X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FsimplCore%2FSimplCore.lhs;h=a386a3d6b021e5138480656f480042fff77dbeae;hb=28a464a75e14cece5db40f2765a29348273ff2d2;hp=2356d85055d28a0c09ab69397946cbf3f09f4559;hpb=731f53de7930c38b5023a871146bd0ec066edf3a;p=ghc-hetmet.git diff --git a/ghc/compiler/simplCore/SimplCore.lhs b/ghc/compiler/simplCore/SimplCore.lhs index 2356d85..a386a3d 100644 --- a/ghc/compiler/simplCore/SimplCore.lhs +++ b/ghc/compiler/simplCore/SimplCore.lhs @@ -4,74 +4,55 @@ \section[SimplCore]{Driver for simplifying @Core@ programs} \begin{code} -module SimplCore ( core2core ) where +module SimplCore ( core2core, simplifyExpr ) where #include "HsVersions.h" -import CmdLineOpts ( CoreToDo(..), SimplifierSwitch(..), - SwitchResult(..), switchIsOn, intSwitchSet, - opt_D_dump_occur_anal, opt_D_dump_rules, - opt_D_dump_simpl_iterations, - opt_D_dump_simpl_stats, - opt_D_dump_simpl, opt_D_dump_rules, - opt_D_verbose_core2core, - opt_D_dump_occur_anal, - opt_UsageSPOn, - ) -import CoreLint ( beginPass, endPass ) +import DynFlags ( CoreToDo(..), SimplifierSwitch(..), + SimplifierMode(..), DynFlags, DynFlag(..), dopt, + getCoreToDo ) import CoreSyn +import HscTypes ( HscEnv(..), ModGuts(..), ExternalPackageState(..), + Dependencies( dep_mods ), + hscEPS, hptRules ) import CSE ( cseProgram ) -import Rules ( RuleBase, ProtoCoreRule(..), pprProtoCoreRule, prepareRuleBase, orphanRule ) -import CoreUnfold -import PprCore ( pprCoreBindings ) -import OccurAnal ( occurAnalyseBinds ) -import CoreUtils ( exprIsTrivial, coreExprType ) +import Rules ( RuleBase, emptyRuleBase, mkRuleBase, unionRuleBase, + extendRuleBaseList, pprRuleBase, ruleCheckProgram, + addSpecInfo, addIdSpecialisations ) +import PprCore ( pprCoreBindings, pprCoreExpr, pprRules ) +import OccurAnal ( occurAnalysePgm, occurAnalyseExpr ) +import IdInfo ( setNewStrictnessInfo, newStrictnessInfo, + setWorkerInfo, workerInfo, + setSpecInfo, specInfo, specInfoRules ) +import CoreUtils ( coreBindsSize ) import Simplify ( simplTopBinds, simplExpr ) -import SimplUtils ( etaCoreExpr, findDefault, simplBinders ) +import SimplEnv ( SimplEnv, simplBinders, mkSimplEnv, setInScopeSet ) import SimplMonad -import Const ( Con(..), Literal(..), literalType, mkMachInt ) -import ErrUtils ( dumpIfSet ) +import ErrUtils ( dumpIfSet, dumpIfSet_dyn, showPass ) +import CoreLint ( endPass ) import FloatIn ( floatInwards ) import FloatOut ( floatOutwards ) -import Id ( Id, mkSysLocal, mkVanillaId, isBottomingId, - idType, setIdType, idName, idInfo, setIdNoDiscard - ) -import VarEnv +import Id ( Id, modifyIdInfo, idInfo, isExportedId, isLocalId, + idSpecialisation, idName ) import VarSet -import Module ( Module ) -import Name ( mkLocalName, tidyOccName, tidyTopName, - NamedThing(..), OccName - ) -import TyCon ( TyCon, isDataTyCon ) -import PrimOp ( PrimOp(..) ) -import PrelInfo ( unpackCStringId, unpackCString2Id, addr2IntegerId ) -import Type ( Type, splitAlgTyConApp_maybe, - isUnLiftedType, - tidyType, tidyTypes, tidyTopType, tidyTyVar, tidyTyVars, - Type - ) -import TysWiredIn ( smallIntegerDataCon, isIntegerTy ) +import VarEnv +import NameEnv ( lookupNameEnv ) import LiberateCase ( liberateCase ) import SAT ( doStaticArgs ) import Specialise ( specProgram) -import UsageSPInf ( doUsageSPInf ) -import StrictAnal ( saBinds ) +import SpecConstr ( specConstrProgram) +import DmdAnal ( dmdAnalPgm ) import WorkWrap ( wwTopBinds ) +#ifdef OLD_STRICTNESS +import StrictAnal ( saBinds ) import CprAnalyse ( cprAnalyse ) +#endif -import Unique ( Unique, Uniquable(..), - ratioTyConKey - ) -import UniqSupply ( UniqSupply, mkSplitUniqSupply, splitUniqSupply, uniqFromSupply ) -import Constants ( tARGET_MIN_INT, tARGET_MAX_INT ) -import Util ( mapAccumL ) -import SrcLoc ( noSrcLoc ) -import Bag -import Maybes +import UniqSupply ( UniqSupply, mkSplitUniqSupply, splitUniqSupply ) import IO ( hPutStr, stderr ) import Outputable - -import Ratio ( numerator, denominator ) +import List ( partition ) +import Maybes ( orElse ) \end{code} %************************************************************************ @@ -81,114 +62,304 @@ import Ratio ( numerator, denominator ) %************************************************************************ \begin{code} -core2core :: [CoreToDo] -- Spec of what core-to-core passes to do - -> [CoreBind] -- Binds in - -> [ProtoCoreRule] -- Rules - -> IO ([CoreBind], [ProtoCoreRule]) +core2core :: HscEnv + -> ModGuts + -> IO ModGuts -core2core core_todos binds rules +core2core hsc_env guts = do - us <- mkSplitUniqSupply 's' - let (cp_us, us1) = splitUniqSupply us - (ru_us, ps_us) = splitUniqSupply us1 + let dflags = hsc_dflags hsc_env + core_todos = getCoreToDo dflags - better_rules <- simplRules ru_us rules binds + us <- mkSplitUniqSupply 's' + let (cp_us, ru_us) = splitUniqSupply us - let (binds1, rule_base) = prepareRuleBase binds better_rules + -- COMPUTE THE RULE BASE TO USE + (imp_rule_base, guts') <- prepareRules hsc_env guts