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, idInlinePragma,
24 isDataConId, isGlobalId, idArity,
26 idDemandInfo, idStrictness, idCprInfo, idName,
28 idNewStrictness, idNewStrictness_maybe,
29 setIdNewStrictness, idNewDemandInfo,
30 idNewDemandInfo_maybe,
34 import IdInfo ( newStrictnessFromOld, newDemand )
38 import UniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly,
39 keysUFM, minusUFM, ufmToList, filterUFM )
40 import Type ( isUnLiftedType )
41 import CoreLint ( showPass, endPass )
42 import Util ( mapAndUnzip, mapAccumL, mapAccumR, lengthIs )
43 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive,
45 import Maybes ( orElse, expectJust )
51 * set a noinline pragma on bottoming Ids
53 * Consider f x = x+1 `fatbar` error (show x)
54 We'd like to unbox x, even if that means reboxing it in the error case.
57 instance Outputable TopLevelFlag where
61 %************************************************************************
63 \subsection{Top level stuff}
65 %************************************************************************
68 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
69 dmdAnalPgm dflags binds
71 showPass dflags "Demand analysis" ;
72 let { binds_plus_dmds = do_prog binds } ;
74 endPass dflags "Demand analysis"
75 Opt_D_dump_stranal binds_plus_dmds ;
77 -- Only if OLD_STRICTNESS is on, because only then is the old
78 -- strictness analyser run
79 let { dmd_changes = get_changes binds_plus_dmds } ;
80 printDump (text "Changes in demands" $$ dmd_changes) ;
82 return binds_plus_dmds
85 do_prog :: [CoreBind] -> [CoreBind]
86 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
88 dmdAnalTopBind :: SigEnv
91 dmdAnalTopBind sigs (NonRec id rhs)
93 ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs)
94 (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs1)
95 -- Do two passes to improve CPR information
96 -- See comments with ignore_cpr_info in mk_sig_ty
97 -- and with extendSigsWithLam
99 (sigs2, NonRec id2 rhs2)
101 dmdAnalTopBind sigs (Rec pairs)
103 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
104 -- We get two iterations automatically
105 -- c.f. the NonRec case above
111 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
112 -- Analyse the RHS and return
113 -- a) appropriate strictness info
114 -- b) the unfolding (decorated with stricntess info)
118 arity = exprArity rhs
119 (rhs_ty, rhs') = dmdAnal emptySigEnv (vanillaCall arity) rhs
120 sig = mkTopSigTy rhs rhs_ty
123 %************************************************************************
125 \subsection{The analyser itself}
127 %************************************************************************
130 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
132 dmdAnal sigs Abs e = (topDmdType, e)
135 | not (isStrictDmd dmd)
137 (res_ty, e') = dmdAnal sigs evalDmd e
139 (deferType res_ty, e')
140 -- It's important not to analyse e with a lazy demand because
141 -- a) When we encounter case s of (a,b) ->
142 -- we demand s with U(d1d2)... but if the overall demand is lazy
143 -- that is wrong, and we'd need to reduce the demand on s,
144 -- which is inconvenient
145 -- b) More important, consider
146 -- f (let x = R in x+x), where f is lazy
147 -- We still want to mark x as demanded, because it will be when we
148 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
149 -- just mark x as Lazy
150 -- c) The application rule wouldn't be right either
151 -- Evaluating (f x) in a L demand does *not* cause
152 -- evaluation of f in a C(L) demand!
