X-Git-Url: http://git.megacz.com/?p=ghc-hetmet.git;a=blobdiff_plain;f=compiler%2Fcmm%2FCmmCPS.hs;h=35eabb331704e26e3a83cb8468b661105e3eeb36;hp=9a9f8a9fb22d1697bf7d2b1c0cb8092c7cdb4d63;hb=643397208b83f1654bceeef40c793f11592ef816;hpb=207802589da0d23c3f16195f453b24a1e46e322d diff --git a/compiler/cmm/CmmCPS.hs b/compiler/cmm/CmmCPS.hs index 9a9f8a9..35eabb3 100644 --- a/compiler/cmm/CmmCPS.hs +++ b/compiler/cmm/CmmCPS.hs @@ -1,378 +1,171 @@ +{-# OPTIONS_GHC -XNoMonoLocalBinds #-} +-- Norman likes local bindings +-- If this module lives on I'd like to get rid of this flag in due course + module CmmCPS ( -- | Converts C-- with full proceedures and parameters -- to a CPS transformed C-- with the stack made manifest. - cmmCPS + -- Well, sort of. + protoCmmCPS ) where -#include "HsVersions.h" - +import CLabel import Cmm -import CmmLint -import PprCmm - -import Dataflow -import CmmLive -import CmmBrokenBlock +import CmmDecl +import CmmBuildInfoTables +import CmmCommonBlockElim import CmmProcPoint -import CmmCallConv - -import MachOp -import ForeignCall -import CLabel -import SMRep -import Constants +import CmmSpillReload +import CmmStackLayout +import OptimizationFuel import DynFlags import ErrUtils -import Maybes +import HscTypes +import Data.Maybe +import Control.Monad +import Data.Map (Map) +import qualified Data.Map as Map import Outputable -import UniqSupply -import UniqFM -import UniqSet -import Unique - -import Monad -import IO -import Data.List +import StaticFlags ----------------------------------------------------------------------------- -- |Top level driver for the CPS pass ----------------------------------------------------------------------------- -cmmCPS :: DynFlags -- ^ Dynamic flags: -dcmm-lint -ddump-cps-cmm - -> [Cmm] -- ^ Input C-- with Proceedures - -> IO [Cmm] -- ^ Output CPS transformed C-- -cmmCPS dflags abstractC = do - when (dopt Opt_DoCmmLinting dflags) $ - do showPass dflags "CmmLint" - case firstJust $ map cmmLint abstractC of - Just err -> do printDump err - ghcExit dflags 1 - Nothing -> return () - showPass dflags "CPS" - - -- TODO: more lint checking - -- check for use of branches to non-existant blocks - -- check for use of Sp, SpLim, R1, R2, etc. - - uniqSupply <- mkSplitUniqSupply 'p' - let supplies = listSplitUniqSupply uniqSupply - let doCpsProc s (Cmm c) = - Cmm $ concat $ zipWith cpsProc (listSplitUniqSupply s) c - let continuationC = zipWith doCpsProc supplies abstractC - - dumpIfSet_dyn dflags Opt_D_dump_cps_cmm "CPS Cmm" (pprCmms continuationC) - - -- TODO: add option to dump Cmm to file - - return continuationC - ------------------------------------------------------------------------------ --- |CPS a single CmmTop (proceedure) --- Only 'CmmProc' are transformed 'CmmData' will be left alone. ------------------------------------------------------------------------------ - -cpsProc :: UniqSupply - -> CmmTop -- ^Input proceedure - -> [CmmTop] -- ^Output proceedure and continuations -cpsProc uniqSupply x@(CmmData _ _) = [x] -cpsProc uniqSupply x@(CmmProc info_table ident params blocks) = cps_procs - where - uniqes :: [[Unique]] - uniqes = map uniqsFromSupply $ listSplitUniqSupply uniqSupply - - -- Break the block at each function call. - -- The part after the function call will have to become a continuation. - broken_blocks :: [BrokenBlock] - broken_blocks = - concat $ zipWith3 breakBlock uniqes blocks - (FunctionEntry ident params:repeat ControlEntry) - - -- Calculate live variables for each broken block. - -- - -- Nothing can be live on entry to the first block - -- so we could take the tail, but for now we wont - -- to help