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 DynFlags ( DynFlags, DynFlag(..) )
17 import StaticFlags ( opt_MaxWorkerArgs )
18 import NewDemand -- All of it
21 import CoreUtils ( exprIsHNF, exprIsTrivial, exprArity )
22 import DataCon ( dataConTyCon )
23 import TyCon ( isProductTyCon, isRecursiveTyCon )
24 import Id ( Id, idType, idInlinePragma,
25 isDataConWorkId, isGlobalId, idArity,
27 idDemandInfo, idStrictness, idCprInfo, idName,
29 idNewStrictness, idNewStrictness_maybe,
30 setIdNewStrictness, idNewDemandInfo,
31 idNewDemandInfo_maybe,
35 import IdInfo ( newStrictnessFromOld, newDemand )
39 import TysWiredIn ( unboxedPairDataCon )
40 import TysPrim ( realWorldStatePrimTy )
41 import UniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly,
42 keysUFM, minusUFM, ufmToList, filterUFM )
43 import Type ( isUnLiftedType, coreEqType, splitTyConApp_maybe )
44 import Coercion ( coercionKind )
45 import CoreLint ( showPass, endPass )
46 import Util ( mapAndUnzip, mapAccumL, mapAccumR, lengthIs )
47 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive,
49 import Maybes ( orElse, expectJust )
55 * set a noinline pragma on bottoming Ids
57 * Consider f x = x+1 `fatbar` error (show x)
58 We'd like to unbox x, even if that means reboxing it in the error case.
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 call_dmd = vanillaCall (exprArity rhs)
119 (_, rhs1) = dmdAnal emptySigEnv call_dmd rhs
120 (rhs_ty, rhs2) = dmdAnal emptySigEnv call_dmd rhs1
121 sig = mkTopSigTy rhs rhs_ty
122 -- Do two passes; see notes with extendSigsWithLam
123 -- Otherwise we get bogus CPR info for constructors like
124 -- newtype T a = MkT a
125 -- The constructor looks like (\x::T a -> x), modulo the coerce
126 -- extendSigsWithLam will optimistically give x a CPR tag the
127 -- first time, which is wrong in the end.
130 %************************************************************************
132 \subsection{The analyser itself}
134 %************************************************************************
137 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
139 dmdAnal sigs Abs e = (topDmdType, e)
142 | not (isStrictDmd dmd)
144 (res_ty, e') = dmdAnal sigs evalDmd e
146 (deferType res_ty, e')
147 -- It's important not to analyse e with a lazy demand because
148 -- a) When we encounter case s of (a,b) ->
149 -- we demand s with U(d1d2)... but if the overall demand is lazy
150 -- that is wrong, and we'd need to reduce the demand on s,
151 -- which is inconvenient
152 -- b) More important, consider
153 -- f (let x = R in x+x), where f is lazy
154 -- We still want to mark x as demanded, because it will be when we
155 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
156 -- just mark x as Lazy
157 -- c) The application rule wouldn't be right either
158 -- Evaluating (f x) in a L demand does *not* cause
159 -- evaluation of f in a C(L) demand!
162 dmdAnal sigs dmd (Lit lit)
163 = (topDmdType, Lit lit)
165 dmdAnal sigs dmd (Var var)
166 = (dmdTransform sigs var dmd, Var var)
168 dmdAnal sigs dmd (Cast e co)
169 = (dmd_ty, Cast e' co)
171 (dmd_ty, e') = dmdAnal sigs dmd' e
172 to_co = snd (coercionKind co)
174 | Just (tc, args) <- splitTyConApp_maybe to_co
175 , isRecursiveTyCon tc = evalDmd
177 -- This coerce usually arises from a recursive
178 -- newtype, and we don't want to look inside them
179 -- for exactly the same reason that we don't look
180 -- inside recursive products -- we might not reach
181 -- a fixpoint. So revert to a vanilla Eval demand
183 dmdAnal sigs dmd (Note n e)
184 = (dmd_ty, Note n e')
186 (dmd_ty, e') = dmdAnal sigs dmd e
188 dmdAnal sigs dmd (App fun (Type ty))
189 = (fun_ty, App fun' (Type ty))
191 (fun_ty, fun') = dmdAnal sigs dmd fun
193 -- Lots of the other code is there to make this
194 -- beautiful, compositional, application rule :-)
195 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
196 = let -- [Type arg handled above]
197 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
198 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
199 (arg_dmd, res_ty) = splitDmdTy fun_ty
201 (res_ty `bothType` arg_ty, App fun' arg')
203 dmdAnal sigs dmd (Lam var body)
206 (body_ty, body') = dmdAnal sigs dmd body
208 (body_ty, Lam var body')
210 | Call body_dmd <- dmd -- A call demand: good!
