+{-# OPTIONS_GHC -XGADTs -XNoMonoLocalBinds #-}
+-- Norman likes local bindings
+-- If this module lives on I'd like to get rid of this flag in due course
+
+-- Todo: remove
+{-# OPTIONS_GHC -fno-warn-warnings-deprecations #-}
+
+{-# OPTIONS_GHC -fno-warn-incomplete-patterns #-}
+#if __GLASGOW_HASKELL__ >= 701
+-- GHC 7.0.1 improved incomplete pattern warnings with GADTs
+{-# OPTIONS_GHC -fwarn-incomplete-patterns #-}
+#endif
+
module CmmStackLayout
( SlotEnv, liveSlotAnal, liveSlotTransfers, removeLiveSlotDefs
, layout, manifestSP, igraph, areaBuilder
where
import Constants
-import qualified Prelude as P
-import Prelude hiding (zip, unzip, last)
+import Prelude hiding (succ, zip, unzip, last)
import BlockId
+import Cmm
import CmmExpr
-import CmmProcPointZ
-import CmmTx
-import DFMonad
-import FiniteMap
+import CmmProcPoint
import Maybes
-import MkZipCfg
-import MkZipCfgCmm hiding (CmmBlock, CmmGraph)
-import Monad
+import MkGraph (stackStubExpr)
+import Control.Monad
+import OptimizationFuel
import Outputable
-import Panic
-import ZipCfg
-import ZipCfgCmmRep
-import ZipDataflow
+import SMRep (ByteOff)
+
+import Compiler.Hoopl
+
+import Data.Map (Map)
+import qualified Data.Map as Map
+import qualified FiniteMap as Map
------------------------------------------------------------------------
-- Stack Layout --
-- a single slot, on insertion.
slotLattice :: DataflowLattice SubAreaSet
-slotLattice = DataflowLattice "live slots" emptyFM add False
- where add new old = case foldFM addArea (False, old) new of
- (True, x) -> aTx x
- (False, x) -> noTx x
+slotLattice = DataflowLattice "live slots" Map.empty add
+ where add _ (OldFact old) (NewFact new) = case Map.foldRightWithKey addArea (False, old) new of
+ (change, x) -> (changeIf change, x)
addArea a newSlots z = foldr (addSlot a) z newSlots
addSlot a slot (changed, map) =
- let (c, live) = liveGen slot $ lookupWithDefaultFM map [] a
- in (c || changed, addToFM map a live)
+ let (c, live) = liveGen slot $ Map.findWithDefault [] a map
+ in (c || changed, Map.insert a live map)
+
+slotLatticeJoin :: [SubAreaSet] -> SubAreaSet
+slotLatticeJoin facts = foldr extend (fact_bot slotLattice) facts
+ where extend fact res = snd $ fact_join slotLattice undefined (OldFact fact) (NewFact res)
type SlotEnv = BlockEnv SubAreaSet
-type SlotFix a = FuelMonad (BackwardFixedPoint Middle Last SubAreaSet a)
+ -- The sub-areas live on entry to the block
-liveSlotAnal :: LGraph Middle Last -> FuelMonad SlotEnv
-liveSlotAnal g = liftM zdfFpFacts (res :: SlotFix ())
- where res = zdfSolveFromL emptyBlockEnv "live slot analysis" slotLattice
- liveSlotTransfers (fact_bot slotLattice) g
+liveSlotAnal :: CmmGraph -> FuelUniqSM SlotEnv
+liveSlotAnal g = liftM snd $ dataflowPassBwd g [] $ analBwd slotLattice liveSlotTransfers
-- Add the subarea s to the subareas in the list-set (possibly coalescing it with
-- adjacent subareas), and also return whether s was a new addition.
-- considered live in to the block -- we treat the first node as a definition site.
-- BEWARE?: Am I being a little careless here in failing to check for the
-- entry Id (which would use the CallArea Old).
