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 )
44 import Maybes ( orElse, expectJust )
50 * set a noinline pragma on bottoming Ids
52 * Consider f x = x+1 `fatbar` error (show x)
53 We'd like to unbox x, even if that means reboxing it in the error case.
56 instance Outputable TopLevelFlag where
60 %************************************************************************
62 \subsection{Top level stuff}
64 %************************************************************************
67 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
68 dmdAnalPgm dflags binds
70 showPass dflags "Demand analysis" ;
71 let { binds_plus_dmds = do_prog binds } ;
73 endPass dflags "Demand analysis"
74 Opt_D_dump_stranal binds_plus_dmds ;
76 -- Only if OLD_STRICTNESS is on, because only then is the old
77 -- strictness analyser run
78 let { dmd_changes = get_changes binds_plus_dmds } ;
79 printDump (text "Changes in demands" $$ dmd_changes) ;
81 return binds_plus_dmds
84 do_prog :: [CoreBind] -> [CoreBind]
85 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
87 dmdAnalTopBind :: SigEnv
90 dmdAnalTopBind sigs (NonRec id rhs)
92 ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel sigs (id, rhs)
93 (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel sigs (id, rhs1)
94 -- Do two passes to improve CPR information
95 -- See comments with ignore_cpr_info in mk_sig_ty
96 -- and with extendSigsWithLam
98 (sigs2, NonRec id2 rhs2)
100 dmdAnalTopBind sigs (Rec pairs)
102 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
103 -- We get two iterations automatically
104 -- c.f. the NonRec case above
110 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
111 -- Analyse the RHS and return
112 -- a) appropriate strictness info
113 -- b) the unfolding (decorated with stricntess info)
117 arity = exprArity rhs
118 (rhs_ty, rhs') = dmdAnal emptySigEnv (vanillaCall arity) rhs
119 sig = mkTopSigTy rhs rhs_ty
122 %************************************************************************
124 \subsection{The analyser itself}
126 %************************************************************************
129 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
131 dmdAnal sigs Abs e = (topDmdType, e)
134 | not (isStrictDmd dmd)
136 (res_ty, e') = dmdAnal sigs evalDmd e
138 (deferType res_ty, e')
139 -- It's important not to analyse e with a lazy demand because
140 -- a) When we encounter case s of (a,b) ->
141 -- we demand s with U(d1d2)... but if the overall demand is lazy
142 -- that is wrong, and we'd need to reduce the demand on s,
143 -- which is inconvenient
144 -- b) More important, consider
145 -- f (let x = R in x+x), where f is lazy
146 -- We still want to mark x as demanded, because it will be when we
147 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
148 -- just mark x as Lazy
149 -- c) The application rule wouldn't be right either
150 -- Evaluating (f x) in a L demand does *not* cause
151 -- evaluation of f in a C(L) demand!
154 dmdAnal sigs dmd (Lit lit)
155 = (topDmdType, Lit lit)
157 dmdAnal sigs dmd (Var var)
158 = (dmdTransform sigs var dmd, Var var)
160 dmdAnal sigs dmd (Note n e)
161 = (dmd_ty, Note n e')
163 (dmd_ty, e') = dmdAnal sigs dmd' e
165 Coerce _ _ -> evalDmd -- This coerce usually arises from a recursive
166 other -> dmd -- newtype, and we don't want to look inside them
167 -- for exactly the same reason that we don't look
168 -- inside recursive products -- we might not reach
169 -- a fixpoint. So revert to a vanilla Eval demand
171 dmdAnal sigs dmd (App fun (Type ty))
172 = (fun_ty, App fun' (Type ty))
174 (fun_ty, fun') = dmdAnal sigs dmd fun
176 -- Lots of the other code is there to make this
177 -- beautiful, compositional, application rule :-)
178 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
179 = let -- [Type arg handled above]
180 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
181 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
182 (arg_dmd, res_ty) = splitDmdTy fun_ty
184 (res_ty `bothType` arg_ty, App fun' arg')
186 dmdAnal sigs dmd (Lam var body)
189 (body_ty, body') = dmdAnal sigs dmd body
191 (body_ty, Lam var body')
193 | Call body_dmd <- dmd -- A call demand: good!
