%
-% (c) The GRASP/AQUA Project, Glasgow University, 1993-1996
+% (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
%
\section[WorkWrap]{Worker/wrapper-generating back-end of strictness analyser}
\begin{code}
-#include "HsVersions.h"
-
-module WorkWrap ( workersAndWrappers, getWorkerIdAndCons ) where
+module WorkWrap ( wwTopBinds, mkWrapper ) where
-IMP_Ubiq(){-uitous-}
+#include "HsVersions.h"
import CoreSyn
-import CoreUnfold ( Unfolding, certainlySmallEnoughToInline, calcUnfoldingGuidance )
-import CmdLineOpts ( opt_UnfoldingCreationThreshold )
-
-import CoreUtils ( coreExprType )
-import Id ( getInlinePragma, getIdStrictness, mkWorkerId,
- addIdStrictness, addInlinePragma,
- SYN_IE(IdSet), emptyIdSet, addOneToIdSet,
- GenId, SYN_IE(Id)
+import CoreUnfold ( certainlyWillInline )
+import CoreLint ( showPass, endPass )
+import CoreUtils ( exprType )
+import Id ( Id, idType, idStrictness, idArity, isOneShotLambda,
+ setIdStrictness, idInlinePragma, mkWorkerId,
+ setIdWorkerInfo, idCprInfo, setInlinePragma )
+import Type ( Type, isNewType, splitForAllTys, splitFunTys )
+import IdInfo ( mkStrictnessInfo, noStrictnessInfo, StrictnessInfo(..),
+ CprInfo(..), InlinePragInfo(..), isNeverInlinePrag,
+ WorkerInfo(..)
)
-import IdInfo ( noIdInfo, addUnfoldInfo,
- mkStrictnessInfo, addStrictnessInfo, StrictnessInfo(..)
- )
-import SaLib
-import UniqSupply ( returnUs, thenUs, mapUs, getUnique, SYN_IE(UniqSM) )
+import Demand ( Demand )
+import UniqSupply ( UniqSupply, initUs_, returnUs, thenUs, mapUs, getUniqueUs, UniqSM )
+import CmdLineOpts
import WwLib
-import Pretty ( Doc )
-import Outputable ( ppr, PprStyle(..) )
-import Util ( pprPanic )
+import Outputable
\end{code}
-We take Core bindings whose binders have their strictness attached (by
-the front-end of the strictness analyser), and we return some
-``plain'' bindings which have been worker/wrapper-ified, meaning:
+We take Core bindings whose binders have:
+
\begin{enumerate}
-\item
-Functions have been split into workers and wrappers where appropriate;
-\item
-Binders' @IdInfos@ have been updated to reflect the existence
-of these workers/wrappers (this is where we get STRICTNESS pragma
+
+\item Strictness attached (by the front-end of the strictness
+analyser), and / or
+
+\item Constructed Product Result information attached by the CPR
+analysis pass.
+
+\end{enumerate}
+
+and we return some ``plain'' bindings which have been
+worker/wrapper-ified, meaning:
+
+\begin{enumerate}
+
+\item Functions have been split into workers and wrappers where
+appropriate. If a function has both strictness and CPR properties
+then only one worker/wrapper doing both transformations is produced;
+
+\item Binders' @IdInfos@ have been updated to reflect the existence of
+these workers/wrappers (this is where we get STRICTNESS and CPR pragma
info for exported values).
