X-Git-Url: http://git.megacz.com/?p=ghc-hetmet.git;a=blobdiff_plain;f=compiler%2Fcmm%2FCmmSpillReload.hs;h=a4bedb0baabc871866f9a84371000153b67db921;hp=b588c46af6d43cdbf7ca2ece5e3dbeb0251564ab;hb=ffd3bd85a6febeec05c99d0da7dfdf34cad59caf;hpb=c0a5a5d2e41341046aaf37c1d2155372e7ed3ee8 diff --git a/compiler/cmm/CmmSpillReload.hs b/compiler/cmm/CmmSpillReload.hs index b588c46..a4bedb0 100644 --- a/compiler/cmm/CmmSpillReload.hs +++ b/compiler/cmm/CmmSpillReload.hs @@ -1,54 +1,62 @@ -{-# OPTIONS -Wall -fno-warn-name-shadowing #-} +{-# LANGUAGE GADTs, NoMonoLocalBinds, FlexibleContexts #-} +-- Norman likes local bindings +-- If this module lives on I'd like to get rid of this flag in due course + +{-# OPTIONS_GHC -fno-warn-warnings-deprecations #-} +#if __GLASGOW_HASKELL__ >= 701 +-- GHC 7.0.1 improved incomplete pattern warnings with GADTs +{-# OPTIONS_GHC -fwarn-incomplete-patterns #-} +#endif module CmmSpillReload - ( ExtendWithSpills(..) - , DualLive(..) - , dualLiveLattice, dualLiveness - , insertSpillsAndReloads --- XXX todo check live-in at entry against formals + ( DualLive(..) + , dualLiveLattice, dualLiveTransfers, dualLiveness + --, insertSpillsAndReloads --- XXX todo check live-in at entry against formals , dualLivenessWithInsertion - , spillAndReloadComments - , availRegsLattice - , cmmAvailableReloads + , removeDeadAssignmentsAndReloads ) where + +import BlockId +import Cmm import CmmExpr -import CmmTx -import CmmLiveZ -import DFMonad -import FastString -import Maybe -import MkZipCfg +import CmmLive +import OptimizationFuel + +import Control.Monad import Outputable hiding (empty) import qualified Outputable as PP -import Panic -import PprCmm() import UniqSet -import ZipCfg -import ZipCfgCmmRep -import ZipDataflow - --- The point of this module is to insert spills and reloads to --- establish the invariant that at a call (or at any proc point with --- an established protocol) all live variables not expected in --- registers are sitting on the stack. We use a backward analysis to --- insert spills and reloads. It should some day be followed by a --- forward transformation to sink reloads as deeply as possible, so as --- to reduce register pressure. - -data ExtendWithSpills m - = NotSpillOrReload m - | Spill RegSet - | Reload RegSet - -type M = ExtendWithSpills Middle - --- A variable can be expected to be live in a register, live on the --- stack, or both. This analysis ensures that spills and reloads are --- inserted as needed to make sure that every live variable needed --- after a call is available on the stack. Spills are pushed back to --- their reaching definitions, but reloads are dropped wherever needed --- and will have to be sunk by a later forward transformation. + +import Compiler.Hoopl hiding (Unique) +import Data.Maybe +import Prelude hiding (succ, zip) + +{- Note [Overview of spill/reload] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +The point of this module is to insert spills and reloads to +establish the invariant that at a call (or at any proc point with +an established protocol) all live variables not expected in +registers are sitting on the stack. We use a backward analysis to +insert spills and reloads. It should be followed by a +forward transformation to sink reloads as deeply as possible, so as +to reduce register pressure. + +A variable can be expected to be live in a register, live on the +stack, or both. This analysis ensures that spills and reloads are +inserted as needed to make sure that every live variable needed +after a call is available on the stack. Spills are pushed back to +their reaching definitions, but reloads are dropped immediately after +we return from a call and will have to be sunk by a later forward +transformation. + +Note that we offer no guarantees about the consistency of the value +in memory and the value in the register, except that they are +equal across calls/procpoints. If the variable is changed, this +mapping breaks: but