ru_us - -- Do the main business - (stats, processed_binds) <- doCorePasses zeroSimplCount cp_us binds1 - rule_base core_todos + -- DO THE BUSINESS + (stats, guts'') <- doCorePasses hsc_env imp_rule_base cp_us + (zeroSimplCount dflags) + guts' core_todos - dumpIfSet opt_D_dump_simpl_stats + dumpIfSet_dyn dflags Opt_D_dump_simpl_stats "Grand total simplifier statistics" (pprSimplCount stats) - -- Do the post-simplification business - post_simpl_binds <- doPostSimplification ps_us processed_binds + return guts'' - -- Return results - return (post_simpl_binds, filter orphanRule better_rules) - -doCorePasses stats us binds irs [] - = return (stats, binds) +simplifyExpr :: DynFlags -- includes spec of what core-to-core passes to do + -> CoreExpr + -> IO CoreExpr +-- simplifyExpr is called by the driver to simplify an +-- expression typed in at the interactive prompt +simplifyExpr dflags expr + = do { + ; showPass dflags "Simplify" -doCorePasses stats us binds irs (to_do : to_dos) - = do - let (us1, us2) = splitUniqSupply us - (stats1, binds1) <- doCorePass us1 binds irs to_do - doCorePasses (stats `plusSimplCount` stats1) us2 binds1 irs to_dos - -doCorePass us binds rb (CoreDoSimplify sw_chkr) = _scc_ "Simplify" simplifyPgm rb sw_chkr us binds -doCorePass us binds rb CoreCSE = _scc_ "CommonSubExpr" noStats (cseProgram binds) -doCorePass us binds rb CoreLiberateCase = _scc_ "LiberateCase" noStats (liberateCase binds) -doCorePass us binds rb CoreDoFloatInwards = _scc_ "FloatInwards" noStats (floatInwards binds) -doCorePass us binds rb CoreDoFullLaziness = _scc_ "FloatOutwards" noStats (floatOutwards us binds) -doCorePass us binds rb CoreDoStaticArgs = _scc_ "StaticArgs" noStats (doStaticArgs us binds) -doCorePass us binds rb CoreDoStrictness = _scc_ "Stranal" noStats (saBinds binds) -doCorePass us binds rb CoreDoWorkerWrapper = _scc_ "WorkWrap" noStats (wwTopBinds us binds) -doCorePass us binds rb CoreDoSpecialising = _scc_ "Specialise" noStats (specProgram us binds) -doCorePass us binds rb CoreDoCPResult = _scc_ "CPResult" noStats (cprAnalyse binds) -doCorePass us binds rb CoreDoPrintCore = _scc_ "PrintCore" noStats (printCore binds) -doCorePass us binds rb CoreDoUSPInf - = _scc_ "CoreUsageSPInf" - if opt_UsageSPOn then - noStats (doUsageSPInf us binds) - else - trace "WARNING: ignoring requested -fusagesp pass; requires -fusagesp-on" $ - noStats (return binds) + ; us <- mkSplitUniqSupply 's' + + ; let (expr', _counts) = initSmpl dflags us $ + simplExprGently gentleSimplEnv expr + + ; dumpIfSet_dyn dflags Opt_D_dump_simpl "Simplified expression" + (pprCoreExpr expr') + + ; return expr' + } + +gentleSimplEnv :: SimplEnv +gentleSimplEnv = mkSimplEnv SimplGently + (isAmongSimpl []) + emptyRuleBase + +doCorePasses :: HscEnv + -> RuleBase -- the imported main rule base + -> UniqSupply -- uniques + -> SimplCount -- simplifier stats + -> ModGuts -- local binds in (with rules attached) + -> [CoreToDo] -- which passes to do + -> IO (SimplCount, ModGuts) -printCore binds = do dumpIfSet True "Print Core" - (pprCoreBindings binds) - return binds +doCorePasses hsc_env rb us stats guts [] + = return (stats, guts) -noStats thing = do { result <- thing; return (zeroSimplCount, result) } +doCorePasses hsc_env rb us stats guts (to_do : to_dos) + = do + let (us1, us2) = splitUniqSupply us + (stats1, guts1) <- doCorePass to_do hsc_env us1 rb guts + doCorePasses hsc_env rb us2 (stats `plusSimplCount` stats1) guts1 to_dos + +doCorePass (CoreDoSimplify mode sws) = _scc_ "Simplify" simplifyPgm mode sws +doCorePass CoreCSE = _scc_ "CommonSubExpr" trBinds cseProgram +doCorePass CoreLiberateCase = _scc_ "LiberateCase" trBinds liberateCase +doCorePass CoreDoFloatInwards = _scc_ "FloatInwards" trBinds floatInwards +doCorePass (CoreDoFloatOutwards f) = _scc_ "FloatOutwards" trBindsU (floatOutwards f) +doCorePass CoreDoStaticArgs = _scc_ "StaticArgs" trBinds doStaticArgs +doCorePass CoreDoStrictness = _scc_ "Stranal" trBinds dmdAnalPgm +doCorePass CoreDoWorkerWrapper = _scc_ "WorkWrap" trBindsU wwTopBinds +doCorePass CoreDoSpecialising = _scc_ "Specialise" trBindsU specProgram +doCorePass CoreDoSpecConstr = _scc_ "SpecConstr" trBindsU specConstrProgram +doCorePass CoreDoGlomBinds = trBinds glomBinds +doCorePass CoreDoPrintCore = observe printCore +doCorePass (CoreDoRuleCheck phase pat) = observe (ruleCheck phase pat) +doCorePass CoreDoNothing = observe (\ _ _ -> return ()) +#ifdef OLD_STRICTNESS +doCorePass CoreDoOldStrictness = _scc_ "OldStrictness" trBinds doOldStrictness +#endif + +#ifdef OLD_STRICTNESS +doOldStrictness dfs binds + = do binds1 <- saBinds dfs binds + binds2 <- cprAnalyse dfs binds1 + return binds2 +#endif + +printCore _ binds = dumpIfSet True "Print Core" (pprCoreBindings binds) + +ruleCheck phase pat dflags binds = do showPass dflags "RuleCheck" + printDump (ruleCheckProgram phase pat binds) + +-- Most passes return no stats and don't change rules +trBinds :: (DynFlags -> [CoreBind] -> IO [CoreBind]) + -> HscEnv -> UniqSupply -> RuleBase -> ModGuts + -> IO (SimplCount, ModGuts) +trBinds do_pass hsc_env us rb guts + = do { binds' <- do_pass dflags (mg_binds guts) + ; return (zeroSimplCount dflags, guts { mg_binds = binds' }) } + where + dflags = hsc_dflags hsc_env + +trBindsU :: (DynFlags -> UniqSupply -> [CoreBind] -> IO [CoreBind]) + -> HscEnv -> UniqSupply -> RuleBase -> ModGuts + -> IO (SimplCount, ModGuts) +trBindsU do_pass hsc_env us rb guts + = do { binds' <- do_pass dflags us (mg_binds guts) + ; return (zeroSimplCount dflags, guts { mg_binds = binds' }) } + where + dflags = hsc_dflags hsc_env + +-- Observer passes just peek; don't modify the bindings at all +observe :: (DynFlags -> [CoreBind] -> IO a) + -> HscEnv -> UniqSupply -> RuleBase -> ModGuts + -> IO (SimplCount, ModGuts) +observe do_pass hsc_env us rb guts + = do { binds <- do_pass dflags (mg_binds guts) + ; return (zeroSimplCount dflags, guts) } + where + dflags = hsc_dflags hsc_env \end{code} + %************************************************************************ %* * \subsection{Dealing with rules} %* * %************************************************************************ -We must do some gentle simplifiation on the template (but not the RHS) +-- prepareLocalRuleBase takes the CoreBinds and rules defined in this module. +-- It attaches those rules that are for local Ids to their binders, and +-- returns the remainder attached to Ids in an IdSet. + +\begin{code} +prepareRules :: HscEnv + -> ModGuts + -> UniqSupply + -> IO (RuleBase, -- Rule base for imported things, incl + -- (a) rules defined in this module (orphans) + -- (b) rules from other modules in home package + -- but not things from other packages + + ModGuts) -- Modified fields are + -- (a) Bindings have rules attached, + -- (b) Rules are now just orphan rules + +prepareRules hsc_env@(HscEnv { hsc_dflags = dflags, hsc_HPT = hpt }) + guts@(ModGuts { mg_binds = binds, mg_deps = deps, mg_rules = local_rules }) + us + = do { let -- Simplify the local rules; boringly, we need to make an in-scope set + -- from the local binders, to avoid warnings from Simplify.simplVar + local_ids = mkInScopeSet (mkVarSet (bindersOfBinds binds)) + env = setInScopeSet gentleSimplEnv local_ids + (better_rules,_) = initSmpl dflags us (mapSmpl (simplRule env) local_rules) + home_pkg_rules = hptRules hsc_env (dep_mods deps) + + -- Find the rules for locally-defined Ids; then we can attach them + -- to the binders in the top-level bindings + -- + -- Reason + -- - It makes the rules easier to look up + -- - It means that transformation rules and specialisations for + -- locally defined Ids are handled uniformly + -- - It keeps alive things that are referred to only from a rule + -- (the occurrence analyser knows about rules attached to Ids) + -- - It makes sure that, when we apply a rule, the free vars + -- of the RHS are more likely to be in scope + -- - The imported rules are carried in the in-scope set + -- which is extended on each iteration by the new wave of + -- local binders; any rules which aren't on the binding will + -- thereby get dropped + (rules_for_locals, rules_for_imps) = partition isLocalRule better_rules + local_rule_base = extendRuleBaseList emptyRuleBase rules_for_locals + binds_w_rules = updateBinders local_rule_base binds + + hpt_rule_base = mkRuleBase home_pkg_rules + imp_rule_base = extendRuleBaseList hpt_rule_base rules_for_imps + + ; dumpIfSet_dyn dflags Opt_D_dump_rules "Transformation rules" + (vcat [text "Local rules", pprRules better_rules, + text "", + text "Imported rules", pprRuleBase imp_rule_base]) + + ; return (imp_rule_base, guts { mg_binds = binds_w_rules, + mg_rules = rules_for_imps }) + } + +updateBinders :: RuleBase -> [CoreBind] -> [CoreBind] +updateBinders local_rules binds + = map update_bndrs binds + where + update_bndrs (NonRec b r) = NonRec (update_bndr b) r + update_bndrs (Rec prs) = Rec [(update_bndr b, r) | (b,r) <- prs] + + update_bndr bndr = case lookupNameEnv local_rules (idName bndr) of + Nothing -> bndr + Just rules -> bndr `addIdSpecialisations` rules + -- The binder might have some existing rules, + -- arising from specialisation pragmas +\end{code} + + +We must do some gentle simplification on the template (but not the RHS) of each rule. The case that forced me to add this was the fold/build rule, which without simplification looked like: fold k z (build (/\a. g a)) ==> ... This doesn't match unless you do eta reduction on the build argument. \begin{code} -simplRules :: UniqSupply -> [ProtoCoreRule] -> [CoreBind] -> IO [ProtoCoreRule] -simplRules us rules binds - = do let (better_rules,_) = initSmpl sw_chkr us bind_vars black_list_all (mapSmpl simplRule rules) - - dumpIfSet opt_D_dump_rules - "Transformation rules" - (vcat (map pprProtoCoreRule better_rules)) +simplRule env rule@(BuiltinRule {}) + = returnSmpl rule +simplRule env rule@(Rule { ru_bndrs = bndrs, ru_args = args, ru_rhs = rhs }) + = simplBinders env