-liveSlotTransfers :: BackwardTransfers Middle Last SubAreaSet
-liveSlotTransfers =
- BackwardTransfers first liveInSlots liveLastIn
- where first live id = delFromFM live (CallArea (Young id))
+liveSlotTransfers :: BwdTransfer CmmNode SubAreaSet
+liveSlotTransfers = mkBTransfer3 frt mid lst
+ where frt :: CmmNode C O -> SubAreaSet -> SubAreaSet
+ frt (CmmEntry l) f = Map.delete (CallArea (Young l)) f
+
+ mid :: CmmNode O O -> SubAreaSet -> SubAreaSet
+ mid n f = foldSlotsUsed addSlot (removeLiveSlotDefs f n) n
+ lst :: CmmNode O C -> FactBase SubAreaSet -> SubAreaSet
+ lst n f = liveInSlots n $ case n of
+ CmmCall {cml_cont=Nothing, cml_args=args} -> add_area (CallArea Old) args out
+ CmmCall {cml_cont=Just k, cml_args=args} -> add_area (CallArea Old) args (add_area (CallArea (Young k)) args out)
+ CmmForeignCall {succ=k, updfr=oldend} -> add_area (CallArea Old) oldend (add_area (CallArea (Young k)) wORD_SIZE out)
+ _ -> out
+ where out = joinOutFacts slotLattice n f
+ add_area _ n live | n == 0 = live
+ add_area a n live = Map.insert a (snd $ liveGen (a, n, n) $ Map.findWithDefault [] a live) live
-- Slot sets: adding slots, removing slots, and checking for membership.
liftToArea :: Area -> ([SubArea] -> [SubArea]) -> SubAreaSet -> SubAreaSet
addSlot, removeSlot :: SubAreaSet -> SubArea -> SubAreaSet
elemSlot :: SubAreaSet -> SubArea -> Bool
-liftToArea a f map = addToFM map a $ f (lookupWithDefaultFM map [] a)
+liftToArea a f map = Map.insert a (f (Map.findWithDefault [] a map)) map
addSlot live (a, i, w) = liftToArea a (snd . liveGen (a, i, w)) live
removeSlot live (a, i, w) = liftToArea a (liveKill (a, i, w)) live
elemSlot live (a, i, w) =
- not $ fst $ liveGen (a, i, w) (lookupWithDefaultFM live [] a)
+ not $ fst $ liveGen (a, i, w) (Map.findWithDefault [] a live)
removeLiveSlotDefs :: (DefinerOfSlots s, UserOfSlots s) => SubAreaSet -> s -> SubAreaSet
removeLiveSlotDefs = foldSlotsDefd removeSlot
-liveInSlots :: (DefinerOfSlots s, UserOfSlots s) => SubAreaSet -> s -> SubAreaSet
-liveInSlots live x = foldSlotsUsed addSlot (removeLiveSlotDefs live x) x
+liveInSlots :: (DefinerOfSlots s, UserOfSlots s) => s -> SubAreaSet -> SubAreaSet
+liveInSlots x live = foldSlotsUsed addSlot (removeLiveSlotDefs live x) x
-liveLastIn :: (BlockId -> SubAreaSet) -> Last -> SubAreaSet
-liveLastIn env l = liveInSlots (liveLastOut env l) l
+liveLastIn :: CmmNode O C -> (BlockId -> SubAreaSet) -> SubAreaSet
+liveLastIn l env = liveInSlots l (liveLastOut env l)
-- Don't forget to keep the outgoing parameters in the CallArea live,
-- as well as the update frame.
-- be a return to keep the update frame live. We'd still better keep the
-- info pointer in the update frame live at any call site;
-- otherwise we could screw up the garbage collector.
-liveLastOut :: (BlockId -> SubAreaSet) -> Last -> SubAreaSet
+liveLastOut :: (BlockId -> SubAreaSet) -> CmmNode O C -> SubAreaSet
liveLastOut env l =
case l of
- LastCall _ Nothing n _ ->
+ CmmCall _ Nothing n _ _ ->
add_area (CallArea Old) n out -- add outgoing args (includes upd frame)
- LastCall _ (Just k) n (Just upd_n) ->
+ CmmCall _ (Just k) n _ _ ->
add_area (CallArea Old) n (add_area (CallArea (Young k)) n out)
- LastCall _ (Just k) n Nothing ->
- add_area (CallArea (Young k)) n out
+ CmmForeignCall { succ = k, updfr = oldend } ->
+ add_area (CallArea Old) oldend (add_area (CallArea (Young k)) wORD_SIZE out)
_ -> out
- where out = joinOuts slotLattice env l
+ where out = slotLatticeJoin $ map env $ successors l
add_area _ n live | n == 0 = live
add_area a n live =
- addToFM live a $ snd $ liveGen (a, n, n) $ lookupWithDefaultFM live [] a
+ Map.insert a (snd $ liveGen (a, n, n) $ Map.findWithDefault [] a live) live
-- The liveness analysis must be precise: otherwise, we won't know if a definition
-- should really kill a live-out stack slot.
-- every time, I provide a function to fold over the nodes, which should be a
-- reasonably efficient approach for the implementations we envision.
-- Of course, it will probably be much easier to program if we just return a list...