\end{enumerate}
\begin{code}
-workersAndWrappers :: [CoreBinding] -> UniqSM [CoreBinding]
-workersAndWrappers top_binds
- = mapUs (wwBind True{-top-level-}) top_binds `thenUs` \ top_binds2 ->
- let
- top_binds3 = map make_top_binding top_binds2
- in
- returnUs (concat top_binds3)
- where
- make_top_binding :: WwBinding -> [CoreBinding]
+wwTopBinds :: DynFlags
+ -> UniqSupply
+ -> [CoreBind]
+ -> IO [CoreBind]
+
+wwTopBinds dflags us binds
+ = do {
+ showPass dflags "Worker Wrapper binds";
+
+ -- Create worker/wrappers, and mark binders with their
+ -- "strictness info" [which encodes their worker/wrapper-ness]
+ let { binds' = workersAndWrappers us binds };
- make_top_binding (WwLet binds) = binds
+ endPass dflags "Worker Wrapper binds"
+ Opt_D_dump_worker_wrapper binds'
+ }
+\end{code}
+
+
+\begin{code}
+workersAndWrappers :: UniqSupply -> [CoreBind] -> [CoreBind]
+
+workersAndWrappers us top_binds
+ = initUs_ us $
+ mapUs wwBind top_binds `thenUs` \ top_binds' ->
+ returnUs (concat top_binds')
\end{code}
%************************************************************************
turn. Non-recursive case first, then recursive...
\begin{code}
-wwBind :: Bool -- True <=> top-level binding
- -> CoreBinding
- -> UniqSM WwBinding -- returns a WwBinding intermediate form;
+wwBind :: CoreBind
+ -> UniqSM [CoreBind] -- returns a WwBinding intermediate form;
-- the caller will convert to Expr/Binding,
-- as appropriate.
-wwBind top_level (NonRec binder rhs)
- = wwExpr rhs `thenUs` \ new_rhs ->
- tryWW binder new_rhs `thenUs` \ new_pairs ->
- returnUs (WwLet [NonRec b e | (b,e) <- new_pairs])
+wwBind (NonRec binder rhs)
+ = wwExpr rhs `thenUs` \ new_rhs ->
+ tryWW True {- non-recursive -} binder new_rhs `thenUs` \ new_pairs ->
+ returnUs [NonRec b e | (b,e) <- new_pairs]
-- Generated bindings must be non-recursive
-- because the original binding was.
------------------------------
-wwBind top_level (Rec pairs)
+wwBind (Rec pairs)
= mapUs do_one pairs `thenUs` \ new_pairs ->
- returnUs (WwLet [Rec (concat new_pairs)])
+ returnUs [Rec (concat new_pairs)]
where
do_one (binder, rhs) = wwExpr rhs `thenUs` \ new_rhs ->
- tryWW binder new_rhs
+ tryWW False {- recursive -} binder new_rhs
\end{code}
@wwExpr@ basically just walks the tree, looking for appropriate
annotations that can be used. Remember it is @wwBind@ that does the
matching by looking for strict arguments of the correct type.
@wwExpr@ is a version that just returns the ``Plain'' Tree.
-???????????????? ToDo
\begin{code}
wwExpr :: CoreExpr -> UniqSM CoreExpr
+wwExpr e@(Type _) = returnUs e
wwExpr e@(Var _) = returnUs e
wwExpr e@(Lit _) = returnUs e
-wwExpr e@(Con _ _) = returnUs e
-wwExpr e@(Prim _ _) = returnUs e
wwExpr (Lam binder expr)
= wwExpr expr `thenUs` \ new_expr ->
wwExpr (App f a)
= wwExpr f `thenUs` \ new_f ->
- returnUs (App new_f a)
-
-wwExpr (SCC cc expr)
- = wwExpr expr `thenUs` \ new_expr ->
- returnUs (SCC cc new_expr)
+ wwExpr a `thenUs` \ new_a ->
+ returnUs (App new_f new_a)
-wwExpr (Coerce c ty expr)
+wwExpr (Note note expr)
= wwExpr expr `thenUs` \ new_expr ->
- returnUs (Coerce c ty new_expr)
+ returnUs (Note note new_expr)
wwExpr (Let bind expr)
- = wwBind False{-not top-level-} bind `thenUs` \ intermediate_bind ->
- wwExpr expr `thenUs` \ new_expr ->