195 sigs' = extendSigsWithLam sigs var
196 (body_ty, body') = dmdAnal sigs' body_dmd body
197 (lam_ty, var') = annotateLamIdBndr body_ty var
199 (lam_ty, Lam var' body')
201 | otherwise -- Not enough demand on the lambda; but do the body
202 = let -- anyway to annotate it and gather free var info
203 (body_ty, body') = dmdAnal sigs evalDmd body
204 (lam_ty, var') = annotateLamIdBndr body_ty var
206 (deferType lam_ty, Lam var' body')
208 dmdAnal sigs dmd (Case scrut case_bndr [alt@(DataAlt dc,bndrs,rhs)])
209 | let tycon = dataConTyCon dc,
210 isProductTyCon tycon,
211 not (isRecursiveTyCon tycon)
213 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
214 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
215 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
216 (_, bndrs', _) = alt'
217 case_bndr_sig = cprSig
218 -- Inside the alternative, the case binder has the CPR property.
219 -- Meaning that a case on it will successfully cancel.
221 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
224 -- We want f to have the CPR property:
225 -- f b x = case fw b x of { r -> I# r }
226 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
229 -- Figure out whether the demand on the case binder is used, and use
230 -- that to set the scrut_dmd. This is utterly essential.
231 -- Consider f x = case x of y { (a,b) -> k y a }
232 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
233 -- worker, so the worker will rebuild
234 -- x = (a, absent-error)
235 -- and that'll crash.
236 -- So at one stage I had:
237 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
238 -- keepity | dead_case_bndr = Drop
239 -- | otherwise = Keep
242 -- case x of y { (a,b) -> h y + a }
243 -- where h : U(LL) -> T
244 -- The above code would compute a Keep for x, since y is not Abs, which is silly
245 -- The insight is, of course, that a demand on y is a demand on the
246 -- scrutinee, so we need to `both` it with the scrut demand
248 scrut_dmd = Eval (Prod [idNewDemandInfo b | b <- bndrs', isId b])
250 idNewDemandInfo case_bndr'
252 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
254 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' [alt'])
256 dmdAnal sigs dmd (Case scrut case_bndr alts)
258 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
259 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
260 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
262 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
263 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' alts')
265 dmdAnal sigs dmd (Let (NonRec id rhs) body)
267 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel sigs (id, rhs)
268 (body_ty, body') = dmdAnal sigs' dmd body
269 (body_ty1, id2) = annotateBndr body_ty id1
270 body_ty2 = addLazyFVs body_ty1 lazy_fv
273 -- If the actual demand is better than the vanilla
274 -- demand, we might do better to re-analyse with the
276 (let vanilla_dmd = vanillaCall (idArity id)
277 actual_dmd = idNewDemandInfo id2
279 if actual_dmd `betterDemand` vanilla_dmd && actual_dmd /= vanilla_dmd then
280 pprTrace "dmdLet: better demand" (ppr id <+> vcat [text "vanilla" <+> ppr vanilla_dmd,
281 text "actual" <+> ppr actual_dmd])
284 (body_ty2, Let (NonRec id2 rhs') body')
286 dmdAnal sigs dmd (Let (Rec pairs) body)
288 bndrs = map fst pairs
289 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
290 (body_ty, body') = dmdAnal sigs' dmd body
291 body_ty1 = addLazyFVs body_ty lazy_fv
293 sigs' `seq` body_ty `seq`
295 (body_ty2, _) = annotateBndrs body_ty1 bndrs
296 -- Don't bother to add demand info to recursive
297 -- binders as annotateBndr does;
298 -- being recursive, we can't treat them strictly.
