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,
28 idNewStrictness, idNewStrictness_maybe,
29 setIdNewStrictness, idNewDemandInfo,
30 idNewDemandInfo_maybe,
31 setIdNewDemandInfo, idName
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
274 -- If the actual demand is better than the vanilla
275 -- demand, we might do better to re-analyse with the
277 (let vanilla_dmd = vanillaCall (idArity id)
278 actual_dmd = idNewDemandInfo id2
280 if actual_dmd `betterDemand` vanilla_dmd && actual_dmd /= vanilla_dmd then
281 pprTrace "dmdLet: better demand" (ppr id <+> vcat [text "vanilla" <+> ppr vanilla_dmd,
282 text "actual" <+> ppr actual_dmd])
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)])
337 | all (same_sig sigs sigs') bndrs
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.
343 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
344 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
345 text "env:" <+> ppr (ufmToList sigs),
346 text "binds:" <+> pprCoreBinding (Rec pairs)]))
347 (emptySigEnv, emptyDmdEnv, orig_pairs) -- Safe output
348 | otherwise = loop (n+1) sigs' pairs'
350 -- Use the new signature to do the next pair
351 -- The occurrence analyser has arranged them in a good order
352 -- so this can significantly reduce the number of iterations needed
353 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
355 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
356 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
358 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
359 ((sigs', lazy_fv'), pair')
362 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
363 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
364 -- old_sig = lookup sigs id
365 -- new_sig = lookup sigs' id
367 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
368 lookup sigs var = case lookupVarEnv sigs var of
371 -- Get an initial strictness signature from the Id
372 -- itself. That way we make use of earlier iterations
373 -- of the fixpoint algorithm. (Cunning plan.)
374 -- Note that the cunning plan extends to the DmdEnv too,
375 -- since it is part of the strictness signature
376 initialSig id = idNewStrictness_maybe id `orElse` botSig
378 dmdAnalRhs :: TopLevelFlag -> RecFlag
379 -> SigEnv -> (Id, CoreExpr)
380 -> (SigEnv, DmdEnv, (Id, CoreExpr))
381 -- Process the RHS of the binding, add the strictness signature
382 -- to the Id, and augment the environment with the signature as well.
384 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
385 = (sigs', lazy_fv, (id', rhs'))
387 arity = idArity id -- The idArity should be up to date
388 -- The simplifier was run just beforehand
389 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
390 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty, ppr id )
391 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
392 id' = id `setIdNewStrictness` sig_ty
393 sigs' = extendSigEnv top_lvl sigs id sig_ty
396 %************************************************************************
398 \subsection{Strictness signatures and types}
400 %************************************************************************
403 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
404 -- Take a DmdType and turn it into a StrictSig
405 -- NB: not used for never-inline things; hence False
406 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
408 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
409 mkSigTy top_lvl rec_flag id rhs dmd_ty
410 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
412 never_inline = isNeverActive (idInlinePragma id)
413 maybe_id_dmd = idNewDemandInfo_maybe id
414 -- Is Nothing the first time round
417 | isTopLevel top_lvl = False -- Top level things don't get
418 -- their demandInfo set at all
419 | isRec rec_flag = False -- Ditto recursive things
420 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
421 | otherwise = True -- Optimistic, first time round
425 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
426 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
427 If the rhs is a thunk, we usually forget the CPR info, because
428 it is presumably shared (else it would have been inlined, and
429 so we'd lose sharing if w/w'd it into a function.
431 However, if the strictness analyser has figured out (in a previous
432 iteration) that it's strict, then we DON'T need to forget the CPR info.
433 Instead we can retain the CPR info and do the thunk-splitting transform
434 (see WorkWrap.splitThunk).
436 This made a big difference to PrelBase.modInt, which had something like
437 modInt = \ x -> let r = ... -> I# v in
438 ...body strict in r...
439 r's RHS isn't a value yet; but modInt returns r in various branches, so
440 if r doesn't have the CPR property then neither does modInt
441 Another case I found in practice (in Complex.magnitude), looks like this:
442 let k = if ... then I# a else I# b
443 in ... body strict in k ....