-type Set x = FiniteMap x ()
+type Set x = Map x ()
data IGraphBuilder n =
Builder { foldNodes :: forall z. SubArea -> (n -> z -> z) -> z -> z
, _wordsOccupied :: AreaMap -> AreaMap -> n -> [Int]
areaBuilder = Builder fold words
where fold (a, _, _) f z = f a z
words areaSize areaMap a =
- case lookupFM areaMap a of
- Just addr -> [addr .. addr + (lookupFM areaSize a `orElse`
- pprPanic "wordsOccupied: unknown area" (ppr a))]
+ case Map.lookup a areaMap of
+ Just addr -> [addr .. addr + (Map.lookup a areaSize `orElse`
+ pprPanic "wordsOccupied: unknown area" (ppr areaSize <+> ppr a))]
Nothing -> []
--slotBuilder :: IGraphBuilder (Area, Int)
-- Now, we can build the interference graph.
-- The usual story: a definition interferes with all live outs and all other
-- definitions.
-type IGraph x = FiniteMap x (Set x)
+type IGraph x = Map x (Set x)
type IGPair x = (IGraph x, IGraphBuilder x)
-igraph :: (Ord x) => IGraphBuilder x -> SlotEnv -> LGraph Middle Last -> IGraph x
-igraph builder env g = foldr interfere emptyFM (postorder_dfs g)
+igraph :: (Ord x) => IGraphBuilder x -> SlotEnv -> CmmGraph -> IGraph x
+igraph builder env g = foldr interfere Map.empty (postorderDfs g)
where foldN = foldNodes builder
- interfere block igraph =
- let (h, l) = goto_end (unzip block)
- --heads :: ZHead Middle -> (IGraph x, SubAreaSet) -> IGraph x
- heads (ZFirst _ _) (igraph, _) = igraph
- heads (ZHead h m) (igraph, liveOut) =
- heads h (addEdges igraph m liveOut, liveInSlots liveOut m)
- -- add edges between a def and the other defs and liveouts
- addEdges igraph i out = fst $ foldSlotsDefd addDef (igraph, out) i
- addDef (igraph, out) def@(a, _, _) =
- (foldN def (addDefN out) igraph,
- addToFM out a (snd $ liveGen def (lookupWithDefaultFM out [] a)))
- addDefN out n igraph =
- let addEdgeNO o igraph = foldN o addEdgeNN igraph
- addEdgeNN n' igraph = addEdgeNN' n n' $ addEdgeNN' n' n igraph
- addEdgeNN' n n' igraph = addToFM igraph n (addToFM set n' ())
- where set = lookupWithDefaultFM igraph emptyFM n
- in foldFM (\ _ os igraph -> foldr addEdgeNO igraph os) igraph out
- env' bid = lookupBlockEnv env bid `orElse` panic "unknown blockId in igraph"
- in heads h $ case l of LastExit -> (igraph, emptyFM)
- LastOther l -> (addEdges igraph l $ liveLastOut env' l,
- liveLastIn env' l)
+ interfere block igraph = foldBlockNodesB3 (first, middle, last) block igraph
+ where first _ (igraph, _) = igraph
+ middle node (igraph, liveOut) =
+ (addEdges igraph node liveOut, liveInSlots node liveOut)
+ last node igraph =
+ (addEdges igraph node $ liveLastOut env' node, liveLastIn node env')
+
+ -- add edges between a def and the other defs and liveouts
+ addEdges igraph i out = fst $ foldSlotsDefd addDef (igraph, out) i
+ addDef (igraph, out) def@(a, _, _) =
+ (foldN def (addDefN out) igraph,
+ Map.insert a (snd $ liveGen def (Map.findWithDefault [] a out)) out)
+ addDefN out n igraph =
+ let addEdgeNO o igraph = foldN o addEdgeNN igraph
+ addEdgeNN n' igraph = addEdgeNN' n n' $ addEdgeNN' n' n igraph
+ addEdgeNN' n n' igraph = Map.insert n (Map.insert n' () set) igraph
+ where set = Map.findWithDefault Map.empty n igraph
+ in Map.foldRightWithKey (\ _ os igraph -> foldr addEdgeNO igraph os) igraph out
+ env' bid = mapLookup bid env `orElse` panic "unknown blockId in igraph"
-- Before allocating stack slots, we need to collect one more piece of information:
-- what's the highest offset (in bytes) used in each Area?
-- We'll need to allocate that much space for each Area.
