2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
10 module DmdAnal ( dmdAnalPgm, dmdAnalTopRhs,
11 both {- needed by WwLib -}
14 #include "HsVersions.h"
16 import CmdLineOpts ( DynFlags, DynFlag(..), opt_MaxWorkerArgs )
17 import NewDemand -- All of it
20 import CoreUtils ( exprIsValue, exprArity )
21 import DataCon ( dataConTyCon )
22 import TyCon ( isProductTyCon, isRecursiveTyCon )
23 import Id ( Id, idType, idDemandInfo,
24 isDataConId, isImplicitId, isGlobalId,
25 idNewStrictness, idNewStrictness_maybe, getNewStrictness, setIdNewStrictness,
26 idNewDemandInfo, setIdNewDemandInfo, newStrictnessFromOld )
27 import IdInfo ( newDemand )
30 import UniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly,
31 keysUFM, minusUFM, ufmToList, filterUFM )
32 import Type ( isUnLiftedType )
33 import CoreLint ( showPass, endPass )
34 import Util ( mapAndUnzip, mapAccumL, mapAccumR )
35 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel )
36 import Maybes ( orElse, expectJust )
42 * set a noinline pragma on bottoming Ids
44 * Consider f x = x+1 `fatbar` error (show x)
45 We'd like to unbox x, even if that means reboxing it in the error case.
48 instance Outputable TopLevelFlag where
52 %************************************************************************
54 \subsection{Top level stuff}
56 %************************************************************************
59 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
60 dmdAnalPgm dflags binds
62 showPass dflags "Demand analysis" ;
63 let { binds_plus_dmds = do_prog binds ;
64 dmd_changes = get_changes binds_plus_dmds } ;
65 endPass dflags "Demand analysis"
66 Opt_D_dump_stranal binds_plus_dmds ;
68 -- Only if DEBUG is on, because only then is the old strictness analyser run
69 printDump (text "Changes in demands" $$ dmd_changes) ;
71 return binds_plus_dmds
74 do_prog :: [CoreBind] -> [CoreBind]
75 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
77 dmdAnalTopBind :: SigEnv
80 dmdAnalTopBind sigs (NonRec id rhs)
81 | isImplicitId id -- Don't touch the info on constructors, selectors etc
82 = (sigs, NonRec id rhs) -- It's pre-computed in MkId.lhs
85 (sigs', _, (id', rhs')) = downRhs TopLevel sigs (id, rhs)
87 (sigs', NonRec id' rhs')
89 dmdAnalTopBind sigs (Rec pairs)
91 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
97 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
98 -- Analyse the RHS and return
99 -- a) appropriate strictness info
100 -- b) the unfolding (decorated with stricntess info)
104 arity = exprArity rhs
105 (rhs_ty, rhs') = dmdAnal emptySigEnv (vanillaCall arity) rhs
106 (_, sig) = mkSigTy rhs rhs_ty
109 %************************************************************************
111 \subsection{The analyser itself}
113 %************************************************************************
116 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
118 dmdAnal sigs Abs e = (topDmdType, e)
119 dmdAnal sigs Bot e = (botDmdType, e)
121 dmdAnal sigs Lazy e = let
122 (res_ty, e') = dmdAnal sigs Eval e
124 (deferType res_ty, e')
125 -- It's important not to analyse e with a lazy demand because
126 -- a) When we encounter case s of (a,b) ->
127 -- we demand s with U(d1d2)... but if the overall demand is lazy
128 -- that is wrong, and we'd need to reduce the demand on s,
129 -- which is inconvenient
130 -- b) More important, consider
131 -- f (let x = R in x+x), where f is lazy
132 -- We still want to mark x as demanded, because it will be when we
133 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
134 -- just mark x as Lazy
137 dmdAnal sigs dmd (Lit lit)
138 = (topDmdType, Lit lit)
140 dmdAnal sigs dmd (Var var)
141 = (dmdTransform sigs var dmd, Var var)
143 dmdAnal sigs dmd (Note n e)
144 = (dmd_ty, Note n e')
146 (dmd_ty, e') = dmdAnal sigs dmd' e
148 Coerce _ _ -> Eval -- This coerce usually arises from a recursive
149 other -> dmd -- newtype, and we don't want to look inside them
150 -- for exactly the same reason that we don't look
151 -- inside recursive products -- we might not reach
152 -- a fixpoint. So revert to a vanilla Eval demand
154 dmdAnal sigs dmd (App fun (Type ty))
155 = (fun_ty, App fun' (Type ty))
157 (fun_ty, fun') = dmdAnal sigs dmd fun
159 dmdAnal sigs dmd (App fun arg) -- Non-type arguments
160 = let -- [Type arg handled above]
161 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
162 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
163 (arg_dmd, res_ty) = splitDmdTy fun_ty
165 (res_ty `bothType` arg_ty, App fun' arg')
167 dmdAnal sigs dmd (Lam var body)
170 (body_ty, body') = dmdAnal sigs dmd body
172 (body_ty, Lam var body')
174 | Call body_dmd <- dmd -- A call demand: good!