212 sigs' = extendSigsWithLam sigs var
213 (body_ty, body') = dmdAnal sigs' body_dmd body
214 (lam_ty, var') = annotateLamIdBndr body_ty var
216 (lam_ty, Lam var' body')
218 | otherwise -- Not enough demand on the lambda; but do the body
219 = let -- anyway to annotate it and gather free var info
220 (body_ty, body') = dmdAnal sigs evalDmd body
221 (lam_ty, var') = annotateLamIdBndr body_ty var
223 (deferType lam_ty, Lam var' body')
225 dmdAnal sigs dmd (Case scrut case_bndr ty [alt@(DataAlt dc,bndrs,rhs)])
226 | let tycon = dataConTyCon dc,
227 isProductTyCon tycon,
228 not (isRecursiveTyCon tycon)
230 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
231 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
232 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
233 (_, bndrs', _) = alt'
234 case_bndr_sig = cprSig
235 -- Inside the alternative, the case binder has the CPR property.
236 -- Meaning that a case on it will successfully cancel.
238 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
241 -- We want f to have the CPR property:
242 -- f b x = case fw b x of { r -> I# r }
243 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
246 -- Figure out whether the demand on the case binder is used, and use
247 -- that to set the scrut_dmd. This is utterly essential.
248 -- Consider f x = case x of y { (a,b) -> k y a }
249 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
250 -- worker, so the worker will rebuild
251 -- x = (a, absent-error)
252 -- and that'll crash.
253 -- So at one stage I had:
254 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
255 -- keepity | dead_case_bndr = Drop
256 -- | otherwise = Keep
259 -- case x of y { (a,b) -> h y + a }
260 -- where h : U(LL) -> T
261 -- The above code would compute a Keep for x, since y is not Abs, which is silly
262 -- The insight is, of course, that a demand on y is a demand on the
263 -- scrutinee, so we need to `both` it with the scrut demand
265 alt_dmd = Eval (Prod [idNewDemandInfo b | b <- bndrs', isId b])
266 scrut_dmd = alt_dmd `both`
267 idNewDemandInfo case_bndr'
269 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
271 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' ty [alt'])
273 dmdAnal sigs dmd (Case scrut case_bndr ty alts)
275 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
276 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
277 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
279 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
280 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' ty alts')
282 dmdAnal sigs dmd (Let (NonRec id rhs) body)
284 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
285 (body_ty, body') = dmdAnal sigs' dmd body
286 (body_ty1, id2) = annotateBndr body_ty id1
287 body_ty2 = addLazyFVs body_ty1 lazy_fv
289 -- If the actual demand is better than the vanilla call
290 -- demand, you might think that we might do better to re-analyse
291 -- the RHS with the stronger demand.
292 -- But (a) That seldom happens, because it means that *every* path in
293 -- the body of the let has to use that stronger demand
294 -- (b) It often happens temporarily in when fixpointing, because
295 -- the recursive function at first seems to place a massive demand.
296 -- But we don't want to go to extra work when the function will
297 -- probably iterate to something less demanding.
298 -- In practice, all the times the actual demand on id2 is more than
299 -- the vanilla call demand seem to be due to (b). So we don't
300 -- bother to re-analyse the RHS.
301 (body_ty2, Let (NonRec id2 rhs') body')
303 dmdAnal sigs dmd (Let (Rec pairs) body)
305 bndrs = map fst pairs
306 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
307 (body_ty, body') = dmdAnal sigs' dmd body
308 body_ty1 = addLazyFVs body_ty lazy_fv
310 sigs' `seq` body_ty `seq`
312 (body_ty2, _) = annotateBndrs body_ty1 bndrs
313 -- Don't bother to add demand info to recursive
314 -- binders as annotateBndr does;
315 -- being recursive, we can't treat them strictly.