bndrs `thenSmpl` \ (env, bndrs') -> + mapSmpl (simplExprGently env) args `thenSmpl` \ args' -> + simplExprGently env rhs `thenSmpl` \ rhs' -> + returnSmpl (rule { ru_bndrs = bndrs', ru_args = args', ru_rhs = rhs' }) + +-- It's important that simplExprGently does eta reduction. +-- For example, in a rule like: +-- augment g (build h) +-- we do not want to get +-- augment (\a. g a) (build h) +-- otherwise we don't match when given an argument like +-- (\a. h a a) +-- +-- The simplifier does indeed do eta reduction (it's in +-- Simplify.completeLam) but only if -O is on. +\end{code} - return better_rules - where - black_list_all v = True -- This stops all inlining - sw_chkr any = SwBool False -- A bit bogus - - -- Boringly, we need to gather the in-scope set. - -- Typically this thunk won't even be force, but the test in - -- simpVar fails if it isn't right, and it might conceivably matter - bind_vars = foldr (unionVarSet . mkVarSet . bindersOf) emptyVarSet binds - - -simplRule rule@(ProtoCoreRule is_local id (Rule name bndrs args rhs)) - | not is_local - = returnSmpl rule -- No need to fiddle with imported rules - | otherwise - = simplBinders bndrs $ \ bndrs' -> - mapSmpl simplExpr args `thenSmpl` \ args' -> - simplExpr rhs `thenSmpl` \ rhs' -> - returnSmpl (ProtoCoreRule is_local id (Rule name bndrs' args' rhs')) +\begin{code} +simplExprGently :: SimplEnv -> CoreExpr -> SimplM CoreExpr +-- Simplifies an expression +-- does occurrence analysis, then simplification +-- and repeats (twice currently) because one pass +-- alone leaves tons of crud. +-- Used (a) for user expressions typed in at the interactive prompt +-- (b) the LHS and RHS of a RULE +-- +-- The name 'Gently' suggests that the SimplifierMode is SimplGently, +-- and in fact that is so.... but the 'Gently' in simplExprGently doesn't +-- enforce that; it just simplifies the expression twice + +simplExprGently env expr + = simplExpr env (occurAnalyseExpr expr) `thenSmpl` \ expr1 -> + simplExpr env (occurAnalyseExpr expr1) \end{code} + +%************************************************************************ +%* * +\subsection{Glomming} +%* * +%************************************************************************ + +\begin{code} +glomBinds :: DynFlags -> [CoreBind] -> IO [CoreBind] +-- Glom all binds together in one Rec, in case any +-- transformations have introduced any new dependencies +-- +-- NB: the global invariant is this: +-- *** the top level bindings are never cloned, and are always unique *** +-- +-- We sort them into dependency order, but applying transformation rules may +-- make something at the top refer to something at the bottom: +-- f = \x -> p (q x) +-- h = \y -> 3 +-- +-- RULE: p (q x) = h x +-- +-- Applying this rule makes f refer to h, +-- although it doesn't appear to in the source program. +-- This pass lets us control where it happens. +-- +-- NOTICE that this cannot happen for rules whose head is a locally-defined +-- function. It only happens for rules whose head is an imported function +-- (p in the example above). So, for example, the rule had been +-- RULE: f (p x) = h x +-- then the rule for f would be attached to f itself (in its IdInfo) +-- by prepareLocalRuleBase and h would be regarded by the occurrency +-- analyser as free in f. + +glomBinds dflags binds + = do { showPass dflags "GlomBinds" ; + let { recd_binds = [Rec (flattenBinds binds)] } ; + return recd_binds } + -- Not much point in printing the result... + -- just consumes output bandwidth +\end{code} + + %************************************************************************ %* * \subsection{The driver for the simplifier} @@ -196,110 +367,117 @@ simplRule rule@(ProtoCoreRule is_local id (Rule name bndrs args rhs)) %************************************************************************ \begin{code} -simplifyPgm :: RuleBase - -> (SimplifierSwitch -> SwitchResult) +simplifyPgm :: SimplifierMode + -> [SimplifierSwitch] + -> HscEnv -> UniqSupply - -> [CoreBind] -- Input - -> IO (SimplCount, [CoreBind]) -- New bindings + -> RuleBase + -> ModGuts + -> IO (SimplCount, ModGuts) -- New bindings -simplifyPgm (imported_rule_ids, rule_lhs_fvs) - sw_chkr us binds +simplifyPgm mode switches hsc_env us imp_rule_base guts = do { - beginPass "Simplify"; - - -- Glom all binds together in one Rec, in case any - -- transformations have introduced any new dependencies - -- - -- NB: the global invariant is this: - -- *** the top level bindings are never cloned, and are always unique *** - -- - -- We sort them into dependency order, but applying transformation rules may - -- make something at the top refer to something at the bottom: - -- f = \x -> p (q x) - -- h = \y -> 3 - -- - -- RULE: p (q x) = h x - -- - -- Applying this rule makes f refer to h, although it doesn't appear to in the - -- source program. Our solution is to do this occasional glom-together step, - -- just once per overall simplfication step. - - let { recd_binds = [Rec (flattenBinds