176 (body_ty, body') = dmdAnal sigs body_dmd body
177 (lam_ty, var') = annotateLamIdBndr body_ty var
179 (lam_ty, Lam var' body')
181 | otherwise -- Not enough demand on the lambda; but do the body
182 = let -- anyway to annotate it and gather free var info
183 (body_ty, body') = dmdAnal sigs Eval body
184 (lam_ty, var') = annotateLamIdBndr body_ty var
186 (deferType lam_ty, Lam var' body')
188 dmdAnal sigs dmd (Case scrut case_bndr [alt@(DataAlt dc,bndrs,rhs)])
189 | let tycon = dataConTyCon dc,
190 isProductTyCon tycon,
191 not (isRecursiveTyCon tycon)
193 (alt_ty, alt') = dmdAnalAlt sigs dmd alt
194 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
195 (_, bndrs', _) = alt'
197 -- Figure out whether the demand on the case binder is used, and use
198 -- that to set the scrut_dmd. This is utterly essential.
199 -- Consider f x = case x of y { (a,b) -> k y a }
200 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
201 -- worker, so the worker will rebuild
202 -- x = (a, absent-error)
203 -- and that'll crash.
204 -- So at one stage I had:
205 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
206 -- keepity | dead_case_bndr = Drop
207 -- | otherwise = Keep
210 -- case x of y { (a,b) -> h y + a }
211 -- where h : U(LL) -> T
212 -- The above code would compute a Keep for x, since y is not Abs, which is silly
213 -- The insight is, of course, that a demand on y is a demand on the
214 -- scrutinee, so we need to `both` it with the scrut demand
216 scrut_dmd = mkSeq Drop [idNewDemandInfo b | b <- bndrs', isId b]
218 idNewDemandInfo case_bndr'
220 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
222 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' [alt'])
224 dmdAnal sigs dmd (Case scrut case_bndr alts)
226 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
227 (scrut_ty, scrut') = dmdAnal sigs Eval scrut
228 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
230 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
231 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' alts')
233 dmdAnal sigs dmd (Let (NonRec id rhs) body)
235 (sigs', lazy_fv, (id1, rhs')) = downRhs NotTopLevel sigs (id, rhs)
236 (body_ty, body') = dmdAnal sigs' dmd body
237 (body_ty1, id2) = annotateBndr body_ty id1
238 body_ty2 = addLazyFVs body_ty1 lazy_fv
240 -- pprTrace "dmdLet" (ppr id <+> ppr (sig,rhs_env))
241 (body_ty2, Let (NonRec id2 rhs') body')
243 dmdAnal sigs dmd (Let (Rec pairs) body)
245 bndrs = map fst pairs
246 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
247 (body_ty, body') = dmdAnal sigs' dmd body
248 body_ty1 = addLazyFVs body_ty lazy_fv
250 sigs' `seq` body_ty `seq`
252 (body_ty2, _) = annotateBndrs body_ty1 bndrs
253 -- Don't bother to add demand info to recursive
254 -- binders as annotateBndr does;
255 -- being recursive, we can't treat them strictly.