444 (For this example, it doesn't matter whether k is returned as part of
445 the overall result; but it does matter that k's RHS has the CPR property.)
446 Left to itself, the simplifier will make a join point thus:
447 let $j k = ...body strict in k...
448 if ... then $j (I# a) else $j (I# b)
449 With thunk-splitting, we get instead
450 let $j x = let k = I#x in ...body strict in k...
451 in if ... then $j a else $j b
452 This is much better; there's a good chance the I# won't get allocated.
454 The difficulty with this is that we need the strictness type to
455 look at the body... but we now need the body to calculate the demand
456 on the variable, so we can decide whether its strictness type should
457 have a CPR in it or not. Simple solution:
458 a) use strictness info from the previous iteration
459 b) make sure we do at least 2 iterations, by doing a second
460 round for top-level non-recs. Top level recs will get at
461 least 2 iterations except for totally-bottom functions
462 which aren't very interesting anyway.
464 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
466 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
467 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
468 Demand info now has a 'Nothing' state, just like strictness info.
469 The analysis works from 'dangerous' towards a 'safe' state; so we
470 start with botSig for 'Nothing' strictness infos, and we start with
471 "yes, it's demanded" for 'Nothing' in the demand info. The
472 fixpoint iteration will sort it all out.
474 We can't start with 'not-demanded' because then consider
478 if ... then t else I# y else f x'
480 In the first iteration we'd have no demand info for x, so assume
481 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
482 we'd see that t was demanded, and so give it the CPR property, but
483 by now f has TopRes, so it will stay TopRes.
485 Instead, with the Nothing setting the first time round, we say
486 'yes t is demanded' the first time.
488 However, this does mean that for non-recursive bindings we must
489 iterate twice to be sure of not getting over-optimistic CPR info,
490 in the case where t turns out to be not-demanded. This is handled
495 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
496 | never_inline && not (isBotRes res)
498 -- Don't strictness-analyse NOINLINE things. Why not? Because
499 -- the NOINLINE says "don't expose any of the inner workings at the call
500 -- site" and the strictness is certainly an inner working.
502 -- More concretely, the demand analyser discovers the following strictness
503 -- for unsafePerformIO: C(U(AV))
505 -- unsafePerformIO (\s -> let r = f x in
506 -- case writeIORef v r s of (# s1, _ #) ->
508 -- The strictness analyser will find that the binding for r is strict,
509 -- (becuase of uPIO's strictness sig), and so it'll evaluate it before
510 -- doing the writeIORef. This actually makes tests/lib/should_run/memo002
513 -- Solution: don't expose the strictness of unsafePerformIO.
515 -- But we do want to expose the strictness of error functions,
516 -- which are also often marked NOINLINE
517 -- {-# NOINLINE foo #-}
518 -- foo x = error ("wubble buggle" ++ x)
519 -- So (hack, hack) we only drop the strictness for non-bottom things
520 -- This is all very unsatisfactory.
521 = (deferEnv fv, topSig)
524 = (lazy_fv, mkStrictSig dmd_ty)
526 dmd_ty = DmdType strict_fv final_dmds res'
528 lazy_fv = filterUFM (not . isStrictDmd) fv
529 strict_fv = filterUFM isStrictDmd fv
530 -- We put the strict FVs in the DmdType of the Id, so
531 -- that at its call sites we unleash demands on its strict fvs.
532 -- An example is 'roll' in imaginary/wheel-sieve2
533 -- Something like this:
535 -- go y = if ... then roll (x-1) else x+1
538 -- We want to see that roll is strict in x, which is because
539 -- go is called. So we put the DmdEnv for x in go's DmdType.
542 -- f :: Int -> Int -> Int
543 -- f x y = let t = x+1
544 -- h z = if z==0 then t else
545 -- if z==1 then x+1 else
549 -- Calling h does indeed evaluate x, but we can only see
550 -- that if we unleash a demand on x at the call site for t.