as the original value of the register may still +be useful in a different context, the memory location is not updated. +-} data DualLive = DualLive { on_stack :: RegSet, in_regs :: RegSet } @@ -61,227 +69,141 @@ dualUnionList ls = DualLive ss rs where ss = unionManyUniqSets $ map on_stack ls rs = unionManyUniqSets $ map in_regs ls -_changeStack, changeRegs :: (RegSet -> RegSet) -> DualLive -> DualLive -_changeStack f live = live { on_stack = f (on_stack live) } -changeRegs f live = live { in_regs = f (in_regs live) } +changeStack, changeRegs :: (RegSet -> RegSet) -> DualLive -> DualLive +changeStack f live = live { on_stack = f (on_stack live) } +changeRegs f live = live { in_regs = f (in_regs live) } dualLiveLattice :: DataflowLattice DualLive -dualLiveLattice = - DataflowLattice "variables live in registers and on stack" empty add False +dualLiveLattice = DataflowLattice "variables live in registers and on stack" empty add where empty = DualLive emptyRegSet emptyRegSet - -- | compute in the Tx monad to track whether anything has changed - add new old = do stack <- add1 (on_stack new) (on_stack old) - regs <- add1 (in_regs new) (in_regs old) - return $ DualLive stack regs - add1 = fact_add_to liveLattice - -dualLivenessWithInsertion :: BPass M Last DualLive -dualLivenessWithInsertion = a_ft_b_unlimited dualLiveness insertSpillsAndReloads - - -dualLiveness :: BAnalysis M Last DualLive -dualLiveness = BComp "dual liveness" exit last middle first - where exit = empty - last = lastDualLiveness - middle = middleDualLiveness - first live _id = live - empty = fact_bot dualLiveLattice - - -- ^ could take a proc-point set and choose to spill here, - -- but it's probably better to run this pass, choose - -- proc-point protocols, insert more CopyIn nodes, and run - -- this pass again - -middleDualLiveness :: DualLive -> M -> DualLive -middleDualLiveness live m@(Spill regs) = - -- live-in on-stack requirements are satisfied; - -- live-out in-regs obligations are created - my_trace "before" (f4sep [ppr m, text "liveness is", ppr live']) $ - live' - where live' = DualLive { on_stack = on_stack live `minusRegSet` regs - , in_regs = in_regs live `plusRegSet` regs } - -middleDualLiveness live m@(Reload regs) = - -- live-in in-regs requirements are satisfied; - -- live-out on-stack obligations are created - my_trace "before" (f4sep [ppr m, text "liveness is", ppr live']) $ - live' - where live' = DualLive { on_stack = on_stack live `plusRegSet` regs - , in_regs = in_regs live `minusRegSet` regs } - -middleDualLiveness live (NotSpillOrReload m) = changeRegs (middleLiveness m) live - -lastDualLiveness :: (BlockId -> DualLive) -> Last -> DualLive -lastDualLiveness env l = last l - where last (LastReturn ress) = changeRegs (gen ress) empty - last (LastJump e args) = changeRegs (gen e . gen args) empty - last (LastBranch id args) = changeRegs (gen args) $ env id - last (LastCall tgt args Nothing) = changeRegs (gen tgt. gen args) empty - last (LastCall tgt args (Just k)) = - -- nothing can be live in registers at this point - -- only 'formals' can be in regs at this point - let live = env k in - if isEmptyUniqSet (in_regs live) then - DualLive (on_stack live) (gen tgt $ gen args emptyRegSet) - else - panic "live values in registers at call continuation" - last (LastCondBranch e t f) = changeRegs (gen e) $ dualUnion (env t) (env f) - last (LastSwitch e tbl) = changeRegs (gen e) $ dualUnionList $ - map env (catMaybes tbl) - empty = fact_bot dualLiveLattice - -gen, kill :: UserOfLocalRegs a => a -> RegSet -> RegSet -gen a live = foldRegsUsed extendRegSet live a -kill a live = foldRegsUsed delOneFromUniqSet live a - -insertSpillsAndReloads :: BFunctionalTransformation M Last DualLive -insertSpillsAndReloads = BComp "CPS spiller" exit last middle first - where exit = Nothing - last = \_ _ -> Nothing - middle = middleInsertSpillsAndReloads - first _ _ = Nothing - -- ^ could take a proc-point set and choose to spill