[commit: ghc] ghc-7.10: Speed up elimCommonBlocks by grouping blocks also by outgoing labels (bac8717)
git at git.haskell.org
git at git.haskell.org
Mon May 18 13:37:43 UTC 2015
Repository : ssh://git@git.haskell.org/ghc
On branch : ghc-7.10
Link : http://ghc.haskell.org/trac/ghc/changeset/bac8717c68ef4907908f80b23dc9dd9e88dfa987/ghc
>---------------------------------------------------------------
commit bac8717c68ef4907908f80b23dc9dd9e88dfa987
Author: Joachim Breitner <mail at joachim-breitner.de>
Date: Sat May 16 01:22:06 2015 +0200
Speed up elimCommonBlocks by grouping blocks also by outgoing labels
This is an attempt to improve the situation described in #10397, where
the linear scan of possible candidates for commoning up is far too
expensive. There is (ever) more room for improvement, but this is a
start.
Differential Revision: https://phabricator.haskell.org/D892
(cherry picked from commit c256357242ee2dd282fd0516260edccbb7617244)
>---------------------------------------------------------------
bac8717c68ef4907908f80b23dc9dd9e88dfa987
compiler/cmm/CmmCommonBlockElim.hs | 143 +++++++++++++++++++++++++++++--------
1 file changed, 112 insertions(+), 31 deletions(-)
diff --git a/compiler/cmm/CmmCommonBlockElim.hs b/compiler/cmm/CmmCommonBlockElim.hs
index 95910d1..83b2841 100644
--- a/compiler/cmm/CmmCommonBlockElim.hs
+++ b/compiler/cmm/CmmCommonBlockElim.hs
@@ -1,4 +1,4 @@
-{-# LANGUAGE GADTs #-}
+{-# LANGUAGE GADTs, BangPatterns #-}
module CmmCommonBlockElim
( elimCommonBlocks
)
@@ -19,9 +19,8 @@ import Data.Word
import qualified Data.Map as M
import Outputable
import UniqFM
-
-my_trace :: String -> SDoc -> a -> a
-my_trace = if False then pprTrace else \_ _ a -> a
+import Unique
+import Control.Arrow (first, second)
-- -----------------------------------------------------------------------------
-- Eliminate common blocks
@@ -37,40 +36,72 @@ my_trace = if False then pprTrace else \_ _ a -> a
-- is made redundant by the old block.
-- Otherwise, it is added to the useful blocks.
+-- 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.
+--
+-- The list of outgoing labels is updated as we merge blocks, and only blocks
+-- that had different labels before are compared.
+--
+-- 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
+
-- TODO: Use optimization fuel
elimCommonBlocks :: CmmGraph -> CmmGraph
elimCommonBlocks g = replaceLabels env $ copyTicks env g
where
- env = iterate hashed_blocks mapEmpty
- hashed_blocks = map (\b -> (hash_block b, b)) $ postorderDfs g
-
--- Iterate over the blocks until convergence
-iterate :: [(HashCode,CmmBlock)] -> BlockEnv BlockId -> BlockEnv BlockId
-iterate blocks subst =
- case foldl common_block (False, emptyUFM, subst) blocks of
- (changed, _, subst)
- | changed -> iterate blocks subst
- | otherwise -> subst
+ env = iterate mapEmpty blocks_with_key
+ groups = groupBy hash_block (postorderDfs g)
+ 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
-type State = (ChangeFlag, UniqFM [CmmBlock], BlockEnv BlockId)
+ 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
-type ChangeFlag = Bool
-type HashCode = Int
+ subst' = subst `mapUnion` new_substs
+ updated_blocks = map (map (first (map (lookupBid subst')))) merged_blocks
--- Try to find a block that is equal (or ``common'') to b.
