X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FsimplCore%2FOccurAnal.lhs;h=021ee87fdb2dc62294163e4068eb3d7eb2220e10;hb=f2054812b868d752c75dd0b148cc884b00f5e5d6;hp=b04eb4b031c74d852e9624b4b28c70da4a5130ce;hpb=6c381e873e222417d9a67aeec77b9555eca7b7a8;p=ghc-hetmet.git diff --git a/ghc/compiler/simplCore/OccurAnal.lhs b/ghc/compiler/simplCore/OccurAnal.lhs index b04eb4b..021ee87 100644 --- a/ghc/compiler/simplCore/OccurAnal.lhs +++ b/ghc/compiler/simplCore/OccurAnal.lhs @@ -1,5 +1,5 @@ % -% (c) The GRASP/AQUA Project, Glasgow University, 1992-1995 +% (c) The GRASP/AQUA Project, Glasgow University, 1992-1998 % %************************************************************************ %* * @@ -11,192 +11,207 @@ The occurrence analyser re-typechecks a core expression, returning a new core expression with (hopefully) improved usage information. \begin{code} -#include "HsVersions.h" - module OccurAnal ( - occurAnalyseBinds, occurAnalyseExpr, occurAnalyseGlobalExpr - - -- and to make the interface self-sufficient... + occurAnalyseBinds, occurAnalyseGlobalExpr, occurAnalyseRule ) where -import Type -import BinderInfo -import CmdLineOpts ( GlobalSwitch(..), SimplifierSwitch(..) ) -import Digraph ( stronglyConnComp ) -import Id ( eqId, idWantsToBeINLINEd, isConstMethodId, - isSpecPragmaId_maybe, SpecInfo ) -import Maybes -import UniqSet -import Util +#include "HsVersions.h" + +import CoreSyn +import CoreFVs ( idRuleVars ) +import CoreUtils ( exprIsTrivial ) +import Id ( isDataConId, isOneShotLambda, setOneShotLambda, + idOccInfo, setIdOccInfo, + isExportedId, modifyIdInfo, idInfo, idArity, + idSpecialisation, isLocalId, + idType, idUnique, Id + ) +import IdInfo ( shortableIdInfo, copyIdInfo ) +import BasicTypes ( OccInfo(..), isOneOcc ) + +import VarSet +import VarEnv + +import Type ( splitFunTy_maybe, splitForAllTys ) +import Maybes ( maybeToBool, orElse ) +import Digraph ( stronglyConnCompR, SCC(..) ) +import PrelNames ( buildIdKey, foldrIdKey, runSTRepIdKey, augmentIdKey ) +import Unique ( Unique ) +import UniqFM ( keysUFM ) +import Util ( zipWithEqual, mapAndUnzip ) +import Outputable \end{code} %************************************************************************ %* * -\subsection[OccurAnal-types]{Data types} +\subsection[OccurAnal-main]{Counting occurrences: main function} %* * %************************************************************************ -\begin{code} -data OccEnv = - OccEnv - Bool -- Keep-unused-bindings flag - -- False <=> OK to chuck away binding - -- and ignore occurrences within it - Bool -- Keep-spec-pragma-ids flag - -- False <=> OK to chuck away spec pragma bindings - -- and ignore occurrences within it - Bool -- Keep-conjurable flag - -- False <=> OK to throw away *dead* - -- "conjurable" Ids; at the moment, that - -- *only* means constant methods, which - -- are top-level. A use of a "conjurable" - -- Id may appear out of thin air -- e.g., - -- specialiser conjuring up refs to const methods. - Bool -- IgnoreINLINEPragma flag - -- False <=> OK to use INLINEPragma information - -- True <=> ignore INLINEPragma information - (UniqSet Id) -- Candidates - -addNewCands :: OccEnv -> [Id] -> OccEnv -addNewCands (OccEnv kd ks kc ip cands) ids - = OccEnv kd ks kc ip (cands `unionUniqSets` mkUniqSet ids) - -addNewCand :: OccEnv -> Id -> OccEnv -addNewCand (OccEnv ks kd kc ip cands) id - = OccEnv kd ks kc ip (cands `unionUniqSets` singletonUniqSet id) - -isCandidate :: OccEnv -> Id -> Bool -isCandidate (OccEnv _ _ _ _ cands) id = id `elementOfUniqSet` cands - -ignoreINLINEPragma :: OccEnv -> Bool -ignoreINLINEPragma (OccEnv _ _ _ ip _) = ip - -keepUnusedBinding :: OccEnv -> Id -> Bool -keepUnusedBinding (OccEnv keep_dead keep_spec keep_conjurable _ _) binder - = keep_dead || (keep_spec && maybeToBool (isSpecPragmaId_maybe binder)) - -keepBecauseConjurable :: OccEnv -> Id -> Bool -keepBecauseConjurable (OccEnv _ _ keep_conjurable _ _) binder - = keep_conjurable && isConstMethodId binder - -type UsageDetails = IdEnv BinderInfo -- A finite map from ids to their usage - -combineUsageDetails, combineAltsUsageDetails - :: UsageDetails -> UsageDetails -> UsageDetails - -combineUsageDetails usage1 usage2 - = --BSCC("combineUsages") - combineIdEnvs combineBinderInfo usage1 usage2 - --ESCC - -combineAltsUsageDetails usage1 usage2 - = --BSCC("combineUsages") - combineIdEnvs combineAltsBinderInfo usage1 usage2 - --ESCC - -addOneOcc :: UsageDetails -> Id -> BinderInfo -> UsageDetails -addOneOcc usage id info = combineIdEnvs combineBinderInfo usage (unitIdEnv id info) - -- ToDo: make this more efficient - -emptyDetails = (nullIdEnv :: UsageDetails) - -unitDetails id info = (unitIdEnv id info :: UsageDetails) - -tagBinders :: UsageDetails -- Of scope - -> [Id] -- Binders - -> (UsageDetails, -- Details with binders removed - [(Id,BinderInfo)]) -- Tagged binders - -tagBinders usage binders - = (usage `delManyFromIdEnv` binders, - [(binder, usage_of usage binder) | binder <- binders] - ) +Here's the externally-callable interface: -tagBinder :: UsageDetails -- Of scope - -> Id -- Binders - -> (UsageDetails, -- Details with binders removed - (Id,BinderInfo)) -- Tagged binders +\begin{code} +occurAnalyseGlobalExpr :: CoreExpr -> CoreExpr +occurAnalyseGlobalExpr expr + = -- Top level expr, so no interesting free vars, and + -- discard occurence info returned + snd (occAnal (initOccEnv emptyVarSet) expr) -tagBinder usage binder - = (usage `delOneFromIdEnv` binder, - (binder, usage_of usage binder) - ) - -usage_of usage binder - | isExported binder = ManyOcc 0 -- Exported things count as many - | otherwise - = case lookupIdEnv usage binder of - Nothing -> DeadCode - Just info -> info - -isNeeded env usage binder - = case usage_of usage binder of - DeadCode -> keepUnusedBinding env binder -- Maybe keep it anyway - other -> True +occurAnalyseRule :: CoreRule -> CoreRule +occurAnalyseRule rule@(BuiltinRule _ _) = rule +occurAnalyseRule (Rule str act tpl_vars tpl_args rhs) + -- Add occ info to tpl_vars, rhs + = Rule str act tpl_vars' tpl_args rhs' + where + (rhs_uds, rhs') = occAnal (initOccEnv (mkVarSet tpl_vars)) rhs + (_, tpl_vars') = tagBinders rhs_uds tpl_vars \end{code} %************************************************************************ %* * -\subsection[OccurAnal-main]{Counting occurrences: main function} +\subsection{Top level stuff} %* * %************************************************************************ -Here's the externally-callable interface: - -\begin{code} -occurAnalyseBinds - :: [CoreBinding] -- input - -> (GlobalSwitch -> Bool) - -> (SimplifierSwitch -> Bool) - -> [SimplifiableCoreBinding] -- output - -occurAnalyseBinds binds global_sw_chkr simplifier_sw_chkr - | global_sw_chkr D_dump_occur_anal = pprTrace "OccurAnal:" (ppr PprDebug binds') binds' - | otherwise = binds' - where - (_, binds') = do initial_env binds - - initial_env = OccEnv (simplifier_sw_chkr KeepUnusedBindings) - (simplifier_sw_chkr KeepSpecPragmaIds) - (not (simplifier_sw_chkr SimplMayDeleteConjurableIds)) - (simplifier_sw_chkr IgnoreINLINEPragma) - emptyUniqSet - - do env [] = (emptyDetails, []) - do env (bind:binds) - = (final_usage, new_binds ++ the_rest) - where - new_env = env `addNewCands` (bindersOf bind) - (binds_usage, the_rest) = do new_env binds - (final_usage, new_binds) = --BSCC("occAnalBind1") - occAnalBind env bind binds_usage - --ESCC -\end{code} +In @occAnalTop@ we do indirection-shorting. That is, if we have this: + + x_local = + ... + x_exported = loc + +where exp is exported, and loc is not, then we replace it with this: + + x_local = x_exported + x_exported = + ... + +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. + +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} -occurAnalyseExpr :: UniqSet Id -- Set of interesting free vars - -> CoreExpr - -> (IdEnv BinderInfo, -- Occ info for interesting free vars - SimplifiableCoreExpr) +occurAnalyseBinds :: [CoreBind] -> [CoreBind] -occurAnalyseExpr candidates expr - = occAnal initial_env expr +occurAnalyseBinds binds + = binds' where - initial_env = OccEnv False {- Drop unused bindings -} - False {- Drop SpecPragmaId bindings -} - True {- Keep conjurable Ids -} - False {- Do not ignore INLINE Pragma -} - candidates - -occurAnalyseGlobalExpr :: CoreExpr -> SimplifiableCoreExpr -occurAnalyseGlobalExpr expr - = -- Top level expr, so no interesting free vars, and - -- discard occurence info returned - expr' where (_, expr') = occurAnalyseExpr emptyUniqSet expr + (_, _, binds') = go (initOccEnv emptyVarSet) binds + + go :: OccEnv -> [CoreBind] + -> (UsageDetails, -- Occurrence info + IdEnv Id, -- Indirection elimination info + -- Maps local-id -> exported-id, but it embodies + -- bindings of the form exported-id = local-id in + -- the argument to go + [CoreBind]) -- Occ-analysed bindings, less the exported-id=local-id ones + + go env [] = (emptyDetails, emptyVarEnv, []) + + go env (bind : binds) + = let + new_env = env `addNewCands` (bindersOf bind) + (scope_usage, ind_env, binds') = go new_env binds + (final_usage, new_binds) = occAnalBind env (zapBind ind_env bind) scope_usage + -- NB: I zap before occur-analysing, so + -- I don't need to worry about getting the + -- occ info on the new bindings right. + in + case bind of + NonRec exported_id (Var local_id) + | shortMeOut ind_env exported_id local_id + -- Special case for eliminating indirections + -- Note: it's a shortcoming that this only works for + -- non-recursive bindings. Elminating indirections + -- makes perfect sense for recursive bindings too, but + -- it's more complicated to implement, so I haven't done so + -> (scope_usage, ind_env', binds') + where + ind_env' = extendVarEnv ind_env local_id exported_id + + other -> -- Ho ho! The normal case + (final_usage, ind_env, new_binds ++ binds') + + +-- Deal with any indirections +zapBind ind_env (NonRec bndr rhs) + | bndr `elemVarEnv` ind_env = Rec (zap ind_env (bndr,rhs)) + -- The Rec isn't strictly necessary, but it's convenient +zapBind ind_env (Rec pairs) + | or [id `elemVarEnv` ind_env | (id,_) <- pairs] = Rec (concat (map (zap ind_env) pairs)) + +zapBind ind_env bind = bind + +zap ind_env pair@(local_id,rhs) + = case lookupVarEnv ind_env local_id of + Nothing -> [pair] + Just exported_id -> [(local_id, Var