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 9
import GhcPrelude hiding (iterate, succ, unzip, zip)

10
import BlockId
11
import Cmm
12
import CmmUtils
13
import CmmSwitch (eqSwitchTargetWith)
Simon Marlow's avatar
Simon Marlow committed
14
import CmmContFlowOpt
15
-- import PprCmm ()
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
16

17 18 19 20
import Hoopl.Block
import Hoopl.Graph
import Hoopl.Label
import Hoopl.Collections
21
import Data.Bits
Peter Wortmann's avatar
Peter Wortmann committed
22
import Data.Maybe (mapMaybe)
Ian Lynagh's avatar
Ian Lynagh committed
23
import qualified Data.List as List
24
import Data.Word
Peter Wortmann's avatar
Peter Wortmann committed
25
import qualified Data.Map as M
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
26 27
import Outputable
import UniqFM
niteria's avatar
niteria committed
28 29
import UniqDFM
import qualified TrieMap as TM
30 31
import Unique
import Control.Arrow (first, second)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
32

33 34 35
-- -----------------------------------------------------------------------------
-- Eliminate common blocks

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
36 37 38 39 40 41 42 43 44 45 46
-- 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.

47 48 49 50 51 52 53
-- 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.
--
54 55
-- 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).
56 57 58 59 60
--
-- 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
61 62
-- TODO: Use optimization fuel
elimCommonBlocks :: CmmGraph -> CmmGraph
Peter Wortmann's avatar
Peter Wortmann committed
63
elimCommonBlocks g = replaceLabels env $ copyTicks env g
64
  where
65
     env = iterate mapEmpty blocks_with_key
66
     groups = groupByInt hash_block (postorderDfs g)
67 68 69 70 71 72
     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]
73
type Subst = LabelMap BlockId
74 75 76 77 78 79 80 81 82

-- 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
83

84 85 86 87 88 89
    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
90

91 92
    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
93

94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
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
113 114 115 116 117 118 119


-- -----------------------------------------------------------------------------
-- 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
120

121
-- To speed up comparisons, we hash each basic block modulo jump labels.
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
122 123
-- The hashing is a bit arbitrary (the numbers are completely arbitrary),
-- but it should be fast and good enough.
124

125 126 127 128
-- 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.

129 130
type HashCode = Int

131
hash_block :: CmmBlock -> HashCode
132 133
hash_block block =
  fromIntegral (foldBlockNodesB3 (hash_fst, hash_mid, hash_lst) block (0 :: Word32) .&. (0x7fffffff :: Word32))
134
  -- UniqFM doesn't like negative Ints
135 136 137 138 139
  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
140
        hash_node n | dont_care n = 0 -- don't care
141 142 143
        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
144
        hash_node (CmmBranch _) = 23 -- NB. ignore the label
145
        hash_node (CmmCondBranch p _ _ _) = hash_e p
146
        hash_node (CmmCall e _ _ _ _ _) = hash_e e
147
        hash_node (CmmForeignCall t _ _ _ _ _ _) = hash_tgt t
148
        hash_node (CmmSwitch e _) = hash_e e
Peter Wortmann's avatar
Peter Wortmann committed
149
        hash_node _ = error "hash_node: unknown Cmm node!"
150

151
        hash_reg :: CmmReg -> Word32
152
        hash_reg   (CmmLocal localReg) = hash_unique localReg -- important for performance, see #10397
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
153
        hash_reg   (CmmGlobal _)    = 19
154

155
        hash_e :: CmmExpr -> Word32
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
156 157 158
        hash_e (CmmLit l) = hash_lit l
        hash_e (CmmLoad e _) = 67 + hash_e e
        hash_e (CmmReg r) = hash_reg r
159
        hash_e (CmmMachOp _ es) = hash_list hash_e es -- pessimal - no operator check
160
        hash_e (CmmRegOff r i) = hash_reg r + cvt i
161
        hash_e (CmmStackSlot _ _) = 13
162

163
        hash_lit :: CmmLit -> Word32
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
164 165
        hash_lit (CmmInt i _) = fromInteger i
        hash_lit (CmmFloat r _) = truncate r
166
        hash_lit (CmmVec ls) = hash_list hash_lit ls
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
167
        hash_lit (CmmLabel _) = 119 -- ugh
168 169
        hash_lit (CmmLabelOff _ i) = cvt $ 199 + i
        hash_lit (CmmLabelDiffOff _ _ i) = cvt $ 299 + i
170
        hash_lit (CmmBlock _) = 191 -- ugh
171
        hash_lit (CmmHighStackMark) = cvt 313
172

173 174
        hash_tgt (ForeignTarget e _) = hash_e e
        hash_tgt (PrimTarget _) = 31 -- lots of these
175 176 177