256 -- But we do need to remove the binders from the result demand env
258 (body_ty2, Let (Rec pairs') body')
261 dmdAnalAlt sigs dmd (con,bndrs,rhs)
263 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
264 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
266 (alt_ty, (con, bndrs', rhs'))
269 %************************************************************************
271 \subsection{Bindings}
273 %************************************************************************
276 dmdFix :: TopLevelFlag
277 -> SigEnv -- Does not include bindings for this binding
280 [(Id,CoreExpr)]) -- Binders annotated with stricness info
282 dmdFix top_lvl sigs orig_pairs
283 = loop 1 initial_sigs orig_pairs
285 bndrs = map fst orig_pairs
286 initial_sigs = extendSigEnvList sigs [(id, (initial_sig id, top_lvl)) | id <- bndrs]
289 -> SigEnv -- Already contains the current sigs
291 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
293 | all (same_sig sigs sigs') bndrs
294 = (sigs', lazy_fv, pairs')
295 -- Note: use pairs', not pairs. pairs' is the result of
296 -- processing the RHSs with sigs (= sigs'), whereas pairs
297 -- is the result of processing the RHSs with the *previous*
298 -- iteration of sigs.
299 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
300 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
301 text "env:" <+> ppr (ufmToList sigs),
302 text "binds:" <+> pprCoreBinding (Rec pairs)]))
303 (emptySigEnv, emptyDmdEnv, orig_pairs) -- Safe output
304 | otherwise = loop (n+1) sigs' pairs'
306 -- Use the new signature to do the next pair
307 -- The occurrence analyser has arranged them in a good order
308 -- so this can significantly reduce the number of iterations needed
309 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
311 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
312 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
314 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
315 ((sigs', lazy_fv'), pair')
318 (sigs', lazy_fv1, pair') = downRhs top_lvl sigs (id,rhs)
319 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
320 -- old_sig = lookup sigs id
321 -- new_sig = lookup sigs' id
323 -- Get an initial strictness signature from the Id
324 -- itself. That way we make use of earlier iterations
325 -- of the fixpoint algorithm. (Cunning plan.)
326 -- Note that the cunning plan extends to the DmdEnv too,
327 -- since it is part of the strictness signature
328 initial_sig id = idNewStrictness_maybe id `orElse` botSig
330 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
331 lookup sigs var = case lookupVarEnv sigs var of
334 downRhs :: TopLevelFlag
335 -> SigEnv -> (Id, CoreExpr)
336 -> (SigEnv, DmdEnv, (Id, CoreExpr))
337 -- Process the RHS of the binding, add the strictness signature
338 -- to the Id, and augment the environment with the signature as well.
340 downRhs top_lvl sigs (id, rhs)
341 = (sigs', lazy_fv, (id', rhs'))
343 arity = exprArity rhs -- The idArity may not be up to date
344 (rhs_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
345 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_ty, ppr id )
347 id' = id `setIdNewStrictness` sig_ty
348 sigs' = extendSigEnv top_lvl sigs id sig_ty
351 %************************************************************************
353 \subsection{Strictness signatures and types}
355 %************************************************************************
358 mkSigTy :: CoreExpr -> DmdType -> (DmdEnv, StrictSig)
359 -- Take a DmdType and turn it into a StrictSig
360 mkSigTy rhs (DmdType fv dmds res)
361 = (lazy_fv, mkStrictSig dmd_ty)
363 dmd_ty = DmdType strict_fv final_dmds res'
365 lazy_fv = filterUFM (not . isStrictDmd) fv
366 strict_fv = filterUFM isStrictDmd fv
367 -- We put the strict FVs in the DmdType of the Id, so
368 -- that at its call sites we unleash demands on its strict fvs.
369 -- An example is 'roll' in imaginary/wheel-sieve2
370 -- Something like this:
372 -- go y = if ... then roll (x-1) else x+1
375 -- We want to see that roll is strict in x, which is because
376 -- go is called. So we put the DmdEnv for x in go's DmdType.
379 -- f :: Int -> Int -> Int
380 -- f x y = let t = x+1
381 -- h z = if z==0 then t else
382 -- if z==1 then x+1 else
386 -- Calling h does indeed evaluate x, but we can only see
387 -- that if we unleash a demand on x at the call site for t.