exported_id), + (exported_id', rhs)] + where + exported_id' = modifyIdInfo (copyIdInfo (idInfo local_id)) exported_id + +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 shortableIdInfo (idInfo exported_id) -- Only if its IdInfo is 'shortable' + -- (see the defn of IdInfo.shortableIdInfo) + then True + else +#ifdef DEBUG + pprTrace "shortMeOut:" (ppr exported_id) +#endif + False +-} + else + False \end{code} + %************************************************************************ %* * \subsection[OccurAnal-main]{Counting occurrences: main function} @@ -207,23 +222,31 @@ Bindings ~~~~~~~~ \begin{code} +type IdWithOccInfo = Id -- An Id with fresh PragmaInfo attached + +type Node details = (details, Unique, [Unique]) -- The Ints are gotten from the Unique, + -- which is gotten from the Id. +type Details1 = (Id, UsageDetails, CoreExpr) +type Details2 = (IdWithOccInfo, CoreExpr) + + occAnalBind :: OccEnv - -> CoreBinding + -> CoreBind -> UsageDetails -- Usage details of scope -> (UsageDetails, -- Of the whole let(rec) - [SimplifiableCoreBinding]) + [CoreBind]) occAnalBind env (NonRec binder rhs) body_usage - | isNeeded env body_usage binder -- It's mentioned in body + | not (binder `usedIn` body_usage) -- It's not mentioned + = (body_usage, []) + + | otherwise -- It's mentioned in the body = (final_body_usage `combineUsageDetails` rhs_usage, [NonRec tagged_binder rhs']) - | otherwise - = (body_usage, []) - where - (rhs_usage, rhs') = occAnalRhs env binder rhs (final_body_usage, tagged_binder) = tagBinder body_usage binder + (rhs_usage, rhs') = occAnalRhs env binder rhs \end{code} Dropping dead code for recursive bindings is done in a very simple way: @@ -263,68 +286,208 @@ It isn't easy to do a perfect job in one blow. Consider \begin{code} occAnalBind env (Rec pairs) body_usage - = foldr do_final_bind (body_usage, []) sccs + = foldr (_scc_ "occAnalBind.dofinal" do_final_bind) (body_usage, []) sccs where + binders = map fst pairs + rhs_env = env `addNewCands` binders - (binders, rhss) = unzip pairs - new_env = env `addNewCands` binders - - analysed_pairs :: [(Id, (UsageDetails, SimplifiableCoreExpr))] - analysed_pairs = [(id, occAnalRhs new_env id rhs) | (id,rhs) <- pairs] + analysed_pairs :: [Details1] + analysed_pairs = [ (bndr, rhs_usage, rhs') + | (bndr, rhs) <- pairs, + let (rhs_usage, rhs') = occAnalRhs rhs_env bndr rhs + ] - lookup :: Id -> (UsageDetails, SimplifiableCoreExpr) - lookup id = assoc "occAnalBind:lookup" analysed_pairs id + sccs :: [SCC (Node Details1)] + sccs = _scc_ "occAnalBind.scc" stronglyConnCompR edges ---- stuff for dependency analysis of binds ------------------------------- + edges :: [Node Details1] + edges = _scc_ "occAnalBind.assoc" + [ (details, idUnique id, edges_from rhs_usage) + | details@(id, rhs_usage, rhs) <- analysed_pairs + ] + + -- (a -> b) means a mentions b + -- Given the usage details (a UFM that gives occ info for each free var of + -- the RHS) we can get the list of free vars -- or rather their Int keys -- + -- by just extracting the keys from the finite map. Grimy, but fast. + -- Previously we had this: + -- [ bndr | bndr <- bndrs, + -- maybeToBool (lookupVarEnv rhs_usage bndr)] + -- which has n**2 cost, and this meant that edges_from alone + -- consumed 10% of total runtime! + edges_from :: UsageDetails -> [Unique] + edges_from rhs_usage = _scc_ "occAnalBind.edges_from" + keysUFM rhs_usage - edges :: [(Id,Id)] -- (a,b) means a mentions b - edges = concat [ edges_from binder rhs_usage - | (binder, (rhs_usage, _)) <- analysed_pairs] + ---- stuff to "re-constitute" bindings from dependency-analysis info ------ - edges_from :: Id -> UsageDetails -> [(Id,Id)] - edges_from id its_rhs_usage - = [(id,mentioned) | mentioned <- binders, - maybeToBool (lookupIdEnv its_rhs_usage mentioned) - ] + -- Non-recursive SCC + do_final_bind (AcyclicSCC ((bndr, rhs_usage, rhs'), _, _)) (body_usage, binds_so_far) + | not (bndr `usedIn` body_usage) + = (body_usage, binds_so_far) -- Dead code + | otherwise + = (combined_usage, new_bind : binds_so_far) + where + total_usage = combineUsageDetails body_usage rhs_usage + (combined_usage, tagged_bndr) = tagBinder total_usage bndr + new_bind = NonRec tagged_bndr rhs' + + -- Recursive SCC + do_final_bind (CyclicSCC cycle) (body_usage, binds_so_far) + | not (any (`usedIn` body_usage) bndrs) -- NB: look at body_usage, not total_usage + = (body_usage, binds_so_far) -- Dead code + | otherwise + = (combined_usage, final_bind:binds_so_far) + where + details = [details | (details, _, _) <- cycle] + bndrs = [bndr | (bndr, _, _) <- details] + rhs_usages = [rhs_usage | (_, rhs_usage, _) <- details] + total_usage = foldr combineUsageDetails body_usage rhs_usages + (combined_usage, tagged_bndrs) = tagBinders total_usage bndrs + final_bind = Rec (reOrderRec env new_cycle) + + new_cycle = CyclicSCC (zipWithEqual "occAnalBind" mk_new_bind tagged_bndrs cycle) + mk_new_bind tagged_bndr ((_, _, rhs'), key, keys) = ((tagged_bndr, rhs'), key, keys) +\end{code} - sccs :: [[Id]] - sccs = case binders of - [_] -> [binders] -- Singleton; no need to analyse - other -> stronglyConnComp eqId edges binders +@reOrderRec@ is applied to the list of (binder,rhs) pairs for a cyclic +strongly connected component (there's guaranteed to be a cycle). It returns the +same pairs, but + a) in a better order, + b) with some of the Ids having a IMustNotBeINLINEd pragma - ---- stuff to "re-constitute" bindings from dependency-analysis info ------ +The "no-inline" Ids are sufficient to break all cycles in the SCC. This means +that the simplifier can guarantee not to loop provided it never records an inlining +for these no-inline guys. - do_final_bind sCC@[binder] (body_usage, binds_so_far) - | isNeeded env body_usage binder - = (combined_usage, new_bind:binds_so_far) +Furthermore, the order of the binds is such that if we neglect dependencies +on the no-inline Ids then the binds are topologically sorted. This means +that the simplifier will generally do a good job if it works from top bottom, +recording inlinings for any Ids which aren't marked as "no-inline" as it goes. - | otherwise -- Dead - = (body_usage, binds_so_far) - where - total_usage = combineUsageDetails body_usage rhs_usage - (rhs_usage, rhs') = lookup binder - (combined_usage, tagged_binder) = tagBinder total_usage binder - - new_bind - | mentions_itself binder rhs_usage = Rec [(tagged_binder,rhs')] - | otherwise = NonRec tagged_binder rhs' - where - mentions_itself binder usage - = maybeToBool (lookupIdEnv usage binder) - - do_final_bind sCC (body_usage, binds_so_far) - | any (isNeeded env body_usage) sCC - = (combined_usage, new_bind:binds_so_far) - - | otherwise -- Dead - = (body_usage, binds_so_far) - where - (rhs_usages, rhss') = unzip (map lookup sCC) - total_usage = foldr combineUsageDetails body_usage rhs_usages - (combined_usage, tagged_binders) = tagBinders total_usage sCC +============== +[June 98: I don't understand the following paragraphs, and I've + changed the a=b case again so that it isn't a special case any more.] + +Here's a case that bit me: + + letrec + a = b + b = \x. BIG + in + ...a...a...a.... + +Re-ordering doesn't change the order of bindings, but there was no loop-breaker. + +My solution was to make a=b bindings record b as Many, rather like INLINE bindings. +Perhaps something cleverer would suffice. +=============== + +You might think that you can prevent non-termination simply by making +sure that we simplify a recursive binding's RHS in an environment that +simply clones the recursive Id. But no. Consider - new_bind = Rec (tagged_binders `zip` rhss') + letrec f = \x -> let z = f x' in ... + + in + let n = f y + in + case n of { ... } + +We bind n to its *simplified* RHS, we then *re-simplify* it when +we inline n. Then we may well inline f; and then the same thing +happens with z! + +I don't think it's possible to prevent non-termination by environment +manipulation in this way. Apart from anything else, successive +iterations of the simplifier may unroll recursive loops in cases like +that above. The idea of beaking every recursive loop with an +IMustNotBeINLINEd pragma is much much better. + + +\begin{code} +reOrderRec + :: OccEnv + -> SCC (Node Details2) + -> [Details2] + -- Sorted into a plausible order. Enough of the Ids have + -- dontINLINE pragmas that there are no loops left. + + -- Non-recursive case +reOrderRec env (AcyclicSCC (bind, _, _)) = [bind] + + -- Common case of simple self-recursion +reOrderRec env (CyclicSCC [bind]) + = [(setIdOccInfo tagged_bndr IAmALoopBreaker, rhs)] + where + ((tagged_bndr, rhs), _, _) = bind + +reOrderRec env (CyclicSCC (bind : binds)) + = -- Choose a loop breaker, mark it no-inline, + -- do SCC analysis on the rest, and recursively sort them out + concat (map (reOrderRec env) (stronglyConnCompR unchosen)) + ++ + [(setIdOccInfo tagged_bndr IAmALoopBreaker, rhs)] + + where + (chosen_pair, unchosen) = choose_loop_breaker bind (score bind) [] binds + (tagged_bndr, rhs) = chosen_pair + + -- This loop looks for the bind with the lowest score + -- to pick as the loop breaker. The rest accumulate in + choose_loop_breaker (details,_,_) loop_sc acc [] + = (details, acc) -- Done + + choose_loop_breaker loop_bind loop_sc acc (bind : binds) + | sc < loop_sc -- Lower score so pick this new one + = choose_loop_breaker bind sc (loop_bind : acc) binds + + | otherwise -- No lower so don't pick it + = choose_loop_breaker loop_bind loop_sc (bind : acc) binds + where + sc = score bind + + score :: Node Details2 -> Int -- Higher score => less likely to be picked as loop breaker + score ((bndr, rhs), _, _) + | exprIsTrivial rhs = 4 -- Practically certain to be inlined + -- Used to have also: && not (isExportedId bndr) + -- But I found this sometimes cost an extra iteration when we have + -- rec { d = (a,b); a = ...df...; b = ...df...; df = d } + -- where df is the exported dictionary. Then df makes a really + -- bad choice for loop breaker + + | not_fun_ty (idType bndr) = 3 -- Data types help with cases + -- This used to have a lower score than inlineCandidate, but + -- it's *really* helpful if dictionaries get inlined fast, + -- so I'm experimenting with giving higher priority to data-typed things + + | inlineCandidate bndr rhs = 2 -- Likely