299 -- But we do need to remove the binders from the result demand env
301 (body_ty2, Let (Rec pairs') body')
304 dmdAnalAlt sigs dmd (con,bndrs,rhs)
306 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
307 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
309 (alt_ty, (con, bndrs', rhs'))
312 %************************************************************************
314 \subsection{Bindings}
316 %************************************************************************
319 dmdFix :: TopLevelFlag
320 -> SigEnv -- Does not include bindings for this binding
323 [(Id,CoreExpr)]) -- Binders annotated with stricness info
325 dmdFix top_lvl sigs orig_pairs
326 = loop 1 initial_sigs orig_pairs
328 bndrs = map fst orig_pairs
329 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
332 -> SigEnv -- Already contains the current sigs
334 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
336 | all (same_sig sigs sigs') bndrs
337 = (sigs', lazy_fv, pairs')
338 -- Note: use pairs', not pairs. pairs' is the result of
339 -- processing the RHSs with sigs (= sigs'), whereas pairs
340 -- is the result of processing the RHSs with the *previous*
341 -- iteration of sigs.
342 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
343 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
344 text "env:" <+> ppr (ufmToList sigs),
345 text "binds:" <+> pprCoreBinding (Rec pairs)]))
346 (emptySigEnv, emptyDmdEnv, orig_pairs) -- Safe output
347 | otherwise = loop (n+1) sigs' pairs'
349 -- Use the new signature to do the next pair
350 -- The occurrence analyser has arranged them in a good order
351 -- so this can significantly reduce the number of iterations needed
352 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
354 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
355 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
357 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
358 ((sigs', lazy_fv'), pair')
361 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl sigs (id,rhs)
362 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
363 -- old_sig = lookup sigs id
364 -- new_sig = lookup sigs' id
366 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
367 lookup sigs var = case lookupVarEnv sigs var of
370 -- Get an initial strictness signature from the Id
371 -- itself. That way we make use of earlier iterations
372 -- of the fixpoint algorithm. (Cunning plan.)
373 -- Note that the cunning plan extends to the DmdEnv too,
374 -- since it is part of the strictness signature
375 initialSig id = idNewStrictness_maybe id `orElse` botSig
377 dmdAnalRhs :: TopLevelFlag
378 -> SigEnv -> (Id, CoreExpr)
379 -> (SigEnv, DmdEnv, (Id, CoreExpr))
380 -- Process the RHS of the binding, add the strictness signature
381 -- to the Id, and augment the environment with the signature as well.
383 dmdAnalRhs top_lvl sigs (id, rhs)
384 = (sigs', lazy_fv, (id', rhs'))
386 arity = idArity id -- The idArity should be up to date
387 -- The simplifier was run just beforehand
388 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
389 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty, ppr id )
390 mkSigTy id rhs rhs_dmd_ty
391 id' = id `setIdNewStrictness` sig_ty
392 sigs' = extendSigEnv top_lvl sigs id sig_ty
395 %************************************************************************
397 \subsection{Strictness signatures and types}
399 %************************************************************************
402 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
403 -- Take a DmdType and turn it into a StrictSig
404 -- NB: not used for never-inline things; hence False
405 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
407 mkSigTy :: Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
408 mkSigTy id rhs dmd_ty = mk_sig_ty (isNeverActive (idInlinePragma id))
412 ok_to_keep_cpr_info = case idNewDemandInfo_maybe id of
413 Nothing -> True -- Is the case the first time round
414 Just dmd -> isStrictDmd dmd
417 The ok_to_keep_cpr_info stuff [CPR-AND-STRICTNESS]
418 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
419 If the rhs is a thunk, we usually forget the CPR info, because
420 it is presumably shared (else it would have been inlined, and
421 so we'd lose sharing if w/w'd it into a function.
423 However, if the strictness analyser has figured out (in a previous
424 iteration) that it's strict, then we DON'T need to forget the CPR info.
425 Instead we can retain the CPR info and do the thunk-splitting transform
426 (see WorkWrap.splitThunk).
428 This made a big difference to PrelBase.modInt, which had something like
429 modInt = \ x -> let r = ... -> I# v in
430 ...body strict in r...