- returnUs (mash_ww_bind intermediate_bind new_expr)
- where
- mash_ww_bind (WwLet binds) body = mkCoLetsNoUnboxed binds body
- mash_ww_bind (WwCase case_fn) body = case_fn body
+ = wwBind bind `thenUs` \ intermediate_bind ->
+ wwExpr expr `thenUs` \ new_expr ->
+ returnUs (mkLets intermediate_bind new_expr)
-wwExpr (Case expr alts)
+wwExpr (Case expr binder alts)
= wwExpr expr `thenUs` \ new_expr ->
- ww_alts alts `thenUs` \ new_alts ->
- returnUs (Case new_expr new_alts)
+ mapUs ww_alt alts `thenUs` \ new_alts ->
+ returnUs (Case new_expr binder new_alts)
where
- ww_alts (AlgAlts alts deflt)
- = mapUs ww_alg_alt alts `thenUs` \ new_alts ->
- ww_deflt deflt `thenUs` \ new_deflt ->
- returnUs (AlgAlts new_alts new_deflt)
-
- ww_alts (PrimAlts alts deflt)
- = mapUs ww_prim_alt alts `thenUs` \ new_alts ->
- ww_deflt deflt `thenUs` \ new_deflt ->
- returnUs (PrimAlts new_alts new_deflt)
-
- ww_alg_alt (con, binders, rhs)
+ ww_alt (con, binders, rhs)
= wwExpr rhs `thenUs` \ new_rhs ->
returnUs (con, binders, new_rhs)
-
- ww_prim_alt (lit, rhs)
- = wwExpr rhs `thenUs` \ new_rhs ->
- returnUs (lit, new_rhs)
-
- ww_deflt NoDefault
- = returnUs NoDefault
-
- ww_deflt (BindDefault binder rhs)
- = wwExpr rhs `thenUs` \ new_rhs ->
- returnUs (BindDefault binder new_rhs)
\end{code}
%************************************************************************
The only reason this is monadised is for the unique supply.
\begin{code}
-tryWW :: Id -- The fn binder
+tryWW :: Bool -- True <=> a non-recursive binding
+ -> Id -- The fn binder
-> CoreExpr -- The bound rhs; its innards
-- are already ww'd
-> UniqSM [(Id, CoreExpr)] -- either *one* or *two* pairs;
-- the orig "wrapper" lives on);
-- if two, then a worker and a
-- wrapper.
-tryWW fn_id rhs
- | (certainlySmallEnoughToInline $
- calcUnfoldingGuidance (getInlinePragma fn_id)
- opt_UnfoldingCreationThreshold
- rhs
- )
- -- No point in worker/wrappering something that is going to be
- -- INLINEd wholesale anyway. If the strictness analyser is run
- -- twice, this test also prevents wrappers (which are INLINEd)
- -- from being re-done.
-
- || not has_strictness_info
- || not (worthSplitting revised_wrap_args_info)
+tryWW non_rec fn_id rhs
+ | isNeverInlinePrag inline_prag || arity == 0
+ = -- Don't split things that will never be inlined
+ returnUs [ (fn_id, rhs) ]
+
+ | non_rec && certainlyWillInline fn_id
+ -- No point in worker/wrappering a function that is going to be
+ -- INLINEd wholesale anyway. If the strictness analyser is run
+ -- twice, this test also prevents wrappers (which are INLINEd)
+ -- from being re-done.
+ --
+ -- It's very important to refrain from w/w-ing an INLINE function
+ -- If we do so by mistake we transform
+ -- f = __inline (\x -> E)
+ -- into
+ -- f = __inline (\x -> case x of (a,b) -> fw E)
+ -- fw = \ab -> (__inline (\x -> E)) (a,b)
+ -- and the original __inline now vanishes, so E is no longer
+ -- inside its __inline wrapper. Death! Disaster!
+ --
+ -- OUT OF DATE NOTE:
+ -- [There used to be "&& not do_coerce_ww" in the above test.
+ -- No longer necessary because SimplUtils.tryEtaExpansion
+ -- now deals with coerces.]
+ -- The do_coerce_ww test is so that
+ -- a function with a coerce should w/w to get rid
+ -- of the coerces, which can significantly improve its arity.