-getAreaSize :: LGraph Middle Last -> AreaMap
-getAreaSize g@(LGraph _ off _) =
- fold_blocks (fold_fwd_block first add_regslots last)
- (unitFM (CallArea Old) off) g
- where first id (StackInfo {argBytes = Just off}) z = add z (CallArea (Young id)) off
- first _ _ z = z
+
+-- JD: WHY CAN'T THIS COME FROM THE slot-liveness info?
+getAreaSize :: ByteOff -> CmmGraph -> AreaMap
+ -- The domain of the returned mapping consists only of Areas
+ -- used for (a) variable spill slots, and (b) parameter passing ares for calls
+getAreaSize entry_off g =
+ foldGraphBlocks (foldBlockNodesF3 (first, add_regslots, last))
+ (Map.singleton (CallArea Old) entry_off) g
+ where first _ z = z
+ last :: CmmNode O C -> Map Area Int -> Map Area Int
+ last l@(CmmCall _ Nothing args res _) z = add_regslots l (add (add z area args) area res)
+ where area = CallArea Old
+ last l@(CmmCall _ (Just k) args res _) z = add_regslots l (add (add z area args) area res)
+ where area = CallArea (Young k)
+ last l@(CmmForeignCall {succ = k}) z = add_regslots l (add z area wORD_SIZE)
+ where area = CallArea (Young k)
+ last l z = add_regslots l z
add_regslots i z = foldSlotsUsed addSlot (foldSlotsDefd addSlot z i) i
- last l@(LastOther (LastCall _ Nothing off _)) z =
- add_regslots l (add z (CallArea Old) off)
- last l@(LastOther (LastCall _ (Just k) off _)) z =
- add_regslots l (add z (CallArea (Young k)) off)
- last l z = add_regslots l z
- addSlot z (a@(RegSlot _), off, _) = add z a off
+ addSlot z (a@(RegSlot (LocalReg _ ty)), _, _) =
+ add z a $ widthInBytes $ typeWidth ty
addSlot z _ = z
- add z a off = addToFM z a (max off (lookupWithDefaultFM z 0 a))
+ add z a off = Map.insert a (max off (Map.findWithDefault 0 a z)) z
+ -- The 'max' is important. Two calls, to f and g, might share a common
+ -- continuation (and hence a common CallArea), but their number of overflow
+ -- parameters might differ.
-- Find the Stack slots occupied by the subarea's conflicts
conflictSlots :: Ord x => IGPair x -> AreaMap -> AreaMap -> SubArea -> Set Int
conflictSlots (ig, Builder foldNodes wordsOccupied) areaSize areaMap subarea =
- foldNodes subarea foldNode emptyFM
- where foldNode n set = foldFM conflict set $ lookupWithDefaultFM ig emptyFM n
+ foldNodes subarea foldNode Map.empty
+ where foldNode n set = Map.foldRightWithKey conflict set $ Map.findWithDefault Map.empty n ig
conflict n' () set = liveInSlots areaMap n' set
-- Add stack slots occupied by igraph node n
liveInSlots areaMap n set = foldr setAdd set (wordsOccupied areaSize areaMap n)
- setAdd w s = addToFM s w ()
+ setAdd w s = Map.insert w () s
-- Find any open space on the stack, starting from the offset.
-- If the area is a CallArea or a spill slot for a pointer, then it must
-- be word-aligned.
freeSlotFrom :: Ord x => IGPair x -> AreaMap -> Int -> AreaMap -> Area -> Int
freeSlotFrom ig areaSize offset areaMap area =
- let size = lookupFM areaSize area `orElse` 0
+ let size = Map.lookup area areaSize `orElse` 0
conflicts = conflictSlots ig areaSize areaMap (area, size, size)
-- CallAreas and Ptrs need to be word-aligned (round up!)
align = case area of CallArea _ -> align'
-- Find a space big enough to hold the area
findSpace curr 0 = curr
findSpace curr cnt = -- part of target slot, # of bytes left to check
- if elemFM curr conflicts then
+ if Map.member curr conflicts then
findSpace (align (curr + size)) size -- try the next (possibly) open space
else findSpace (curr - 1) (cnt - 1)
in findSpace (align (offset + size)) size
-- Find an open space on the stack, and assign it to the area.
allocSlotFrom :: Ord x => IGPair x -> AreaMap -> Int -> AreaMap -> Area -> AreaMap
allocSlotFrom ig areaSize from areaMap area =
- if elemFM area areaMap then areaMap
- else addToFM areaMap area $ freeSlotFrom ig areaSize from areaMap area
+ if Map.member area areaMap then areaMap
+ else Map.insert area (freeSlotFrom ig areaSize from areaMap area) areaMap
-- | Greedy stack layout.