431 r's RHS isn't a value yet; but modInt returns r in various branches, so
432 if r doesn't have the CPR property then neither does modInt
433 Another case I found in practice (in Complex.magnitude), looks like this:
434 let k = if ... then I# a else I# b
435 in ... body strict in k ....
436 (For this example, it doesn't matter whether k is returned as part of
437 the overall result; but it does matter that k's RHS has the CPR property.)
438 Left to itself, the simplifier will make a join point thus:
439 let $j k = ...body strict in k...
440 if ... then $j (I# a) else $j (I# b)
441 With thunk-splitting, we get instead
442 let $j x = let k = I#x in ...body strict in k...
443 in if ... then $j a else $j b
444 This is much better; there's a good chance the I# won't get allocated.
446 The difficulty with this is that we need the strictness type to
447 look at the body... but we now need the body to calculate the demand
448 on the variable, so we can decide whether its strictness type should
449 have a CPR in it or not. Simple solution:
450 a) use strictness info from the previous iteration
451 b) make sure we do at least 2 iterations, by doing a second
452 round for top-level non-recs. Top level recs will get at
453 least 2 iterations except for totally-bottom functions
454 which aren't very interesting anyway.
456 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
458 The Nothing case in ok_to_keep_cpr_info [CPR-AND-STRICTNESS]
459 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
460 Demand info now has a 'Nothing' state, just like strictness info.
461 The analysis works from 'dangerous' towards a 'safe' state; so we
462 start with botSig for 'Nothing' strictness infos, and we start with
463 "yes, it's demanded" for 'Nothing' in the demand info. The
464 fixpoint iteration will sort it all out.
466 We can't start with 'not-demanded' because then consider
470 if ... then t else I# y else f x'
472 In the first iteration we'd have no demand info for x, so assume
473 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
474 we'd see that t was demanded, and so give it the CPR property, but
475 by now f has TopRes, so it will stay TopRes.
477 Instead, with the Nothing setting the first time round, we say
478 'yes t is demanded' the first time.
480 However, this does mean that for non-recursive bindings we must
481 iterate twice to be sure of not getting over-optimistic CPR info,
482 in the case where t turns out to be not-demanded. This is handled
487 mk_sig_ty never_inline ok_to_keep_cpr_info rhs (DmdType fv dmds res)
488 | never_inline && not (isBotRes res)
490 -- Don't strictness-analyse NOINLINE things. Why not? Because
491 -- the NOINLINE says "don't expose any of the inner workings at the call
492 -- site" and the strictness is certainly an inner working.
494 -- More concretely, the demand analyser discovers the following strictness
495 -- for unsafePerformIO: C(U(AV))
497 -- unsafePerformIO (\s -> let r = f x in
498 -- case writeIORef v r s of (# s1, _ #) ->
500 -- The strictness analyser will find that the binding for r is strict,
501 -- (becuase of uPIO's strictness sig), and so it'll evaluate it before
502 -- doing the writeIORef. This actually makes tests/lib/should_run/memo002
505 -- Solution: don't expose the strictness of unsafePerformIO.
507 -- But we do want to expose the strictness of error functions,
508 -- which are also often marked NOINLINE
509 -- {-# NOINLINE foo #-}
510 -- foo x = error ("wubble buggle" ++ x)
511 -- So (hack, hack) we only drop the strictness for non-bottom things
512 -- This is all very unsatisfactory.
513 = (deferEnv fv, topSig)
516 = (lazy_fv, mkStrictSig dmd_ty)
518 dmd_ty = DmdType strict_fv final_dmds res'
520 lazy_fv = filterUFM (not . isStrictDmd) fv
521 strict_fv = filterUFM isStrictDmd fv
522 -- We put the strict FVs in the DmdType of the Id, so
523 -- that at its call sites we unleash demands on its strict fvs.
524 -- An example is 'roll' in imaginary/wheel-sieve2
525 -- Something like this:
527 -- go y = if ... then roll (x-1) else x+1
530 -- We want to see that roll is strict in x, which is because
531 -- go is called. So we put the DmdEnv for x in go's DmdType.