552 -- Incidentally, here's a place where lambda-lifting h would
553 -- lose the cigar --- we couldn't see the joint strictness in t/x
556 -- We don't want to put *all* the fv's from the RHS into the
557 -- DmdType, because that makes fixpointing very slow --- the
558 -- DmdType gets full of lazy demands that are slow to converge.
560 final_dmds = setUnpackStrategy dmds
561 -- Set the unpacking strategy
564 RetCPR | ignore_cpr_info -> TopRes
566 ignore_cpr_info = not (exprIsValue rhs || thunk_cpr_ok)
569 The unpack strategy determines whether we'll *really* unpack the argument,
570 or whether we'll just remember its strictness. If unpacking would give
571 rise to a *lot* of worker args, we may decide not to unpack after all.
574 setUnpackStrategy :: [Demand] -> [Demand]
576 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
578 go :: Int -- Max number of args available for sub-components of [Demand]
580 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
582 go n (Eval (Prod cs) : ds)
583 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
584 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
587 n' = n + 1 - non_abs_args
588 -- Add one to the budget 'cos we drop the top-level arg
589 non_abs_args = nonAbsentArgs cs
590 -- Delete # of non-absent args to which we'll now be committed
592 go n (d:ds) = d `cons` go n ds
595 cons d (n,ds) = (n, d:ds)
597 nonAbsentArgs :: [Demand] -> Int
599 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
600 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
604 %************************************************************************
606 \subsection{Strictness signatures and types}
608 %************************************************************************
611 splitDmdTy :: DmdType -> (Demand, DmdType)
612 -- Split off one function argument
613 -- We already have a suitable demand on all
614 -- free vars, so no need to add more!
615 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
616 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
620 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
622 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
623 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
624 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
626 addLazyFVs (DmdType fv ds res) lazy_fvs
627 = DmdType both_fv1 ds res
629 both_fv = (plusUFM_C both fv lazy_fvs)
630 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
631 -- This modifyEnv is vital. Consider
632 -- let f = \x -> (x,y)
634 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
635 -- demand with the bottom coming up from 'error'
637 -- I got a loop in the fixpointer without this, due to an interaction
638 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
640 -- = letrec g y = x `fatbar`
641 -- letrec h z = z + ...g...
644 -- In the initial iteration for f, f=Bot
645 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
646 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
647 -- places on its free variables. Suppose it places none. Then the
648 -- x `fatbar` ...call to h...
649 -- will give a x->V demand for x. That turns into a L demand for x,
650 -- which floats out of the defn for h. Without the modifyEnv, that
651 -- L demand doesn't get both'd with the Bot coming up from the inner
652 -- call to f. So we just get an L demand for x for g.
654 -- A better way to say this is that the lazy-fv filtering should give the
655 -- same answer as putting the lazy fv demands in the function's type.