389 -- Incidentally, here's a place where lambda-lifting h would
390 -- lose the cigar --- we couldn't see the joint strictness in t/x
393 -- We don't want to put *all* the fv's from the RHS into the
394 -- DmdType, because that makes fixpointing very slow --- the
395 -- DmdType gets full of lazy demands that are slow to converge.
397 lazified_dmds = map funArgDemand dmds
398 -- Get rid of defers in the arguments
399 final_dmds = setUnpackStrategy lazified_dmds
400 -- Set the unpacking strategy
403 RetCPR | not (exprIsValue rhs) -> TopRes
405 -- If the rhs is a thunk, we forget the CPR info, because
406 -- it is presumably shared (else it would have been inlined, and
407 -- so we'd lose sharing if w/w'd it into a function.
409 -- DONE IN OLD CPR ANALYSER, BUT NOT YET HERE
410 -- Also, if the strictness analyser has figured out that it's strict,
411 -- the let-to-case transformation will happen, so again it's good.
412 -- (CPR analysis runs before the simplifier has had a chance to do
413 -- the let-to-case transform.)
414 -- This made a big difference to PrelBase.modInt, which had something like
415 -- modInt = \ x -> let r = ... -> I# v in
416 -- ...body strict in r...
417 -- r's RHS isn't a value yet; but modInt returns r in various branches, so
418 -- if r doesn't have the CPR property then neither does modInt
421 The unpack strategy determines whether we'll *really* unpack the argument,
422 or whether we'll just remember its strictness. If unpacking would give
423 rise to a *lot* of worker args, we may decide not to unpack after all.
426 setUnpackStrategy :: [Demand] -> [Demand]
428 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
430 go :: Int -- Max number of args available for sub-components of [Demand]
432 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
434 go n (Seq keep cs : ds)
435 | n' >= 0 = Seq keep cs' `cons` go n'' ds
436 | otherwise = Eval `cons` go n ds
439 n' = n + box - non_abs_args
442 Drop -> 1 -- Add one to the budget if we drop the top-level arg
443 non_abs_args = nonAbsentArgs cs
444 -- Delete # of non-absent args to which we'll now be committed
446 go n (d:ds) = d `cons` go n ds
449 cons d (n,ds) = (n, d:ds)
451 nonAbsentArgs :: [Demand] -> Int
453 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
454 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
458 %************************************************************************
460 \subsection{Strictness signatures and types}
462 %************************************************************************
465 splitDmdTy :: DmdType -> (Demand, DmdType)
466 -- Split off one function argument
467 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
468 splitDmdTy ty@(DmdType fv [] TopRes) = (Lazy, ty)
469 splitDmdTy ty@(DmdType fv [] BotRes) = (Bot, ty)
471 splitDmdTy (DmdType fv [] RetCPR) = panic "splitDmdTy"
472 -- We already have a suitable demand on all
473 -- free vars, so no need to add more!
477 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
479 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
480 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
481 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
483 addLazyFVs (DmdType fv ds res) lazy_fvs
484 = DmdType both_fv1 ds res
486 both_fv = (plusUFM_C both fv lazy_fvs)
487 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
488 -- This modifyEnv is vital. Consider
489 -- let f = \x -> (x,y)
491 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
492 -- demand with the bottom coming up from 'error'
494 -- I got a loop in the fixpointer without this, due to an interaction
495 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
497 -- = letrec g y = x `fatbar`
498 -- letrec h z = z + ...g...
501 -- In the initial iteration for f, f=Bot
502 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
503 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
504 -- places on its free variables. Suppose it places none. Then the
505 -- x `fatbar` ...call to h...
506 -- will give a x->V demand for x. That turns into a L demand for x,
507 -- which floats out of the defn for h. Without the modifyEnv, that
508 -- L demand doesn't get both'd with the Bot coming up from the inner
509 -- call to f. So we just get an L demand for x for g.
511 -- A better way to say this is that the lazy-fv filtering should give the
512 -- same answer as putting the lazy fv demands in the function's type.
