diff options
Diffstat (limited to 'compiler/nativeGen/CFG.hs')
| -rw-r--r-- | compiler/nativeGen/CFG.hs | 748 |
1 files changed, 684 insertions, 64 deletions
diff --git a/compiler/nativeGen/CFG.hs b/compiler/nativeGen/CFG.hs index 44ddecd216..e1251b76f2 100644 --- a/compiler/nativeGen/CFG.hs +++ b/compiler/nativeGen/CFG.hs @@ -6,31 +6,40 @@ {-# LANGUAGE TupleSections #-} {-# LANGUAGE GeneralizedNewtypeDeriving #-} {-# LANGUAGE CPP #-} +{-# LANGUAGE Rank2Types #-} +{-# LANGUAGE BangPatterns #-} +{-# LANGUAGE DataKinds #-} module CFG ( CFG, CfgEdge(..), EdgeInfo(..), EdgeWeight(..) , TransitionSource(..) --Modify the CFG - , addWeightEdge, addEdge, delEdge + , addWeightEdge, addEdge + , delEdge, delNode , addNodesBetween, shortcutWeightMap , reverseEdges, filterEdges , addImmediateSuccessor - , mkWeightInfo, adjustEdgeWeight + , mkWeightInfo, adjustEdgeWeight, setEdgeWeight --Query the CFG , infoEdgeList, edgeList , getSuccessorEdges, getSuccessors - , getSuccEdgesSorted, weightedEdgeList + , getSuccEdgesSorted , getEdgeInfo , getCfgNodes, hasNode - , loopMembers + + -- Loop Information + , loopMembers, loopLevels, loopInfo --Construction/Misc , getCfg, getCfgProc, pprEdgeWeights, sanityCheckCfg --Find backedges and update their weight - , optimizeCFG ) + , optimizeCFG + , mkGlobalWeights + + ) where #include "HsVersions.h" @@ -38,9 +47,8 @@ where import GhcPrelude import BlockId -import Cmm ( RawCmmDecl, GenCmmDecl( .. ), CmmBlock, succ, g_entry - , CmmGraph ) -import CmmNode +import Cmm + import CmmUtils import CmmSwitch import Hoopl.Collections @@ -50,10 +58,24 @@ import qualified Hoopl.Graph as G import Util import Digraph +import Maybes + +import Unique +import qualified Dominators as Dom +import Data.IntMap.Strict (IntMap) +import Data.IntSet (IntSet) + +import qualified Data.IntMap.Strict as IM +import qualified Data.Map as M +import qualified Data.IntSet as IS +import qualified Data.Set as S +import Data.Tree +import Data.Bifunctor import Outputable -- DEBUGGING ONLY --import Debug +-- import Debug.Trace --import OrdList --import Debug.Trace import PprCmm () -- For Outputable instances @@ -61,17 +83,28 @@ import qualified DynFlags as D import Data.List --- import qualified Data.IntMap.Strict as M --TODO: LabelMap +import Data.STRef.Strict +import Control.Monad.ST + +import Data.Array.MArray +import Data.Array.ST +import Data.Array.IArray +import Data.Array.Unsafe (unsafeFreeze) +import Data.Array.Base (unsafeRead, unsafeWrite) + +import Control.Monad + +type Prob = Double type Edge = (BlockId, BlockId) type Edges = [Edge] newtype EdgeWeight - = EdgeWeight Int - deriving (Eq,Ord,Enum,Num,Real,Integral) + = EdgeWeight { weightToDouble :: Double } + deriving (Eq,Ord,Enum,Num,Real,Fractional) instance Outputable EdgeWeight where - ppr (EdgeWeight w) = ppr w + ppr (EdgeWeight w) = doublePrec 5 w type EdgeInfoMap edgeInfo = LabelMap (LabelMap edgeInfo) @@ -108,15 +141,28 @@ instance Outputable CfgEdge where = parens (ppr from1 <+> text "-(" <> ppr edgeInfo <> text ")->" <+> ppr to1) -- | Can we trace back a edge to a specific Cmm Node --- or has it been introduced for codegen. We use this to maintain +-- or has it been introduced during assembly codegen. We use this to maintain -- some information which would otherwise be lost during the -- Cmm <-> asm transition. -- See also Note [Inverting Conditional Branches] data TransitionSource - = CmmSource (CmmNode O C) + = CmmSource { trans_cmmNode :: (CmmNode O C) + , trans_info :: BranchInfo } | AsmCodeGen deriving (Eq) +data BranchInfo = NoInfo -- ^ Unknown, but not heap or stack check. + | HeapStackCheck -- ^ Heap or stack check + deriving Eq + +instance Outputable BranchInfo where + ppr NoInfo = text "regular" + ppr HeapStackCheck = text "heap/stack" + +isHeapOrStackCheck :: TransitionSource -> Bool +isHeapOrStackCheck (CmmSource { trans_info = HeapStackCheck}) = True +isHeapOrStackCheck _ = False + -- | Information about edges data EdgeInfo = EdgeInfo @@ -127,12 +173,10 @@ data EdgeInfo instance Outputable EdgeInfo where ppr edgeInfo = text "weight:" <+> ppr (edgeWeight edgeInfo) --- Allow specialization -{-# INLINEABLE mkWeightInfo #-} -- | Convenience function, generate edge info based -- on weight not originating from cmm. -mkWeightInfo :: Integral n => n -> EdgeInfo -mkWeightInfo = EdgeInfo