-- Compute liveness, build the interference graph, and allocate slots for the areas.
-- We visit each basic block in a (generally) forward order.
+
-- At each instruction that names a register subarea r, we immediately allocate
-- any available slot on the stack by the following procedure:
--- 1. Find the nodes N' that conflict with r
--- 2. Find the stack slots used for N'
--- 3. Choose a contiguous stack space s not in N' (s must be large enough to hold r)
+-- 1. Find the sub-areas S that conflict with r
+-- 2. Find the stack slots used for S
+-- 3. Choose a contiguous stack space s not in S (s must be large enough to hold r)
+
-- For a CallArea, we allocate the stack space only when we reach a function
-- call that returns to the CallArea's blockId.
--- We use a similar procedure, with one exception: the stack space
--- must be allocated below the youngest stack slot that is live out.
+-- Then, we allocate the Area subject to the following constraints:
+-- a) It must be younger than all the sub-areas that are live on entry to the block
+-- This constraint is only necessary for the successor of a call
+-- b) It must not overlap with any already-allocated Area with which it conflicts
+-- (ie at some point, not necessarily now, is live at the same time)
+-- Part (b) is just the 1,2,3 part above
-- Note: The stack pointer only has to be younger than the youngest live stack slot
-- at proc points. Otherwise, the stack pointer can point anywhere.
-layout :: ProcPointSet -> SlotEnv -> LGraph Middle Last -> AreaMap
-layout procPoints env g@(LGraph _ entrySp _) =
- let builder = areaBuilder
- ig = (igraph builder env g, builder)
- env' bid = lookupBlockEnv env bid `orElse` panic "unknown blockId in igraph"
- areaSize = getAreaSize g
- -- Find the slots that are live-in to the block
- live_in (ZTail m l) = liveInSlots (live_in l) m
- live_in (ZLast (LastOther l)) = liveLastIn env' l
- live_in (ZLast LastExit) = emptyFM
- -- Find the youngest live stack slot
+
+layout :: ProcPointSet -> SlotEnv -> ByteOff -> CmmGraph -> AreaMap
+-- The domain of the returned map includes an Area for EVERY block
+-- including each block that is not the successor of a call (ie is not a proc-point)
+-- That's how we return the info of what the SP should be at the entry of every block
+
+layout procPoints env entry_off g =
+ let ig = (igraph areaBuilder env g, areaBuilder)
+ env' bid = mapLookup bid env `orElse` panic "unknown blockId in igraph"
+ areaSize = getAreaSize entry_off g
+
+ -- Find the youngest live stack slot that has already been allocated
+ youngest_live :: AreaMap -- Already allocated
+ -> SubAreaSet -- Sub-areas live here
+ -> ByteOff -- Offset of the youngest byte of any
+ -- already-allocated, live sub-area
youngest_live areaMap live = fold_subareas young_slot live 0
- where young_slot (a, o, _) z = case lookupFM areaMap a of
+ where young_slot (a, o, _) z = case Map.lookup a areaMap of
Just top -> max z $ top + o
Nothing -> z
- fold_subareas :: (SubArea -> z -> z) -> SubAreaSet -> z -> z
- fold_subareas f m z = foldFM (\_ s z -> foldr f z s) z m
+ fold_subareas f m z = Map.foldRightWithKey (\_ s z -> foldr f z s) z m
+
-- Allocate space for spill slots and call areas
allocVarSlot = allocSlotFrom ig areaSize 0
- allocCallSlot areaMap (Block id stackInfo t)
- | elemBlockSet id procPoints =
- let young = youngest_live areaMap $ live_in t
- start = case returnOff stackInfo of Just b -> max b young
- Nothing -> young
- z = allocSlotFrom ig areaSize start areaMap (CallArea (Young id))
- in -- pprTrace "allocCallSlot for" (ppr id <+> ppr young <+> ppr (live_in t) <+> ppr z)
- z
- allocCallSlot areaMap _ = areaMap
- -- mid foreign calls need to have info tables placed on the stack
- allocMidCall m@(MidForeignCall (Safe bid _) _ _ _) t areaMap =
- let young = youngest_live areaMap $ removeLiveSlotDefs (live_in t) m
- area = CallArea (Young bid)
- areaSize' = addToFM areaSize area (widthInBytes (typeWidth gcWord))
- in allocSlotFrom ig areaSize' young areaMap area
- allocMidCall _ _ areaMap = areaMap
- alloc m t areaMap =
- foldSlotsDefd alloc' (foldSlotsUsed alloc' (allocMidCall m t areaMap) m) m
- where alloc' areaMap (a@(RegSlot _), _, _) = allocVarSlot areaMap a
- alloc' areaMap _ = areaMap
- layoutAreas areaMap b@(Block _ _ t) = layout areaMap t
- where layout areaMap (ZTail m t) = layout (alloc m t areaMap) t
- layout areaMap (ZLast _) = allocCallSlot areaMap b
- areaMap = foldl layoutAreas (addToFM emptyFM (CallArea Old) 0) (postorder_dfs g)
+
+ -- Update the successor's incoming SP.