155 dmdAnal sigs dmd (Lit lit)
156 = (topDmdType, Lit lit)
158 dmdAnal sigs dmd (Var var)
159 = (dmdTransform sigs var dmd, Var var)
161 dmdAnal sigs dmd (Note n e)
162 = (dmd_ty, Note n e')
164 (dmd_ty, e') = dmdAnal sigs dmd' e
166 Coerce _ _ -> evalDmd -- This coerce usually arises from a recursive
167 other -> dmd -- newtype, and we don't want to look inside them
168 -- for exactly the same reason that we don't look
169 -- inside recursive products -- we might not reach
170 -- a fixpoint. So revert to a vanilla Eval demand
172 dmdAnal sigs dmd (App fun (Type ty))
173 = (fun_ty, App fun' (Type ty))
175 (fun_ty, fun') = dmdAnal sigs dmd fun
177 -- Lots of the other code is there to make this
178 -- beautiful, compositional, application rule :-)
179 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
180 = let -- [Type arg handled above]
181 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
182 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
183 (arg_dmd, res_ty) = splitDmdTy fun_ty
185 (res_ty `bothType` arg_ty, App fun' arg')
187 dmdAnal sigs dmd (Lam var body)
190 (body_ty, body') = dmdAnal sigs dmd body
192 (body_ty, Lam var body')
194 | Call body_dmd <- dmd -- A call demand: good!
196 sigs' = extendSigsWithLam sigs var
197 (body_ty, body') = dmdAnal sigs' body_dmd body
198 (lam_ty, var') = annotateLamIdBndr body_ty var
200 (lam_ty, Lam var' body')
202 | otherwise -- Not enough demand on the lambda; but do the body
203 = let -- anyway to annotate it and gather free var info
204 (body_ty, body') = dmdAnal sigs evalDmd body
205 (lam_ty, var') = annotateLamIdBndr body_ty var
207 (deferType lam_ty, Lam var' body')
209 dmdAnal sigs dmd (Case scrut case_bndr [alt@(DataAlt dc,bndrs,rhs)])
210 | let tycon = dataConTyCon dc,
211 isProductTyCon tycon,
212 not (isRecursiveTyCon tycon)
214 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
215 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
216 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
217 (_, bndrs', _) = alt'
218 case_bndr_sig = cprSig
219 -- Inside the alternative, the case binder has the CPR property.
220 -- Meaning that a case on it will successfully cancel.
222 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
225 -- We want f to have the CPR property:
226 -- f b x = case fw b x of { r -> I# r }
227 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
230 -- Figure out whether the demand on the case binder is used, and use
231 -- that to set the scrut_dmd. This is utterly essential.
232 -- Consider f x = case x of y { (a,b) -> k y a }
233 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
234 -- worker, so the worker will rebuild
235 -- x = (a, absent-error)
236 -- and that'll crash.
237 -- So at one stage I had:
238 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
239 -- keepity | dead_case_bndr = Drop
240 -- | otherwise = Keep
243 -- case x of y { (a,b) -> h y + a }
244 -- where h : U(LL) -> T
245 -- The above code would compute a Keep for x, since y is not Abs, which is silly
246 -- The insight is, of course, that a demand on y is a demand on the
247 -- scrutinee, so we need to `both` it with the scrut demand
249 scrut_dmd = Eval (Prod [idNewDemandInfo b | b <- bndrs', isId b])
251 idNewDemandInfo case_bndr'
253 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
255 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' [alt'])
257 dmdAnal sigs dmd (Case scrut case_bndr alts)
259 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
260 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
261 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
263 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
264 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' alts')
266 dmdAnal sigs dmd (Let (NonRec id rhs) body)
268 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
269 (body_ty, body') = dmdAnal sigs' dmd body
270 (body_ty1, id2) = annotateBndr body_ty id1
271 body_ty2 = addLazyFVs body_ty1 lazy_fv
273 -- If the actual demand is better than the vanilla call
274 -- demand, you might think that we might do better to re-analyse
275 -- the RHS with the stronger demand.
276 -- But (a) That seldom happens, because it means that *every* path in
277 -- the body of the let has to use that stronger demand
278 -- (b) It often happens temporarily in when fixpointing, because
279 -- the recursive function at first seems to place a massive demand.
280 -- But we don't want to go to extra work when the function will
281 -- probably iterate to something less demanding.
282 -- In practice, all the times the actual demand on id2 is more than
283 -- the vanilla call demand seem to be due to (b). So we don't
284 -- bother to re-analyse the RHS.