-common_block :: State -> (HashCode, CmmBlock) -> State
-common_block (old_change, bmap, subst) (hash, b) =
- case lookupUFM bmap hash of
- Just bs -> case (List.find (eqBlockBodyWith (eqBid subst) b) bs,
- mapLookup bid subst) of
- (Just b', Nothing) -> addSubst b'
- (Just b', Just b'') | entryLabel b' /= b'' -> addSubst b'
- | otherwise -> (old_change, bmap, subst)
- _ -> (old_change, addToUFM bmap hash (b : bs), subst)
- Nothing -> (old_change, addToUFM bmap hash [b], subst)
- where bid = entryLabel b
- addSubst b' = my_trace "found new common block" (ppr bid <> char '=' <> ppr (entryLabel b')) $
- (True, bmap, mapInsert bid (entryLabel b') subst)
+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
-- -----------------------------------------------------------------------------
@@ -82,6 +113,9 @@ common_block (old_change, bmap, subst) (hash, b) =
-- To speed up comparisons, we hash each basic block modulo labels.
-- The hashing is a bit arbitrary (the numbers are completely arbitrary),
-- but it should be fast and good enough.
+
+type HashCode = Int
+
hash_block :: CmmBlock -> HashCode
hash_block block =
fromIntegral (foldBlockNodesB3 (hash_fst, hash_mid, hash_lst) block (0 :: Word32) .&. (0x7fffffff :: Word32))
@@ -237,3 +271,50 @@ copyTicks env g
(CmmEntry lbl scp1, code) = blockSplitHead to
in CmmEntry lbl (combineTickScopes scp0 scp1) `blockJoinHead`
foldr blockCons code (map CmmTick ticks)
+
+-- Group by [Label]
+groupByLabel :: [(Key, a)] -> [(Key, [a])]
+groupByLabel = go emptyILM
+ where
+ go !m [] = elemsILM m
+ go !m ((k,v) : entries) = go (alterILM adjust m k') entries
+ where k' = map getUnique k
+ adjust Nothing = Just (k,[v])
+ adjust (Just (_,vs)) = Just (k,v:vs)
+
+groupBy :: (a -> Int) -> [a] -> [[a]]
+groupBy f xs = eltsUFM $ List.foldl' go emptyUFM xs
+ where go m x = alterUFM (Just . maybe [x] (x:)) m (f x)
+
+-- Efficient lookup into [([Unique], a)]
+data IntListMap a = ILM (Maybe a) (UniqFM (IntListMap a))
+
+emptyILM :: IntListMap a
+emptyILM = ILM Nothing emptyUFM
+
+unitILM :: [Unique] -> a -> IntListMap a
+unitILM [] a = ILM (Just a) emptyUFM
+unitILM (l:ls) a = ILM Nothing (unitUFM l (unitILM ls a))
+
+
+alterILM :: (Maybe a -> Maybe a) -> IntListMap a -> [Unique] -> IntListMap a
+alterILM f (ILM ma m) [] = ILM (f ma) m
+alterILM f (ILM ma m) (l:ls) = ILM ma (alterUFM go m l)
+ where go Nothing = fmap (unitILM ls) (f Nothing)
+ go (Just ilm) = Just $ alterILM f ilm ls
+
+{- currently unused
+addToILM :: IntListMap a -> [Unique] -> a -> IntListMap a
+addToILM (ILM _ m) [] a = ILM (Just a) m
+addToILM (ILM ma m) (l:ls) a = ILM ma $ alterUFM go m l
+ where go Nothing = Just $ unitILM ls a
+ go (Just ilm) = Just $ addToILM ilm ls a
+
+lookupILM :: IntListMap a -> [Unique] -> Maybe a
+lookupILM (ILM ma _) [] = ma
+lookupILM (ILM _ m) (l:ls) = lookupUFM m l >>= (\m -> lookupILM m ls)
+-}
+
+elemsILM :: IntListMap a -> [a]
+elemsILM (ILM ma m) = maybe id (:) ma $ concatMap elemsILM $ eltsUFM m
+
More information about the ghc-commits
mailing list