657 annotateBndr :: DmdType -> Var -> (DmdType, Var)
658 -- The returned env has the var deleted
659 -- The returned var is annotated with demand info
660 -- No effect on the argument demands
661 annotateBndr dmd_ty@(DmdType fv ds res) var
662 | isTyVar var = (dmd_ty, var)
663 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
665 (fv', dmd) = removeFV fv var res
667 annotateBndrs = mapAccumR annotateBndr
669 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
670 -- For lambdas we add the demand to the argument demands
671 -- Only called for Ids
673 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
675 (fv', dmd) = removeFV fv id res
676 hacked_dmd = argDemand dmd
677 -- This call to argDemand is vital, because otherwise we label
678 -- a lambda binder with demand 'B'. But in terms of calling
679 -- conventions that's Abs, because we don't pass it. But
680 -- when we do a w/w split we get
681 -- fw x = (\x y:B -> ...) x (error "oops")
682 -- And then the simplifier things the 'B' is a strict demand
683 -- and evaluates the (error "oops"). Sigh
685 removeFV fv id res = (fv', zapUnlifted id dmd)
687 fv' = fv `delVarEnv` id
688 dmd = lookupVarEnv fv id `orElse` deflt
689 deflt | isBotRes res = Bot
692 -- For unlifted-type variables, we are only
693 -- interested in Bot/Abs/Box Abs
694 zapUnlifted is Bot = Bot
695 zapUnlifted id Abs = Abs
696 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
700 %************************************************************************
702 \subsection{Strictness signatures}
704 %************************************************************************
707 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
708 -- We use the SigEnv to tell us whether to
709 -- record info about a variable in the DmdEnv
710 -- We do so if it's a LocalId, but not top-level
712 -- The DmdEnv gives the demand on the free vars of the function
713 -- when it is given enough args to satisfy the strictness signature
715 emptySigEnv = emptyVarEnv
717 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
718 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
720 extendSigEnvList = extendVarEnvList
722 extendSigsWithLam :: SigEnv -> Id -> SigEnv
723 -- Extend the SigEnv when we meet a lambda binder
724 -- If the binder is marked demanded with a product demand, then give it a CPR
725 -- signature, because in the likely event that this is a lambda on a fn defn
726 -- [we only use this when the lambda is being consumed with a call demand],
727 -- it'll be w/w'd and so it will be CPR-ish.
729 -- NOTE: see notes [CPR-AND-STRICTNESS]
731 -- Also note that we only want to do this for something that
732 -- definitely has product type, else we may get over-optimistic
733 -- CPR results (e.g. from \x -> x!).
735 extendSigsWithLam sigs id
736 = case idNewDemandInfo_maybe id of
737 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
738 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
742 cprSig = StrictSig (mkDmdType emptyVarEnv [] RetCPR)
745 dmdTransform :: SigEnv -- The strictness environment
746 -> Id -- The function
747 -> Demand -- The demand on the function
748 -> DmdType -- The demand type of the function in this context
749 -- Returned DmdEnv includes the demand on
750 -- this function plus demand on its free variables
752 dmdTransform sigs var dmd
754 ------ DATA CONSTRUCTOR
755 | isDataConId var -- Data constructor
757 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
758 DmdType _ _ con_res = dmd_ty
761 if arity == call_depth then -- Saturated, so unleash the demand
763 -- Important! If we Keep the constructor application, then
764 -- we need the demands the constructor places (always lazy)
765 -- If not, we don't need to. For example:
766 -- f p@(x,y) = (p,y) -- S(AL)
768 -- It's vital that we don't calculate Absent for a!
769 dmd_ds = case res_dmd of
770 Box (Eval ds) -> mapDmds box ds
774 -- ds can be empty, when we are just seq'ing the thing
775 -- If so we must make up a suitable bunch of demands
776 arg_ds = case dmd_ds of
777 Poly d -> replicate arity d
778 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
781 mkDmdType emptyDmdEnv arg_ds con_res
782 -- Must remember whether it's a product, hence con_res, not TopRes
786 ------ IMPORTED FUNCTION
787 | isGlobalId var, -- Imported function
788 let StrictSig dmd_ty = idNewStrictness var
789 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
794 ------ LOCAL LET/REC BOUND THING
795 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
797 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
798 | otherwise = deferType dmd_ty
799 -- NB: it's important to use deferType, and not just return topDmdType
800 -- Consider let { f x y = p + x } in f 1
801 -- The application isn't saturated, but we must nevertheless propagate
802 -- a lazy demand for p!
804 addVarDmd top_lvl fn_ty var dmd
806 ------ LOCAL NON-LET/REC BOUND THING
807 | otherwise -- Default case
811 (call_depth, res_dmd) = splitCallDmd dmd
815 %************************************************************************
819 %************************************************************************
822 splitCallDmd :: Demand -> (Int, Demand)
823 splitCallDmd (Call d) = case splitCallDmd d of
825 splitCallDmd d = (0, d)
827 vanillaCall :: Arity -> Demand
828 vanillaCall 0 = evalDmd
829 vanillaCall n = Call (vanillaCall (n-1))
831 deferType :: DmdType -> DmdType
832 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
833 -- Notice that we throw away info about both arguments and results
834 -- For example, f = let ... in \x -> x
835 -- We don't want to get a stricness type V->T for f.
