X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FsimplCore%2FOccurAnal.lhs;h=750e2b6cde3430a3475e727bb406a0dac1ecb2f7;hb=0282f593071144bb319226118d36561cab55d723;hp=5796cd4e96765d9698bc4586370d6f52fcb0beae;hpb=33c0b416327d85d06763e91cbc1de937e742d0e5;p=ghc-hetmet.git diff --git a/ghc/compiler/simplCore/OccurAnal.lhs b/ghc/compiler/simplCore/OccurAnal.lhs index 5796cd4..750e2b6 100644 --- a/ghc/compiler/simplCore/OccurAnal.lhs +++ b/ghc/compiler/simplCore/OccurAnal.lhs @@ -1,5 +1,5 @@ % -% (c) The GRASP/AQUA Project, Glasgow University, 1992-1996 +% (c) The GRASP/AQUA Project, Glasgow University, 1992-1998 % %************************************************************************ %* * @@ -11,258 +11,213 @@ 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 + occurAnalyseBinds, occurAnalyseGlobalExpr, occurAnalyseRule ) where -IMP_Ubiq(){-uitous-} -IMPORT_1_3(List(partition)) +#include "HsVersions.h" -import BinderInfo -import CmdLineOpts ( opt_D_dump_occur_anal, SimplifierSwitch(..) ) import CoreSyn -import Digraph ( stronglyConnComp, stronglyConnCompR, SCC(..) ) -import Id ( idWantsToBeINLINEd, addNoInlinePragma, nukeNoInlinePragma, - idType, idUnique, SYN_IE(Id), - emptyIdSet, unionIdSets, mkIdSet, - unitIdSet, elementOfIdSet, - addOneToIdSet, SYN_IE(IdSet), - nullIdEnv, unitIdEnv, combineIdEnvs, - delOneFromIdEnv, delManyFromIdEnv, isNullIdEnv, - mapIdEnv, lookupIdEnv, SYN_IE(IdEnv), - GenId{-instance Eq-} +import CoreFVs ( idRuleVars ) +import CoreUtils ( exprIsTrivial ) +import Id ( isDataConId, isOneShotLambda, setOneShotLambda, + idOccInfo, setIdOccInfo, + isExportedId, modifyIdInfo, idInfo, + idSpecialisation, isLocalId, + idType, idUnique, Id ) -import Name ( isExported, isLocallyDefined ) -import Type ( getFunTy_maybe, splitForAllTy ) -import Maybes ( maybeToBool ) -import Outputable ( PprStyle(..), Outputable(..){-instance * (,) -} ) -import PprCore -import PprType ( GenType{-instance Outputable-}, GenTyVar{-ditto-} ) -import Pretty ( Doc, vcat, ptext, nest, punctuate, comma, hcat, text ) -import TyVar ( GenTyVar{-instance Eq-} ) -import Unique ( Unique{-instance Eq-}, u2i ) -import UniqFM ( keysUFM ) -import Util ( assoc, zipEqual, zipWithEqual, Ord3(..) - , pprTrace, panic -#ifdef DEBUG - , assertPanic -#endif - ) - -isSpecPragmaId_maybe x = Nothing -- ToDo:!trace "OccurAnal.isSpecPragmaId_maybe" +import IdInfo ( OccInfo(..), shortableIdInfo, copyIdInfo ) + +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 ( u2i ) +import UniqFM ( keysUFM ) +import Util ( zipWithEqual, mapAndUnzip ) +import FastTypes +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 - - (Id -> IdSet -> Bool) -- Tells whether an Id occurrence is interesting, - -- given the set of in-scope variables - - IdSet -- In-scope Ids - - -addNewCands :: OccEnv -> [Id] -> OccEnv -addNewCands (OccEnv kd ks kc ip ifun cands) ids - = OccEnv kd ks kc ip ifun (cands `unionIdSets` mkIdSet ids) - -addNewCand :: OccEnv -> Id -> OccEnv -addNewCand (OccEnv ks kd kc ip ifun cands) id - = OccEnv kd ks kc ip ifun (addOneToIdSet cands id) +Here's the externally-callable interface: -isCandidate :: OccEnv -> Id -> Bool -isCandidate (OccEnv _ _ _ _ ifun cands) id = ifun id cands +\begin{code} +occurAnalyseExpr :: (Id -> Bool) -- Tells if a variable is interesting + -> CoreExpr + -> (IdEnv OccInfo, -- Occ info for interesting free vars + CoreExpr) -inlineMe :: OccEnv -> Id -> Bool -inlineMe env id - = {- See comments with simplIdWantsToBeINLINEd in SimplUtils.lhs - not ignore_inline_prag && - -} - idWantsToBeINLINEd id +occurAnalyseExpr interesting