binds)] }; - - (termination_msg, it_count, counts_out, binds') <- iteration us 1 zeroSimplCount recd_binds; - - dumpIfSet (opt_D_verbose_core2core && opt_D_dump_simpl_stats) + showPass dflags "Simplify"; + + (termination_msg, it_count, counts_out, binds') + <- do_iteration us 1 (zeroSimplCount dflags) (mg_binds guts) ; + + dumpIfSet (dopt Opt_D_verbose_core2core dflags + && dopt Opt_D_dump_simpl_stats dflags) "Simplifier statistics" (vcat [text termination_msg <+> text "after" <+> ppr it_count <+> text "iterations", text "", pprSimplCount counts_out]); - endPass "Simplify" - (opt_D_verbose_core2core && not opt_D_dump_simpl_iterations) - binds' ; + endPass dflags "Simplify" Opt_D_verbose_core2core binds'; - return (counts_out, binds') + return (counts_out, guts { mg_binds = binds' }) } where - max_iterations = getSimplIntSwitch sw_chkr MaxSimplifierIterations - black_list_fn = blackListed rule_lhs_fvs (intSwitchSet sw_chkr SimplInlinePhase) - - core_iter_dump binds | opt_D_verbose_core2core = pprCoreBindings binds - | otherwise = empty - - iteration us iteration_no counts binds + dflags = hsc_dflags hsc_env + phase_info = case mode of + SimplGently -> "gentle" + SimplPhase n -> show n + + sw_chkr = isAmongSimpl switches + max_iterations = intSwitchSet sw_chkr MaxSimplifierIterations `orElse` 2 + + do_iteration us iteration_no counts binds + -- iteration_no is the number of the iteration we are + -- about to begin, with '1' for the first + | iteration_no > max_iterations -- Stop if we've run out of iterations + = do { +#ifdef DEBUG + if max_iterations > 2 then + hPutStr stderr ("NOTE: Simplifier still going after " ++ + show max_iterations ++ + " iterations; bailing out.\n") + else + return (); +#endif + -- Subtract 1 from iteration_no to get the + -- number of iterations we actually completed + return ("Simplifier baled out", iteration_no - 1, counts, binds) + } + + -- Try and force thunks off the binds; significantly reduces + -- space usage, especially with -O. JRS, 000620. + | let sz = coreBindsSize binds in sz == sz = do { -- Occurrence analysis - let { tagged_binds = _scc_ "OccAnal" occurAnalyseBinds binds } ; - - dumpIfSet opt_D_dump_occur_anal "Occurrence analysis" + let { tagged_binds = _scc_ "OccAnal" occurAnalysePgm binds } ; + dumpIfSet_dyn dflags Opt_D_dump_occur_anal "Occurrence analysis" (pprCoreBindings tagged_binds); - -- Simplify - let { (binds', counts') = initSmpl sw_chkr us1 imported_rule_ids - black_list_fn - (simplTopBinds tagged_binds); - -- The imported_rule_ids are used by initSmpl to initialise - -- the in-scope set. That way, the simplifier will change any - -- occurrences of the imported id to the one in the imported_rule_ids - -- set, which are decorated with their rules. - - all_counts = counts `plusSimplCount` counts' - } ; + -- Get any new rules, and extend the rule base + -- We need to do this regularly, because simplification can + -- poke on IdInfo thunks, which in turn brings in new rules + -- behind the scenes. Otherwise there's a danger we'll simply + -- miss the rules for Ids hidden inside imported inlinings + eps <- hscEPS hsc_env ; + let { rule_base' = unionRuleBase imp_rule_base (eps_rule_base eps) + ; simpl_env = mkSimplEnv mode sw_chkr rule_base' } ; + + -- Simplify the program + -- We do this with a *case* not a *let* because lazy pattern + -- matching bit us with bad space leak! + -- With a let, we ended up with + -- let + -- t = initSmpl ... + -- counts' = snd t + -- in + -- case t of {(_,counts') -> if counts'=0 then ... } + -- So the conditional didn't force counts', because the + -- selection got duplicated. Sigh! + case initSmpl dflags us1 (_scc_ "SimplTopBinds" simplTopBinds simpl_env tagged_binds) of { + (binds', counts') -> do { + + let { all_counts = counts `plusSimplCount` counts' + ; herald = "Simplifier phase " ++ phase_info ++ + ", iteration " ++ show iteration_no ++ + " out of " ++ show max_iterations + } ; -- Stop if nothing happened; don't dump output if isZeroSimplCount counts' then - return ("Simplifier reached fixed point", iteration_no, all_counts, binds') + return ("Simplifier reached fixed point", iteration_no, + all_counts, binds') else do { + -- Short out indirections + -- We do this *after* at least one run of the simplifier + -- because indirection-shorting uses the export flag on *occurrences* + -- and that isn't guaranteed to be ok until after the first run propagates + -- stuff from the binding site to its occurrences + let { binds'' = _scc_ "ZapInd" shortOutIndirections binds' } ; -- Dump the result of this iteration - dumpIfSet opt_D_dump_simpl_iterations - ("Simplifier iteration " ++ show iteration_no - ++ " out of " ++ show max_iterations) - (pprSimplCount counts') ; - - if opt_D_dump_simpl_iterations then - endPass ("Simplifier iteration " ++ show iteration_no ++ " result") - opt_D_verbose_core2core - binds' - else - return [] ; - - -- Stop if we've run out of iterations - if iteration_no == max_iterations then - do { - if max_iterations > 2 then - hPutStr stderr ("NOTE: Simplifier still going after " ++ - show max_iterations ++ - " iterations; bailing out.