838 deferEnv :: DmdEnv -> DmdEnv
839 deferEnv fv = mapVarEnv defer fv
843 argDemand :: Demand -> Demand
844 -- The 'Defer' demands are just Lazy at function boundaries
845 -- Ugly! Ask John how to improve it.
846 argDemand Top = lazyDmd
847 argDemand (Defer d) = lazyDmd
848 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
849 argDemand (Box Bot) = evalDmd
850 argDemand (Box d) = box (argDemand d)
851 argDemand Bot = Abs -- Don't pass args that are consumed by bottom/err
856 betterStrictness :: StrictSig -> StrictSig -> Bool
857 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
859 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
861 betterDemand :: Demand -> Demand -> Bool
862 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
863 betterDemand d1 d2 = (d1 `lub` d2) == d2
867 -------------------------
868 -- Consider (if x then y else []) with demand V
869 -- Then the first branch gives {y->V} and the second
870 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
871 -- in the result env.
872 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
873 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
875 lub_fv = plusUFM_C lub fv1 fv2
876 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
877 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
878 -- lub is the identity for Bot
880 -- Extend the shorter argument list to match the longer
881 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
883 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
884 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
886 -----------------------------------
887 -- (t1 `bothType` t2) takes the argument/result info from t1,
888 -- using t2 just for its free-var info
889 -- NB: Don't forget about r2! It might be BotRes, which is
890 -- a bottom demand on all the in-scope variables.
891 -- Peter: can this be done more neatly?
892 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
893 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
895 both_fv = plusUFM_C both fv1 fv2
896 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
897 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
898 -- both is the identity for Abs
905 lubRes RetCPR RetCPR = RetCPR
906 lubRes r1 r2 = TopRes
908 -- If either diverges, the whole thing does
909 -- Otherwise take CPR info from the first
910 bothRes r1 BotRes = BotRes
915 modifyEnv :: Bool -- No-op if False
916 -> (Demand -> Demand) -- The zapper
917 -> DmdEnv -> DmdEnv -- Env1 and Env2
918 -> DmdEnv -> DmdEnv -- Transform this env
919 -- Zap anything in Env1 but not in Env2
920 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
922 modifyEnv need_to_modify zapper env1 env2 env
923 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
926 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
928 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
932 %************************************************************************
934 \subsection{LUB and BOTH}
936 %************************************************************************
939 lub :: Demand -> Demand -> Demand
942 lub Abs d2 = absLub d2
944 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
946 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
947 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
948 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
949 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
951 -- For the Eval case, we use these approximation rules
952 -- Box Bot <= Eval (Box Bot ...)
953 -- Box Top <= Defer (Box Bot ...)
954 -- Box (Eval ds) <= Eval (map Box ds)
955 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
956 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
957 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
958 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
959 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
961 lub (Box d1) (Box d2) = box (d1 `lub` d2)
962 lub d1@(Box _) d2 = d2 `lub` d1
964 lubs = zipWithDmds lub
966 ---------------------
967 -- box is the smart constructor for Box
968 -- It computes <B,bot> & d
969 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
970 -- Seems to be no point in allowing (Box (Call d))
971 box (Call d) = Call d -- The odd man out. Why?
973 box (Defer _) = lazyDmd
974 box Top = lazyDmd -- Box Abs and Box Top
975 box Abs = lazyDmd -- are the same <B,L>
976 box d = Box d -- Bot, Eval
979 defer :: Demand -> Demand
981 -- defer is the smart constructor for Defer
982 -- The idea is that (Defer ds) = <U(ds), L>
984 -- It specifies what happens at a lazy function argument
985 -- or a lambda; the L* operator
986 -- Set the strictness part to L, but leave
987 -- the boxity side unaffected
988 -- It also ensures that Defer (Eval [LLLL]) = L
993 defer (Call _) = lazyDmd -- Approximation here?