534 -- f :: Int -> Int -> Int
535 -- f x y = let t = x+1
536 -- h z = if z==0 then t else
537 -- if z==1 then x+1 else
541 -- Calling h does indeed evaluate x, but we can only see
542 -- that if we unleash a demand on x at the call site for t.
544 -- Incidentally, here's a place where lambda-lifting h would
545 -- lose the cigar --- we couldn't see the joint strictness in t/x
548 -- We don't want to put *all* the fv's from the RHS into the
549 -- DmdType, because that makes fixpointing very slow --- the
550 -- DmdType gets full of lazy demands that are slow to converge.
552 final_dmds = setUnpackStrategy dmds
553 -- Set the unpacking strategy
556 RetCPR | ignore_cpr_info -> TopRes
558 ignore_cpr_info = not (exprIsValue rhs || ok_to_keep_cpr_info)
561 The unpack strategy determines whether we'll *really* unpack the argument,
562 or whether we'll just remember its strictness. If unpacking would give
563 rise to a *lot* of worker args, we may decide not to unpack after all.
566 setUnpackStrategy :: [Demand] -> [Demand]
568 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
570 go :: Int -- Max number of args available for sub-components of [Demand]
572 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
574 go n (Eval (Prod cs) : ds)
575 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
576 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
579 n' = n + 1 - non_abs_args
580 -- Add one to the budget 'cos we drop the top-level arg
581 non_abs_args = nonAbsentArgs cs
582 -- Delete # of non-absent args to which we'll now be committed
584 go n (d:ds) = d `cons` go n ds
587 cons d (n,ds) = (n, d:ds)
589 nonAbsentArgs :: [Demand] -> Int
591 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
592 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
596 %************************************************************************
598 \subsection{Strictness signatures and types}
600 %************************************************************************
603 splitDmdTy :: DmdType -> (Demand, DmdType)
604 -- Split off one function argument
605 -- We already have a suitable demand on all
606 -- free vars, so no need to add more!
607 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
608 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
612 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
614 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
615 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
616 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
618 addLazyFVs (DmdType fv ds res) lazy_fvs
619 = DmdType both_fv1 ds res
621 both_fv = (plusUFM_C both fv lazy_fvs)
622 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
623 -- This modifyEnv is vital. Consider
624 -- let f = \x -> (x,y)
626 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
627 -- demand with the bottom coming up from 'error'
629 -- I got a loop in the fixpointer without this, due to an interaction
630 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
632 -- = letrec g y = x `fatbar`
633 -- letrec h z = z + ...g...
636 -- In the initial iteration for f, f=Bot
637 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
638 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
639 -- places on its free variables. Suppose it places none. Then the
640 -- x `fatbar` ...call to h...
641 -- will give a x->V demand for x. That turns into a L demand for x,
642 -- which floats out of the defn for h. Without the modifyEnv, that
643 -- L demand doesn't get both'd with the Bot coming up from the inner
644 -- call to f. So we just get an L demand for x for g.
646 -- A better way to say this is that the lazy-fv filtering should give the
647 -- same answer as putting the lazy fv demands in the function's type.