+ setSuccSPs inSp bid areaMap =
+ case (Map.lookup area areaMap , mapLookup bid (toBlockMap g)) of
+ (Just _, _) -> areaMap -- succ already knows incoming SP
+ (Nothing, Just _) ->
+ if setMember bid procPoints then
+ let young = youngest_live areaMap $ env' bid
+ -- start = case returnOff stackInfo of Just b -> max b young
+ -- Nothing -> young
+ start = young -- maybe wrong, but I don't understand
+ -- why the preceding is necessary...
+ in allocSlotFrom ig areaSize start areaMap area
+ else Map.insert area inSp areaMap
+ (_, Nothing) -> panic "Block not found in cfg"
+ where area = CallArea (Young bid)
+
+ layoutAreas areaMap block = foldBlockNodesF3 (flip const, allocMid, allocLast (entryLabel block)) block areaMap
+ allocMid m areaMap = foldSlotsDefd alloc' (foldSlotsUsed alloc' areaMap m) m
+ allocLast bid l areaMap =
+ foldr (setSuccSPs inSp) areaMap' (successors l)
+ where inSp = expectJust "sp in" $ Map.lookup (CallArea (Young bid)) areaMap
+ areaMap' = foldSlotsDefd alloc' (foldSlotsUsed alloc' areaMap l) l
+ alloc' areaMap (a@(RegSlot _), _, _) = allocVarSlot areaMap a
+ alloc' areaMap _ = areaMap
+
+ initMap = Map.insert (CallArea (Young (g_entry g))) 0 $
+ Map.insert (CallArea Old) 0 Map.empty
+
+ areaMap = foldl layoutAreas initMap (postorderDfs g)
in -- pprTrace "ProcPoints" (ppr procPoints) $
- -- pprTrace "Area SizeMap" (ppr areaSize) $
+ -- pprTrace "Area SizeMap" (ppr areaSize) $
-- pprTrace "Entry SP" (ppr entrySp) $
-- pprTrace "Area Map" (ppr areaMap) $
areaMap
-- stack pointer to be younger than the live values on the stack at proc points.
-- 3. Compute the maximum stack offset used in the procedure and replace
-- the stack high-water mark with that offset.
-manifestSP :: ProcPointSet -> BlockEnv Status -> AreaMap ->
- LGraph Middle Last -> FuelMonad (LGraph Middle Last)
-manifestSP procPoints procMap areaMap g@(LGraph entry args blocks) =
- liftM (LGraph entry args) blocks'
- where blocks' = foldl replB (return emptyBlockEnv) (postorder_dfs g)
- slot a = -- pprTrace "slot" (ppr a) $
- lookupFM areaMap a `orElse` panic "unallocated Area"
+manifestSP :: AreaMap -> ByteOff -> CmmGraph -> FuelUniqSM CmmGraph
+manifestSP areaMap entry_off g@(CmmGraph {g_entry=entry}) =
+ ofBlockMap entry `liftM` foldl replB (return mapEmpty) (postorderDfs g)
+ where slot a = -- pprTrace "slot" (ppr a) $
+ Map.lookup a areaMap `orElse` panic "unallocated Area"
slot' (Just id) = slot $ CallArea (Young id)
slot' Nothing = slot $ CallArea Old
sp_high = maxSlot slot g
- proc_entry_sp = slot (CallArea Old) + args
+ proc_entry_sp = slot (CallArea Old) + entry_off
+
+ add_sp_off :: CmmBlock -> BlockEnv Int -> BlockEnv Int
+ add_sp_off b env =
+ case lastNode b of
+ CmmCall {cml_cont=Just succ, cml_ret_args=off} -> mapInsert succ off env
+ CmmForeignCall {succ=succ} -> mapInsert succ wORD_SIZE env
+ _ -> env
+ spEntryMap = foldGraphBlocks add_sp_off (mapInsert entry entry_off emptyBlockMap) g
+ spOffset id = mapLookup id spEntryMap `orElse` 0
+
sp_on_entry id | id == entry = proc_entry_sp
- sp_on_entry id =
- case lookupBlockEnv blocks id of
- Just (Block _ (StackInfo {argBytes = Just o}) _) -> slot' (Just id) + o
- _ ->
- case expectJust "sp_on_entry" (lookupBlockEnv procMap id) of
- ReachedBy pp ->
- case blockSetToList pp of
- [id] -> sp_on_entry id
- _ -> panic "block not reached by one proc point"
- ProcPoint -> pprPanic "procpoint doesn't take any arguments?"