+ -- Example: f [] = return [] :: IO [Int]
+ -- f (x:xs) = return (x:xs)
+ -- If we aren't careful we end up with
+ -- f = \ x -> case x of {
+ -- x:xs -> __coerce (IO [Int]) (\ s -> (# s, x:xs #)
+ -- [] -> lvl_sJ8
+ --
+ -- OUT OF DATE NOTE:
+ -- [Out of date because the size calculation in CoreUnfold now
+ -- makes wrappers look very cheap even when they are inlined.]
+ -- In this case we add an INLINE pragma to the RHS. Why?
+ -- Because consider
+ -- f = \x -> g x x
+ -- g = \yz -> ... -- And g is strict
+ -- Then f is small, so we don't w/w it. But g is big, and we do, so
+ -- g's wrapper will get inlined in f's RHS, which makes f look big now.
+ -- So f doesn't get inlined, but it is strict and we have failed to w/w it.
+ = returnUs [ (fn_id, rhs) ]
+
+ | not (do_strict_ww || do_cpr_ww || do_coerce_ww)
= returnUs [ (fn_id, rhs) ]
| otherwise -- Do w/w split
- = let
- (uvars, tyvars, wrap_args, body) = collectBinders rhs
- in
- mkWwBodies tyvars wrap_args
- (coreExprType body)
- revised_wrap_args_info `thenUs` \ (wrap_fn, work_fn, work_demands) ->
- getUnique `thenUs` \ work_uniq ->
+ = mkWwBodies fun_ty arity wrap_dmds result_bot one_shots cpr_info `thenUs` \ (work_demands, wrap_fn, work_fn) ->
+ getUniqueUs `thenUs` \ work_uniq ->
let
- work_rhs = work_fn body
- work_id = mkWorkerId work_uniq fn_id (coreExprType work_rhs) work_info
- work_info = noIdInfo `addStrictnessInfo` mkStrictnessInfo work_demands False
+ work_rhs = work_fn rhs
+ proto_work_id = mkWorkerId work_uniq fn_id (exprType work_rhs)
+ `setInlinePragma` inline_prag
+
+ work_id | has_strictness = proto_work_id `setIdStrictness` mkStrictnessInfo (work_demands, result_bot)
+ | otherwise = proto_work_id
wrap_rhs = wrap_fn work_id
- wrap_id = addInlinePragma (fn_id `addIdStrictness`
- mkStrictnessInfo revised_wrap_args_info True)
+ wrap_id = fn_id `setIdStrictness` wrapper_strictness
+ `setIdWorkerInfo` HasWorker work_id arity
+ `setInlinePragma` NoInlinePragInfo -- Put it on the worker instead
-- Add info to the wrapper:
- -- (a) we want to inline it everywhere
- -- (b) we want to pin on its revised stricteness info
- -- (c) we pin on its worker id and the list of constructors mentioned in the wrapper
+ -- (a) we want to set its arity
+ -- (b) we want to pin on its revised strictness info
+ -- (c) we pin on its worker id
in
returnUs ([(work_id, work_rhs), (wrap_id, wrap_rhs)])
-- Worker first, because wrapper mentions it
+ -- mkWwBodies has already built a wrap_rhs with an INLINE pragma wrapped around it
where
- strictness_info = getIdStrictness fn_id
- has_strictness_info = case strictness_info of
- StrictnessInfo _ _ -> True
- other -> False
-
- wrap_args_info = case strictness_info of
- StrictnessInfo args_info _ -> args_info
- revised_wrap_args_info = setUnpackStrategy wrap_args_info
-
--- This rather crude function looks at a wrapper function, and
--- snaffles out (a) the worker Id and (b) constructors needed to
--- make the wrapper.
--- These are needed when we write an interface file.