to be inlined + + | not (isEmptyCoreRules (idSpecialisation bndr)) = 1 + -- Avoid things with specialisations; we'd like + -- to take advantage of them in the subsequent bindings + + | otherwise = 0 + + inlineCandidate :: Id -> CoreExpr -> Bool + inlineCandidate id (Note InlineMe _) = True + inlineCandidate id rhs = isOneOcc (idOccInfo id) + + -- Real example (the Enum Ordering instance from PrelBase): + -- rec f = \ x -> case d of (p,q,r) -> p x + -- g = \ x -> case d of (p,q,r) -> q x + -- d = (v, f, g) + -- + -- Here, f and g occur just once; but we can't inline them into d. + -- On the other hand we *could* simplify those case expressions if + -- we didn't stupidly choose d as the loop breaker. + -- But we won't because constructor args are marked "Many". + + not_fun_ty ty = not (maybeToBool (splitFunTy_maybe rho_ty)) + where + (_, rho_ty) = splitForAllTys ty \end{code} @occAnalRhs@ deals with the question of bindings where the Id is marked @@ -334,21 +497,44 @@ inlined binder also occurs many times in its scope, but if it doesn't we'll catch it next time round. At worst this costs an extra simplifier pass. ToDo: try using the occurrence info for the inline'd binder. +[March 97] We do the same for atomic RHSs. Reason: see notes with reOrderRec. +[June 98, SLPJ] I've undone this change; I don't understand it. See notes with reOrderRec. + + \begin{code} occAnalRhs :: OccEnv - -> Id -- Binder - -> CoreExpr -- Rhs - -> (UsageDetails, SimplifiableCoreExpr) + -> Id -> CoreExpr -- Binder and rhs + -> (UsageDetails, CoreExpr) occAnalRhs env id rhs - | idWantsToBeINLINEd id && not (ignoreINLINEPragma env) - = (mapIdEnv markMany rhs_usage, rhs') - - | otherwise - = (rhs_usage, rhs') - + = (final_usage, rhs') where - (rhs_usage, rhs') = occAnal env rhs + (rhs_usage, rhs') = occAnal (rhsCtxt env) rhs + -- Note that we use an rhsCtxt. This tells the occ anal that it's + -- looking at an RHS, which has an effect in occAnalApp + -- + -- But there's a problem. Consider + -- x1 = a0 : [] + -- x2 = a1 : x1 + -- x3 = a2 : x2 + -- g = f x2 + -- First time round, it looks as if x1 and x2 occur as an arg of a + -- let-bound constructor ==> give them a many-occurrence. + -- But then x3 is inlined (unconditionally as it happens) and + -- next time round, x2 will be, and the next time round x1 will be + -- Result: multiple simplifier iterations. Sigh. + -- Possible solution: use rhsCtxt for things that occur just once... + + -- [March 98] A new wrinkle is that if the binder has specialisations inside + -- it then we count the specialised Ids as "extra rhs's". That way + -- the "parent" keeps the specialised "children" alive. If the parent + -- dies (because it isn't referenced any more), then the children will + -- die too unless they are already referenced directly. + + final_usage = foldVarSet add rhs_usage (idRuleVars id) + add v u = addOneOcc u v NoOccInfo -- Give a non-committal binder info + -- (i.e manyOcc) because many copies + -- of the specialised thing can appear \end{code} Expressions @@ -356,125 +542,439 @@ Expressions \begin{code} occAnal :: OccEnv -> CoreExpr - -> (UsageDetails, -- Gives info only about the "interesting" Ids - SimplifiableCoreExpr) + -> (UsageDetails, -- Gives info only about the "interesting" Ids + CoreExpr) -occAnal env (Var v) - | isCandidate env v - = (unitIdEnv v (funOccurrence 0), Var v) +occAnal env (Type t) = (emptyDetails, Type t) - | otherwise - = (emptyDetails, Var v) +occAnal env (Var v) + = (var_uds, Var v) + where + var_uds | isCandidate env v = unitVarEnv v oneOcc + | otherwise = emptyDetails -occAnal env (Lit lit) = (emptyDetails, Lit lit) -occAnal env (Con con tys args) = (occAnalAtoms env args, Con con tys args) -occAnal env (Prim op tys args) = (occAnalAtoms env args, Prim op tys args) + -- At one stage, I gathered the idRuleVars for v here too, + -- which in a way is the right thing to do. + -- But that went wrong right after specialisation, when + -- the *occurrences* of the overloaded function didn't have any + -- rules in them, so the *specialised* versions looked as if they + -- weren't used at all. -occAnal env (SCC cc body) - = (mapIdEnv markInsideSCC usage, SCC cc body') - where - (usage, body') = occAnal env body +\end{code} -occAnal env (App fun arg) - = (fun_usage `combineUsageDetails` arg_usage, App fun' arg) - where - (fun_usage, fun') = occAnal env fun - arg_usage = occAnalAtom env arg +We regard variables that occur as constructor arguments as "dangerousToDup": -occAnal env (CoTyApp fun ty) - = (fun_usage, CoTyApp fun' ty) - where - (fun_usage, fun') = occAnal env fun +\begin{verbatim} +module A where +f x = let y = expensive x in + let z = (True,y) in + (case z of {(p,q)->q}, case z of {(p,q)->q}) +\end{verbatim} + +We feel free to duplicate the WHNF (True,y), but that means +that y may be duplicated thereby. + +If we aren't careful we duplicate the (expensive x) call! +Constructors are rather like lambdas in this way. -occAnal env (Lam binder body) - = (mapIdEnv markDangerousToDup final_usage, Lam tagged_binder body') +\begin{code} +occAnal env expr@(Lit lit) = (emptyDetails, expr) +\end{code} + +\begin{code} +occAnal