285 (body_ty2, Let (NonRec id2 rhs') body')
287 dmdAnal sigs dmd (Let (Rec pairs) body)
289 bndrs = map fst pairs
290 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
291 (body_ty, body') = dmdAnal sigs' dmd body
292 body_ty1 = addLazyFVs body_ty lazy_fv
294 sigs' `seq` body_ty `seq`
296 (body_ty2, _) = annotateBndrs body_ty1 bndrs
297 -- Don't bother to add demand info to recursive
298 -- binders as annotateBndr does;
299 -- being recursive, we can't treat them strictly.
300 -- But we do need to remove the binders from the result demand env
302 (body_ty2, Let (Rec pairs') body')
305 dmdAnalAlt sigs dmd (con,bndrs,rhs)
307 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
308 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
310 (alt_ty, (con, bndrs', rhs'))
313 %************************************************************************
315 \subsection{Bindings}
317 %************************************************************************
320 dmdFix :: TopLevelFlag
321 -> SigEnv -- Does not include bindings for this binding
324 [(Id,CoreExpr)]) -- Binders annotated with stricness info
326 dmdFix top_lvl sigs orig_pairs
327 = loop 1 initial_sigs orig_pairs
329 bndrs = map fst orig_pairs
330 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
333 -> SigEnv -- Already contains the current sigs
335 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
338 = (sigs', lazy_fv, pairs')
339 -- Note: use pairs', not pairs. pairs' is the result of
340 -- processing the RHSs with sigs (= sigs'), whereas pairs
341 -- is the result of processing the RHSs with the *previous*
342 -- iteration of sigs.
344 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
345 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
346 text "env:" <+> ppr (ufmToList sigs),
347 text "binds:" <+> pprCoreBinding (Rec pairs)]))
348 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
349 -- The lazy_fv part is really important! orig_pairs has no strictness
350 -- info, including nothing about free vars. But if we have
351 -- letrec f = ....y..... in ...f...
352 -- where 'y' is free in f, we must record that y is mentioned,
353 -- otherwise y will get recorded as absent altogether
355 | otherwise = loop (n+1) sigs' pairs'
357 found_fixpoint = all (same_sig sigs sigs') bndrs
358 -- Use the new signature to do the next pair
359 -- The occurrence analyser has arranged them in a good order
360 -- so this can significantly reduce the number of iterations needed
361 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
363 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
364 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
366 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
367 ((sigs', lazy_fv'), pair')
370 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
371 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
372 -- old_sig = lookup sigs id
373 -- new_sig = lookup sigs' id
375 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
376 lookup sigs var = case lookupVarEnv sigs var of
379 -- Get an initial strictness signature from the Id
380 -- itself. That way we make use of earlier iterations
381 -- of the fixpoint algorithm. (Cunning plan.)
382 -- Note that the cunning plan extends to the DmdEnv too,
383 -- since it is part of the strictness signature
384 initialSig id = idNewStrictness_maybe id `orElse` botSig
386 dmdAnalRhs :: TopLevelFlag -> RecFlag
387 -> SigEnv -> (Id, CoreExpr)
388 -> (SigEnv, DmdEnv, (Id, CoreExpr))
389 -- Process the RHS of the binding, add the strictness signature
390 -- to the Id, and augment the environment with the signature as well.
392 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
393 = (sigs', lazy_fv, (id', rhs'))
395 arity = idArity id -- The idArity should be up to date
396 -- The simplifier was run just beforehand
397 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
398 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty, ppr id )
399 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
400 id' = id `setIdNewStrictness` sig_ty
401 sigs' = extendSigEnv top_lvl sigs id sig_ty
404 %************************************************************************
406 \subsection{Strictness signatures and types}
408 %************************************************************************
411 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
412 -- Take a DmdType and turn it into a StrictSig
413 -- NB: not used for never-inline things; hence False
414 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
416 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
417 mkSigTy top_lvl rec_flag id rhs dmd_ty
418 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
420 never_inline = isNeverActive (idInlinePragma id)
421 maybe_id_dmd = idNewDemandInfo_maybe id
422 -- Is Nothing the first time round
425 | isTopLevel top_lvl = False -- Top level things don't get
426 -- their demandInfo set at all
427 | isRec rec_flag = False -- Ditto recursive things
428 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
429 | otherwise = True -- Optimistic, first time round
433 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
434 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
435 If the rhs is a thunk, we usually forget the CPR info, because
436 it is presumably shared (else it would have been inlined, and
437 so we'd lose sharing if w/w'd it into a function.