expr + = occAnal initial_env expr + where + initial_env = OccEnv interesting emptyVarSet [] -keepUnusedBinding :: OccEnv -> Id -> Bool -keepUnusedBinding (OccEnv keep_dead keep_spec keep_conjurable _ _ _) binder - = keep_dead || (keep_spec && maybeToBool (isSpecPragmaId_maybe binder)) +occurAnalyseGlobalExpr :: CoreExpr -> CoreExpr +occurAnalyseGlobalExpr expr + = -- Top level expr, so no interesting free vars, and + -- discard occurence info returned + snd (occurAnalyseExpr (\_ -> False) expr) -keepBecauseConjurable :: OccEnv -> Id -> Bool -keepBecauseConjurable (OccEnv _ _ keep_conjurable _ _ _) binder - = False - {- keep_conjurable && isConstMethodId binder -} +occurAnalyseRule :: CoreRule -> CoreRule +occurAnalyseRule rule@(BuiltinRule _) = rule +occurAnalyseRule (Rule str tpl_vars tpl_args rhs) + -- Add occ info to tpl_vars, rhs + = Rule str tpl_vars' tpl_args rhs' + where + (rhs_uds, rhs') = occurAnalyseExpr isLocalId rhs + (_, tpl_vars') = tagBinders rhs_uds tpl_vars +\end{code} -type UsageDetails = IdEnv BinderInfo -- A finite map from ids to their usage -combineUsageDetails, combineAltsUsageDetails - :: UsageDetails -> UsageDetails -> UsageDetails +%************************************************************************ +%* * +\subsection{Top level stuff} +%* * +%************************************************************************ -combineUsageDetails usage1 usage2 - = combineIdEnvs addBinderInfo usage1 usage2 +In @occAnalTop@ we do indirection-shorting. That is, if we have this: -combineAltsUsageDetails usage1 usage2 - = combineIdEnvs orBinderInfo usage1 usage2 + x_local = + ... + x_exported = loc -addOneOcc :: UsageDetails -> Id -> BinderInfo -> UsageDetails -addOneOcc usage id info - = combineIdEnvs addBinderInfo usage (unitIdEnv id info) - -- ToDo: make this more efficient +where exp is exported, and loc is not, then we replace it with this: -emptyDetails = (nullIdEnv :: UsageDetails) + x_local = x_exported + x_exported = + ... -unitDetails id info = (unitIdEnv id info :: UsageDetails) +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. -tagBinders :: UsageDetails -- Of scope - -> [Id] -- Binders - -> (UsageDetails, -- Details with binders removed - [(Id,BinderInfo)]) -- Tagged binders - -tagBinders usage binders = - let - usage' = usage `delManyFromIdEnv` binders - uss = [ (binder, usage_of usage binder) | binder <- binders ] - in - if isNullIdEnv usage' then - (usage', uss) - else - (usage', uss) -{- - = (usage `delManyFromIdEnv` binders, - [ (binder, usage_of usage binder) | binder <- binders ] - ) --} -tagBinder :: UsageDetails -- Of scope - -> Id -- Binders - -> (UsageDetails, -- Details with binders removed - (Id,BinderInfo)) -- Tagged binders - -tagBinder usage binder = - let - usage' = usage `delOneFromIdEnv` binder - us = usage_of usage binder - cont = - if isNullIdEnv usage' then -- Bogus test to force evaluation. - (usage', (binder, us)) - else - (usage', (binder, us)) - in - if isDeadOcc us then -- Ditto - cont - else - cont - - -usage_of usage binder - | isExported binder = noBinderInfo -- Visible-elsewhere things count as many - | otherwise - = case (lookupIdEnv usage binder) of - Nothing -> deadOccurrence - Just info -> info - -isNeeded env usage binder - = if isDeadOcc (usage_of usage binder) then - keepUnusedBinding env binder -- Maybe keep it anyway - else - True -\end{code} - - -%************************************************************************ -%* * -\subsection[OccurAnal-main]{Counting occurrences: main function} -%* * -%************************************************************************ +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 = .... -Here's the externally-callable interface: + x_exported2 = x_exported1 +\end{verbatim} -\begin{code} -occurAnalyseBinds - :: [CoreBinding] -- input - -> (SimplifierSwitch -> Bool) - -> [SimplifiableCoreBinding] -- output - -occurAnalyseBinds binds simplifier_sw_chkr - | opt_D_dump_occur_anal = pprTrace "OccurAnal:" - (vcat (map ppr_bind binds')) - binds' - | otherwise = binds' - where - (_, binds') = doo initial_env binds - - initial_env = OccEnv (simplifier_sw_chkr KeepUnusedBindings) - (simplifier_sw_chkr KeepSpecPragmaIds) - (not (simplifier_sw_chkr SimplMayDeleteConjurableIds)) - (simplifier_sw_chkr IgnoreINLINEPragma) - (\id in_scope -> isLocallyDefined id) -- Anything local is interesting - emptyIdSet -- Not actually used - - doo env [] = (emptyDetails, []) - doo env (bind:binds) - = (final_usage, new_binds ++ the_rest) - where - new_env = env `addNewCands` (bindersOf bind) - (binds_usage, the_rest) = doo new_env binds - (final_usage, new_binds) = occAnalBind env bind binds_usage - - -- This really ought to be done properly by PprCore, but - -- it isn't. pprCoreBinding only works on Id binders, and - -- the general case is complicated by the fact that it has to work - -- for interface files too. Sigh - -ppr_bind bind@(NonRec binder expr) - = ppr PprDebug bind - -ppr_bind bind@(Rec binds) - = vcat [ptext SLIT("Rec {"), - nest 2 (ppr PprDebug bind), - ptext SLIT("end Rec }")] -\end{code} +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 :: (Id -> Bool) -- Tells if a variable is interesting - -> CoreExpr - -> (IdEnv BinderInfo, -- Occ info for interesting free vars - SimplifiableCoreExpr) +occurAnalyseBinds :: [CoreBind] -> [CoreBind] -occurAnalyseExpr interesting 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 -} - (\id locals -> interesting id || elementOfIdSet id locals) - emptyIdSet - -occurAnalyseGlobalExpr :: CoreExpr -> SimplifiableCoreExpr -occurAnalyseGlobalExpr expr - = -- Top level expr, so no interesting free vars, and - -- discard occurence info returned - snd (occurAnalyseExpr (\_ -> False) expr) + (_, _, binds') = go initialTopEnv 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') + +initialTopEnv = OccEnv isLocalId -- Anything local is interesting + emptyVarSet + [] + + +-- 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 + if shortableIdInfo (idInfo exported_id) -- Only if its IdInfo is 'shortable' + -- (see the defn of IdInfo.shortableIdInfo) + then True + else pprTrace "shortMeOut:" (ppr exported_id) False + else + False \end{code} + %************************************************************************ %* * \subsection[OccurAnal-main]{Counting occurrences: main function} @@ -273,30 +228,31 @@ Bindings ~~~~~~~~ \begin{code} +type IdWithOccInfo = Id -- An Id with fresh PragmaInfo attached + type Node details = (details, Int, [Int]) -- The Ints are gotten from the Unique, -- which is gotten from the Id. -type Details1 = (Id, (UsageDetails, SimplifiableCoreExpr)) -type Details2 = ((Id, BinderInfo), SimplifiableCoreExpr) +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 -- Not mentioned, so drop dead code - = (body_usage, []) - where - binder' = nukeNoInlinePragma binder - (rhs_usage, rhs') = occAnalRhs env binder' rhs - (final_body_usage, tagged_binder) = tagBinder body_usage binder' + (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: @@ -338,15 +294,14 @@ It isn't easy to do a perfect job in one blow. Consider occAnalBind env (Rec pairs) body_usage = foldr (_scc_ "occAnalBind.dofinal" do_final_bind) (body_usage, []) sccs where - pp_scc (CyclicSCC cycle) = hcat [text "Cyclic ", hcat (punctuate comma (map pp_item cycle))] - pp_scc (AcyclicSCC item) = hcat [text "Acyclic ", pp_item item] - pp_item (_, bndr, _) = ppr PprDebug bndr - binders = map fst pairs - new_env = env `addNewCands` binders + rhs_env = env `addNewCands` binders analysed_pairs :: [Details1] - analysed_pairs = [(nukeNoInlinePragma bndr, occAnalRhs new_env bndr rhs) | (bndr,rhs) <- pairs] + analysed_pairs = [ (bndr, rhs_usage, rhs') + | (bndr, rhs) <- pairs, + let (rhs_usage, rhs') = occAnalRhs rhs_env bndr rhs + ] sccs :: [SCC (Node Details1)] sccs = _scc_ "occAnalBind.scc" stronglyConnCompR edges @@ -355,8 +310,8 @@ occAnalBind env (Rec pairs) body_usage ---- stuff for dependency analysis of binds ------------------------------- edges :: [Node Details1] edges = _scc_ "occAnalBind.assoc" - [ (pair, IBOX(u2i (idUnique id)), edges_from rhs_usage) - | pair@(id, (rhs_usage, rhs)) <- analysed_pairs + [ (details, iBox (u2i (idUnique id)), edges_from rhs_usage) + | details@(id, rhs_usage, rhs) <- analysed_pairs ] -- (a -> b) means a mentions b @@ -365,7 +320,7 @@ occAnalBind env (Rec pairs) body_usage -- by just extracting the keys from the finite map. Grimy, but fast. -- Previously we had this: -- [ bndr | bndr <- bndrs, - -- maybeToBool (lookupIdEnv rhs_usage bndr)] + -- 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 -> [Int] @@ -375,11 +330,11 @@ occAnalBind env (Rec pairs) body_usage ---- stuff to "re-constitute" bindings from dependency-analysis info ------ -- Non-recursive SCC - do_final_bind (AcyclicSCC ((bndr, (rhs_usage, rhs')), _, _)) (body_usage, binds_so_far) - | isNeeded env body_usage bndr - = (combined_usage, new_bind : binds_so_far) - | otherwise + 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 @@ -387,20 +342,20 @@ occAnalBind env (Rec pairs) body_usage -- Recursive SCC do_final_bind (CyclicSCC cycle) (body_usage, binds_so_far) - | any (isNeeded env body_usage) bndrs - = (combined_usage, final_bind:binds_so_far) - | otherwise + | 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 - pairs = [pair | (pair, _, _) <- cycle] - bndrs = [bndr | (bndr, _) <- pairs] - rhs_usages = [rhs_usage | (_, (rhs_usage, _)) <- pairs] - total_usage = foldr combineUsageDetails body_usage rhs_usages - (combined_usage, tagged_binders) = tagBinders total_usage bndrs - final_bind = Rec (reOrderRec env new_cycle) - - new_cycle = CyclicSCC (zipWithEqual "occAnalBind" mk_new_bind tagged_binders cycle) - mk_new_bind tagged_bndr ((_, (_, rhs')), key, keys) = ((tagged_bndr, rhs'), key, keys) + 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} @reOrderRec@ is applied to the list of (binder,rhs) pairs for a cyclic @@ -418,6 +373,10 @@ 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. +============== +[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 @@ -430,6 +389,7 @@ 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 @@ -462,46 +422,68 @@ reOrderRec -- dontINLINE pragmas that there are no loops left. -- Non-recursive case -reOrderRec env (AcyclicSCC (pair, _, _)) = [pair] +reOrderRec env (AcyclicSCC (bind, _, _)) = [bind] -- Common case of simple self-recursion reOrderRec env (CyclicSCC [bind]) - = [((addNoInlinePragma bndr, occ_info), rhs)] + = [(setIdOccInfo tagged_bndr IAmALoopBreaker, rhs)] where - (((bndr,occ_info), rhs), _, _) = bind + ((tagged_bndr, rhs), _, _) = bind -reOrderRec env (CyclicSCC binds) +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)) ++ - [((addNoInlinePragma