env (Note InlineMe body) + = case occAnal env body of { (usage, body') -> + (mapVarEnv markMany usage, Note InlineMe body') + } + +occAnal env (Note note@(SCC cc) body) + = case occAnal env body of { (usage, body') -> + (mapVarEnv markInsideSCC usage, Note note body') + } + +occAnal env (Note note body) + = case occAnal env body of { (usage, body') -> + (usage, Note note body') + } +\end{code} + +\begin{code} +occAnal env app@(App fun arg) + = occAnalApp env (collectArgs app) False + +-- Ignore type variables altogether +-- (a) occurrences inside type lambdas only not marked as InsideLam +-- (b) type variables not in environment + +occAnal env expr@(Lam x body) | isTyVar x + = case occAnal env body of { (body_usage, body') -> + (body_usage, Lam x body') + } + +-- For value lambdas we do a special hack. Consider +-- (\x. \y. ...x...) +-- If we did nothing, x is used inside the \y, so would be marked +-- as dangerous to dup. But in the common case where the abstraction +-- is applied to two arguments this is over-pessimistic. +-- So instead, we just mark each binder with its occurrence +-- info in the *body* of the multiple lambda. +-- Then, the simplifier is careful when partially applying lambdas. + +occAnal env expr@(Lam _ _) + = case occAnal env_body body of { (body_usage, body') -> + let + (final_usage, tagged_binders) = tagBinders body_usage binders + -- URGH! Sept 99: we don't seem to be able to use binders' here, because + -- we get linear-typed things in the resulting program that we can't handle yet. + -- (e.g. PrelShow) TODO + + really_final_usage = if linear then + final_usage + else + mapVarEnv markInsideLam final_usage + in + (really_final_usage, + mkLams tagged_binders body') } + where + (binders, body) = collectBinders expr + (linear, env1, _) = oneShotGroup env binders + env2 = env1 `addNewCands` binders -- Add in-scope binders + env_body = vanillaCtxt env2 -- Body is (no longer) an RhsContext + +occAnal env (Case scrut bndr alts) + = case mapAndUnzip (occAnalAlt alt_env bndr) alts of { (alts_usage_s, alts') -> + case occAnal (vanillaCtxt env) scrut of { (scrut_usage, scrut') -> + -- No need for rhsCtxt + let + alts_usage = foldr1 combineAltsUsageDetails alts_usage_s + alts_usage' = addCaseBndrUsage alts_usage + (alts_usage1, tagged_bndr) = tagBinder alts_usage' bndr + total_usage = scrut_usage `combineUsageDetails` alts_usage1 + in + total_usage `seq` (total_usage, Case scrut' tagged_bndr alts') }} where - (body_usage, body') = occAnal (env `addNewCand` binder) body - (final_usage, tagged_binder) = tagBinder body_usage binder + alt_env = env `addNewCand` bndr + + -- The case binder gets a usage of either "many" or "dead", never "one". + -- Reason: we like to inline single occurrences, to eliminate a binding, + -- but inlining a case binder *doesn't* eliminate a binding. + -- We *don't* want to transform + -- case x of w { (p,q) -> f w } + -- into + -- case x of w { (p,q) -> f (p,q) } + addCaseBndrUsage usage = case lookupVarEnv usage bndr of + Nothing -> usage + Just occ -> extendVarEnv usage bndr (markMany occ) --- ANDY: WE MUST THINK ABOUT THIS! (ToDo) -occAnal env (CoTyLam tyvar body) - = (mapIdEnv markDangerousToDup body_usage, CoTyLam tyvar body') +occAnal env (Let bind body) + = case occAnal new_env body of { (body_usage, body') -> + case occAnalBind env bind body_usage of { (final_usage, new_binds) -> + (final_usage, mkLets new_binds body') }} where - (body_usage, body') = occAnal env body + new_env = env `addNewCands` (bindersOf bind) -occAnal env (Case scrut alts) - = (scrut_usage `combineUsageDetails` alts_usage, - Case scrut' alts') +occAnalArgs env args + = case mapAndUnzip (occAnal arg_env) args of { (arg_uds_s, args') -> + (foldr combineUsageDetails emptyDetails arg_uds_s, args')} where - (scrut_usage, scrut') = occAnal env scrut - (alts_usage, alts') = occAnalAlts env alts + arg_env = vanillaCtxt env +\end{code} -occAnal env (Let bind body) - = (final_usage, foldr Let body' new_binds) -- mkCoLet* wants Core... (sigh) +Applications are dealt with specially because we want +the "build hack" to work. + +\begin{code} +-- Hack for build, fold, runST +occAnalApp env (Var fun, args) is_rhs + = case args_stuff of { (args_uds, args') -> + let + -- We mark the free vars of the argument of a constructor or PAP + -- as "many", if it is the RHS of a let(rec). + -- This means that nothing gets inlined into a constructor argument + -- position, which is what we want. Typically those constructor + -- arguments are just variables, or trivial expressions. + -- + -- This is the *whole point* of the isRhsEnv predicate + final_args_uds + | isRhsEnv env, + isDataConId fun || valArgCount args < idArity fun + = mapVarEnv markMany args_uds + | otherwise = args_uds + in + (fun_uds `combineUsageDetails` final_args_uds, mkApps (Var fun) args') } + where + fun_uniq = idUnique fun + + fun_uds | isCandidate env fun = unitVarEnv fun oneOcc + | otherwise = emptyDetails + + args_stuff | fun_uniq == buildIdKey = appSpecial env 2 [True,True] args + | fun_uniq == augmentIdKey = appSpecial env 2 [True,True] args + | fun_uniq == foldrIdKey = appSpecial env 3 [False,True] args + | fun_uniq == runSTRepIdKey = appSpecial env 2 [True] args + -- (foldr k z xs) may call k many times, but it never + -- shares a partial application of k; hence [False,True] + -- This means we can optimise + -- foldr (\x -> let v = ...x... in \y -> ...v...) z xs + -- by floating in the v + + | otherwise = occAnalArgs env args + + +occAnalApp env (fun, args) is_rhs + = case occAnal (addAppCtxt env args) fun of { (fun_uds, fun') -> + -- The addAppCtxt is a bit cunning. One iteration of the simplifier + -- often leaves behind beta redexs like + -- (\x y -> e) a1 a2 + -- Here we would like to mark x,y as one-shot, and treat the whole + -- thing much like a let. We do this by pushing some True items + -- onto the context stack. + + case occAnalArgs env args of { (args_uds, args') -> + let + final_uds = fun_uds `combineUsageDetails` args_uds + in + (final_uds, mkApps fun' args') }} + +appSpecial :: OccEnv + -> Int -> CtxtTy -- Argument number, and context to use for it + -> [CoreExpr] + -> (UsageDetails, [CoreExpr]) +appSpecial env n ctxt args + = go n args where - new_env = env `addNewCands` (bindersOf bind) - (body_usage, body') = occAnal new_env body - (final_usage, new_binds) = --BSCC("occAnalBind2") - occAnalBind env bind body_usage - --ESCC + arg_env = vanillaCtxt env + + go n [] = (emptyDetails, []) -- Too few args + + go 1 (arg:args) -- The magic arg + = case occAnal (setCtxt arg_env ctxt) arg of { (arg_uds, arg') -> + case occAnalArgs env args of { (args_uds, args') -> + (combineUsageDetails arg_uds args_uds, arg':args') }} + + go n (arg:args) + = case occAnal arg_env arg of { (arg_uds, arg') -> + case go (n-1) args of { (args_uds, args') -> + (combineUsageDetails arg_uds args_uds, arg':args') }} \end{code} + Case alternatives ~~~~~~~~~~~~~~~~~ +If the case binder occurs at all, the other binders effectively do too. +For example + case e of x { (a,b) -> rhs } +is rather like + let x = (a,b) in rhs +If e turns out to be (e1,e2) we indeed get something like + let a = e1; b = e2; x = (a,b) in rhs + \begin{code} -occAnalAlts env (AlgAlts alts deflt) - = (foldr combineAltsUsageDetails deflt_usage alts_usage, - -- Note: combine*Alts*UsageDetails... - AlgAlts alts' deflt') - where - (alts_usage, alts') = unzip (map do_alt alts) - (deflt_usage, deflt') = occAnalDeflt env deflt +occAnalAlt env case_bndr (con, bndrs, rhs) + = case occAnal (env `addNewCands` bndrs) rhs of { (rhs_usage, rhs') -> + let + (final_usage, tagged_bndrs) = tagBinders rhs_usage bndrs + final_bndrs | case_bndr `elemVarEnv` final_usage = bndrs + | otherwise = tagged_bndrs + -- Leave the binders untagged if the case + -- binder occurs at all; see note above + in + (final_usage, (con, final_bndrs, rhs')) } +\end{code} - do_alt (con, args, rhs) - = (final_usage, (con, tagged_args, rhs')) - where - new_env = env `addNewCands` args - (rhs_usage, rhs') = occAnal new_env rhs - (final_usage, tagged_args) = tagBinders rhs_usage args - -occAnalAlts env (PrimAlts alts deflt) - = (foldr combineAltsUsageDetails deflt_usage alts_usage, - -- Note: combine*Alts*UsageDetails... - PrimAlts alts' deflt') - where - (alts_usage, alts') = unzip (map do_alt alts) - (deflt_usage, deflt') = occAnalDeflt env deflt - do_alt (lit, rhs) - = (rhs_usage, (lit, rhs')) - where - (rhs_usage, rhs') = occAnal env rhs +%************************************************************************ +%* * +\subsection[OccurAnal-types]{OccEnv} +%* * +%************************************************************************ + +\begin{code} +data OccEnv + = OccEnv IdSet -- In-scope Ids; we gather info about these only + OccEncl -- Enclosing context information + CtxtTy -- Tells about linearity + +-- OccEncl is used to control whether to inline into constructor arguments +-- For example: +-- x = (p,q) -- Don't inline p or q +-- y = /\a -> (p a, q a) -- Still don't inline p or q +-- z = f (p,q) -- Do inline p,q; it may make a rule fire +-- So OccEncl tells enought about the context to know what to do when +-- we encounter a contructor application or PAP. + +data OccEncl + = OccRhs -- RHS of let(rec), albeit perhaps inside a type lambda + -- Don't inline into constructor args here + | OccVanilla -- Argument of function, body of lambda, scruintee of case etc. + -- Do inline into constructor args here + +type CtxtTy = [Bool] + -- [] No info + -- + -- True:ctxt Analysing a function-valued expression that will be + -- applied just once + -- + -- False:ctxt Analysing a function-valued expression that may + -- be applied many times; but when it is, + -- the CtxtTy inside applies + +initOccEnv :: VarSet -> OccEnv +initOccEnv vars = OccEnv vars OccRhs [] + +isRhsEnv (OccEnv _ OccRhs _) = True +isRhsEnv (OccEnv _ OccVanilla _) = False + +isCandidate :: OccEnv -> Id -> Bool +isCandidate (OccEnv cands encl _) id = id `elemVarSet` cands + +addNewCands :: OccEnv -> [Id] -> OccEnv +addNewCands (OccEnv cands encl ctxt) ids + = OccEnv (cands `unionVarSet` mkVarSet ids) encl ctxt + +addNewCand :: OccEnv -> Id -> OccEnv +addNewCand (OccEnv cands encl ctxt) id + = OccEnv (extendVarSet cands