514 annotateBndr :: DmdType -> Var -> (DmdType, Var)
515 -- The returned env has the var deleted
516 -- The returned var is annotated with demand info
517 -- No effect on the argument demands
518 annotateBndr dmd_ty@(DmdType fv ds res) var
519 | isTyVar var = (dmd_ty, var)
520 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var hacked_dmd)
522 (fv', dmd) = removeFV fv var res
523 hacked_dmd | isUnLiftedType (idType var) = unliftedDemand dmd
526 annotateBndrs = mapAccumR annotateBndr
528 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
529 -- For lambdas we add the demand to the argument demands
530 -- Only called for Ids
532 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
534 (fv', dmd) = removeFV fv id res
535 hacked_dmd | isUnLiftedType (idType id) = unliftedDemand dmd
536 | otherwise = funArgDemand dmd
537 -- This call to funArgDemand is vital, because otherwise we label
538 -- a lambda binder with demand 'B'. But in terms of calling
539 -- conventions that's Abs, because we don't pass it. But
540 -- when we do a w/w split we get
541 -- fw x = (\x y:B -> ...) x (error "oops")
542 -- And then the simplifier things the 'B' is a strict demand
543 -- and evaluates the (error "oops"). Sigh
545 removeFV fv var res = (fv', dmd)
547 fv' = fv `delVarEnv` var
548 dmd = lookupVarEnv fv var `orElse` deflt
549 deflt | isBotRes res = Bot
553 %************************************************************************
555 \subsection{Strictness signatures}
557 %************************************************************************
560 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
561 -- We use the SigEnv to tell us whether to
562 -- record info about a variable in the DmdEnv
563 -- We do so if it's a LocalId, but not top-level
565 -- The DmdEnv gives the demand on the free vars of the function
566 -- when it is given enough args to satisfy the strictness signature
568 emptySigEnv = emptyVarEnv
570 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
571 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
573 extendSigEnvList = extendVarEnvList
575 dmdTransform :: SigEnv -- The strictness environment
576 -> Id -- The function
577 -> Demand -- The demand on the function
578 -> DmdType -- The demand type of the function in this context
579 -- Returned DmdEnv includes the demand on
580 -- this function plus demand on its free variables
582 dmdTransform sigs var dmd
584 ------ DATA CONSTRUCTOR
585 | isDataConId var, -- Data constructor
586 Seq k ds <- res_dmd, -- and the demand looks inside its fields
587 let StrictSig dmd_ty = idNewStrictness var, -- It must have a strictness sig
588 let DmdType _ con_ds con_res = dmd_ty
589 = if length con_ds == length ds then -- Saturated, so unleash the demand
590 -- ds can be empty, when we are just seq'ing the thing
593 Keep -> zipWith lub ds con_ds
596 -- Important! If we Keep the constructor application, then
597 -- we need the demands the constructor places (usually lazy)
598 -- If not, we don't need to. For example:
599 -- f p@(x,y) = (p,y) -- S(AL)
601 -- It's vital that we don't calculate Absent for a!
603 mkDmdType emptyDmdEnv arg_ds con_res
604 -- Must remember whether it's a product, hence con_res, not TopRes
608 ------ IMPORTED FUNCTION
609 | isGlobalId var, -- Imported function
610 let StrictSig dmd_ty = getNewStrictness var
611 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
616 ------ LOCAL LET/REC BOUND THING
617 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
619 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
620 | otherwise = deferType dmd_ty
621 -- NB: it's important to use deferType, and not just return topDmdType
622 -- Consider let { f x y = p + x } in f 1
623 -- The application isn't saturated, but we must nevertheless propagate
624 -- a lazy demand for p!
626 addVarDmd top_lvl fn_ty var dmd
628 ------ LOCAL NON-LET/REC BOUND THING
629 | otherwise -- Default case
633 (call_depth, res_dmd) = splitCallDmd dmd
637 %************************************************************************
641 %************************************************************************
644 splitCallDmd :: Demand -> (Int, Demand)
645 splitCallDmd (Call d) = case splitCallDmd d of
647 splitCallDmd d = (0, d)
649 vanillaCall :: Arity -> Demand
651 vanillaCall n = Call (vanillaCall (n-1))
653 deferType :: DmdType -> DmdType
654 deferType (DmdType fv _ _) = DmdType (mapVarEnv defer fv) [] TopRes
655 -- Notice that we throw away info about both arguments and results
656 -- For example, f = let ... in \x -> x
657 -- We don't want to get a stricness type V->T for f.