bndr, occ_info), rhs)] + [(setIdOccInfo tagged_bndr IAmALoopBreaker, rhs)] where - (chosen_pair, unchosen) = choose_loop_breaker binds - ((bndr,occ_info), rhs) = chosen_pair - - -- Choosing the loop breaker; heursitic - choose_loop_breaker (bind@(pair, _, _) : rest) - | not (null rest) && - bad_choice pair - = (chosen, bind : unchosen) -- Don't pick it - | otherwise -- Pick it - = (pair,rest) + (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 - (chosen, unchosen) = choose_loop_breaker rest - - bad_choice ((bndr, occ_info), rhs) - = var_rhs rhs -- Dont pick var RHS - || inlineMe env bndr -- Dont pick INLINE thing - || isOneFunOcc occ_info -- Dont pick single-occ thing - || not_fun_ty (idType bndr) -- Dont pick data-ty thing - - -- isOneFunOcc looks for one textual occurrence, whether inside lambda or whatever. - -- We stick to just FunOccs because if we're not going to be able - -- to inline the thing on this round it might be better to pick - -- this one as the loop breaker. Real example (the Enum Ordering instance - -- from PrelBase): + 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 = case idOccInfo id of + OneOcc _ _ -> True + other -> False + + -- 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) @@ -509,14 +491,11 @@ reOrderRec env (CyclicSCC binds) -- 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 (getFunTy_maybe rho_ty)) + not_fun_ty ty = not (maybeToBool (splitFunTy_maybe rho_ty)) where - (_, rho_ty) = splitForAllTy ty - - -- A variable RHS - var_rhs (Var v) = True - var_rhs other_rhs = False + (_, rho_ty) = splitForAllTys ty \end{code} @occAnalRhs@ deals with the question of bindings where the Id is marked @@ -527,28 +506,29 @@ 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 -> CoreExpr -- Binder and rhs - -> (UsageDetails, SimplifiableCoreExpr) - -occAnalRhs env id (Var v) - | isCandidate env v - = (unitIdEnv v (markMany (funOccurrence 0)), Var v) - - | otherwise - = (emptyDetails, Var v) + -> (UsageDetails, CoreExpr) occAnalRhs env id rhs - | inlineMe env id - = (mapIdEnv markMany rhs_usage, rhs') - - | otherwise - = (rhs_usage, rhs') - + = (final_usage, rhs') where - (rhs_usage, rhs') = occAnal env rhs + (rhs_usage, rhs') = occAnal (zapCtxt env) rhs + + -- [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 @@ -557,17 +537,23 @@ Expressions occAnal :: OccEnv -> CoreExpr -> (UsageDetails, -- Gives info only about the "interesting" Ids - SimplifiableCoreExpr) + CoreExpr) + +occAnal env (Type t) = (emptyDetails, Type t) -occAnal env (Var v) - | isCandidate env v - = (unitIdEnv v (funOccurrence 0), Var v) +occAnal env (Var v) + = (var_uds, Var v) + where + var_uds | isCandidate env v = unitVarEnv v oneOcc + | otherwise = emptyDetails - | otherwise - = (emptyDetails, Var v) + -- 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 (Lit lit) = (emptyDetails, Lit lit) -occAnal env (Prim op args) = (occAnalArgs env args, Prim op args) \end{code} We regard variables that occur as constructor arguments as "dangerousToDup": @@ -586,139 +572,340 @@ If we aren't careful we duplicate the (expensive x) call! Constructors are rather like lambdas in this way. \begin{code} -occAnal env (Con con args) = (mapIdEnv markDangerousToDup (occAnalArgs env args), - Con con args) +occAnal env expr@(Lit lit) = (emptyDetails, expr) +\end{code} -occAnal env (SCC cc body) - = (mapIdEnv markInsideSCC usage, SCC cc body') - where - (usage, body') = occAnal env body +\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} -occAnal env (Coerce c ty body) - = (usage, Coerce c ty body') - where - (usage, body') = occAnal env