AsmCodeGen . fromIntegral +mkWeightInfo :: EdgeWeight -> EdgeInfo +mkWeightInfo = EdgeInfo AsmCodeGen -- | Adjust the weight between the blocks using the given function. -- If there is no such edge returns the original map. @@ -140,12 +184,25 @@ adjustEdgeWeight :: CFG -> (EdgeWeight -> EdgeWeight) -> BlockId -> BlockId -> CFG adjustEdgeWeight cfg f from to | Just info <- getEdgeInfo from to cfg - , weight <- edgeWeight info - = addEdge from to (info { edgeWeight = f weight}) cfg + , !weight <- edgeWeight info + , !newWeight <- f weight + = addEdge from to (info { edgeWeight = newWeight}) cfg + | otherwise = cfg + +-- | Set the weight between the blocks to the given weight. +-- If there is no such edge returns the original map. +setEdgeWeight :: CFG -> EdgeWeight + -> BlockId -> BlockId -> CFG +setEdgeWeight cfg !weight from to + | Just info <- getEdgeInfo from to cfg + = addEdge from to (info { edgeWeight = weight}) cfg | otherwise = cfg + + getCfgNodes :: CFG -> LabelSet -getCfgNodes m = mapFoldMapWithKey (\k v -> setFromList (k:mapKeys v)) m +getCfgNodes m = + mapFoldlWithKey (\s k toMap -> mapFoldlWithKey (\s k _ -> setInsert k s) (setInsert k s) toMap ) setEmpty m hasNode :: CFG -> BlockId -> Bool hasNode m node = mapMember node m || any (mapMember node) m @@ -294,6 +351,11 @@ delEdge from to m = remDest Nothing = Nothing remDest (Just wm) = Just $ mapDelete to wm +delNode :: BlockId -> CFG -> CFG +delNode node cfg = + fmap (mapDelete node) -- < Edges to the node + (mapDelete node cfg) -- < Edges from the node + -- | Destinations from bid ordered by weight (descending) getSuccEdgesSorted :: CFG -> BlockId -> [(BlockId,EdgeInfo)] getSuccEdgesSorted m bid = @@ -315,36 +377,54 @@ getEdgeInfo from to m | otherwise = Nothing +getEdgeWeight :: CFG -> BlockId -> BlockId -> EdgeWeight +getEdgeWeight cfg from to = + edgeWeight $ expectJust "Edgeweight for noexisting block" $ + getEdgeInfo from to cfg + +getTransitionSource :: BlockId -> BlockId -> CFG -> TransitionSource +getTransitionSource from to cfg = transitionSource $ expectJust "Source info for noexisting block" $ + getEdgeInfo from to cfg + reverseEdges :: CFG -> CFG -reverseEdges cfg = foldr add mapEmpty flatElems +reverseEdges cfg = mapFoldlWithKey (\cfg from toMap -> go (addNode cfg from) from toMap) mapEmpty cfg where - elems = mapToList $ fmap mapToList cfg :: [(BlockId,[(BlockId,EdgeInfo)])] - flatElems = - concatMap (\(from,ws) -> map (\(to,info) -> (to,from,info)) ws ) elems - add (to,from,info) m = addEdge to from info m + -- We preserve nodes without outgoing edges! + addNode :: CFG -> BlockId -> CFG + addNode cfg b = mapInsertWith mapUnion b mapEmpty cfg + go :: CFG -> BlockId -> (LabelMap EdgeInfo) -> CFG + go cfg from toMap = mapFoldlWithKey (\cfg to info -> addEdge to from info cfg) cfg toMap :: CFG + -- | Returns a unordered list of all edges with info infoEdgeList :: CFG -> [CfgEdge] infoEdgeList m = - mapFoldMapWithKey - (\from toMap -> - map (\(to,info) -> CfgEdge from to info) (mapToList toMap)) - m - --- | Unordered list of edges with weight as Tuple (from,to,weight) -weightedEdgeList :: CFG -> [(BlockId,BlockId,EdgeWeight)] -weightedEdgeList m = - mapFoldMapWithKey - (\from toMap -> - map (\(to,info) -> - (from,to, edgeWeight info)) (mapToList toMap)) - m - -- (\(from, tos) -> map (\(to,info) -> (from,to, edgeWeight info)) tos ) + go (mapToList m) [] + where + -- We avoid foldMap to avoid thunk buildup + go :: [(BlockId,LabelMap EdgeInfo)] -> [CfgEdge] -> [CfgEdge] + go [] acc = acc + go ((from,toMap):xs) acc + = go' xs from (mapToList toMap) acc + go' :: [(BlockId,LabelMap EdgeInfo)] -> BlockId -> [(BlockId,EdgeInfo)] -> [CfgEdge] -> [CfgEdge] + go' froms _ [] acc = go froms acc + go' froms from ((to,info):tos) acc + = go' froms from tos (CfgEdge from to info : acc) -- | Returns a unordered list of all edges without weights edgeList :: CFG -> [Edge] edgeList m = - mapFoldMapWithKey (\from toMap -> fmap (from,) (mapKeys toMap)) m + go (mapToList m) [] + where + -- We avoid foldMap to avoid thunk buildup + go :: [(BlockId,LabelMap EdgeInfo)] -> [Edge] -> [Edge] + go [] acc = acc + go ((from,toMap):xs) acc + = go' xs from (mapKeys toMap) acc + go' :: [(BlockId,LabelMap EdgeInfo)] -> BlockId -> [BlockId] -> [Edge] -> [Edge] + go' froms _ [] acc = go froms acc + go' froms from (to:tos) acc + = go' froms from tos ((from,to) : acc) -- | Get successors of a given node without edge weights. getSuccessors :: CFG -> BlockId -> [BlockId] @@ -355,8 +435,8 @@ getSuccessors m bid pprEdgeWeights :: CFG -> SDoc pprEdgeWeights m = - let edges = sort $ weightedEdgeList m - printEdge (from, to, weight) + let edges = sort $ infoEdgeList m :: [CfgEdge] + printEdge (CfgEdge from to (EdgeInfo { edgeWeight = weight })) = text "\t" <> ppr from <+> text "->" <+> ppr to <> text "[label=\"" <> ppr weight <> text "\",weight=\"" <> ppr weight <> text "\"];\n" @@ -365,7 +445,7 @@ pprEdgeWeights m = --to immediately see it when it does. printNode node = text "\t" <> ppr node <> text ";\n" - getEdgeNodes (from, to, _weight) = [from,to] + getEdgeNodes (CfgEdge from to _) = [from,to] edgeNodes = setFromList $ concatMap getEdgeNodes edges :: LabelSet nodes = filter (\n -> (not . setMember n) edgeNodes) . mapKeys $ mapFilter null m in @@ -378,8 +458,8 @@ pprEdgeWeights m = updateEdgeWeight :: (EdgeWeight -> EdgeWeight) -> Edge -> CFG -> CFG updateEdgeWeight f (from, to) cfg | Just oldInfo <- getEdgeInfo from to cfg - = let oldWeight = edgeWeight oldInfo - newWeight = f oldWeight + = let !oldWeight = edgeWeight oldInfo + !newWeight = f oldWeight in addEdge from to (oldInfo {edgeWeight = newWeight}) cfg | otherwise = panic "Trying to update invalid edge" @@ -447,9 +527,7 @@ addNodesBetween m updates = Should A or B be placed in front of C? The block layout algorithm decides this based on which edge (A,C)/(B,C) is heavier. So we - make a educated guess how often execution will transer control - along each edge as well as how much we gain by placing eg A before - C. + make a educated guess on which branch should be preferred. We rank edges in this order: * Unconditional Control Transfer - They will always @@ -478,7 +556,6 @@ addNodesBetween m updates = address. This reduces the chance that we return to the same cache line further. - -} -- | Generate weights for a Cmm proc based on some simple heuristics. getCfgProc :: D.CfgWeights -> RawCmmDecl -> CFG @@ -514,13 +591,24 @@ getCfg weights graph = getBlockEdges block = case branch of CmmBranch dest -> [mkEdge dest uncondWeight] - CmmCondBranch _c t f l + CmmCondBranch cond t f l | l == Nothing -> [mkEdge f condBranchWeight, mkEdge t condBranchWeight] | l == Just True -> [mkEdge f unlikelyCondWeight, mkEdge t likelyCondWeight] | l == Just False -> [mkEdge f likelyCondWeight, mkEdge t unlikelyCondWeight] + where + mkEdgeInfo = -- pprTrace "Info" (ppr branchInfo <+> ppr cond) + EdgeInfo (CmmSource branch branchInfo) . fromIntegral + mkEdge target weight = ((bid,target), mkEdgeInfo weight) + branchInfo = + foldRegsUsed + (panic "foldRegsDynFlags") + (\info r -> if r == SpLim || r == HpLim || r == BaseReg + then HeapStackCheck else info) + NoInfo cond + (CmmSwitch _e ids) -> let switchTargets = switchTargetsToList ids --Compiler performance hack - for very wide switches don't @@ -538,7 +626,7 @@ getCfg weights graph = map (\x -> ((bid,x),mkEdgeInfo 0)) $ G.successors other where bid = G.entryLabel block - mkEdgeInfo = EdgeInfo (CmmSource branch) . fromIntegral + mkEdgeInfo = EdgeInfo (CmmSource branch NoInfo) . fromIntegral mkEdge target weight = ((bid,target), mkEdgeInfo weight) branch = lastNode block :: CmmNode O C @@ -560,6 +648,11 @@ findBackEdges root cfg = optimizeCFG :: D.CfgWeights -> RawCmmDecl -> CFG -> CFG optimizeCFG _ (CmmData {}) cfg = cfg optimizeCFG weights (CmmProc info _lab _live graph) cfg = + {-# SCC optimizeCFG #-} + -- pprTrace "Initial:" (pprEdgeWeights cfg) $ + -- pprTrace "Initial:" (ppr $ mkGlobalWeights (g_entry graph) cfg) $ + + -- pprTrace "LoopInfo:" (ppr $ loopInfo cfg (g_entry graph)) $ favourFewerPreds . penalizeInfoTables info . increaseBackEdgeWeight (g_entry graph) $ cfg @@ -589,12 +682,8 @@ optimizeCFG weights (CmmProc info _lab _live graph) cfg = = weight - (fromIntegral $ D.infoTablePenalty weights) | otherwise = weight - -{- Note [Optimize for Fallthrough] - --} -- | If a block has two successors, favour the one with fewer - -- predecessors. (As that one is more likely to become a fallthrough) + -- predecessors and/or the one allowing fall through. favourFewerPreds :: CFG -> CFG favourFewerPreds cfg = let @@ -611,16 +700,17 @@ optimizeCFG weights (CmmProc info _lab _live