future proof the code. - live :: BlockEntryLiveness - live = cmmLiveness $ map cmmBlockFromBrokenBlock broken_blocks - - -- Calculate which blocks must be made into full fledged procedures. - proc_points :: UniqSet BlockId - proc_points = calculateProcPoints broken_blocks - - -- Construct a map so we can lookup a broken block by its 'BlockId'. - block_env :: BlockEnv BrokenBlock - block_env = blocksToBlockEnv broken_blocks - - -- Group the blocks into continuations based on the set of proc-points. - continuations :: [Continuation] - continuations = map (gatherBlocksIntoContinuation proc_points block_env) - (uniqSetToList proc_points) - - -- Select the stack format on entry to each continuation. - -- - -- This is an association list instead of a UniqFM because - -- CLabel's don't have a 'Uniqueable' instance. - formats :: [(CLabel, StackFormat)] - formats = selectStackFormat live continuations - - -- Do the actual CPS transform. - cps_procs :: [CmmTop] - cps_procs = map (continuationToProc formats) continuations - --------------------------------------------------------------------------------- - --- The format for the call to a continuation --- The fst is the arguments that must be passed to the continuation --- by the continuation's caller. --- The snd is the live values that must be saved on stack. --- A Nothing indicates an ignored slot. --- The head of each list is the stack top or the first parameter. - --- The format for live values for a particular continuation --- All on stack for now. --- Head element is the top of the stack (or just under the header). --- Nothing means an empty slot. --- Future possibilities include callee save registers (i.e. passing slots in register) --- and heap memory (not sure if that's usefull at all though, but it may --- be worth exploring the design space). - -continuationLabel (Continuation _ _ l _ _) = l -data Continuation = - Continuation - Bool -- True => Function entry, False => Continuation/return point - [CmmStatic] -- Info table, may be empty - CLabel -- Used to generate both info & entry labels - CmmFormals -- Argument locals live on entry (C-- procedure params) - [BrokenBlock] -- Code, may be empty. The first block is - -- the entry point. The order is otherwise initially - -- unimportant, but at some point the code gen will - -- fix the order. - - -- the BlockId of the first block does not give rise - -- to a label. To jump to the first block in a Proc, - -- use the appropriate CLabel. - --- Describes the layout of a stack frame for a continuation -data StackFormat - = StackFormat { - stack_label :: Maybe CLabel, -- The label occupying the top slot - stack_frame_size :: WordOff, -- Total frame size in words (not including arguments) - stack_live :: [(LocalReg, WordOff)] -- local reg offsets from stack top - -- TODO: see if the above can be LocalReg - } - --- A block can be a continuation of a call --- A block can be a continuation of another block (w/ or w/o joins) --- A block can be an entry to a function - ------------------------------------------------------------------------------ - -collectNonProcPointTargets :: - UniqSet BlockId -> BlockEnv BrokenBlock - -> UniqSet BlockId -> BlockId -> UniqSet BlockId -collectNonProcPointTargets proc_points blocks current_targets block = - if sizeUniqSet current_targets == sizeUniqSet new_targets - then current_targets - else foldl (collectNonProcPointTargets proc_points blocks) new_targets targets - where - block' = lookupWithDefaultUFM blocks (panic "TODO") block - targets = - -- Note the subtlety that since the extra branch after a call - -- will always be to a block that