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

178
        cvt = fromInteger . toInteger
Peter Wortmann's avatar
Peter Wortmann committed
179

180 181 182
        hash_unique :: Uniquable a => a -> Word32
        hash_unique = cvt . getKey . getUnique

Peter Wortmann's avatar
Peter Wortmann committed
183 184 185 186
-- | Ignore these node types for equality
dont_care :: CmmNode O x -> Bool
dont_care CmmComment {}  = True
dont_care CmmTick {}     = True
187
dont_care CmmUnwind {}   = True
Peter Wortmann's avatar
Peter Wortmann committed
188 189
dont_care _other         = False

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
190 191 192
-- Utilities: equality and substitution on the graph.

-- Given a map ``subst'' from BlockID -> BlockID, we define equality.
193
eqBid :: LabelMap BlockId -> BlockId -> BlockId -> Bool
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
194
eqBid subst bid bid' = lookupBid subst bid == lookupBid subst bid'
195
lookupBid :: LabelMap BlockId -> BlockId -> BlockId
196
lookupBid subst bid = case mapLookup bid subst of
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
197 198 199
                        Just bid  -> lookupBid subst bid
                        Nothing -> bid

200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235
-- 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
236 237
-- 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
238
eqBlockBodyWith :: (BlockId -> BlockId -> Bool) -> CmmBlock -> CmmBlock -> Bool
Simon Marlow's avatar
Simon Marlow committed
239
eqBlockBodyWith eqBid block block'
240 241 242 243 244
  {-
  | 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
245
  where (_,m,l)   = blockSplit block
Peter Wortmann's avatar
Peter Wortmann committed
246
        nodes     = filter (not . dont_care) (blockToList m)
Simon Marlow's avatar
Simon Marlow committed
247
        (_,m',l') = blockSplit block'
Peter Wortmann's avatar
Peter Wortmann committed
248
        nodes'    = filter (not . dont_care) (blockToList m')
249

250 251
        equal = and (zipWith (eqMiddleWith eqBid) nodes nodes') &&
                eqLastWith eqBid l l'
252 253


254 255
eqLastWith :: (BlockId -> BlockId -> Bool) -> CmmNode O C -> CmmNode O C -> Bool
eqLastWith eqBid (CmmBranch bid1) (CmmBranch bid2) = eqBid bid1 bid2
256 257
eqLastWith eqBid (CmmCondBranch c1 t1 f1 l1) (CmmCondBranch c2 t2 f2 l2) =
  c1 == c2 && l1 == l2 && eqBid t1 t2 && eqBid f1 f2
258 259
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
260 261
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
262 263 264 265 266 267
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
268 269 270 271 272

-- | 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.
273
copyTicks :: LabelMap BlockId -> CmmGraph -> CmmGraph
Peter Wortmann's avatar
Peter Wortmann committed
274 275
copyTicks env g
  | mapNull env = g
Peter Wortmann's avatar
Peter Wortmann committed
276 277 278
  | otherwise   = ofBlockMap (g_entry g) $ mapMap copyTo blockMap
  where -- Reverse block merge map
        blockMap = toBlockMap g
Peter Wortmann's avatar
Peter Wortmann committed
279 280
        revEnv = mapFoldWithKey insertRev M.empty env
        insertRev k x = M.insertWith (const (k:)) x [k]
Peter Wortmann's avatar
Peter Wortmann committed
281 282
        -- Copy ticks and scopes into the given block
        copyTo block = case M.lookup (entryLabel block) revEnv of
Peter Wortmann's avatar
Peter Wortmann committed
283
          Nothing -> block
Peter Wortmann's avatar
Peter Wortmann committed
284 285 286 287 288 289 290
          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)
291 292 293

-- Group by [Label]
groupByLabel :: [(Key, a)] -> [(Key, [a])]
niteria's avatar
niteria committed
294
groupByLabel = go (TM.emptyTM :: TM.ListMap UniqDFM a)
295
  where
niteria's avatar
niteria committed
296 297
    go !m [] = TM.foldTM (:) m []
    go !m ((k,v) : entries) = go (TM.alterTM k' adjust m) entries
298 299 300 301 302
      where k' = map getUnique k
            adjust Nothing       = Just (k,[v])
            adjust (Just (_,vs)) = Just (k,v:vs)


303
groupByInt :: (a -> Int) -> [a] -> [[a]]
niteria's avatar
niteria committed
304 305
groupByInt f xs = nonDetEltsUFM $ List.foldl' go emptyUFM xs
  -- See Note [Unique Determinism and code generation]
306
  where go m x = alterUFM (Just . maybe [x] (x:)) m (f x)