\n") - else return (); - - return ("Simplifier baled out", iteration_no, all_counts, binds') - } - - -- Else loop - else iteration us2 (iteration_no + 1) all_counts binds' - } } + dumpIfSet_dyn dflags Opt_D_dump_simpl_iterations herald + (pprSimplCount counts') ; + endPass dflags herald Opt_D_dump_simpl_iterations binds'' ; + + -- Loop + do_iteration us2 (iteration_no + 1) all_counts binds'' + } } } } where (us1, us2) = splitUniqSupply us \end{code} @@ -307,283 +485,190 @@ simplifyPgm (imported_rule_ids, rule_lhs_fvs) %************************************************************************ %* * -\subsection{PostSimplification} + Shorting out indirections %* * %************************************************************************ -Several tasks are performed by the post-simplification pass - -1. Make the representation of NoRep literals explicit, and - float their bindings to the top level. We only do the floating - part for NoRep lits inside a lambda (else no gain). We need to - take care with let x = "foo" in e - that we don't end up with a silly binding - let x = y in e - with a floated "foo". What a bore. - -4. Do eta reduction for lambda abstractions appearing in: - - the RHS of case alternatives - - the body of a let - - These will otherwise turn into local bindings during Core->STG; - better to nuke them if possible. (In general the simplifier does - eta expansion not eta reduction, up to this point. It does eta - on the RHSs of bindings but not the RHSs of case alternatives and - let bodies) +If we have this: - -------------------- NOT DONE ANY MORE ------------------------ -[March 98] Indirections are now elimianted by the occurrence analyser -1. Eliminate indirections. The point here is to transform - x_local = E + x_local = + ...bindings... x_exported = x_local - ==> - x_exported = E - -[Dec 98] [Not now done because there is no penalty in the code - generator for using the former form] -2. Convert - case x of {...; x' -> ...x'...} - ==> - case x of {...; _ -> ...x... } - See notes in SimplCase.lhs, near simplDefault for the reasoning here. --------------------------------------------------------------- - -Special case -~~~~~~~~~~~~ - -NOT ENABLED AT THE MOMENT (because the floated Ids are global-ish -things, and we need local Ids for non-floated stuff): - Don't float stuff out of a binder that's marked as a bottoming Id. - Reason: it doesn't do any good, and creates more CAFs that increase - the size of SRTs. +where x_exported is exported, and x_local is not, then we replace it with this: -eg. + x_exported = + x_local = x_exported + ...bindings... - f = error "string" +Without this we never get rid of the x_exported = x_local thing. This +save a gratuitous jump (from \tr{x_exported} to \tr{x_local}), and +makes strictness information propagate better. This used to happen in +the final phase, but it's tidier to do it here. -is translated to +STRICTNESS: if we have done strictness analysis, we want the strictness info on +x_local to transfer to x_exported. Hence the copyIdInfo call. - f' = unpackCString# "string" - f = error f' +RULES: we want to *add* any RULES for x_local to x_exported. -hence f' and f become CAFs. Instead, the special case for -tidyTopBinding below makes sure this comes out as +Note [Rules and indirection-zapping] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +Problem: what if x_exported has a RULE that mentions something in ...bindings...? +Then the things mentioned can be out of scope! Solution + a) Make sure that in this pass the usage-info from x_exported is + available for ...bindings... + b) If there are any such RULES, rec-ify the entire top-level. + It'll get sorted out next time round - f = let f' = unpackCString# "string" in error f' - -and we can safely ignore f as a CAF, since it can only ever be entered once. +Messing up the rules +~~~~~~~~~~~~~~~~~~~~ +The example that went bad on me at one stage was this one: + + iterate :: (a -> a) -> a -> [a] + [Exported] + iterate = iterateList + + iterateFB c f x = x `c` iterateFB c f (f x) + iterateList f x = x : iterateList f (f x) + [Not exported] + + {-# RULES + "iterate" forall f x. iterate f x = build (\c _n -> iterateFB c f x) + "iterateFB" iterateFB (:) = iterateList + #-} +This got shorted out to: + iterateList :: (a -> a) -> a -> [a] + iterateList = iterate + + iterateFB c f x = x `c` iterateFB c f (f x) + iterate f x = x : iterate f (f x) + + {-# RULES + "iterate" forall f x. iterate f x = build (\c _n -> iterateFB c f x) + "iterateFB" iterateFB (:) = iterate + #-} + +And now we get an infinite loop in the rule system + iterate f x -> build (\cn -> iterateFB c f x) + -> iterateFB (:) f x + -> iterate f x + +Tiresome old solution: + don't do shorting out if f has rewrite rules (see shortableIdInfo) + +New solution (I think): + use