id) encl ctxt -occAnalDeflt env NoDefault = (emptyDetails, NoDefault) +setCtxt :: OccEnv -> CtxtTy -> OccEnv +setCtxt (OccEnv cands encl _) ctxt = OccEnv cands encl ctxt -occAnalDeflt env (BindDefault binder rhs) - = (final_usage, BindDefault tagged_binder rhs') +oneShotGroup :: OccEnv -> [CoreBndr] -> (Bool, OccEnv, [CoreBndr]) + -- True <=> this is a one-shot linear lambda group + -- The [CoreBndr] are the binders. + + -- The result binders have one-shot-ness set that they might not have had originally. + -- This happens in (build (\cn -> e)). Here the occurrence analyser + -- linearity context knows that c,n are one-shot, and it records that fact in + -- the binder. This is useful to guide subsequent float-in/float-out tranformations + +oneShotGroup (OccEnv cands encl ctxt) bndrs + = case go ctxt bndrs [] of + (new_ctxt, new_bndrs) -> (all is_one_shot new_bndrs, OccEnv cands encl new_ctxt, new_bndrs) where - new_env = env `addNewCand` binder - (rhs_usage, rhs') = occAnal new_env rhs - (final_usage, tagged_binder) = tagBinder rhs_usage binder + is_one_shot b = isId b && isOneShotLambda b + + go ctxt [] rev_bndrs = (ctxt, reverse rev_bndrs) + + go (lin_ctxt:ctxt) (bndr:bndrs) rev_bndrs + | isId bndr = go ctxt bndrs (bndr':rev_bndrs) + where + bndr' | lin_ctxt = setOneShotLambda bndr + | otherwise = bndr + + go ctxt (bndr:bndrs) rev_bndrs = go ctxt bndrs (bndr:rev_bndrs) + + +vanillaCtxt (OccEnv cands _ _) = OccEnv cands OccVanilla [] +rhsCtxt (OccEnv cands _ _) = OccEnv cands OccRhs [] + +addAppCtxt (OccEnv cands encl ctxt) args + = OccEnv cands encl (replicate (valArgCount args) True ++ ctxt) \end{code} +%************************************************************************ +%* * +\subsection[OccurAnal-types]{OccEnv} +%* * +%************************************************************************ -Atoms -~~~~~ \begin{code} -occAnalAtoms :: OccEnv -> [CoreArg] -> UsageDetails +type UsageDetails = IdEnv OccInfo -- A finite map from ids to their usage + +combineUsageDetails, combineAltsUsageDetails + :: UsageDetails -> UsageDetails -> UsageDetails + +combineUsageDetails usage1 usage2 + = plusVarEnv_C addOccInfo usage1 usage2 + +combineAltsUsageDetails usage1 usage2 + = plusVarEnv_C orOccInfo usage1 usage2 + +addOneOcc :: UsageDetails -> Id -> OccInfo -> UsageDetails +addOneOcc usage id info + = plusVarEnv_C addOccInfo usage (unitVarEnv id info) + -- ToDo: make this more efficient + +emptyDetails = (emptyVarEnv :: UsageDetails) -occAnalAtoms env atoms - = foldr do_one_atom emptyDetails atoms +usedIn :: Id -> UsageDetails -> Bool +v `usedIn` details = isExportedId v || v `elemVarEnv` details + +tagBinders :: UsageDetails -- Of scope + -> [Id] -- Binders + -> (UsageDetails, -- Details with binders removed + [IdWithOccInfo]) -- Tagged binders + +tagBinders usage binders + = let + usage' = usage `delVarEnvList` binders + uss = map (setBinderOcc usage) binders + in + usage' `seq` (usage', uss) + +tagBinder :: UsageDetails -- Of scope + -> Id -- Binders + -> (UsageDetails, -- Details with binders removed + IdWithOccInfo) -- Tagged binders + +tagBinder usage binder + = let + usage' = usage `delVarEnv` binder + binder' = setBinderOcc usage binder + in + usage' `seq` (usage', binder') + +setBinderOcc :: UsageDetails -> CoreBndr -> CoreBndr +setBinderOcc usage bndr + | isTyVar bndr = bndr + | isExportedId bndr = case idOccInfo bndr of + NoOccInfo -> bndr + other -> setIdOccInfo bndr NoOccInfo + -- Don't use local usage info for visible-elsewhere things + -- BUT *do* erase any IAmALoopBreaker annotation, because we're + -- about to re-generate it and it shouldn't be "sticky" + + | otherwise = setIdOccInfo bndr occ_info where - do_one_atom (LitArg lit) usage = usage - do_one_atom (VarArg v) usage - | isCandidate env v = addOneOcc usage v (argOccurrence 0) - | otherwise = usage + occ_info = lookupVarEnv usage bndr `orElse` IAmDead +\end{code} + + +%************************************************************************ +%* * +\subsection{Operations over OccInfo} +%* * +%************************************************************************ + +\begin{code} +oneOcc :: OccInfo +oneOcc = OneOcc False True + +markMany, markInsideLam, markInsideSCC :: OccInfo -> OccInfo + +markMany IAmDead = IAmDead +markMany other = NoOccInfo + +markInsideSCC occ = markMany occ + +markInsideLam (OneOcc _ one_br) = OneOcc True one_br +markInsideLam occ = occ + +addOccInfo, orOccInfo :: OccInfo -> OccInfo -> OccInfo + +addOccInfo IAmDead info2 = info2 +addOccInfo info1 IAmDead = info1 +addOccInfo info1 info2 = NoOccInfo +-- (orOccInfo orig new) is used +-- when combining occurrence info from branches of a case -occAnalAtom :: OccEnv -> CoreArg -> UsageDetails +orOccInfo IAmDead info2 = info2 +orOccInfo info1 IAmDead = info1 +orOccInfo (OneOcc in_lam1 one_branch1) + (OneOcc in_lam2 one_branch2) + = OneOcc (in_lam1 || in_lam2) + False -- False, because it occurs in both branches -occAnalAtom env (LitArg lit) = emptyDetails -occAnalAtom env (VarArg v) - | isCandidate env v = unitDetails v (argOccurrence 0) - | otherwise = emptyDetails +orOccInfo info1 info2 = NoOccInfo \end{code}