649 annotateBndr :: DmdType -> Var -> (DmdType, Var)
650 -- The returned env has the var deleted
651 -- The returned var is annotated with demand info
652 -- No effect on the argument demands
653 annotateBndr dmd_ty@(DmdType fv ds res) var
654 | isTyVar var = (dmd_ty, var)
655 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
657 (fv', dmd) = removeFV fv var res
659 annotateBndrs = mapAccumR annotateBndr
661 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
662 -- For lambdas we add the demand to the argument demands
663 -- Only called for Ids
665 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
667 (fv', dmd) = removeFV fv id res
668 hacked_dmd = argDemand dmd
669 -- This call to argDemand is vital, because otherwise we label
670 -- a lambda binder with demand 'B'. But in terms of calling
671 -- conventions that's Abs, because we don't pass it. But
672 -- when we do a w/w split we get
673 -- fw x = (\x y:B -> ...) x (error "oops")
674 -- And then the simplifier things the 'B' is a strict demand
675 -- and evaluates the (error "oops"). Sigh
677 removeFV fv id res = (fv', zapUnlifted id dmd)
679 fv' = fv `delVarEnv` id
680 dmd = lookupVarEnv fv id `orElse` deflt
681 deflt | isBotRes res = Bot
684 -- For unlifted-type variables, we are only
685 -- interested in Bot/Abs/Box Abs
686 zapUnlifted is Bot = Bot
687 zapUnlifted id Abs = Abs
688 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
692 %************************************************************************
694 \subsection{Strictness signatures}
696 %************************************************************************
699 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
700 -- We use the SigEnv to tell us whether to
701 -- record info about a variable in the DmdEnv
702 -- We do so if it's a LocalId, but not top-level
704 -- The DmdEnv gives the demand on the free vars of the function
705 -- when it is given enough args to satisfy the strictness signature
707 emptySigEnv = emptyVarEnv
709 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
710 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
712 extendSigEnvList = extendVarEnvList
714 extendSigsWithLam :: SigEnv -> Id -> SigEnv
715 -- Extend the SigEnv when we meet a lambda binder
716 -- If the binder is marked demanded with a product demand, then give it a CPR
717 -- signature, because in the likely event that this is a lambda on a fn defn
718 -- [we only use this when the lambda is being consumed with a call demand],
719 -- it'll be w/w'd and so it will be CPR-ish.
721 -- NOTE: see notes [CPR-AND-STRICTNESS]
723 -- Also note that we only want to do this for something that
724 -- definitely has product type, else we may get over-optimistic
725 -- CPR results (e.g. from \x -> x!).
727 extendSigsWithLam sigs id
728 = case idNewDemandInfo_maybe id of
729 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
730 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
734 cprSig = StrictSig (mkDmdType emptyVarEnv [] RetCPR)
737 dmdTransform :: SigEnv -- The strictness environment
738 -> Id -- The function
739 -> Demand -- The demand on the function
740 -> DmdType -- The demand type of the function in this context
741 -- Returned DmdEnv includes the demand on
742 -- this function plus demand on its free variables
744 dmdTransform sigs var dmd
746 ------ DATA CONSTRUCTOR
747 | isDataConId var -- Data constructor
749 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
750 DmdType _ _ con_res = dmd_ty
753 if arity == call_depth then -- Saturated, so unleash the demand
755 -- Important! If we Keep the constructor application, then
756 -- we need the demands the constructor places (always lazy)
757 -- If not, we don't need to. For example:
758 -- f p@(x,y) = (p,y) -- S(AL)
760 -- It's vital that we don't calculate Absent for a!
761 dmd_ds = case res_dmd of
762 Box (Eval ds) -> mapDmds box ds
766 -- ds can be empty, when we are just seq'ing the thing
767 -- If so we must make up a suitable bunch of demands
768 arg_ds = case dmd_ds of
769 Poly d -> replicate arity d
770 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
773 mkDmdType emptyDmdEnv arg_ds con_res
774 -- Must remember whether it's a product, hence con_res, not TopRes
778 ------ IMPORTED FUNCTION
779 | isGlobalId var, -- Imported function
780 let StrictSig dmd_ty = idNewStrictness var
781 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
786 ------ LOCAL LET/REC BOUND THING
787 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
789 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
790 | otherwise = deferType dmd_ty
791 -- NB: it's important to use deferType, and not just return topDmdType
792 -- Consider let { f x y = p + x } in f 1
793 -- The application isn't saturated, but we must nevertheless propagate
794 -- a lazy demand for p!
796 addVarDmd top_lvl fn_ty var dmd
798 ------ LOCAL NON-LET/REC BOUND THING
799 | otherwise -- Default case
803 (call_depth, res_dmd) = splitCallDmd dmd
807 %************************************************************************
811 %************************************************************************
814 splitCallDmd :: Demand -> (Int, Demand)
815 splitCallDmd (Call d) = case splitCallDmd d of
817 splitCallDmd d = (0, d)
819 vanillaCall :: Arity -> Demand
820 vanillaCall 0 = evalDmd
821 vanillaCall n = Call (vanillaCall (n-1))
823 deferType :: DmdType -> DmdType
824 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
825 -- Notice that we throw away info about both arguments and results
826 -- For example, f = let ... in \x -> x
827 -- We don't want to get a stricness type V->T for f.
