  -> lookupBid subst bid
                        Nothing -> bid

196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231
-- Middle nodes and expressions can contain BlockIds, in particular in
-- CmmStackSlot and CmmBlock, so we have to use a special equality for
-- these.
--
eqMiddleWith :: (BlockId -> BlockId -> Bool)
             -> CmmNode O O -> CmmNode O O -> Bool
eqMiddleWith eqBid (CmmAssign r1 e1) (CmmAssign r2 e2)
  = r1 == r2 && eqExprWith eqBid e1 e2
eqMiddleWith eqBid (CmmStore l1 r1) (CmmStore l2 r2)
  = eqExprWith eqBid l1 l2 && eqExprWith eqBid r1 r2
eqMiddleWith eqBid (CmmUnsafeForeignCall t1 r1 a1)
                   (CmmUnsafeForeignCall t2 r2 a2)
  = t1 == t2 && r1 == r2 && and (zipWith (eqExprWith eqBid) a1 a2)
eqMiddleWith _ _ _ = False

eqExprWith :: (BlockId -> BlockId -> Bool)
           -> CmmExpr -> CmmExpr -> Bool
eqExprWith eqBid = eq
 where
  CmmLit l1          `eq` CmmLit l2          = eqLit l1 l2
  CmmLoad e1 _       `eq` CmmLoad e2 _       = e1 `eq` e2
  CmmReg r1          `eq` CmmReg r2          = r1==r2
  CmmRegOff r1 i1    `eq` CmmRegOff r2 i2    = r1==r2 && i1==i2
  CmmMachOp op1 es1  `eq` CmmMachOp op2 es2  = op1==op2 && es1 `eqs` es2
  CmmStackSlot a1 i1 `eq` CmmStackSlot a2 i2 = eqArea a1 a2 && i1==i2
  _e1                `eq` _e2                = False

  xs `eqs` ys = and (zipWith eq xs ys)

  eqLit (CmmBlock id1) (CmmBlock id2) = eqBid id1 id2
  eqLit l1 l2 = l1 == l2

  eqArea Old Old = True
  eqArea (Young id1) (Young id2) = eqBid id1 id2
  eqArea _ _ = False

Simon Marlow's avatar
Simon Marlow committed
232 233
-- Equality on the body of a block, modulo a function mapping block
-- IDs to block IDs.
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
234
eqBlockBodyWith :: (BlockId -> BlockId -> Bool) -> CmmBlock -> CmmBlock -> Bool
Simon Marlow's avatar
Simon Marlow committed
235
eqBlockBodyWith eqBid block block'
236 237 238 239 240
  {-
  | equal     = pprTrace "equal" (vcat [ppr block, ppr block']) True
  | otherwise = pprTrace "not equal" (vcat [ppr block, ppr block']) False
  -}
  = equal
Simon Marlow's avatar
Simon Marlow committed
241
  where (_,m,l)   = blockSplit block
Peter Wortmann's avatar
Peter Wortmann committed
242
        nodes     = filter (not . dont_care) (blockToList m)
Simon Marlow's avatar
Simon Marlow committed
243
        (_,m',l') = blockSplit block'
Peter Wortmann's avatar
Peter Wortmann committed
244
        nodes'    = filter (not . dont_care) (blockToList m')
245

246 247
        equal = and (zipWith (eqMiddleWith eqBid) nodes nodes') &&
                eqLastWith eqBid l l'
248 249


250 251
eqLastWith :: (BlockId -> BlockId -> Bool) -> CmmNode O C -> CmmNode O C -> Bool
eqLastWith eqBid (CmmBranch bid1) (CmmBranch bid2) = eqBid bid1 bid2
252 253
eqLastWith eqBid (CmmCondBranch c1 t1 f1 l1) (CmmCondBranch c2 t2 f2 l2) =
  c1 == c2 && l1 == l2 && eqBid t1 t2 && eqBid f1 f2
254 255
eqLastWith eqBid (CmmCall t1 c1 g1 a1 r1 u1) (CmmCall t2 c2 g2 a2 r2 u2) =
  t1 == t2 && eqMaybeWith eqBid c1 c2 && a1 == a2 && r1 == r2 && u1 == u2 && g1 == g2
256 257
eqLastWith eqBid (CmmSwitch e1 ids1) (CmmSwitch e2 ids2) =
  e1 == e2 && eqSwitchTargetWith eqBid ids1 ids2
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
258 259 260 261 262 263
eqLastWith _ _ _ = False

eqMaybeWith :: (a -> b -> Bool) -> Maybe a -> Maybe b -> Bool
eqMaybeWith eltEq (Just e) (Just e') = eltEq e e'
eqMaybeWith _ Nothing Nothing = True
eqMaybeWith _ _ _ = False
Peter Wortmann's avatar
Peter Wortmann committed
264 265 266 267 268 269 270 271

-- | Given a block map, ensure that all "target" blocks are covered by
-- the same ticks as the respective "source" blocks. This not only
-- means copying ticks, but also adjusting tick scopes where
-- necessary.
copyTicks :: BlockEnv BlockId -> CmmGraph -> CmmGraph
copyTicks env g
  | mapNull env = g
Peter Wortmann's avatar
Peter Wortmann committed
272 273 274
  | otherwise   = ofBlockMap (g_entry g) $ mapMap copyTo blockMap
  where -- Reverse block merge map
        blockMap = toBlockMap g
Peter Wortmann's avatar
Peter Wortmann committed
275 276
        revEnv = mapFoldWithKey insertRev M.empty env
        insertRev k x = M.insertWith (const (k:)) x [k]
Peter Wortmann's avatar
Peter Wortmann committed
277 278
        -- Copy ticks and scopes into the given block
        copyTo block = case M.lookup (entryLabel block) revEnv of
Peter Wortmann's avatar
Peter Wortmann committed
279
          Nothing -> block
Peter Wortmann's avatar
Peter Wortmann committed
280 281 282 283 284 285 286
          Just ls -> foldr copy block $ mapMaybe (flip mapLookup blockMap) ls
        copy from to =
          let ticks = blockTicks from
              CmmEntry  _   scp0        = firstNode from
              (CmmEntry lbl scp1, code) = blockSplitHead to
          in CmmEntry lbl (combineTickScopes scp0 scp1) `blockJoinHead`
             foldr blockCons code (map CmmTick ticks)
287 288 289

-- Group by [Label]
groupByLabel :: [(Key, a)] -> [(Key, [a])]
niteria's avatar
niteria committed
290
groupByLabel = go (TM.emptyTM :: TM.ListMap UniqDFM a)
291
  where
niteria's avatar
niteria committed
292 293
    go !m [] = TM.foldTM (:) m []
    go !m ((k,v) : entries) = go (TM.alterTM k' adjust m) entries
294 295 296 297 298
      where k' = map getUnique k
            adjust Nothing       = Just (k,[v])
            adjust (Just (_,vs)) = Just (k,v:vs)


299
groupByInt :: (a -> Int) -> [a] -> [[a]]
niteria's avatar
niteria committed
300 301
groupByInt f xs = nonDetEltsUFM $ List.foldl' go emptyUFM xs
  -- See Note [Unique Determinism and code generation]
302
  where go m x = alterUFM (Just . maybe [x] (x:)) m (f x)