439 However, if the strictness analyser has figured out (in a previous
440 iteration) that it's strict, then we DON'T need to forget the CPR info.
441 Instead we can retain the CPR info and do the thunk-splitting transform
442 (see WorkWrap.splitThunk).
444 This made a big difference to PrelBase.modInt, which had something like
445 modInt = \ x -> let r = ... -> I# v in
446 ...body strict in r...
447 r's RHS isn't a value yet; but modInt returns r in various branches, so
448 if r doesn't have the CPR property then neither does modInt
449 Another case I found in practice (in Complex.magnitude), looks like this:
450 let k = if ... then I# a else I# b
451 in ... body strict in k ....
452 (For this example, it doesn't matter whether k is returned as part of
453 the overall result; but it does matter that k's RHS has the CPR property.)
454 Left to itself, the simplifier will make a join point thus:
455 let $j k = ...body strict in k...
456 if ... then $j (I# a) else $j (I# b)
457 With thunk-splitting, we get instead
458 let $j x = let k = I#x in ...body strict in k...
459 in if ... then $j a else $j b
460 This is much better; there's a good chance the I# won't get allocated.
462 The difficulty with this is that we need the strictness type to
463 look at the body... but we now need the body to calculate the demand
464 on the variable, so we can decide whether its strictness type should
465 have a CPR in it or not. Simple solution:
466 a) use strictness info from the previous iteration
467 b) make sure we do at least 2 iterations, by doing a second
468 round for top-level non-recs. Top level recs will get at
469 least 2 iterations except for totally-bottom functions
470 which aren't very interesting anyway.
472 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
474 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
475 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
476 Demand info now has a 'Nothing' state, just like strictness info.
477 The analysis works from 'dangerous' towards a 'safe' state; so we
478 start with botSig for 'Nothing' strictness infos, and we start with
479 "yes, it's demanded" for 'Nothing' in the demand info. The
480 fixpoint iteration will sort it all out.
482 We can't start with 'not-demanded' because then consider
486 if ... then t else I# y else f x'
488 In the first iteration we'd have no demand info for x, so assume
489 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
490 we'd see that t was demanded, and so give it the CPR property, but by
491 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
492 setting the first time round, we say 'yes t is demanded' the first
495 However, this does mean that for non-recursive bindings we must
496 iterate twice to be sure of not getting over-optimistic CPR info,
497 in the case where t turns out to be not-demanded. This is handled
502 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
503 | never_inline && not (isBotRes res)
505 -- Don't strictness-analyse NOINLINE things. Why not? Because
506 -- the NOINLINE says "don't expose any of the inner workings at the call
507 -- site" and the strictness is certainly an inner working.
509 -- More concretely, the demand analyser discovers the following strictness
510 -- for unsafePerformIO: C(U(AV))
512 -- unsafePerformIO (\s -> let r = f x in
513 -- case writeIORef v r s of (# s1, _ #) ->
515 -- The strictness analyser will find that the binding for r is strict,
516 -- (becuase of uPIO's strictness sig), and so it'll evaluate it before
517 -- doing the writeIORef. This actually makes tests/lib/should_run/memo002
520 -- Solution: don't expose the strictness of unsafePerformIO.
522 -- But we do want to expose the strictness of error functions,
523 -- which are also often marked NOINLINE
524 -- {-# NOINLINE foo #-}
525 -- foo x = error ("wubble buggle" ++ x)
526 -- So (hack, hack) we only drop the strictness for non-bottom things
527 -- This is all very unsatisfactory.
528 = (deferEnv fv, topSig)
531 = (lazy_fv, mkStrictSig dmd_ty)
533 dmd_ty = DmdType strict_fv final_dmds res'
535 lazy_fv = filterUFM (not . isStrictDmd) fv
536 strict_fv = filterUFM isStrictDmd fv
537 -- We put the strict FVs in the DmdType of the Id, so
538 -- that at its call sites we unleash demands on its strict fvs.