830 deferEnv :: DmdEnv -> DmdEnv
831 deferEnv fv = mapVarEnv defer fv
835 argDemand :: Demand -> Demand
836 -- The 'Defer' demands are just Lazy at function boundaries
837 -- Ugly! Ask John how to improve it.
838 argDemand Top = lazyDmd
839 argDemand (Defer d) = lazyDmd
840 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
841 argDemand (Box Bot) = evalDmd
842 argDemand (Box d) = box (argDemand d)
843 argDemand Bot = Abs -- Don't pass args that are consumed by bottom/err
848 betterStrictness :: StrictSig -> StrictSig -> Bool
849 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
851 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
853 betterDemand :: Demand -> Demand -> Bool
854 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
855 betterDemand d1 d2 = (d1 `lub` d2) == d2
859 -------------------------
860 -- Consider (if x then y else []) with demand V
861 -- Then the first branch gives {y->V} and the second
862 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
863 -- in the result env.
864 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
865 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
867 lub_fv = plusUFM_C lub fv1 fv2
868 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
869 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
870 -- lub is the identity for Bot
872 -- Extend the shorter argument list to match the longer
873 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
875 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
876 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
878 -----------------------------------
879 -- (t1 `bothType` t2) takes the argument/result info from t1,
880 -- using t2 just for its free-var info
881 -- NB: Don't forget about r2! It might be BotRes, which is
882 -- a bottom demand on all the in-scope variables.
883 -- Peter: can this be done more neatly?
884 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
885 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
887 both_fv = plusUFM_C both fv1 fv2
888 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
889 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
890 -- both is the identity for Abs
897 lubRes RetCPR RetCPR = RetCPR
898 lubRes r1 r2 = TopRes
900 -- If either diverges, the whole thing does
901 -- Otherwise take CPR info from the first
902 bothRes r1 BotRes = BotRes
907 modifyEnv :: Bool -- No-op if False
908 -> (Demand -> Demand) -- The zapper
909 -> DmdEnv -> DmdEnv -- Env1 and Env2
910 -> DmdEnv -> DmdEnv -- Transform this env
911 -- Zap anything in Env1 but not in Env2
912 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
914 modifyEnv need_to_modify zapper env1 env2 env
915 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
918 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
920 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
924 %************************************************************************
926 \subsection{LUB and BOTH}
928 %************************************************************************
931 lub :: Demand -> Demand -> Demand
934 lub Abs d2 = absLub d2
936 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
938 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
939 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
940 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
941 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
943 -- For the Eval case, we use these approximation rules
944 -- Box Bot <= Eval (Box Bot ...)
945 -- Box Top <= Defer (Box Bot ...)
946 -- Box (Eval ds) <= Eval (map Box ds)
947 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
948 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
949 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
950 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
951 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
953 lub (Box d1) (Box d2) = box (d1 `lub` d2)
954 lub d1@(Box _) d2 = d2 `lub` d1
956 lubs = zipWithDmds lub
958 ---------------------
959 -- box is the smart constructor for Box
960 -- It computes <B,bot> & d
961 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
962 -- Seems to be no point in allowing (Box (Call d))
963 box (Call d) = Call d -- The odd man out. Why?
965 box (Defer _) = lazyDmd
966 box Top = lazyDmd -- Box Abs and Box Top
967 box Abs = lazyDmd -- are the same <B,L>
968 box d = Box d -- Bot, Eval
971 defer :: Demand -> Demand
973 -- defer is the smart constructor for Defer
974 -- The idea is that (Defer ds) = <U(ds), L>
976 -- It specifies what happens at a lazy function argument
977 -- or a lambda; the L* operator
978 -- Set the strictness part to L, but leave
979 -- the boxity side unaffected
980 -- It also ensures that Defer (Eval [LLLL]) = L
985 defer (Call _) = lazyDmd -- Approximation here?