graph) cfg = | preds1 == preds2 = ( 0, 0) | otherwise = (-1, 1) + update :: CFG -> BlockId -> CFG update cfg node | [(s1,e1),(s2,e2)] <- getSuccessorEdges cfg node - , w1 <- edgeWeight e1 - , w2 <- edgeWeight e2 + , !w1 <- edgeWeight e1 + , !w2 <- edgeWeight e2 --Only change the weights if there isn't already a ordering. , w1 == w2 , (mod1,mod2) <- modifiers (predCount s1) (predCount s2) = (\cfg' -> (adjustEdgeWeight cfg' (+mod2) node s2)) - (adjustEdgeWeight cfg (+mod1) node s1) + (adjustEdgeWeight cfg (+mod1) node s1) | otherwise = cfg in setFoldl update cfg nodes @@ -629,13 +719,12 @@ optimizeCFG weights (CmmProc info _lab _live graph) cfg = fallthroughTarget to (EdgeInfo source _weight) | mapMember to info = False | AsmCodeGen <- source = True - | CmmSource (CmmBranch {}) <- source = True - | CmmSource (CmmCondBranch {}) <- source = True + | CmmSource { trans_cmmNode = CmmBranch {} } <- source = True + | CmmSource { trans_cmmNode = CmmCondBranch {} } <- source = True | otherwise = False -- | Determine loop membership of blocks based on SCC analysis --- Ideally we would replace this with a variant giving us loop --- levels instead but the SCC code will do for now. +-- This is faster but only gives yes/no answers. loopMembers :: CFG -> LabelMap Bool loopMembers cfg = foldl' (flip setLevel) mapEmpty sccs @@ -649,3 +738,534 @@ loopMembers cfg = setLevel :: SCC BlockId -> LabelMap Bool -> LabelMap Bool setLevel (AcyclicSCC bid) m = mapInsert bid False m setLevel (CyclicSCC bids) m = foldl' (\m k -> mapInsert k True m) m bids + +loopLevels :: CFG -> BlockId -> LabelMap Int +loopLevels cfg root = liLevels $ loopInfo cfg root + +data LoopInfo = LoopInfo + { liBackEdges :: [(Edge)] -- ^ List of back edges + , liLevels :: LabelMap Int -- ^ BlockId -> LoopLevel mapping + , liLoops :: [(Edge, LabelSet)] -- ^ (backEdge, loopBody), body includes header + } + +instance Outputable LoopInfo where + ppr (LoopInfo _ _lvls loops) = + text "Loops:(backEdge, bodyNodes)" $$ + (vcat $ map ppr loops) + +-- | Determine loop membership of blocks based on Dominator analysis. +-- This is slower but gives loop levels instead of just loop membership. +-- However it only detects natural loops. Irreducible control flow is not +-- recognized even if it loops. But that is rare enough that we don't have +-- to care about that special case. +loopInfo :: CFG -> BlockId -> LoopInfo +loopInfo cfg root = LoopInfo { liBackEdges = backEdges + , liLevels = mapFromList loopCounts + , liLoops = loopBodies } + where + revCfg = reverseEdges cfg + graph = fmap (setFromList . mapKeys ) cfg :: LabelMap LabelSet + + --TODO - This should be a no op: Export constructors? Use unsafeCoerce? ... + rooted = ( fromBlockId root + , toIntMap $ fmap toIntSet graph) :: (Int, IntMap IntSet) + -- rooted = unsafeCoerce (root, graph) + tree = fmap toBlockId $ Dom.domTree rooted :: Tree BlockId + + -- Map from Nodes to their dominators + domMap :: LabelMap LabelSet + domMap = mkDomMap tree + + edges = edgeList cfg :: [(BlockId, BlockId)] + -- We can't recompute this from the edges, there might be blocks not connected via edges. + nodes = getCfgNodes cfg :: LabelSet + + -- identify back edges + isBackEdge (from,to) + | Just doms <- mapLookup from domMap + , setMember to doms + = True + | otherwise = False + + -- determine the loop body for a back edge + findBody edge@(tail, head) + = ( edge, setInsert head $ go (setSingleton tail) (setSingleton tail) ) + where + -- The reversed cfg makes it easier to look up predecessors + cfg' = delNode head revCfg + go :: LabelSet -> LabelSet -> LabelSet + go found current + | setNull current = found + | otherwise = go (setUnion newSuccessors found) + newSuccessors + where + newSuccessors = setFilter (\n -> not $ setMember n found) successors :: LabelSet + successors = setFromList $ concatMap + (getSuccessors cfg') + (setElems current) :: LabelSet + + backEdges = filter isBackEdge edges + loopBodies = map findBody backEdges :: [(Edge, LabelSet)] + + -- Block b is part of n loop bodies => loop nest level of n + loopCounts = + let bodies = map (first snd) loopBodies -- [(Header, Body)] + loopCount n = length $ nub . map fst . filter (setMember n . snd) $ bodies + in map (\n -> (n, loopCount n)) $ setElems nodes :: [(BlockId, Int)] + + toIntSet :: LabelSet -> IntSet + toIntSet s = IS.fromList . map fromBlockId . setElems $ s + toIntMap :: LabelMap a -> IntMap a + toIntMap m = IM.fromList $ map (\(x,y) -> (fromBlockId x,y)) $ mapToList m + + mkDomMap :: Tree BlockId -> LabelMap LabelSet + mkDomMap root = mapFromList $ go setEmpty root + where + go :: LabelSet -> Tree BlockId -> [(Label,LabelSet)] + go parents (Node lbl []) + = [(lbl, parents)] + go parents (Node _ leaves) + = let nodes = map rootLabel leaves + entries = map (\x -> (x,parents)) nodes + in entries ++ concatMap + (\n -> go (setInsert (rootLabel n) parents) n) + leaves + + fromBlockId :: BlockId -> Int + fromBlockId = getKey . getUnique + + toBlockId :: Int -> BlockId + toBlockId = mkBlockId . mkUniqueGrimily + +-- We make the CFG a Hoopl Graph, so we can reuse revPostOrder. +newtype BlockNode (e :: Extensibility) (x :: Extensibility) = BN (BlockId,[BlockId]) + +instance G.NonLocal (BlockNode) where + entryLabel (BN (lbl,_)) = lbl + successors (BN (_,succs)) = succs + +revPostorderFrom :: CFG -> BlockId -> [BlockId] +revPostorderFrom cfg root = + map fromNode $ G.revPostorderFrom hooplGraph root + where + nodes = getCfgNodes cfg + hooplGraph = setFoldl (\m n -> mapInsert n (toNode n) m) mapEmpty nodes + + fromNode :: BlockNode C C -> BlockId + fromNode (BN x) = fst x + + toNode :: BlockId -> BlockNode C C + toNode bid = + BN (bid,getSuccessors cfg $ bid) + + +-- | We take in a CFG which has on its edges weights which are +-- relative only to other edges originating from the same node. +-- +-- We return a CFG for which each edge represents a GLOBAL weight. +-- This means edge weights are comparable across the whole graph. +-- +-- For irreducible control flow results might be imprecise, otherwise they +-- are reliable. +-- +-- The algorithm is based on the Paper +-- "Static Branch Prediction and Program Profile Analysis" by Y Wu, JR Larus +-- The only big change is that we go over the nodes in the body of loops in +-- reverse post order. Which is required for diamond control flow to work probably. +-- +-- We also apply a few prediction heuristics (based on the same paper) + +{-# SCC mkGlobalWeights #-} +mkGlobalWeights :: BlockId -> CFG -> (LabelMap Double, LabelMap (LabelMap Double)) +mkGlobalWeights root localCfg + | null localCfg = panic "Error - Empty CFG" + | otherwise + = --pprTrace "revOrder" (ppr revOrder) $ + -- undefined --propagate (mapSingleton root 1) (revOrder) + (blockFreqs', edgeFreqs') + where + -- Calculate fixpoints + (blockFreqs, edgeFreqs) = calcFreqs nodeProbs backEdges' bodies' revOrder' + blockFreqs' = mapFromList $ map (first fromVertex) (assocs blockFreqs) :: LabelMap Double + edgeFreqs' = fmap fromVertexMap $ fromVertexMap edgeFreqs + + fromVertexMap :: IM.IntMap x -> LabelMap x + fromVertexMap m = mapFromList . map (first fromVertex) $ IM.toList m + + revOrder = revPostorderFrom localCfg root :: [BlockId] + loopinfo@(LoopInfo backedges _levels bodies) = loopInfo localCfg root + + revOrder' = map toVertex revOrder + backEdges' = map (bimap toVertex toVertex) backedges + bodies' = map calcBody bodies + + estimatedCfg = staticBranchPrediction root loopinfo localCfg + -- Normalize the weights to probabilities and apply heuristics + nodeProbs = cfgEdgeProbabilities estimatedCfg toVertex + + -- By mapping vertices to numbers in reverse post order we can bring any subset into reverse post + -- order simply by sorting. + -- TODO: The sort is redundant if we can guarantee that setElems returns elements ascending + calcBody (backedge, blocks) = + (toVertex $ snd backedge, sort . map toVertex $ (setElems blocks)) + + vertexMapping = mapFromList $ zip revOrder [0..] :: LabelMap Int + blockMapping = listArray (0,mapSize vertexMapping - 1) revOrder :: Array Int BlockId + -- Map from blockId to indicies starting at zero + toVertex :: BlockId -> Int + toVertex blockId = expectJust "mkGlobalWeights" $ mapLookup blockId vertexMapping + -- Map from indicies starting at zero to blockIds + fromVertex :: Int -> BlockId + fromVertex vertex = blockMapping ! vertex + +{- Note [Static Branch Prediction] + ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +The work here has been based on the paper +"Static Branch Prediction and Program Profile Analysis" by Y Wu, JR Larus. + +The primary differences are that if we branch on the result of a heap +check we do not