here, - -- but it's probably better to run this pass, choose - -- proc-point protocols, insert more CopyIn nodes, and run - -- this pass again - - -middleInsertSpillsAndReloads :: DualLive -> M -> Maybe (Graph M Last) -middleInsertSpillsAndReloads _ (Spill _) = Nothing -middleInsertSpillsAndReloads _ (Reload _) = Nothing -middleInsertSpillsAndReloads live (NotSpillOrReload m) = middle m - where middle (MidAssign (CmmLocal reg') _) = - if reg' `elemRegSet` on_stack live then -- must spill - my_trace "Spilling" (f4sep [text "spill" <+> ppr reg', - text "after", ppr m]) $ - Just $ graphOfMiddles [NotSpillOrReload m, Spill $ mkRegSet [reg']] - else - Nothing - middle (CopyIn _ formals _) = - -- only 'formals' can be in regs at this point - let regs' = kill formals (in_regs live) -- live in regs; must reload - is_stack_var r = elemRegSet r (on_stack live) - needs_spilling = -- a formal that is expected on the stack; must spill - foldRegsUsed (\rs r -> if is_stack_var r then extendRegSet rs r - else rs) emptyRegSet formals - in if isEmptyUniqSet regs' && isEmptyUniqSet needs_spilling then - Nothing - else - let reload = if isEmptyUniqSet regs' then [] - else [Reload regs'] - spill_reload = if isEmptyUniqSet needs_spilling then reload - else Spill needs_spilling : reload - middles = NotSpillOrReload m : spill_reload - in - my_trace "At CopyIn" (f4sep [text "Triggered by ", ppr live, - ppr (Reload regs' :: M), - ppr (Spill needs_spilling :: M), - text "after", ppr m]) $ - Just $ graphOfMiddles middles - middle _ = Nothing - --- | For conversion back to vanilla C-- -spillAndReloadComments :: M -> Middle -spillAndReloadComments (NotSpillOrReload m) = m -spillAndReloadComments (Spill regs) = show_regs "Spill" regs -spillAndReloadComments (Reload regs) = show_regs "Reload" regs - -show_regs :: String -> RegSet -> Middle -show_regs s regs = MidComment $ mkFastString $ showSDoc $ ppr_regs s regs - - ----------------------------------------------------------------- ---- sinking reloads - --- The idea is to compute at each point the set of registers such that --- on every path to the point, the register is defined by a Reload --- instruction. Then, if a use appears at such a point, we can safely --- insert a Reload right before the use. Finally, we can eliminate --- the early reloads along with other dead assignments. - -data AvailRegs = UniverseMinus RegSet - | AvailRegs RegSet - - -availRegsLattice :: DataflowLattice AvailRegs -availRegsLattice = DataflowLattice "register gotten from reloads" empty add True - where empty = UniverseMinus emptyRegSet - -- | compute in the Tx monad to track whether anything has changed - add new old = - let join = interAvail new old in - if join `smallerAvail` old then aTx join else noTx join - - -interAvail :: AvailRegs -> AvailRegs -> AvailRegs -interAvail (UniverseMinus s) (UniverseMinus s') = UniverseMinus (s `plusRegSet` s') -interAvail (AvailRegs s) (AvailRegs s') = AvailRegs (s `timesRegSet` s') -interAvail (AvailRegs s) (UniverseMinus s') = AvailRegs (s `minusRegSet` s') -interAvail (UniverseMinus s) (AvailRegs s') = AvailRegs (s' `minusRegSet` s ) - -smallerAvail :: AvailRegs -> AvailRegs -> Bool -smallerAvail (AvailRegs _) (UniverseMinus _) = True -smallerAvail (UniverseMinus _) (AvailRegs _) = False -smallerAvail (AvailRegs s) (AvailRegs s') = sizeUniqSet s < sizeUniqSet s' -smallerAvail (UniverseMinus s) (UniverseMinus s') = sizeUniqSet s > sizeUniqSet s' - -extendAvail :: AvailRegs -> LocalReg -> AvailRegs -extendAvail (UniverseMinus s) r = UniverseMinus (deleteFromRegSet s r) -extendAvail (AvailRegs s) r = AvailRegs (extendRegSet s r) - -deleteFromAvail :: AvailRegs -> LocalReg -> AvailRegs -deleteFromAvail (UniverseMinus s) r = UniverseMinus (extendRegSet s r) -deleteFromAvail (AvailRegs s) r = AvailRegs (deleteFromRegSet s r) - -cmmAvailableReloads :: LGraph M Last -> BlockEnv AvailRegs -cmmAvailableReloads g = env - where env = runDFA availRegsLattice $ - do run_f_anal transfer (fact_bot availRegsLattice) g - allFacts - transfer :: FAnalysis M Last AvailRegs - transfer = FComp "available-reloads analysis" first middle last exit - exit _ = LastOutFacts [] - first avail _ = avail - middle = flip middleAvail - last = lastAvail - - --- | The transfer equations use the traditional 'gen' and 'kill' --- notations, which should be familiar from the dragon book. -agen, akill :: UserOfLocalRegs a => a -> AvailRegs -> AvailRegs -agen a live = foldRegsUsed extendAvail live a -akill a live = foldRegsUsed deleteFromAvail live a - -middleAvail :: M -> AvailRegs -> AvailRegs -middleAvail (Spill _) = id -middleAvail (Reload regs) = agen regs -middleAvail (NotSpillOrReload m) = middle m - where middle (MidNop) = id - middle (MidComment {}) = id - middle (MidAssign lhs _expr) = akill lhs - middle (MidStore {}) = id - middle (MidUnsafeCall _tgt ress _args) = akill ress - middle (CopyIn _ formals _) = akill formals - middle (CopyOut {}) = id - -lastAvail :: AvailRegs -> Last -> LastOutFacts AvailRegs -lastAvail _ (LastCall _ _ (Just k)) = LastOutFacts [(k, AvailRegs emptyRegSet)] -lastAvail avail l = LastOutFacts $ map (\id -> (id, avail)) $ succs l - + add _ (OldFact old) (NewFact new) = (changeIf $ change1 || change2, DualLive stack regs) + where (change1, stack) = add1 (on_stack old) (on_stack new) + (change2, regs) = add1 (in_regs old) (in_regs new) + add1 old new = if sizeUniqSet join > sizeUniqSet old then (True, join) else (False, old) + where join = unionUniqSets old new + +dualLivenessWithInsertion :: BlockSet -> CmmGraph -> FuelUniqSM CmmGraph +dualLivenessWithInsertion procPoints g = + liftM fst $ dataflowPassBwd g [] $ analRewBwd dualLiveLattice + (dualLiveTransfers (g_entry g) procPoints) + (insertSpillAndReloadRewrites g procPoints) + +dualLiveness :: BlockSet -> CmmGraph -> FuelUniqSM (BlockEnv DualLive) +dualLiveness procPoints g = + liftM snd $ dataflowPassBwd g [] $ analBwd dualLiveLattice $ dualLiveTransfers (g_entry g) procPoints + +dualLiveTransfers :: BlockId -> BlockSet -> (BwdTransfer CmmNode DualLive) +dualLiveTransfers entry procPoints = mkBTransfer3 first middle last + where first :: CmmNode C O -> DualLive -> DualLive + first (CmmEntry id) live = check live id $ -- live at procPoint => spill + if id /= entry && setMember id procPoints + then DualLive { on_stack = on_stack live `plusRegSet` in_regs live + , in_regs = emptyRegSet } + else live + where check live id x = if id == entry then noLiveOnEntry id (in_regs live) x else x + + middle :: CmmNode O O -> DualLive -> DualLive + middle m = changeStack updSlots + . changeRegs updRegs + where -- Reuse middle of liveness analysis from CmmLive + updRegs = case getBTransfer3 xferLive of (_, middle, _) -> middle m + + updSlots live = foldSlotsUsed reload (foldSlotsDefd spill live m) m + spill live s@(RegSlot r, _, _) = check s $ deleteFromRegSet live r + spill live _ = live + reload live s@(RegSlot r, _, _) = check s $ extendRegSet live r + reload live _ = live + check (RegSlot (LocalReg _ ty), o, w) x + | o == w && w == widthInBytes (typeWidth ty) = x + check _ _ = panic "middleDualLiveness unsupported: slices" + last :: CmmNode O C -> FactBase DualLive -> DualLive + last l fb = case l of + CmmBranch id -> lkp id + l@(CmmCall {cml_cont=Nothing}) -> changeRegs (gen l . kill l) empty + l@(CmmCall {cml_cont=Just k}) -> call l k + l@(CmmForeignCall {succ=k}) -> call l k + l@(CmmCondBranch _ t f) -> changeRegs (gen l . kill l) $ dualUnion (lkp t) (lkp f) + l@(CmmSwitch _ tbl) -> changeRegs (gen l . kill l) $ dualUnionList $ map lkp (catMaybes tbl) + where empty = fact_bot dualLiveLattice + lkp id = empty `fromMaybe` lookupFact id fb + call l k = DualLive (on_stack (lkp k)) (gen l emptyRegSet) + +gen :: UserOfLocalRegs a => a -> RegSet -> RegSet +gen a live = foldRegsUsed extendRegSet live a +kill :: DefinerOfLocalRegs a => a -> RegSet -> RegSet +kill a live = foldRegsDefd deleteFromRegSet live