994 defer (Box _) = lazyDmd
995 defer (Defer ds) = Defer ds
996 defer (Eval ds) = deferEval ds
998 -- deferEval ds = defer (Eval ds)
999 deferEval ds | allTop ds = Top
1000 | otherwise = Defer ds
1002 ---------------------
1003 absLub :: Demand -> Demand
1004 -- Computes (Abs `lub` d)
1005 -- For the Bot case consider
1006 -- f x y = if ... then x else error x
1007 -- Then for y we get Abs `lub` Bot, and we really
1012 absLub (Call _) = Top
1013 absLub (Box _) = Top
1014 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1015 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1017 absLubs = mapDmds absLub
1020 both :: Demand -> Demand -> Demand
1026 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1029 -- From 'error' itself we get demand Bot on x
1030 -- From the arg demand on x we get
1031 -- x :-> evalDmd = Box (Eval (Poly Abs))
1032 -- So we get Bot `both` Box (Eval (Poly Abs))
1033 -- = Seq Keep (Poly Bot)
1036 -- f x = if ... then error (fst x) else fst x
1037 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1039 -- which is what we want.
1042 both Top Bot = errDmd
1045 both Top (Box d) = Box d
1046 both Top (Call d) = Call d
1047 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1048 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1049 -- = defer (Eval (mapDmds (`both` Top) ds))
1050 = deferEval (mapDmds (`both` Top) ds)
1053 both (Box d1) (Box d2) = box (d1 `both` d2)
1054 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1055 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1056 both (Box d1) (Defer d2) = Box d1
1057 both d1@(Box _) d2 = d2 `both` d1
1059 both (Call d1) (Call d2) = Call (d1 `both` d2)
1060 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1061 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1062 both d1@(Call _) d2 = d1 `both` d1
1064 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1065 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1066 both d1@(Eval ds1) d2 = d2 `both` d1
1068 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1069 both d1@(Defer ds1) d2 = d2 `both` d1
1071 boths = zipWithDmds both
1076 %************************************************************************
1078 \subsection{Miscellaneous
1080 %************************************************************************
1084 #ifdef OLD_STRICTNESS
1085 get_changes binds = vcat (map get_changes_bind binds)
1087 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1088 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1090 get_changes_pr (id,rhs)
1091 = get_changes_var id $$ get_changes_expr rhs
1094 | isId var = get_changes_str var $$ get_changes_dmd var
1097 get_changes_expr (Type t) = empty
1098 get_changes_expr (Var v) = empty
1099 get_changes_expr (Lit l) = empty
1100 get_changes_expr (Note n e) = get_changes_expr e
1101 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1102 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1103 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1104 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1106 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1109 | new_better && old_better = empty
1110 | new_better = message "BETTER"
1111 | old_better = message "WORSE"
1112 | otherwise = message "INCOMPARABLE"
1114 message word = text word <+> text "strictness for" <+> ppr id <+> info
1115 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1116 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1117 -- strictness analyser can't track
1118 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1119 old_better = old `betterStrictness` new
1120 new_better = new `betterStrictness` old
1123 | isUnLiftedType (idType id) = empty -- Not useful
1124 | new_better && old_better = empty
1125 | new_better = message "BETTER"
1126 | old_better = message "WORSE"
1127 | otherwise = message "INCOMPARABLE"
1129 message word = text word <+> text "demand for" <+> ppr id <+> info
1130 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1131 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1133 old = newDemand (idDemandInfo id)
1134 new_better = new `betterDemand` old
1135 old_better = old `betterDemand` new
1138 squashSig (StrictSig (DmdType fv ds res))
1139 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1141 -- squash just gets rid of call demands
1142 -- which the old analyser doesn't track
1143 squashDmd (Call d) = evalDmd
1144 squashDmd (Box d) = Box (squashDmd d)
1145 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1146 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)