body +\begin{code} +occAnal env app@(App fun arg) + = occAnalApp env (collectArgs app) -occAnal env (App fun arg) - = (fun_usage `combineUsageDetails` arg_usage, App fun' arg) - where - (fun_usage, fun') = occAnal env fun - arg_usage = occAnalArg env arg +-- 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 don't take account of the \y when dealing with x's usage; --- instead, the simplifier is careful when partially applying lambdas - -occAnal env expr@(Lam (ValBinder binder) body) - = (mapIdEnv markDangerousToDup final_usage, - foldr ( \v b -> Lam (ValBinder v) b) body' tagged_binders) +-- 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 `addNewCands` binders) 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) = collectValBinders expr - (body_usage, body') = occAnal (env `addNewCands` binders) body - (final_usage, tagged_binders) = tagBinders body_usage binders - --- ANDY: WE MUST THINK ABOUT THIS! (ToDo) -occAnal env (Lam (TyBinder tyvar) body) - = case occAnal env body of { (body_usage, body') -> - (mapIdEnv markDangerousToDup body_usage, - Lam (TyBinder tyvar) body') } --- where --- (body_usage, body') = occAnal env body - -occAnal env (Lam (UsageBinder _) _) = panic "OccurAnal.occAnal Lam UsageBinder" - -occAnal env (Case scrut alts) - = case occAnalAlts env alts of { (alts_usage, alts') -> - case occAnal env scrut of { (scrut_usage, scrut') -> - let - det = scrut_usage `combineUsageDetails` alts_usage - in - if isNullIdEnv det then - (det, Case scrut' alts') - else - (det, Case scrut' alts') }} -{- - (scrut_usage `combineUsageDetails` alts_usage, - Case scrut' alts') + (binders, body) = collectBinders expr + (linear, env_body, _) = oneShotGroup env binders + +occAnal env (Case scrut bndr alts) + = case mapAndUnzip (occAnalAlt alt_env) alts of { (alts_usage_s, alts') -> + case occAnal (zapCtxt env) scrut of { (scrut_usage, scrut') -> + 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 - (scrut_usage, scrut') = occAnal env scrut - (alts_usage, alts') = occAnalAlts env alts --} + 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) 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, foldr Let body' new_binds) }} -- mkCoLet* wants Core... (sigh) + (final_usage, mkLets new_binds body') }} + where + new_env = env `addNewCands` (bindersOf bind) + +occAnalArgs env args + = case mapAndUnzip (occAnal arg_env) args of { (arg_uds_s, args') -> + (foldr combineUsageDetails emptyDetails arg_uds_s, args')} where - new_env = env `addNewCands` (bindersOf bind) --- (body_usage, body') = occAnal new_env body --- (final_usage, new_binds) = occAnalBind env bind body_usage + arg_env = zapCtxt env \end{code} +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) + = case args_stuff of { (args_uds, args') -> + let + final_uds = fun_uds `combineUsageDetails` args_uds + in + (final_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 + + | isDataConId fun = case occAnalArgs env args of + (arg_uds, args') -> (mapVarEnv markMany arg_uds, args') + -- We mark the free vars of the argument of a constructor as "many" + -- 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. + + | otherwise = occAnalArgs env args + + +occAnalApp env (fun, args) + = case occAnal (zapCtxt env) fun of { (fun_uds, fun') -> + 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 -> [CoreExpr] -> (UsageDetails, [CoreExpr]) +appSpecial env n ctxt args + = go n args + where + go n [] = (emptyDetails, []) -- Too few args + + go 1 (arg:args) -- The magic arg + = case occAnal (setCtxt 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 