apply any of the heuristics. +The reason is simple: They look like loops in the control flow graph +but are usually never entered, and if at most once. + +Currently implemented is a heuristic to predict that we do not exit +loops (lehPredicts) and one to predict that backedges are more likely +than any other edge. + +The back edge case is special as it superceeds any other heuristic if it +applies. + +Do NOT rely solely on nofib results for benchmarking this. I recommend at least +comparing megaparsec and container benchmarks. Nofib does not seeem to have +many instances of "loopy" Cmm where these make a difference. + +TODO: +* The paper containers more benchmarks which should be implemented. +* If we turn the likelyhood on if/else branches into a probability + instead of true/false we could implement this as a Cmm pass. + + The complete Cmm code still exists and can be accessed by the heuristics + + There is no chance of register allocation/codegen inserting branches/blocks + + making the TransitionSource info wrong. + + potential to use this information in CmmPasses. + - Requires refactoring of all the code relying on the binary nature of likelyhood. + - Requires refactoring `loopInfo` to work on both, Cmm Graphs and the backend CFG. +-} + +-- | Combination of target node id and information about the branch +-- we are looking at. +type TargetNodeInfo = (BlockId, EdgeInfo) + + +-- | Update branch weights based on certain heuristics. +-- See Note [Static Branch Prediction] +-- TODO: This should be combined with optimizeCFG +{-# SCC staticBranchPrediction #-} +staticBranchPrediction :: BlockId -> LoopInfo -> CFG -> CFG +staticBranchPrediction _root (LoopInfo l_backEdges loopLevels l_loops) cfg = + -- pprTrace "staticEstimatesOn" (ppr (cfg)) $ + setFoldl update cfg nodes + where + nodes = getCfgNodes cfg + backedges = S.fromList $ l_backEdges + -- Loops keyed by their back edge + loops = M.fromList $ l_loops :: M.Map Edge LabelSet + loopHeads = S.fromList $ map snd $ M.keys loops + + update :: CFG -> BlockId -> CFG + update cfg node + -- No successors, nothing to do. + | null successors = cfg + + -- Mix of backedges and others: + -- Always predict the backedges. + | not (null m) && length m < length successors + -- Heap/Stack checks "loop", but only once. + -- So we simply exclude any case involving them. + , not $ any (isHeapOrStackCheck . transitionSource . snd) successors + = let loopChance = repeat $! pred_LBH / (fromIntegral $ length m) + exitChance = repeat $! (1 - pred_LBH) / fromIntegral (length not_m) + updates = zip (map fst m) loopChance ++ zip (map fst not_m) exitChance + in -- pprTrace "mix" (ppr (node,successors)) $ + foldl' (\cfg (to,weight) -> setEdgeWeight cfg weight node to) cfg updates + + -- For (regular) non-binary branches we keep the weights from the STG -> Cmm translation. + | length successors /= 2 + = cfg + + -- Only backedges - no need to adjust + | length m > 0 + = cfg + + -- A regular binary branch, we can plug addition predictors in here. + | [(s1,s1_info),(s2,s2_info)] <- successors + , not $ any (isHeapOrStackCheck . transitionSource . snd) successors + = -- Normalize weights to total of 1 + let !w1 = max (edgeWeight s1_info) (0) + !w2 = max (edgeWeight s2_info) (0) + -- Of both weights are <= 0 we set both to 0.5 + normalizeWeight w = if w1 + w2 == 0 then 0.5 else w/(w1+w2) + !cfg' = setEdgeWeight cfg (normalizeWeight w1) node s1 + !cfg'' = setEdgeWeight cfg' (normalizeWeight w2) node s2 + + -- Figure out which heuristics apply to these successors + heuristics = map ($ ((s1,s1_info),(s2,s2_info))) + [lehPredicts, phPredicts, ohPredicts, ghPredicts, lhhPredicts, chPredicts + , shPredicts, rhPredicts] + -- Apply result of a heuristic. Argument is the likelyhood + -- predicted for s1. + applyHeuristic :: CFG -> Maybe Prob -> CFG + applyHeuristic cfg Nothing = cfg + applyHeuristic cfg (Just (s1_pred :: Double)) + | s1_old == 0 || s2_old == 0 || + isHeapOrStackCheck (transitionSource s1_info) || + isHeapOrStackCheck (transitionSource s2_info) + = cfg + | otherwise = + let -- Predictions from heuristic + s1_prob = EdgeWeight s1_pred :: EdgeWeight + s2_prob = 1.0 - s1_prob + -- Update + d = (s1_old * s1_prob) + (s2_old * s2_prob) :: EdgeWeight + s1_prob' = s1_old * s1_prob / d + !s2_prob' = s2_old * s2_prob / d + !cfg_s1 = setEdgeWeight cfg s1_prob' node s1 + in -- pprTrace "Applying heuristic!" (ppr (node,s1,s2) $$ ppr (s1_prob', s2_prob')) $ + setEdgeWeight cfg_s1 s2_prob' node s2 + where + -- Old weights + s1_old = getEdgeWeight cfg node s1 + s2_old = getEdgeWeight cfg node s2 + + in + -- pprTraceIt "RegularCfgResult" $ + foldl' applyHeuristic cfg'' heuristics + + -- Branch on heap/stack check + | otherwise = cfg + + where + -- Chance that loops are taken. + pred_LBH = 0.875 + -- successors + successors = getSuccessorEdges cfg node + -- backedges + (m,not_m) = partition (\succ -> S.member (node, fst succ) backedges) successors + + -- Heuristics return nothing if they don't say anything about this branch + -- or Just (prob_s1) where prob_s1 is the likelyhood for s1 to be the + -- taken branch. s1 is the branch in the true case. + + -- Loop exit heuristic. + -- We are unlikely to leave a loop unless it's to enter another one. + pred_LEH = 0.75 + -- If and only if no successor is a loopheader, + -- then we will likely not exit the current loop body. + lehPredicts :: (TargetNodeInfo,TargetNodeInfo) -> Maybe Prob + lehPredicts ((s1,_s1_info),(s2,_s2_info)) + | S.member s1 loopHeads || S.member s2 loopHeads + = Nothing + + | otherwise + = --pprTrace "lehPredict:" (ppr $ compare s1Level s2Level) $ + case compare s1Level s2Level of + EQ -> Nothing + LT -> Just (1-pred_LEH) --s1 exits to a shallower loop level (exits loop) + GT -> Just (pred_LEH) --s1 exits to a deeper loop level + where + s1Level = mapLookup s1 loopLevels + s2Level = mapLookup s2 loopLevels + + -- Comparing to a constant is unlikely to be equal. + ohPredicts (s1,_s2) + | CmmSource { trans_cmmNode = src1 } <- getTransitionSource node (fst s1) cfg + , CmmCondBranch cond ltrue _lfalse likely <- src1 + , likely == Nothing + , CmmMachOp mop args <- cond + , MO_Eq {} <- mop + , not (null [x | x@CmmLit{} <- args]) + = if fst s1 == ltrue then Just 0.3 else Just 0.7 + + | otherwise + = Nothing + + -- TODO: These are all the other heuristics from the paper. + -- Not all will apply, for now we just stub them out as Nothing. + phPredicts = const Nothing + ghPredicts = const Nothing + lhhPredicts = const Nothing + chPredicts = const Nothing + shPredicts = const Nothing + rhPredicts = const Nothing + +-- We normalize all edge weights as probabilities between 0 and 1. +-- Ignoring rounding errors all outgoing edges sum up to 1. +cfgEdgeProbabilities :: CFG -> (BlockId -> Int) -> IM.IntMap (IM.IntMap Prob) +cfgEdgeProbabilities cfg toVertex + = mapFoldlWithKey foldEdges IM.empty cfg + where + foldEdges = (\m from toMap -> IM.insert (toVertex from) (normalize toMap) m) + + normalize :: (LabelMap EdgeInfo) -> (IM.IntMap Prob) + normalize weightMap + | edgeCount <= 1 = mapFoldlWithKey (\m k _ -> IM.insert (toVertex k) 1.0 m) IM.empty weightMap + | otherwise = mapFoldlWithKey (\m k _ -> IM.insert (toVertex k) (normalWeight k) m) IM.empty weightMap + where + edgeCount = mapSize weightMap + -- Negative weights are generally allowed but are mapped to zero. + -- We then check if there is at least one non-zero edge and if not + -- assign uniform weights to all branches. + minWeight = 0 :: Prob + weightMap' = fmap (\w -> max (weightToDouble . edgeWeight $ w) minWeight) weightMap + totalWeight = sum weightMap' + + normalWeight :: BlockId -> Prob + normalWeight bid + | totalWeight == 0 + = 1.0 / fromIntegral edgeCount + | Just w <- mapLookup bid weightMap' + = w/totalWeight + | otherwise = panic "impossible" + +-- This is the fixpoint algorithm from +-- "Static Branch Prediction and Program Profile Analysis" by Y Wu, JR Larus +-- The adaption to Haskell is my own. +calcFreqs :: IM.IntMap (IM.IntMap Prob) -> [(Int,Int)] -> [(Int, [Int])] -> [Int] + -> (Array Int Double, IM.IntMap (IM.IntMap Prob)) +calcFreqs graph backEdges loops revPostOrder = runST $ do + visitedNodes <- newArray (0,nodeCount-1) False :: ST s (STUArray s Int Bool) + blockFreqs <- newArray (0,nodeCount-1) 0.0 :: ST s (STUArray s Int Double) + edgeProbs <- newSTRef graph + edgeBackProbs <- newSTRef graph + + -- let traceArray a = do + -- vs <- forM [0..nodeCount-1] $ \i -> readArray a i >>= (\v -> return (i,v)) + -- trace ("array: " ++ show vs) $ return () + + let -- See #1600, we need to inline or unboxing makes perf worse. + -- {-# INLINE getFreq #-} + {-# INLINE