rule switching-off pragmas to get rid + of iterateList in the first place + + +Other remarks +~~~~~~~~~~~~~ +If more than one exported thing is equal to a local thing (i.e., the +local thing really is shared), then we do one only: +\begin{verbatim} + x_local = .... + x_exported1 = x_local + x_exported2 = x_local +==> + x_exported1 = .... + + x_exported2 = x_exported1 +\end{verbatim} + +We rely on prior eta reduction to simplify things like +\begin{verbatim} + x_exported = /\ tyvars -> x_local tyvars +==> + x_exported = x_local +\end{verbatim} +Hence,there's a possibility of leaving unchanged something like this: +\begin{verbatim} + x_local = .... + x_exported1 = x_local Int +\end{verbatim} +By the time we've thrown away the types in STG land this +could be eliminated. But I don't think it's very common +and it's dangerous to do this fiddling in STG land +because we might elminate a binding that's mentioned in the +unfolding for something. \begin{code} -doPostSimplification :: UniqSupply -> [CoreBind] -> IO [CoreBind] -doPostSimplification us binds_in - = do - beginPass "Post-simplification pass" - let binds_out = initPM us (postSimplTopBinds binds_in) - endPass "Post-simplification pass" opt_D_verbose_core2core binds_out - -postSimplTopBinds :: [CoreBind] -> PostM [CoreBind] -postSimplTopBinds binds - = mapPM postSimplTopBind binds `thenPM` \ binds' -> - returnPM (bagToList (unionManyBags binds')) - -postSimplTopBind :: CoreBind -> PostM (Bag CoreBind) -postSimplTopBind (NonRec bndr rhs) - | isBottomingId bndr -- Don't lift out floats for bottoming Ids - -- See notes above - = getFloatsPM (postSimplExpr rhs) `thenPM` \ (rhs', floats) -> - returnPM (unitBag (NonRec bndr (foldrBag Let rhs' floats))) - -postSimplTopBind bind - = getFloatsPM (postSimplBind bind) `thenPM` \ (bind', floats) -> - returnPM (floats `snocBag` bind') - -postSimplBind (NonRec bndr rhs) - = postSimplExpr rhs `thenPM` \ rhs' -> - returnPM (NonRec bndr rhs') - -postSimplBind (Rec pairs) - = mapPM postSimplExpr rhss `thenPM` \ rhss' -> - returnPM (Rec (bndrs `zip` rhss')) - where - (bndrs, rhss) = unzip pairs -\end{code} - +type IndEnv = IdEnv Id -- Maps local_id -> exported_id -Expressions -~~~~~~~~~~~ -\begin{code} -postSimplExpr (Var v) = returnPM (Var v) -postSimplExpr (Type ty) = returnPM (Type ty) - -postSimplExpr (App fun arg) - = postSimplExpr fun `thenPM` \ fun' -> - postSimplExpr arg `thenPM` \ arg' -> - returnPM (App fun' arg') - -postSimplExpr (Con (Literal lit) args) - = ASSERT( null args ) - litToRep lit `thenPM` \ (lit_ty, lit_expr) -> - getInsideLambda `thenPM` \ in_lam -> - if in_lam && not (exprIsTrivial lit_expr) then - -- It must have been a no-rep literal with a - -- non-trivial representation; and we're inside a lambda; - -- so float it to the top - addTopFloat lit_ty lit_expr `thenPM` \ v -> - returnPM (Var v) - else - returnPM lit_expr - -postSimplExpr (Con con args) - = mapPM postSimplExpr args `thenPM` \ args' -> - returnPM (Con con args') - -postSimplExpr (Lam bndr body) - = insideLambda bndr $ - postSimplExpr body `thenPM` \ body' -> - returnPM (Lam bndr body') - -postSimplExpr (Let bind body) - = postSimplBind bind `thenPM` \ bind' -> - postSimplExprEta body `thenPM` \ body' -> - returnPM (Let bind' body') - -postSimplExpr (Note note body) - = postSimplExpr body `thenPM` \ body' -> - -- Do *not* call postSimplExprEta here - -- We don't want to turn f = \x -> coerce t (\y -> f x y) - -- into f = \x -> coerce t (f x) - -- because then f has a lower arity. - -- This is not only bad in general, it causes the arity to - -- not match the [Demand] on an Id, - -- which confuses the importer of this module. - returnPM (Note note body') - -postSimplExpr (Case scrut case_bndr alts) - = postSimplExpr scrut `thenPM` \ scrut' -> - mapPM ps_alt alts `thenPM` \ alts' -> - returnPM (Case scrut' case_bndr alts') +shortOutIndirections :: [CoreBind] -> [CoreBind] +shortOutIndirections binds + | isEmptyVarEnv ind_env = binds + | no_need_to_flatten = binds' + | otherwise = [Rec (flattenBinds binds')] -- See Note [Rules and indirect-zapping] where - ps_alt (con,bndrs,rhs) = postSimplExprEta rhs `thenPM` \ rhs' -> - returnPM (con, bndrs, rhs') - -postSimplExprEta e = postSimplExpr e `thenPM` \ e' -> - returnPM (etaCoreExpr e') -\end{code} - - -%************************************************************************ -%* * -\subsection[coreToStg-lits]{Converting literals} -%* * -%************************************************************************ - -Literals: the NoRep kind need to be de-no-rep'd. -We always replace them with a simple variable, and float a suitable -binding out to the top level. - -\begin{code} -litToRep :: Literal -> PostM (Type, CoreExpr) - -litToRep (NoRepStr s ty) - = returnPM (ty, rhs) + ind_env = makeIndEnv binds + exp_ids = varSetElems ind_env -- These exported Ids are the subjects + exp_id_set = mkVarSet exp_ids -- of the indirection-elimination + no_need_to_flatten = all (null . specInfoRules . idSpecialisation) exp_ids + binds' = concatMap zap binds + + zap (NonRec bndr rhs) = [NonRec b r | (b,r) <- zapPair (bndr,rhs)] + zap (Rec pairs) = [Rec (concatMap zapPair pairs)] + + zapPair (bndr, rhs) + | bndr `elemVarSet` exp_id_set = [] + | Just exp_id <- lookupVarEnv ind_env bndr = [(transferIdInfo exp_id bndr, rhs), + (bndr, Var exp_id)] + | otherwise = [(bndr,rhs)] + +makeIndEnv :: [CoreBind] -> IndEnv +makeIndEnv binds + = foldr add_bind emptyVarEnv binds where - rhs = if (any is_NUL (_UNPK_ s)) - - then -- Must cater for NULs in literal string - mkApps (Var unpackCString2Id) - [mkLit (MachStr s), - mkLit (mkMachInt (toInteger (_LENGTH_ s)))] - - else -- No NULs in the string - App (Var unpackCStringId) (mkLit (MachStr s)) - - is_NUL c = c == '\0' -\end{code} + add_bind :: CoreBind -> IndEnv -> IndEnv + add_bind (NonRec exported_id rhs) env = add_pair (exported_id, rhs) env + add_bind (Rec pairs) env = foldr add_pair env pairs + + add_pair :: (Id,CoreExpr) -> IndEnv -> IndEnv + add_pair (exported_id, Var local_id) env + | shortMeOut env exported_id local_id = extendVarEnv env local_id exported_id + add_pair (exported_id, rhs) env + = env + +shortMeOut ind_env exported_id local_id +-- The if-then-else stuff is just so I can get a pprTrace to see +-- how often I don't get shorting out becuase of IdInfo stuff + = if isExportedId exported_id && -- Only if this is exported + + isLocalId local_id && -- Only if this one is defined in this + -- module, so that we *can* change its + -- binding to be the exported thing! + + not (isExportedId local_id) && -- Only if this one is not itself exported, + -- since the transformation will nuke it + + not (local_id `elemVarEnv` ind_env) -- Only if not already substituted for + then + True + +{- No longer needed + if isEmptySpecInfo (specInfo (idInfo exported_id)) -- Only if no rules + then True -- See note on "Messing up rules" + else +#ifdef DEBUG + pprTrace "shortMeOut:" (ppr exported_id) +#endif + False +-} + else + False -If an Integer is small enough (Haskell implementations must support -Ints in the range $[-2^29+1, 2^29-1]$), wrap it up in @int2Integer@; -otherwise, wrap with @addr2Integer@. -\begin{code} -litToRep (NoRepInteger i integer_ty) - = returnPM (integer_ty, rhs) - where - rhs | i > tARGET_MIN_INT && -- Small enough, so start from an Int - i < tARGET_MAX_INT - = Con (DataCon smallIntegerDataCon) [Con (Literal (mkMachInt i)) []] - - | otherwise -- Big, so start from a string - = App (Var addr2IntegerId) (Con (Literal (MachStr (_PK_ (show i)))) []) - - -litToRep (NoRepRational r rational_ty) - = postSimplExpr (mkLit (NoRepInteger (numerator r) integer_ty)) `thenPM` \ num_arg -> - postSimplExpr (mkLit (NoRepInteger (denominator r) integer_ty)) `thenPM` \ denom_arg -> - returnPM (rational_ty, mkConApp ratio_data_con [Type integer_ty, num_arg, denom_arg]) +----------------- +transferIdInfo :: Id -> Id -> Id +transferIdInfo exported_id local_id + = modifyIdInfo transfer exported_id where - (ratio_data_con, integer_ty) - = case (splitAlgTyConApp_maybe rational_ty) of - Just (tycon, [i_ty], [con]) - -> ASSERT(isIntegerTy i_ty && getUnique tycon == ratioTyConKey) - (con, i_ty) - - _ -> (panic "ratio_data_con", panic "integer_ty") - -litToRep other_lit = returnPM (literalType other_lit, mkLit other_lit) -\end{code} - - -%************************************************************************ -%* * -\subsection{The monad} -%* * -%************************************************************************ - -\begin{code} -type PostM a = Bool -- True <=> inside a *value* lambda - -> (UniqSupply, Bag CoreBind) -- Unique supply and Floats in - -> (a, (UniqSupply, Bag CoreBind)) - -initPM :: UniqSupply -> PostM a -> a -initPM us m - = case m False {- not inside lambda -} (us, emptyBag) of - (result, _) -> result - -returnPM v in_lam usf = (v, usf) -thenPM m k in_lam usf = case m in_lam usf of - (r, usf') -> k r in_lam usf' - -mapPM f [] = returnPM [] -mapPM f (x:xs) = f x `thenPM` \ r -> - mapPM f xs `thenPM` \ rs -> - returnPM (r:rs) - -insideLambda :: CoreBndr -> PostM a -> PostM a -insideLambda bndr m in_lam usf | isId bndr = m True usf - | otherwise = m in_lam usf - -getInsideLambda :: PostM Bool -getInsideLambda in_lam usf = (in_lam, usf) - -getFloatsPM :: PostM a -> PostM (a, Bag CoreBind) -getFloatsPM m in_lam (us, floats) - = let - (a, (us', floats')) = m in_lam (us, emptyBag) - in - ((a, floats'), (us', floats)) - -addTopFloat :: Type -> CoreExpr -> PostM Id -addTopFloat lit_ty lit_rhs in_lam (us, floats) - = let - (us1, us2) = splitUniqSupply us - uniq = uniqFromSupply us1 - lit_id = mkSysLocal SLIT("lf") uniq lit_ty - in - (lit_id, (us2, floats `snocBag` NonRec lit_id lit_rhs)) + local_info = idInfo local_id + transfer exp_info = exp_info `setNewStrictnessInfo` newStrictnessInfo local_info + `setWorkerInfo` workerInfo local_info + `setSpecInfo` addSpecInfo (specInfo exp_info) + (specInfo local_info) \end{code} - -