539 -- An example is 'roll' in imaginary/wheel-sieve2
540 -- Something like this:
542 -- go y = if ... then roll (x-1) else x+1
545 -- We want to see that roll is strict in x, which is because
546 -- go is called. So we put the DmdEnv for x in go's DmdType.
549 -- f :: Int -> Int -> Int
550 -- f x y = let t = x+1
551 -- h z = if z==0 then t else
552 -- if z==1 then x+1 else
556 -- Calling h does indeed evaluate x, but we can only see
557 -- that if we unleash a demand on x at the call site for t.
559 -- Incidentally, here's a place where lambda-lifting h would
560 -- lose the cigar --- we couldn't see the joint strictness in t/x
563 -- We don't want to put *all* the fv's from the RHS into the
564 -- DmdType, because that makes fixpointing very slow --- the
565 -- DmdType gets full of lazy demands that are slow to converge.
567 final_dmds = setUnpackStrategy dmds
568 -- Set the unpacking strategy
571 RetCPR | ignore_cpr_info -> TopRes
573 ignore_cpr_info = not (exprIsValue rhs || thunk_cpr_ok)
576 The unpack strategy determines whether we'll *really* unpack the argument,
577 or whether we'll just remember its strictness. If unpacking would give
578 rise to a *lot* of worker args, we may decide not to unpack after all.
581 setUnpackStrategy :: [Demand] -> [Demand]
583 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
585 go :: Int -- Max number of args available for sub-components of [Demand]
587 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
589 go n (Eval (Prod cs) : ds)
590 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
591 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
594 n' = n + 1 - non_abs_args
595 -- Add one to the budget 'cos we drop the top-level arg
596 non_abs_args = nonAbsentArgs cs
597 -- Delete # of non-absent args to which we'll now be committed
599 go n (d:ds) = d `cons` go n ds
602 cons d (n,ds) = (n, d:ds)
604 nonAbsentArgs :: [Demand] -> Int
606 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
607 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
611 %************************************************************************
613 \subsection{Strictness signatures and types}
615 %************************************************************************
618 splitDmdTy :: DmdType -> (Demand, DmdType)
619 -- Split off one function argument
620 -- We already have a suitable demand on all
621 -- free vars, so no need to add more!
622 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
623 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
627 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
629 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
630 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
631 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
633 addLazyFVs (DmdType fv ds res) lazy_fvs
634 = DmdType both_fv1 ds res
636 both_fv = (plusUFM_C both fv lazy_fvs)
637 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
638 -- This modifyEnv is vital. Consider
639 -- let f = \x -> (x,y)
641 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
642 -- demand with the bottom coming up from 'error'
644 -- I got a loop in the fixpointer without this, due to an interaction
645 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
647 -- = letrec g y = x `fatbar`
648 -- letrec h z = z + ...g...
651 -- In the initial iteration for f, f=Bot
652 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
653 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
654 -- places on its free variables. Suppose it places none. Then the
655 -- x `fatbar` ...call to h...
656 -- will give a x->V demand for x. That turns into a L demand for x,
657 -- which floats out of the defn for h. Without the modifyEnv, that
658 -- L demand doesn't get both'd with the Bot coming up from the inner
659 -- call to f. So we just get an L demand for x for g.
661 -- A better way to say this is that the lazy-fv filtering should give the
662 -- same answer as putting the lazy fv demands in the function's type.