- (ppr id <+> ppr g <+> ppr procPoints <+> ppr procMap)
+ sp_on_entry id = slot' (Just id) + spOffset id
-- On entry to procpoints, the stack pointer is conventional;
-- otherwise, we check the SP set by predecessors.
- replB :: FuelMonad (BlockEnv CmmBlock) -> CmmBlock -> FuelMonad (BlockEnv CmmBlock)
- replB blocks (Block id o t) =
- do bs <- replTail (Block id o) spIn t
- -- pprTrace "spIn" (ppr id <+> ppr spIn) $ do
- liftM (flip (foldr insertBlock) bs) blocks
- where spIn = sp_on_entry id
- replTail :: (ZTail Middle Last -> CmmBlock) -> Int -> (ZTail Middle Last) ->
- FuelMonad ([CmmBlock])
- replTail h spOff (ZTail m@(MidForeignCall (Safe bid _) _ _ _) t) =
- replTail (\t' -> h (setSp spOff spOff' (ZTail (middle spOff m) t'))) spOff' t
- where spOff' = slot' (Just bid) + widthInBytes (typeWidth gcWord)
- replTail h spOff (ZTail m t) = replTail (h . ZTail (middle spOff m)) spOff t
- replTail h spOff (ZLast (LastOther l)) = fixSp h spOff l
- replTail h _ l@(ZLast LastExit) = return [h l]
- middle spOff m = mapExpDeepMiddle (replSlot spOff) m
- last spOff l = mapExpDeepLast (replSlot spOff) l
- replSlot spOff (CmmStackSlot a i) = CmmRegOff (CmmGlobal Sp) (spOff - (slot a + i))
- replSlot spOff (CmmLit CmmHighStackMark) = -- replacing the high water mark
- CmmLit (CmmInt (toInteger (max 0 (sp_high - proc_entry_sp))) (typeWidth bWord))
- replSlot _ e = e
- -- The block must establish the SP expected at each successsor.
- fixSp :: (ZTail Middle Last -> CmmBlock) -> Int -> Last -> FuelMonad ([CmmBlock])
- fixSp h spOff l@(LastCall _ k n _) = updSp h spOff (slot' k + n) l
- fixSp h spOff l@(LastBranch k) =
- let succSp = sp_on_entry k in
- if succSp /= spOff then
- -- pprTrace "updSp" (ppr k <> ppr spOff <> ppr (sp_on_entry k)) $
- updSp h spOff succSp l
- else return $ [h (ZLast (LastOther (last spOff l)))]
- fixSp h spOff l = liftM (uncurry (:)) $ fold_succs succ l $ return (b, [])
- where b = h (ZLast (LastOther (last spOff l)))
- succ succId z =
- let succSp = sp_on_entry succId in
- if succSp /= spOff then
- do (b, bs) <- z
- (b', bs') <- insertBetween b [setSpMid spOff succSp] succId
- return (b', bs ++ bs')
- else z
- updSp h old new l = return [h $ setSp old new $ ZLast $ LastOther (last new l)]
- setSpMid sp sp' = MidAssign (CmmGlobal Sp) e
- where e = CmmMachOp (MO_Add wordWidth) [CmmReg (CmmGlobal Sp), off]
- off = CmmLit $ CmmInt (toInteger $ sp - sp') wordWidth
- setSp sp sp' t = if sp == sp' then t else ZTail (setSpMid sp sp') t
+ replB :: FuelUniqSM (BlockEnv CmmBlock) -> CmmBlock -> FuelUniqSM (BlockEnv CmmBlock)
+ replB blocks block =
+ do let (head, middles, JustC tail :: MaybeC C (CmmNode O C)) = blockToNodeList block
+ middles' = map (middle spIn) middles
+ bs <- replLast head middles' tail
+ flip (foldr insertBlock) bs `liftM` blocks
+ where spIn = sp_on_entry (entryLabel block)
+
+ middle spOff m = mapExpDeep (replSlot spOff) m
+ last spOff l = mapExpDeep (replSlot spOff) l
+ replSlot spOff (CmmStackSlot a i) = CmmRegOff (CmmGlobal Sp) (spOff - (slot a + i))
+ replSlot _ (CmmLit CmmHighStackMark) = -- replacing the high water mark