660 bothLazy :: Demand -> Demand
662 bothLazy_s :: [Demand] -> [Demand]
663 bothLazy_s = map bothLazy
665 funArgDemand :: Demand -> Demand
666 -- The 'Defer' demands are just Lazy at function boundaries
667 -- Ugly! Ask John how to improve it.
668 funArgDemand (Seq Defer ds) = Lazy
669 funArgDemand (Seq k ds) = Seq k (map funArgDemand ds)
670 funArgDemand Err = Eval -- Args passed to a bottoming function
671 funArgDemand Bot = Abs -- Don't pass args that are consumed by bottom/err
674 unliftedDemand :: Demand -> Demand
675 -- Same idea, but for unlifted types the domain is much simpler:
676 -- Either we use it (Lazy) or we don't (Abs)
677 unliftedDemand Bot = Abs
678 unliftedDemand Abs = Abs
679 unliftedDemand other = Lazy
683 betterStrictness :: StrictSig -> StrictSig -> Bool
684 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
686 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
688 betterDemand :: Demand -> Demand -> Bool
689 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
690 betterDemand d1 d2 = (d1 `lub` d2) == d2
692 squashDmdEnv (StrictSig (DmdType fv ds res)) = StrictSig (DmdType emptyDmdEnv ds res)
696 -------------------------
697 -- Consider (if x then y else []) with demand V
698 -- Then the first branch gives {y->V} and the second
699 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
700 -- in the result env.
701 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
702 = DmdType lub_fv2 (zipWith lub ds1 ds2) (r1 `lubRes` r2)
704 lub_fv = plusUFM_C lub fv1 fv2
705 lub_fv1 = modifyEnv (not (isBotRes r1)) defer fv2 fv1 lub_fv
706 lub_fv2 = modifyEnv (not (isBotRes r2)) defer fv1 fv2 lub_fv1
707 -- lub is the identity for Bot
709 -----------------------------------
710 -- (t1 `bothType` t2) takes the argument/result info from t1,
711 -- using t2 just for its free-var info
712 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
713 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
715 both_fv = plusUFM_C both fv1 fv2
716 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
717 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
718 -- both is the identity for Abs
725 lubRes RetCPR RetCPR = RetCPR
726 lubRes r1 r2 = TopRes
728 -- If either diverges, the whole thing does
729 -- Otherwise take CPR info from the first
730 bothRes BotRes r2 = BotRes
731 bothRes r1 BotRes = BotRes
736 -- A Seq can have an empty list of demands, in the polymorphic case.
739 lubs ds1 ds2 = ASSERT( length ds1 == length ds2 ) zipWith lub ds1 ds2
741 -----------------------------------
742 -- A Seq can have an empty list of demands, in the polymorphic case.
745 boths ds1 ds2 = ASSERT( length ds1 == length ds2 ) zipWith both ds1 ds2
749 modifyEnv :: Bool -- No-op if False
750 -> (Demand -> Demand) -- The zapper
751 -> DmdEnv -> DmdEnv -- Env1 and Env2
752 -> DmdEnv -> DmdEnv -- Transform this env
753 -- Zap anything in Env1 but not in Env2
754 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
756 modifyEnv need_to_modify zapper env1 env2 env
757 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
760 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
762 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
766 %************************************************************************
768 \subsection{LUB and BOTH}
770 %************************************************************************
774 lub :: Demand -> Demand -> Demand
787 lub Eval (Seq Drop ds) | not (null ds) = Seq Drop [Lazy | d <- ds]
789 -- For the Seq case, consier
791 -- f n (x:xs) = f (n+x) xs
792 -- Here we want to do better than just V for n. It's
793 -- unboxed in the (x:xs) case, and we might be prepared to
794 -- rebox it in the [] case.