664 annotateBndr :: DmdType -> Var -> (DmdType, Var)
665 -- The returned env has the var deleted
666 -- The returned var is annotated with demand info
667 -- No effect on the argument demands
668 annotateBndr dmd_ty@(DmdType fv ds res) var
669 | isTyVar var = (dmd_ty, var)
670 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
672 (fv', dmd) = removeFV fv var res
674 annotateBndrs = mapAccumR annotateBndr
676 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
677 -- For lambdas we add the demand to the argument demands
678 -- Only called for Ids
680 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
682 (fv', dmd) = removeFV fv id res
683 hacked_dmd = argDemand dmd
684 -- This call to argDemand is vital, because otherwise we label
685 -- a lambda binder with demand 'B'. But in terms of calling
686 -- conventions that's Abs, because we don't pass it. But
687 -- when we do a w/w split we get
688 -- fw x = (\x y:B -> ...) x (error "oops")
689 -- And then the simplifier things the 'B' is a strict demand
690 -- and evaluates the (error "oops"). Sigh
692 removeFV fv id res = (fv', zapUnlifted id dmd)
694 fv' = fv `delVarEnv` id
695 dmd = lookupVarEnv fv id `orElse` deflt
696 deflt | isBotRes res = Bot
699 -- For unlifted-type variables, we are only
700 -- interested in Bot/Abs/Box Abs
701 zapUnlifted is Bot = Bot
702 zapUnlifted id Abs = Abs
703 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
707 %************************************************************************
709 \subsection{Strictness signatures}
711 %************************************************************************
714 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
715 -- We use the SigEnv to tell us whether to
716 -- record info about a variable in the DmdEnv
717 -- We do so if it's a LocalId, but not top-level
719 -- The DmdEnv gives the demand on the free vars of the function
720 -- when it is given enough args to satisfy the strictness signature
722 emptySigEnv = emptyVarEnv
724 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
725 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
727 extendSigEnvList = extendVarEnvList
729 extendSigsWithLam :: SigEnv -> Id -> SigEnv
730 -- Extend the SigEnv when we meet a lambda binder
731 -- If the binder is marked demanded with a product demand, then give it a CPR
732 -- signature, because in the likely event that this is a lambda on a fn defn
733 -- [we only use this when the lambda is being consumed with a call demand],
734 -- it'll be w/w'd and so it will be CPR-ish. E.g.
735 -- f = \x::(Int,Int). if ...strict in x... then
739 -- We want f to have the CPR property because x does, by the time f has been w/w'd
741 -- NOTE: see notes [CPR-AND-STRICTNESS]
743 -- Also note that we only want to do this for something that
744 -- definitely has product type, else we may get over-optimistic
745 -- CPR results (e.g. from \x -> x!).
747 extendSigsWithLam sigs id
748 = case idNewDemandInfo_maybe id of
749 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
750 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
754 dmdTransform :: SigEnv -- The strictness environment
755 -> Id -- The function
756 -> Demand -- The demand on the function
757 -> DmdType -- The demand type of the function in this context
758 -- Returned DmdEnv includes the demand on
759 -- this function plus demand on its free variables
761 dmdTransform sigs var dmd
763 ------ DATA CONSTRUCTOR
764 | isDataConId var -- Data constructor
766 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
767 DmdType _ _ con_res = dmd_ty
770 if arity == call_depth then -- Saturated, so unleash the demand
772 -- Important! If we Keep the constructor application, then
773 -- we need the demands the constructor places (always lazy)
774 -- If not, we don't need to. For example:
775 -- f p@(x,y) = (p,y) -- S(AL)
777 -- It's vital that we don't calculate Absent for a!
778 dmd_ds = case res_dmd of
779 Box (Eval ds) -> mapDmds box ds
783 -- ds can be empty, when we are just seq'ing the thing
784 -- If so we must make up a suitable bunch of demands
785 arg_ds = case dmd_ds of
786 Poly d -> replicate arity d
787 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
790 mkDmdType emptyDmdEnv arg_ds con_res
791 -- Must remember whether it's a product, hence con_res, not TopRes
795 ------ IMPORTED FUNCTION
796 | isGlobalId var, -- Imported function
797 let StrictSig dmd_ty = idNewStrictness var
798 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
803 ------ LOCAL LET/REC BOUND THING
804 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
806 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
807 | otherwise = deferType dmd_ty
808 -- NB: it's important to use deferType, and not just return topDmdType
809 -- Consider let { f x y = p + x } in f 1
810 -- The application isn't saturated, but we must nevertheless propagate
811 -- a lazy demand for p!