+ CmmLit (CmmInt (toInteger (max 0 (sp_high - proc_entry_sp))) (typeWidth bWord))
+ replSlot _ e = e
+
+ replLast :: MaybeC C (CmmNode C O) -> [CmmNode O O] -> CmmNode O C -> FuelUniqSM [CmmBlock]
+ replLast h m l@(CmmCall _ k n _ _) = updSp (slot' k + n) h m l
+ -- JD: LastForeignCall probably ought to have an outgoing
+ -- arg size, just like LastCall
+ replLast h m l@(CmmForeignCall {succ=k}) = updSp (slot' (Just k) + wORD_SIZE) h m l
+ replLast h m l@(CmmBranch k) = updSp (sp_on_entry k) h m l
+ replLast h m l = uncurry (:) `liftM` foldr succ (return (b, [])) (successors l)
+ where b :: CmmBlock
+ b = updSp' spIn h m l
+ succ succId z =
+ let succSp = sp_on_entry succId in
+ if succSp /= spIn then
+ do (b, bs) <- z
+ (b', bs') <- insertBetween b (adjustSp succSp) succId
+ return (b', bs' ++ bs)
+ else z
+
+ updSp sp h m l = return [updSp' sp h m l]
+ updSp' sp h m l | sp == spIn = blockOfNodeList (h, m, JustC $ last sp l)
+ | otherwise = blockOfNodeList (h, m ++ adjustSp sp, JustC $ last sp l)
+ adjustSp sp = [CmmAssign (CmmGlobal Sp) e]
+ where e = CmmMachOp (MO_Add wordWidth) [CmmReg (CmmGlobal Sp), off]
+ off = CmmLit $ CmmInt (toInteger $ spIn - sp) wordWidth
-- To compute the stack high-water mark, we fold over the graph and
-- compute the highest slot offset.
maxSlot :: (Area -> Int) -> CmmGraph -> Int
-maxSlot slotOff g = fold_blocks (fold_fwd_block (\ _ _ x -> x) highSlot highSlot) 0 g
+maxSlot slotOff g = foldGraphBlocks (foldBlockNodesF3 (flip const, highSlot, highSlot)) 0 g
where highSlot i z = foldSlotsUsed add (foldSlotsDefd add z i) i
- add z (a, i, w) = max z (slotOff a + i)
+ add z (a, i, _) = max z (slotOff a + i)
-----------------------------------------------------------------------------
-- | Sanity check: stub pointers immediately after they die
-- This will miss stack slots that are last used in a Last node,
-- but it should do pretty well...
-type StubPtrFix = FuelMonad (BackwardFixedPoint Middle Last SubAreaSet CmmGraph)
-
-stubSlotsOnDeath :: (LGraph Middle Last) -> FuelMonad (LGraph Middle Last)
-stubSlotsOnDeath g = liftM zdfFpContents $ (res :: StubPtrFix)
- where res = zdfBRewriteFromL RewriteShallow emptyBlockEnv "stub ptrs" slotLattice
- liveSlotTransfers rewrites (fact_bot slotLattice) g
- rewrites = BackwardRewrites first middle last Nothing
- first _ _ = Nothing
- last _ _ = Nothing
- middle liveSlots m = foldSlotsUsed (stub liveSlots m) Nothing m
+stubSlotsOnDeath :: CmmGraph -> FuelUniqSM CmmGraph
+stubSlotsOnDeath g = liftM fst $ dataflowPassBwd g [] $ analRewBwd slotLattice
+ liveSlotTransfers
+ rewrites
+ where rewrites = mkBRewrite3 frt mid lst
+ frt _ _ = return Nothing
+ mid m liveSlots = return $ foldSlotsUsed (stub liveSlots m) Nothing m
+ lst _ _ = return Nothing
stub liveSlots m rst subarea@(a, off, w) =
if elemSlot liveSlots subarea then rst
- else let store = mkStore (CmmStackSlot a off)
- (stackStubExpr (widthFromBytes w))
+ else let store = mkMiddle $ CmmStore (CmmStackSlot a off)
+ (stackStubExpr (widthFromBytes w))
in case rst of Nothing -> Just (mkMiddle m <*> store)
Just g -> Just (g <*> store)
projects / ghc-hetmet.git / blobdiff