316 -- But we do need to remove the binders from the result demand env
318 (body_ty2, Let (Rec pairs') body')
321 dmdAnalAlt sigs dmd (con,bndrs,rhs)
323 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
324 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
325 final_alt_ty | io_hack_reqd = alt_ty `lubType` topDmdType
328 -- There's a hack here for I/O operations. Consider
329 -- case foo x s of { (# s, r #) -> y }
330 -- Is this strict in 'y'. Normally yes, but what if 'foo' is an I/O
331 -- operation that simply terminates the program (not in an erroneous way)?
332 -- In that case we should not evaluate y before the call to 'foo'.
333 -- Hackish solution: spot the IO-like situation and add a virtual branch,
337 -- other -> return ()
338 -- So the 'y' isn't necessarily going to be evaluated
340 -- A more complete example where this shows up is:
341 -- do { let len = <expensive> ;
342 -- ; when (...) (exitWith ExitSuccess)
345 io_hack_reqd = con == DataAlt unboxedPairDataCon &&
346 idType (head bndrs) `coreEqType` realWorldStatePrimTy
348 (final_alt_ty, (con, bndrs', rhs'))
351 %************************************************************************
353 \subsection{Bindings}
355 %************************************************************************
358 dmdFix :: TopLevelFlag
359 -> SigEnv -- Does not include bindings for this binding
362 [(Id,CoreExpr)]) -- Binders annotated with stricness info
364 dmdFix top_lvl sigs orig_pairs
365 = loop 1 initial_sigs orig_pairs
367 bndrs = map fst orig_pairs
368 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
371 -> SigEnv -- Already contains the current sigs
373 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
376 = (sigs', lazy_fv, pairs')
377 -- Note: use pairs', not pairs. pairs' is the result of
378 -- processing the RHSs with sigs (= sigs'), whereas pairs
379 -- is the result of processing the RHSs with the *previous*
380 -- iteration of sigs.
382 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
383 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
384 text "env:" <+> ppr (ufmToList sigs),
385 text "binds:" <+> pprCoreBinding (Rec pairs)]))
386 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
387 -- The lazy_fv part is really important! orig_pairs has no strictness
388 -- info, including nothing about free vars. But if we have
389 -- letrec f = ....y..... in ...f...
390 -- where 'y' is free in f, we must record that y is mentioned,
391 -- otherwise y will get recorded as absent altogether
393 | otherwise = loop (n+1) sigs' pairs'
395 found_fixpoint = all (same_sig sigs sigs') bndrs
396 -- Use the new signature to do the next pair
397 -- The occurrence analyser has arranged them in a good order
398 -- so this can significantly reduce the number of iterations needed
399 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
401 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
402 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
404 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
405 ((sigs', lazy_fv'), pair')
408 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
409 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
410 -- old_sig = lookup sigs id
411 -- new_sig = lookup sigs' id
413 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
414 lookup sigs var = case lookupVarEnv sigs var of
417 -- Get an initial strictness signature from the Id
418 -- itself. That way we make use of earlier iterations
419 -- of the fixpoint algorithm. (Cunning plan.)
420 -- Note that the cunning plan extends to the DmdEnv too,
421 -- since it is part of the strictness signature
422 initialSig id = idNewStrictness_maybe id `orElse` botSig
424 dmdAnalRhs :: TopLevelFlag -> RecFlag
425 -> SigEnv -> (Id, CoreExpr)
426 -> (SigEnv, DmdEnv, (Id, CoreExpr))
427 -- Process the RHS of the binding, add the strictness signature
428 -- to the Id, and augment the environment with the signature as well.
430 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
431 = (sigs', lazy_fv, (id', rhs'))
433 arity = idArity id -- The idArity should be up to date
434 -- The simplifier was run just beforehand
435 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
436 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs), ppr id )
437 -- The RHS can be eta-reduced to just a variable,
438 -- in which case we should not complain.