visited #-} + visited b = unsafeRead visitedNodes b + getFreq b = unsafeRead blockFreqs b + -- setFreq :: forall s. Int -> Double -> ST s () + setFreq b f = unsafeWrite blockFreqs b f + -- setVisited :: forall s. Node -> ST s () + setVisited b = unsafeWrite visitedNodes b True + -- Frequency/probability that edge is taken. + getProb' arr b1 b2 = readSTRef arr >>= + (\graph -> + return . + fromMaybe (error "getFreq 1") . + IM.lookup b2 . + fromMaybe (error "getFreq 2") $ + (IM.lookup b1 graph) + ) + setProb' arr b1 b2 prob = do + g <- readSTRef arr + let !m = fromMaybe (error "Foo") $ IM.lookup b1 g + !m' = IM.insert b2 prob m + writeSTRef arr $! (IM.insert b1 m' g) + + getEdgeFreq b1 b2 = getProb' edgeProbs b1 b2 + setEdgeFreq b1 b2 = setProb' edgeProbs b1 b2 + getProb b1 b2 = fromMaybe (error "getProb") $ do + m' <- IM.lookup b1 graph + IM.lookup b2 m' + + getBackProb b1 b2 = getProb' edgeBackProbs b1 b2 + setBackProb b1 b2 = setProb' edgeBackProbs b1 b2 + + + let -- calcOutFreqs :: Node -> ST s () + calcOutFreqs bhead block = do + !f <- getFreq block + forM (successors block) $ \bi -> do + let !prob = getProb block bi + let !succFreq = f * prob + setEdgeFreq block bi succFreq + -- traceM $ "SetOut: " ++ show (block, bi, f, prob, succFreq) + when (bi == bhead) $ setBackProb block bi succFreq + + + let propFreq block head = do + -- traceM ("prop:" ++ show (block,head)) + -- traceShowM block + + !v <- visited block + if v then + return () --Dont look at nodes twice + else if block == head then + setFreq block 1.0 -- Loop header frequency is always 1 + else do + let preds = IS.elems $ predecessors block + irreducible <- (fmap or) $ forM preds $ \bp -> do + !bp_visited <- visited bp + let bp_backedge = isBackEdge bp block + return (not bp_visited && not bp_backedge) + + if irreducible + then return () -- Rare we don't care + else do + setFreq block 0 + !cycleProb <- sum <$> (forM preds $ \pred -> do + if isBackEdge pred block + then + getBackProb pred block + else do + !f <- getFreq block + !prob <- getEdgeFreq pred block + setFreq block $! f + prob + return 0) + -- traceM $ "cycleProb:" ++ show cycleProb + let limit = 1 - 1/512 -- Paper uses 1 - epsilon, but this works. + -- determines how large likelyhoods in loops can grow. + !cycleProb <- return $ min cycleProb limit -- <- return $ if cycleProb > limit then limit else cycleProb + -- traceM $ "cycleProb:" ++ show cycleProb + + !f <- getFreq block + setFreq block (f / (1.0 - cycleProb)) + + setVisited block + calcOutFreqs head block + + -- Loops, by nesting, inner to outer + forM_ loops $ \(head, body) -> do + forM_ [0 .. nodeCount - 1] (\i -> unsafeWrite visitedNodes i True) -- Mark all nodes as visited. + forM_ body (\i -> unsafeWrite visitedNodes i False) -- Mark all blocks reachable from head as not visited + forM_ body $ \block -> propFreq block head + + -- After dealing with all loops, deal with non-looping parts of the CFG + forM_ [0 .. nodeCount - 1] (\i -> unsafeWrite visitedNodes i False) -- Everything in revPostOrder is reachable + forM_ revPostOrder $ \block -> propFreq block (head revPostOrder) + + -- trace ("Final freqs:") $ return () + -- let freqString = pprFreqs freqs + -- trace (unlines freqString) $ return () + -- trace (pprFre) $ return () + graph' <- readSTRef edgeProbs + freqs' <- unsafeFreeze blockFreqs + + return (freqs', graph') + where + predecessors :: Int -> IS.IntSet + predecessors b = fromMaybe IS.empty $ IM.lookup b revGraph + successors b = fromMaybe (lookupError "succ" b graph)$ IM.keys <$> IM.lookup b graph + lookupError s b g = pprPanic ("Lookup error " ++ s) $ + ( text "node" <+> ppr b $$ + text "graph" <+> + vcat (map (\(k,m) -> ppr (k,m :: IM.IntMap Double)) $ IM.toList g) + ) + + nodeCount = IM.foldl' (\count toMap -> IM.foldlWithKey' countTargets count toMap) (IM.size graph) graph + where + countTargets = (\count k _ -> countNode k + count ) + countNode n = if IM.member n graph then 0 else 1 + + isBackEdge from to = S.member (from,to) backEdgeSet + backEdgeSet = S.fromList backEdges + + revGraph :: IntMap IntSet + revGraph = IM.foldlWithKey' (\m from toMap -> addEdges m from toMap) IM.empty graph + where + addEdges m0 from toMap = IM.foldlWithKey' (\m k _ -> addEdge m from k) m0 toMap + addEdge m0 from to = IM.insertWith IS.union to (IS.singleton from) m0 |