is a proc-point, - -- this subtraction will always remove that case - uniqSetToList $ (mkUniqSet $ brokenBlockTargets block') `minusUniqSet` proc_points - -- TODO: remove redundant uniqSetToList - new_targets = current_targets `unionUniqSets` (mkUniqSet targets) - --- TODO: insert proc point code here --- * Branches and switches to proc points may cause new blocks to be created --- (or proc points could leave behind phantom blocks that just jump to them) --- * Proc points might get some live variables passed as arguments - -gatherBlocksIntoContinuation :: - UniqSet BlockId -> BlockEnv BrokenBlock - -> BlockId -> Continuation -gatherBlocksIntoContinuation proc_points blocks start = - Continuation is_entry info_table clabel params body - where - children = (collectNonProcPointTargets proc_points blocks (unitUniqSet start) start) `delOneFromUniqSet` start - start_block = lookupWithDefaultUFM blocks (panic "TODO") start - children_blocks = map (lookupWithDefaultUFM blocks (panic "TODO")) (uniqSetToList children) - body = start_block : children_blocks - info_table = [] -- TODO - start_block_entry = brokenBlockEntry start_block - is_entry = case start_block_entry of - FunctionEntry _ _ -> True - _ -> False - clabel = case start_block_entry of - FunctionEntry label _ -> label - _ -> mkReturnPtLabel $ getUnique start - params = case start_block_entry of - FunctionEntry _ args -> args - ContinuationEntry args -> args - ControlEntry -> [] -- TODO: it's a proc-point, we could pass lives in parameter registers - --------------------------------------------------------------------------------- --- For now just select the continuation orders in the order they are in the set with no gaps - -selectStackFormat :: BlockEnv CmmLive -> [Continuation] -> [(CLabel, StackFormat)] -selectStackFormat live continuations = - map (\c -> (continuationLabel c, selectStackFormat' c)) continuations - where - selectStackFormat' (Continuation True info_table label formals blocks) = - StackFormat (Just label) 0 [] - selectStackFormat' (Continuation False info_table label formals blocks) = - -- TODO: assumes the first block is the entry block - let ident = brokenBlockId $ head blocks -- TODO: CLabel isn't a uniquable, but we need a better way than this - in live_to_format label formals $ lookupWithDefaultUFM live unknown_block ident - - live_to_format :: CLabel -> CmmFormals -> CmmLive -> StackFormat - live_to_format label formals live = - foldl extend_format - (StackFormat (Just label) retAddrSizeW []) - (uniqSetToList (live `minusUniqSet` mkUniqSet formals)) - - extend_format :: StackFormat -> LocalReg -> StackFormat - extend_format (StackFormat label size offsets) reg = - StackFormat label (slot_size reg + size) ((reg, size) : offsets) - - slot_size :: LocalReg -> Int - slot_size reg = ((machRepByteWidth (localRegRep reg) - 1) `div` wORD_SIZE) + 1 - - unknown_block = panic "unknown BlockId in selectStackFormat" - -continuationToProc :: [(CLabel, StackFormat)] -> Continuation -> CmmTop -continuationToProc formats (Continuation is_entry info label formals blocks) = - CmmProc info label formals (map (continuationToProc' label formats) blocks) - where - continuationToProc' :: CLabel -> [(CLabel, StackFormat)] -> BrokenBlock - -> CmmBasicBlock - continuationToProc' curr_ident formats (BrokenBlock ident entry stmts _ exit) = - BasicBlock ident (prefix++stmts++postfix) - where - curr_format = maybe unknown_block id $ lookup curr_ident formats - unknown_block = panic "unknown BlockId in continuationToProc" - prefix = case entry of - ControlEntry -> [] - FunctionEntry _ formals -> -- TODO: gc_stack_check - function_entry formals curr_format - ContinuationEntry formals -> - function_entry formals curr_format - postfix = case exit of - FinalBranch next -> [CmmBranch next] - FinalSwitch expr targets -> [CmmSwitch expr targets] - FinalReturn arguments -> - tail_call (stack_frame_size curr_format) - (CmmLoad (CmmReg spReg) wordRep) - arguments - FinalJump target arguments -> - tail_call (stack_frame_size curr_format) target arguments - FinalCall next (CmmForeignCall target CmmCallConv) - results arguments -> - pack_continuation curr_format cont_format ++ - tail_call (stack_frame_size curr_format - stack_frame_size cont_format) - target arguments - where - cont_format = maybe unknown_block id $ - lookup (mkReturnPtLabel $ getUnique next) formats - FinalCall next _ results arguments -> panic "unimplemented CmmCall" - --------------------------------------------------------------------------------- --- Functions that generate CmmStmt sequences --- for packing/unpacking continuations --- and entering/exiting functions - -tail_call :: WordOff -> CmmExpr -> CmmActuals -> [CmmStmt] -tail_call spRel target arguments - = store_arguments ++ adjust_spReg ++ jump where - store_arguments = - [stack_put spRel expr offset - | ((expr, _), StackParam offset) <- argument_formats] ++ - [global_put expr global - | ((expr, _), RegisterParam global) <- argument_formats] - adjust_spReg = - if spRel == 0 - then [] - else [CmmAssign spReg (CmmRegOff spReg (spRel*wORD_SIZE))] - jump = [CmmJump target arguments] - - argument_formats = assignArguments (cmmExprRep . fst) arguments - -gc_stack_check :: WordOff -> [CmmStmt] -gc_stack_check max_frame_size - = check_stack_limit where - check_stack_limit = [ - CmmCondBranch - (CmmMachOp (MO_U_Lt $ cmmRegRep spReg) - [CmmRegOff spReg max_frame_size, CmmReg spLimReg]) - gc_block] - gc_block = panic "gc_check not implemented" -- TODO: get stack and heap checks to go to same - --- TODO: fix branches to proc point (we have to insert a new block to marshel the continuation) -pack_continuation :: StackFormat -> StackFormat -> [CmmStmt] -pack_continuation (StackFormat curr_id curr_frame_size _) - (StackFormat cont_id cont_frame_size cont_offsets) - = store_live_values ++ set_stack_header where - -- TODO: only save variables when actually needed (may be handled by latter pass) - store_live_values = - [stack_put spRel (CmmReg (CmmLocal reg)) offset - | (reg, offset) <- cont_offsets] - set_stack_header = - if not needs_header - then [] - else [stack_put spRel continuation_function 0] - - spRel = curr_frame_size - cont_frame_size - continuation_function = CmmLit $ CmmLabel $ fromJust cont_id - needs_header = - case (curr_id, cont_id) of - (Just x, Just y) -> x /= y - _ -> isJust cont_id - --- Lazy adjustment of stack headers assumes all blocks --- that could branch to eachother (i.e. control blocks) --- have the same stack format (this causes a problem --- only for proc-point). -function_entry :: CmmFormals -> StackFormat -> [CmmStmt] -function_entry formals (StackFormat _ _ curr_offsets) - = load_live_values ++ load_args where - -- TODO: only save variables when actually needed (may be handled by latter pass) - load_live_values = - [stack_get 0 reg offset - | (reg, offset) <- curr_offsets] - load_args = - [stack_get 0 reg offset - | (reg, StackParam offset) <- argument_formats] ++ - [global_get reg global - | (reg, RegisterParam global) <- argument_formats] - - argument_formats = assignArguments (localRegRep) formals - ------------------------------------------------------------------------------ --- Section: Stack and argument register puts and gets ------------------------------------------------------------------------------ --- TODO: document - --- |Construct a 'CmmStmt' that will save a value on the stack -stack_put :: WordOff -- ^ Offset from the real 'Sp' that 'offset' - -- is relative to (added to offset) - -> CmmExpr -- ^ What to store onto the stack - -> WordOff -- ^ Where on the stack to store it - -- (positive <=> higher addresses) - -> CmmStmt -stack_put spRel expr offset = - CmmStore (CmmRegOff spReg (wORD_SIZE*(spRel + offset))) expr - --------------------------------- --- |Construct a -stack_get :: WordOff - -> LocalReg - -> WordOff - -> CmmStmt -stack_get spRel reg offset = - CmmAssign (CmmLocal reg) (CmmLoad (CmmRegOff spReg (wORD_SIZE*(spRel + offset))) (localRegRep reg)) -global_put :: CmmExpr -> GlobalReg -> CmmStmt -global_put expr global = CmmAssign (CmmGlobal global) expr -global_get :: LocalReg -> GlobalReg -> CmmStmt -global_get reg global = CmmAssign (CmmLocal reg) (CmmReg (CmmGlobal global)) - +-- There are two complications here: +-- 1. We need to compile the procedures in two stages because we need +-- an analysis of the procedures to tell us what CAFs they use. +-- The first stage returns a map from procedure labels to CAFs, +-- along with a closure that will compute SRTs and attach them to +-- the compiled procedures. +-- The second stage is to combine the CAF information into a top-level +-- CAF environment mapping non-static closures to the CAFs they keep live, +-- then pass that environment to the closures returned in the first +-- stage of compilation. +-- 2. We need to thread the module's SRT around when the SRT tables +-- are computed for each procedure. +-- The SRT needs to be threaded because it is grown lazily. +protoCmmCPS :: HscEnv -- Compilation env including + -- dynamic flags: -dcmm-lint -ddump-cps-cmm + -> (TopSRT, [Cmm]) -- SRT table and accumulating list of compiled procs + -> Cmm -- Input C-- with Procedures + -> IO (TopSRT, [Cmm]) -- Output CPS transformed C-- +protoCmmCPS hsc_env (topSRT, rst) (Cmm tops) = + do let dflags = hsc_dflags hsc_env + showPass dflags "CPSZ" + (cafEnvs, tops) <- liftM unzip $ mapM (cpsTop hsc_env) tops + let topCAFEnv = mkTopCAFInfo (concat cafEnvs) + (topSRT, tops) <- foldM (toTops hsc_env topCAFEnv) (topSRT, []) tops + let cmms = Cmm (reverse (concat tops)) + dumpIfSet_dyn dflags Opt_D_dump_cps_cmm "Post CPS Cmm" (ppr cmms) + return (topSRT, cmms : rst) + +{- [Note global fuel] +~~~~~~~~~~~~~~~~~~~~~ +The identity and the last pass are stored in +mutable reference cells in an 'HscEnv' and are +global to one compiler session. +-} + +-- EZY: It might be helpful to have an easy way of dumping the "pre" +-- input for any given phase, besides just turning it all on with +-- -ddump-cmmz + +cpsTop :: HscEnv -> CmmTop -> IO ([(CLabel, CAFSet)], [(CAFSet, CmmTop)]) +cpsTop _ p@(CmmData {}) = return ([], [(Map.empty, p)]) +cpsTop hsc_env (CmmProc h@(TopInfo {stack_info=StackInfo {arg_space=entry_off}}) l g) = + do + -- Why bother doing these early: dualLivenessWithInsertion, + -- insertLateReloads, rewriteAssignments? + + ----------- Eliminate common blocks ------------------- + g <- return $ elimCommonBlocks g + dump Opt_D_dump_cmmz_cbe "Post common block elimination" g + -- Any work storing block Labels must be performed _after_ elimCommonBlocks + + ----------- Proc points ------------------- + let callPPs = callProcPoints g + procPoints <- run $ minimalProcPointSet callPPs g + g <- run $ addProcPointProtocols callPPs procPoints g + dump Opt_D_dump_cmmz_proc "Post Proc Points Added" g + + ----------- Spills and reloads ------------------- + g <- run $ dualLivenessWithInsertion procPoints g + dump Opt_D_dump_cmmz_spills "Post spills and reloads" g + + ----------- Sink and inline assignments ------------------- + g <- runOptimization $ rewriteAssignments g + dump Opt_D_dump_cmmz_rewrite "Post rewrite assignments" g + + ----------- Eliminate dead assignments ------------------- + -- Remove redundant reloads (and any other redundant asst) + g <- runOptimization $ removeDeadAssignmentsAndReloads procPoints g + dump Opt_D_dump_cmmz_dead "Post Dead Assignment Elimination" g + + ----------- Zero dead stack slots (Debug only) --------------- + -- Debugging: stubbing slots on death can cause crashes early + g <- if opt_StubDeadValues + then run $ stubSlotsOnDeath g + else return g + dump Opt_D_dump_cmmz_stub "Post stub dead stack slots" g + + --------------- Stack layout ---------------- + slotEnv <- run $ liveSlotAnal g + let spEntryMap = getSpEntryMap entry_off g + mbpprTrace "live slot analysis results: " (ppr slotEnv) $ return () + let areaMap = layout procPoints spEntryMap slotEnv entry_off g + mbpprTrace "areaMap" (ppr areaMap) $ return () + + ------------ Manifest the stack pointer -------- + g <- run $ manifestSP spEntryMap areaMap entry_off g + dump Opt_D_dump_cmmz_sp "Post manifestSP" g + -- UGH... manifestSP can require updates to the procPointMap. + -- We can probably do something quicker here for the update... + + ------------- Split into separate procedures ------------ + procPointMap <- run $ procPointAnalysis procPoints g + dump Opt_D_dump_cmmz_procmap "procpoint map" procPointMap + gs <- run $ splitAtProcPoints l callPPs procPoints procPointMap + (CmmProc h l g) + mapM_ (dump Opt_D_dump_cmmz_split "Post splitting") gs + + ------------- More CAFs and foreign calls ------------ + cafEnv <- run $ cafAnal g + let localCAFs = catMaybes $ map (localCAFInfo cafEnv) gs + mbpprTrace "localCAFs" (ppr localCAFs) $ return () + + gs <- run $ mapM (lowerSafeForeignCalls areaMap) gs + mapM_ (dump Opt_D_dump_cmmz_lower "Post lowerSafeForeignCalls") gs + + -- NO MORE GRAPH TRANSFORMATION AFTER HERE -- JUST MAKING INFOTABLES + gs <- return $ map (setInfoTableStackMap slotEnv areaMap) gs + mapM_ (dump Opt_D_dump_cmmz_info "after setInfoTableStackMap") gs + gs <- return $ map (bundleCAFs cafEnv) gs + mapM_ (dump Opt_D_dump_cmmz_cafs "after bundleCAFs") gs + return (localCAFs, gs) + where dflags = hsc_dflags hsc_env + mbpprTrace x y z = if dopt Opt_D_dump_cmmz dflags then pprTrace x y z else z + dump f txt g = do + -- ToDo: No easy way of say "dump all the cmmz, *and* split + -- them into files." Also, -ddump-cmmz doesn't play nicely + -- with -ddump-to-file, since the headers get omitted. + dumpIfSet_dyn dflags f txt (ppr g) + when (not (dopt f dflags)) $ + dumpIfSet_dyn dflags Opt_D_dump_cmmz txt (ppr g) + -- Runs a required transformation/analysis + run = runInfiniteFuelIO (hsc_OptFuel hsc_env) + -- Runs an optional transformation/analysis (and should + -- thus be subject to optimization fuel) + runOptimization = runFuelIO (hsc_OptFuel hsc_env) + +-- This probably belongs in CmmBuildInfoTables? +-- We're just finishing the job here: once we know what CAFs are defined +-- in non-static closures, we can build the SRTs. +toTops :: HscEnv -> Map CLabel CAFSet -> (TopSRT, [[CmmTop]]) + -> [(CAFSet, CmmTop)] -> IO (TopSRT, [[CmmTop]]) +toTops hsc_env topCAFEnv (topSRT, tops) gs = + do let setSRT (topSRT, rst) g = + do (topSRT, gs) <- setInfoTableSRT topCAFEnv topSRT g + return (topSRT, gs : rst) + (topSRT, gs') <- runFuelIO (hsc_OptFuel hsc_env) $ foldM setSRT (topSRT, []) gs + return (topSRT, concat gs' : tops)