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 ~~~~~~~~~~~~~~~~~ \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 (con, bndrs, rhs) + = case occAnal (env `addNewCands` bndrs) rhs of { (rhs_usage, rhs') -> + let + (final_usage, tagged_bndrs) = tagBinders rhs_usage bndrs + in + (final_usage, (con, tagged_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') + +%************************************************************************ +%* * +\subsection[OccurAnal-types]{Data types} +%* * +%************************************************************************ + +\begin{code} +-- We gather inforamtion for variables that are either +-- (a) in scope or +-- (b) interesting + +data OccEnv = + OccEnv (Id -> Bool) -- Tells whether an Id occurrence is interesting, + IdSet -- In-scope Ids + CtxtTy -- Tells about linearity + +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 + +isCandidate :: OccEnv -> Id -> Bool +isCandidate (OccEnv ifun cands _) id = id `elemVarSet` cands || ifun id + +addNewCands :: OccEnv -> [Id] -> OccEnv +addNewCands (OccEnv ifun cands ctxt) ids + = OccEnv ifun (cands `unionVarSet` mkVarSet ids) ctxt + +addNewCand :: OccEnv -> Id -> OccEnv +addNewCand (OccEnv ifun cands ctxt) id + = OccEnv ifun (extendVarSet cands id) ctxt + +setCtxt :: OccEnv -> CtxtTy -> OccEnv +setCtxt (OccEnv ifun cands _) ctxt = OccEnv ifun cands ctxt + +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 ifun cands ctxt) bndrs + = case go ctxt bndrs [] of + (new_ctxt, new_bndrs) -> (all is_one_shot new_bndrs, OccEnv ifun cands new_ctxt, new_bndrs) where - (alts_usage, alts') = unzip (map do_alt alts) - (deflt_usage, deflt') = occAnalDeflt env deflt + is_one_shot b = isId b && isOneShotLambda b - do_alt (lit, rhs) - = (rhs_usage, (lit, rhs')) - where - (rhs_usage, rhs') = occAnal env rhs + 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) + + +zapCtxt env@(OccEnv ifun cands []) = env +zapCtxt (OccEnv ifun cands _ ) = OccEnv ifun cands [] + +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 -occAnalDeflt env NoDefault = (emptyDetails, NoDefault) +addOneOcc :: UsageDetails -> Id -> OccInfo -> UsageDetails +addOneOcc usage id info + = plusVarEnv_C addOccInfo usage (unitVarEnv id info) + -- ToDo: make this more efficient + +emptyDetails = (emptyVarEnv :: UsageDetails) + +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) -occAnalDeflt env (BindDefault binder rhs) - = (final_usage, BindDefault tagged_binder rhs') +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 - new_env = env `addNewCand` binder - (rhs_usage, rhs') = occAnal new_env rhs - (final_usage, tagged_binder) = tagBinder rhs_usage binder + occ_info = lookupVarEnv usage bndr `orElse` IAmDead \end{code} -Atoms -~~~~~ +%************************************************************************ +%* * +\subsection{Operations over OccInfo} +%* * +%************************************************************************ + \begin{code} -occAnalArgs :: OccEnv -> [CoreArg] -> UsageDetails +oneOcc :: OccInfo +oneOcc = OneOcc False True -occAnalArgs env atoms - = foldr do_one_atom emptyDetails atoms - where - do_one_atom (VarArg v) usage - | isCandidate env v = addOneOcc usage v (argOccurrence 0) - | otherwise = usage - do_one_atom other_arg usage = usage +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 -occAnalArg :: 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 -occAnalArg env (VarArg v) - | isCandidate env v = unitDetails v (argOccurrence 0) - | otherwise = emptyDetails -occAnalArg _ _ = emptyDetails +orOccInfo info1 info2 = NoOccInfo \end{code}