813 addVarDmd top_lvl fn_ty var dmd
815 ------ LOCAL NON-LET/REC BOUND THING
816 | otherwise -- Default case
820 (call_depth, res_dmd) = splitCallDmd dmd
824 %************************************************************************
828 %************************************************************************
831 splitCallDmd :: Demand -> (Int, Demand)
832 splitCallDmd (Call d) = case splitCallDmd d of
834 splitCallDmd d = (0, d)
836 vanillaCall :: Arity -> Demand
837 vanillaCall 0 = evalDmd
838 vanillaCall n = Call (vanillaCall (n-1))
840 deferType :: DmdType -> DmdType
841 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
842 -- Notice that we throw away info about both arguments and results
843 -- For example, f = let ... in \x -> x
844 -- We don't want to get a stricness type V->T for f.
847 deferEnv :: DmdEnv -> DmdEnv
848 deferEnv fv = mapVarEnv defer fv
852 argDemand :: Demand -> Demand
853 -- The 'Defer' demands are just Lazy at function boundaries
854 -- Ugly! Ask John how to improve it.
855 argDemand Top = lazyDmd
856 argDemand (Defer d) = lazyDmd
857 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
858 argDemand (Box Bot) = evalDmd
859 argDemand (Box d) = box (argDemand d)
860 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
865 betterStrictness :: StrictSig -> StrictSig -> Bool
866 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
868 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
870 betterDemand :: Demand -> Demand -> Bool
871 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
872 betterDemand d1 d2 = (d1 `lub` d2) == d2
876 -------------------------
877 -- Consider (if x then y else []) with demand V
878 -- Then the first branch gives {y->V} and the second
879 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
880 -- in the result env.
881 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
882 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
884 lub_fv = plusUFM_C lub fv1 fv2
885 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
886 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
887 -- lub is the identity for Bot
889 -- Extend the shorter argument list to match the longer
890 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
892 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
893 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
895 -----------------------------------
896 -- (t1 `bothType` t2) takes the argument/result info from t1,
897 -- using t2 just for its free-var info
898 -- NB: Don't forget about r2! It might be BotRes, which is
899 -- a bottom demand on all the in-scope variables.
900 -- Peter: can this be done more neatly?
901 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
902 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
904 both_fv = plusUFM_C both fv1 fv2
905 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
906 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
907 -- both is the identity for Abs
914 lubRes RetCPR RetCPR = RetCPR
915 lubRes r1 r2 = TopRes
917 -- If either diverges, the whole thing does
918 -- Otherwise take CPR info from the first
919 bothRes r1 BotRes = BotRes
924 modifyEnv :: Bool -- No-op if False
925 -> (Demand -> Demand) -- The zapper
926 -> DmdEnv -> DmdEnv -- Env1 and Env2
927 -> DmdEnv -> DmdEnv -- Transform this env
928 -- Zap anything in Env1 but not in Env2
929 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
931 modifyEnv need_to_modify zapper env1 env2 env
932 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
935 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
937 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
941 %************************************************************************
943 \subsection{LUB and BOTH}
945 %************************************************************************
948 lub :: Demand -> Demand -> Demand
951 lub Abs d2 = absLub d2
953 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
955 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
956 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
957 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
958 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
960 -- For the Eval case, we use these approximation rules
961 -- Box Bot <= Eval (Box Bot ...)
962 -- Box Top <= Defer (Box Bot ...)
963 -- Box (Eval ds) <= Eval (map Box ds)
964 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
965 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
966 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
967 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
968 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
970 lub (Box d1) (Box d2) = box (d1 `lub` d2)
971 lub d1@(Box _) d2 = d2 `lub` d1
973 lubs = zipWithDmds lub
975 ---------------------
976 -- box is the smart constructor for Box
977 -- It computes <B,bot> & d
978 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
979 -- Seems to be no point in allowing (Box (Call d))
980 box (Call d) = Call d -- The odd man out. Why?
982 box (Defer _) = lazyDmd
983 box Top = lazyDmd -- Box Abs and Box Top
984 box Abs = lazyDmd -- are the same <B,L>
985 box d = Box d -- Bot, Eval
988 defer :: Demand -> Demand
990 -- defer is the smart constructor for Defer
991 -- The idea is that (Defer ds) = <U(ds), L>
993 -- It specifies what happens at a lazy function argument
994 -- or a lambda; the L* operator
995 -- Set the strictness part to L, but leave
996 -- the boxity side unaffected
997 -- It also ensures that Defer (Eval [LLLL]) = L
1002 defer (Call _) = lazyDmd -- Approximation here?