a + +insertSpillAndReloadRewrites :: CmmGraph -> BlockSet -> CmmBwdRewrite DualLive +insertSpillAndReloadRewrites graph procPoints = deepBwdRw3 first middle nothing + -- Beware: deepBwdRw with one polymorphic function seems more reasonable here, + -- but GHC miscompiles it, see bug #4044. + where first :: CmmNode C O -> Fact O DualLive -> CmmReplGraph C O + first e@(CmmEntry id) live = return $ + if id /= (g_entry graph) && setMember id procPoints then + case map reload (uniqSetToList spill_regs) of + [] -> Nothing + is -> Just $ mkFirst e <*> mkMiddles is + else Nothing + where + -- If we are splitting procedures, we need the LastForeignCall + -- to spill its results to the stack because they will only + -- be used by a separate procedure (so they can't stay in LocalRegs). + splitting = True + spill_regs = if splitting then in_regs live + else in_regs live `minusRegSet` defs + defs = case mapLookup id firstDefs of + Just defs -> defs + Nothing -> emptyRegSet + -- A LastForeignCall may contain some definitions, which take place + -- on return from the function call. Therefore, we build a map (firstDefs) + -- from BlockId to the set of variables defined on return to the BlockId. + firstDefs = mapFold addLive emptyBlockMap (toBlockMap graph) + addLive :: CmmBlock -> BlockEnv RegSet -> BlockEnv RegSet + addLive b env = case lastNode b of + CmmForeignCall {succ=k, res=defs} -> add k (mkRegSet defs) env + _ -> env + add bid defs env = mapInsert bid defs'' env + where defs'' = case mapLookup bid env of + Just defs' -> timesRegSet defs defs' + Nothing -> defs + + middle :: CmmNode O O -> Fact O DualLive -> CmmReplGraph O O + middle (CmmAssign (CmmLocal reg) (CmmLoad (CmmStackSlot (RegSlot reg') _) _)) _ | reg == reg' = return Nothing + middle m@(CmmAssign (CmmLocal reg) _) live = return $ + if reg `elemRegSet` on_stack live then -- must spill + my_trace "Spilling" (f4sep [text "spill" <+> ppr reg, + text "after"{-, ppr m-}]) $ + Just $ mkMiddles $ [m, spill reg] + else Nothing + middle _ _ = return Nothing + + nothing _ _ = return Nothing + +spill, reload :: LocalReg -> CmmNode O O +spill r = CmmStore (regSlot r) (CmmReg $ CmmLocal r) +reload r = CmmAssign (CmmLocal r) (CmmLoad (regSlot r) $ localRegType r) + +removeDeadAssignmentsAndReloads :: BlockSet -> CmmGraph -> FuelUniqSM CmmGraph +removeDeadAssignmentsAndReloads procPoints g = + liftM fst $ dataflowPassBwd g [] $ analRewBwd dualLiveLattice + (dualLiveTransfers (g_entry g) procPoints) + rewrites + where rewrites = deepBwdRw3 nothing middle nothing + -- Beware: deepBwdRw with one polymorphic function seems more reasonable here, + -- but GHC panics while compiling, see bug #4045. + middle :: CmmNode O O -> Fact O DualLive -> CmmReplGraph O O + middle (CmmAssign (CmmLocal reg') _) live | not (reg' `elemRegSet` in_regs live) = return $ Just emptyGraph + -- XXX maybe this should be somewhere else... + middle (CmmAssign lhs (CmmReg rhs)) _ | lhs == rhs = return $ Just emptyGraph + middle (CmmStore lhs (CmmLoad rhs _)) _ | lhs == rhs = return $ Just emptyGraph + middle _ _ = return Nothing + + nothing _ _ = return Nothing --------------------- -- prettyprinting -instance Outputable m => Outputable (ExtendWithSpills m) where - ppr (Spill regs) = ppr_regs "Spill" regs - ppr (Reload regs) = ppr_regs "Reload" regs - ppr (NotSpillOrReload m) = ppr m - -instance Outputable (LGraph M Last) where - ppr = pprLgraph - -instance DebugNodes M Last - ppr_regs :: String -> RegSet -> SDoc ppr_regs s regs = text s <+> commafy (map ppr $ uniqSetToList regs) where commafy xs = hsep $ punctuate comma xs @@ -296,10 +218,6 @@ instance Outputable DualLive where if isEmptyUniqSet stack then PP.empty else (ppr_regs "live on stack =" stack)] -instance Outputable AvailRegs where - ppr (UniverseMinus s) = ppr_regs "available = all but" s - ppr (AvailRegs s) = ppr_regs "available = " s - my_trace :: String -> SDoc -> a -> a my_trace = if False then pprTrace else \_ _ a -> a