986 defer (Box _) = lazyDmd
987 defer (Defer ds) = Defer ds
988 defer (Eval ds) = deferEval ds
990 -- deferEval ds = defer (Eval ds)
991 deferEval ds | allTop ds = Top
992 | otherwise = Defer ds
994 ---------------------
995 absLub :: Demand -> Demand
996 -- Computes (Abs `lub` d)
997 -- For the Bot case consider
998 -- f x y = if ... then x else error x
999 -- Then for y we get Abs `lub` Bot, and we really
1004 absLub (Call _) = Top
1005 absLub (Box _) = Top
1006 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1007 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1009 absLubs = mapDmds absLub
1012 both :: Demand -> Demand -> Demand
1018 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1021 -- From 'error' itself we get demand Bot on x
1022 -- From the arg demand on x we get
1023 -- x :-> evalDmd = Box (Eval (Poly Abs))
1024 -- So we get Bot `both` Box (Eval (Poly Abs))
1025 -- = Seq Keep (Poly Bot)
1028 -- f x = if ... then error (fst x) else fst x
1029 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1031 -- which is what we want.
1034 both Top Bot = errDmd
1037 both Top (Box d) = Box d
1038 both Top (Call d) = Call d
1039 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1040 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1041 -- = defer (Eval (mapDmds (`both` Top) ds))
1042 = deferEval (mapDmds (`both` Top) ds)
1045 both (Box d1) (Box d2) = box (d1 `both` d2)
1046 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1047 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1048 both (Box d1) (Defer d2) = Box d1
1049 both d1@(Box _) d2 = d2 `both` d1
1051 both (Call d1) (Call d2) = Call (d1 `both` d2)
1052 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1053 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1054 both d1@(Call _) d2 = d1 `both` d1
1056 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1057 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1058 both d1@(Eval ds1) d2 = d2 `both` d1
1060 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1061 both d1@(Defer ds1) d2 = d2 `both` d1
1063 boths = zipWithDmds both
1068 %************************************************************************
1070 \subsection{Miscellaneous
1072 %************************************************************************
1076 #ifdef OLD_STRICTNESS
1077 get_changes binds = vcat (map get_changes_bind binds)
1079 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1080 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1082 get_changes_pr (id,rhs)
1083 = get_changes_var id $$ get_changes_expr rhs
1086 | isId var = get_changes_str var $$ get_changes_dmd var
1089 get_changes_expr (Type t) = empty
1090 get_changes_expr (Var v) = empty
1091 get_changes_expr (Lit l) = empty
1092 get_changes_expr (Note n e) = get_changes_expr e
1093 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1094 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1095 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1096 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1098 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1101 | new_better && old_better = empty
1102 | new_better = message "BETTER"
1103 | old_better = message "WORSE"
1104 | otherwise = message "INCOMPARABLE"
1106 message word = text word <+> text "strictness for" <+> ppr id <+> info
1107 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1108 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1109 -- strictness analyser can't track
1110 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1111 old_better = old `betterStrictness` new
1112 new_better = new `betterStrictness` old
1115 | isUnLiftedType (idType id) = empty -- Not useful
1116 | new_better && old_better = empty
1117 | new_better = message "BETTER"
1118 | old_better = message "WORSE"
1119 | otherwise = message "INCOMPARABLE"
1121 message word = text word <+> text "demand for" <+> ppr id <+> info
1122 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1123 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1125 old = newDemand (idDemandInfo id)
1126 new_better = new `betterDemand` old
1127 old_better = old `betterDemand` new
1130 squashSig (StrictSig (DmdType fv ds res))
1131 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1133 -- squash just gets rid of call demands
1134 -- which the old analyser doesn't track
1135 squashDmd (Call d) = evalDmd
1136 squashDmd (Box d) = Box (squashDmd d)
1137 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1138 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)