-getWorkerIdAndCons wrap_id wrapper_fn
- = go wrapper_fn
+ fun_ty = idType fn_id
+ arity = idArity fn_id -- The arity is set by the simplifier using exprEtaExpandArity
+ -- So it may be more than the number of top-level-visible lambdas
+
+ inline_prag = idInlinePragma fn_id
+
+ strictness_info = idStrictness fn_id
+ has_strictness = case strictness_info of
+ StrictnessInfo _ _ -> True
+ NoStrictnessInfo -> False
+ (arg_demands, result_bot) = case strictness_info of
+ StrictnessInfo d r -> (d, r)
+ NoStrictnessInfo -> ([], False)
+
+ wrap_dmds = setUnpackStrategy arg_demands
+ do_strict_ww = WARN( has_strictness && not result_bot && arity < length arg_demands && worthSplitting wrap_dmds result_bot,
+ text "Insufficient arity" <+> ppr fn_id <+> ppr arity <+> ppr arg_demands )
+ (result_bot || arity >= length arg_demands) -- Only if there's enough visible arity
+ && -- (else strictness info isn't valid)
+ --
+ worthSplitting wrap_dmds result_bot -- And it's useful
+ -- worthSplitting returns False for an empty list of demands,
+ -- and hence do_strict_ww is False if arity is zero
+ -- Also it's false if there is no strictness (arg_demands is [])
+
+ wrapper_strictness | has_strictness = mkStrictnessInfo (wrap_dmds, result_bot)
+ | otherwise = noStrictnessInfo
+
+ -------------------------------------------------------------
+ cpr_info = idCprInfo fn_id
+ do_cpr_ww = arity > 0 &&
+ case cpr_info of
+ ReturnsCPR -> True
+ other -> False
+
+ -------------------------------------------------------------
+ do_coerce_ww = check_for_coerce arity fun_ty
+ -- We are willing to do a w/w even if the arity is zero.
+ -- x = coerce t E
+ -- ==>
+ -- x' = E
+ -- x = coerce t x'
+
+ -------------------------------------------------------------
+ one_shots = get_one_shots rhs
+
+-- See if there's a Coerce before we run out of arity;
+-- if so, it's worth trying a w/w split. Reason: we find
+-- functions like f = coerce (\s -> e)
+-- and g = \x -> coerce (\s -> e)
+-- and they may have no useful strictness or cpr info, but if we
+-- do the w/w thing we get rid of the coerces.
+
+check_for_coerce arity ty
+ = length arg_tys <= arity && isNewType res_ty
+ -- Don't look further than arity args,
+ -- but if there are arity or fewer, see if there's
+ -- a newtype in the corner
where
- go (Lam _ body) = go body
- go (Case _ (AlgAlts [(con,_,rhs)] _)) = let (wrap_id, cons) = go rhs
- in (wrap_id, cons `addOneToIdSet` con)
- go other = (get_work_id other, emptyIdSet)
-
- get_work_id (App fn _) = get_work_id fn
- get_work_id (Var work_id) = work_id
- get_work_id other = pprPanic "getWorkerIdAndCons" (ppr PprDebug wrap_id)
+ (_, tau) = splitForAllTys ty
+ (arg_tys, res_ty) = splitFunTys tau
+
+-- If the original function has one-shot arguments, it is important to
+-- make the wrapper and worker have corresponding one-shot arguments too.
+-- Otherwise we spuriously float stuff out of case-expression join points,
+-- which is very annoying.
+get_one_shots (Lam b e)
+ | isId b = isOneShotLambda b : get_one_shots e
+ | otherwise = get_one_shots e
+get_one_shots (Note _ e) = get_one_shots e
+get_one_shots other = noOneShotInfo
+\end{code}
+
+
+
+%************************************************************************
+%* *
+\subsection{The worker wrapper core}
+%* *
+%************************************************************************
+
+@mkWrapper@ is called when importing a function. We have the type of
+the function and the name of its worker, and we want to make its body (the wrapper).
+
+\begin{code}
+mkWrapper :: Type -- Wrapper type
+ -> Int -- Arity
+ -> [Demand] -- Wrapper strictness info
+ -> Bool -- Function returns bottom
+ -> CprInfo -- Wrapper cpr info
+ -> UniqSM (Id -> CoreExpr) -- Wrapper body, missing worker Id
+
+mkWrapper fun_ty arity demands res_bot cpr_info
+ = mkWwBodies fun_ty arity demands res_bot noOneShotInfo cpr_info `thenUs` \ (_, wrap_fn, _) ->
+ returnUs wrap_fn
+
+noOneShotInfo = repeat False
\end{code}
+
+