795 -- But if we don't use *any* of the components, give up
798 lub (Call d1) (Call d2) = Call (lub d1 d2)
799 lub d1@(Call _) d2 = d2 `lub` d1
801 lub (Seq k1 ds1) (Seq k2 ds2)
802 = Seq (k1 `lub_keep` k2) (lub_ds k1 ds1 k2 ds2)
805 lub_ds Keep ds1 Keep ds2 = ds1 `lubs` ds2
806 lub_ds Keep ds1 non_keep ds2 | null ds1 = [Lazy | d <- ds2]
807 | otherwise = bothLazy_s ds1 `lubs` ds2
809 lub_ds non_keep ds1 Keep ds2 | null ds2 = [Lazy | d <- ds1]
810 | otherwise = ds1 `lubs` bothLazy_s ds2
812 lub_ds k1 ds1 k2 ds2 = ds1 `lubs` ds2
817 lub_keep Drop Defer = Defer
818 lub_keep Drop k = Drop
820 lub_keep Defer k = Defer
822 lub d1@(Seq _ _) d2 = d2 `lub` d1
825 both :: Demand -> Demand -> Demand
837 both Lazy Eval = Eval
838 both Lazy (Call d) = Call d
839 both Lazy (Seq Defer ds) = Lazy
840 both Lazy (Seq k ds) = Seq Keep ds
843 -- For the (Eval `both` Bot) case, consider
845 -- From 'error' itself we get demand Bot on x
846 -- From the arg demand on x we get Eval
847 -- So we want Eval `both` Bot to be Err.
848 -- That's what Err is *for*
851 both Eval (Seq k ds) = Seq Keep ds
854 both (Call d1) (Call d2) = Call (d1 `both` d2)
855 both d1@(Call _) d2 = d2 `both` d1
857 both (Seq k1 ds1) (Seq k2 ds2)
858 = Seq (k1 `both_keep` k2) (both_ds k1 ds1 k2 ds2)
861 both_keep Keep k2 = Keep
863 both_keep Drop Keep = Keep
864 both_keep Drop k2 = Drop
866 both_keep Defer k2 = k2
869 both_ds Defer ds1 Defer ds2 = ds1 `boths` ds2
870 both_ds Defer ds1 non_defer ds2 = map defer ds1 `boths` ds2
872 both_ds non_defer ds1 Defer ds2 = ds1 `boths` map defer ds2
874 both_ds k1 ds1 k2 ds2 = ds1 `boths` ds2
876 both d1@(Seq _ _) d2 = d2 `both` d1
880 %************************************************************************
882 \subsection{Miscellaneous
884 %************************************************************************
888 get_changes binds = vcat (map get_changes_bind binds)
890 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
891 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
893 get_changes_pr (id,rhs)
894 | isImplicitId id = empty -- We don't look inside these
895 | otherwise = get_changes_var id $$ get_changes_expr rhs
898 | isId var = get_changes_str var $$ get_changes_dmd var
901 get_changes_expr (Type t) = empty
902 get_changes_expr (Var v) = empty
903 get_changes_expr (Lit l) = empty
904 get_changes_expr (Note n e) = get_changes_expr e
905 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
906 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
907 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
908 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
910 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
913 | new_better && old_better = empty
914 | new_better = message "BETTER"
915 | old_better = message "WORSE"
916 | otherwise = message "INCOMPARABLE"
918 message word = text word <+> text "strictness for" <+> ppr id <+> info
919 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
920 new = squashDmdEnv (idNewStrictness id) -- Don't report diffs in the env
921 old = newStrictnessFromOld id
922 old_better = old `betterStrictness` new
923 new_better = new `betterStrictness` old
926 | isUnLiftedType (idType id) = empty -- Not useful
927 | new_better && old_better = empty
928 | new_better = message "BETTER"
929 | old_better = message "WORSE"
930 | otherwise = message "INCOMPARABLE"
932 message word = text word <+> text "demand for" <+> ppr id <+> info
933 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
934 new = funArgDemand (idNewDemandInfo id) -- FunArgDemand to avoid spurious improvements
935 old = newDemand (idDemandInfo id)
936 new_better = new `betterDemand` old
937 old_better = old `betterDemand` new