1003 defer (Box _) = lazyDmd
1004 defer (Defer ds) = Defer ds
1005 defer (Eval ds) = deferEval ds
1007 -- deferEval ds = defer (Eval ds)
1008 deferEval ds | allTop ds = Top
1009 | otherwise = Defer ds
1011 ---------------------
1012 absLub :: Demand -> Demand
1013 -- Computes (Abs `lub` d)
1014 -- For the Bot case consider
1015 -- f x y = if ... then x else error x
1016 -- Then for y we get Abs `lub` Bot, and we really
1021 absLub (Call _) = Top
1022 absLub (Box _) = Top
1023 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1024 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1026 absLubs = mapDmds absLub
1029 both :: Demand -> Demand -> Demand
1035 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1038 -- From 'error' itself we get demand Bot on x
1039 -- From the arg demand on x we get
1040 -- x :-> evalDmd = Box (Eval (Poly Abs))
1041 -- So we get Bot `both` Box (Eval (Poly Abs))
1042 -- = Seq Keep (Poly Bot)
1045 -- f x = if ... then error (fst x) else fst x
1046 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1048 -- which is what we want.
1051 both Top Bot = errDmd
1054 both Top (Box d) = Box d
1055 both Top (Call d) = Call d
1056 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1057 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1058 -- = defer (Eval (mapDmds (`both` Top) ds))
1059 = deferEval (mapDmds (`both` Top) ds)
1062 both (Box d1) (Box d2) = box (d1 `both` d2)
1063 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1064 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1065 both (Box d1) (Defer d2) = Box d1
1066 both d1@(Box _) d2 = d2 `both` d1
1068 both (Call d1) (Call d2) = Call (d1 `both` d2)
1069 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1070 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1071 both d1@(Call _) d2 = d1 `both` d1
1073 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1074 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1075 both d1@(Eval ds1) d2 = d2 `both` d1
1077 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1078 both d1@(Defer ds1) d2 = d2 `both` d1
1080 boths = zipWithDmds both
1085 %************************************************************************
1087 \subsection{Miscellaneous
1089 %************************************************************************
1093 #ifdef OLD_STRICTNESS
1094 get_changes binds = vcat (map get_changes_bind binds)
1096 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1097 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1099 get_changes_pr (id,rhs)
1100 = get_changes_var id $$ get_changes_expr rhs
1103 | isId var = get_changes_str var $$ get_changes_dmd var
1106 get_changes_expr (Type t) = empty
1107 get_changes_expr (Var v) = empty
1108 get_changes_expr (Lit l) = empty
1109 get_changes_expr (Note n e) = get_changes_expr e
1110 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1111 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1112 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1113 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1115 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1118 | new_better && old_better = empty
1119 | new_better = message "BETTER"
1120 | old_better = message "WORSE"
1121 | otherwise = message "INCOMPARABLE"
1123 message word = text word <+> text "strictness for" <+> ppr id <+> info
1124 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1125 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1126 -- strictness analyser can't track
1127 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1128 old_better = old `betterStrictness` new
1129 new_better = new `betterStrictness` old
1132 | isUnLiftedType (idType id) = empty -- Not useful
1133 | new_better && old_better = empty
1134 | new_better = message "BETTER"
1135 | old_better = message "WORSE"
1136 | otherwise = message "INCOMPARABLE"
1138 message word = text word <+> text "demand for" <+> ppr id <+> info
1139 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1140 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1142 old = newDemand (idDemandInfo id)
1143 new_better = new `betterDemand` old
1144 old_better = old `betterDemand` new
1147 squashSig (StrictSig (DmdType fv ds res))
1148 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1150 -- squash just gets rid of call demands
1151 -- which the old analyser doesn't track
1152 squashDmd (Call d) = evalDmd
1153 squashDmd (Box d) = Box (squashDmd d)
1154 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1155 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)