439 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
440 id' = id `setIdNewStrictness` sig_ty
441 sigs' = extendSigEnv top_lvl sigs id sig_ty
444 %************************************************************************
446 \subsection{Strictness signatures and types}
448 %************************************************************************
451 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
452 -- Take a DmdType and turn it into a StrictSig
453 -- NB: not used for never-inline things; hence False
454 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
456 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
457 mkSigTy top_lvl rec_flag id rhs dmd_ty
458 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
460 never_inline = isNeverActive (idInlinePragma id)
461 maybe_id_dmd = idNewDemandInfo_maybe id
462 -- Is Nothing the first time round
465 | isTopLevel top_lvl = False -- Top level things don't get
466 -- their demandInfo set at all
467 | isRec rec_flag = False -- Ditto recursive things
468 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
469 | otherwise = True -- Optimistic, first time round
473 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
474 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
475 If the rhs is a thunk, we usually forget the CPR info, because
476 it is presumably shared (else it would have been inlined, and
477 so we'd lose sharing if w/w'd it into a function). E.g.
479 let r = case expensive of
483 If we marked r as having the CPR property, then we'd w/w into
485 let $wr = \() -> case expensive of
491 But now r is a thunk, which won't be inlined, so we are no further ahead.
494 f x = let r = case expensive of (a,b) -> (b,a)
495 in if foo r then r else (x,x)
497 Does f have the CPR property? Well, no.
499 However, if the strictness analyser has figured out (in a previous
500 iteration) that it's strict, then we DON'T need to forget the CPR info.
501 Instead we can retain the CPR info and do the thunk-splitting transform
502 (see WorkWrap.splitThunk).
504 This made a big difference to PrelBase.modInt, which had something like
505 modInt = \ x -> let r = ... -> I# v in
506 ...body strict in r...
507 r's RHS isn't a value yet; but modInt returns r in various branches, so
508 if r doesn't have the CPR property then neither does modInt
509 Another case I found in practice (in Complex.magnitude), looks like this:
510 let k = if ... then I# a else I# b
511 in ... body strict in k ....
512 (For this example, it doesn't matter whether k is returned as part of
513 the overall result; but it does matter that k's RHS has the CPR property.)
514 Left to itself, the simplifier will make a join point thus:
515 let $j k = ...body strict in k...
516 if ... then $j (I# a) else $j (I# b)
517 With thunk-splitting, we get instead
518 let $j x = let k = I#x in ...body strict in k...
519 in if ... then $j a else $j b
520 This is much better; there's a good chance the I# won't get allocated.
522 The difficulty with this is that we need the strictness type to
523 look at the body... but we now need the body to calculate the demand
524 on the variable, so we can decide whether its strictness type should
525 have a CPR in it or not. Simple solution:
526 a) use strictness info from the previous iteration
527 b) make sure we do at least 2 iterations, by doing a second
528 round for top-level non-recs. Top level recs will get at
529 least 2 iterations except for totally-bottom functions
530 which aren't very interesting anyway.
532 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
534 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
535 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
536 Demand info now has a 'Nothing' state, just like strictness info.
537 The analysis works from 'dangerous' towards a 'safe' state; so we
538 start with botSig for 'Nothing' strictness infos, and we start with
539 "yes, it's demanded" for 'Nothing' in the demand info. The
540 fixpoint iteration will sort it all out.
542 We can't start with 'not-demanded' because then consider
546 if ... then t else I# y else f x'
548 In the first iteration we'd have no demand info for x, so assume
549 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
550 we'd see that t was demanded, and so give it the CPR property, but by
551 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
552 setting the first time round, we say 'yes t is demanded' the first
555 However, this does mean that for non-recursive bindings we must
556 iterate twice to be sure of not getting over-optimistic CPR info,
557 in the case where t turns out to be not-demanded. This is handled
562 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
563 = (lazy_fv, mkStrictSig dmd_ty)
565 dmd_ty = DmdType strict_fv final_dmds res'
567 lazy_fv = filterUFM (not . isStrictDmd) fv
568 strict_fv = filterUFM isStrictDmd fv
569 -- We put the strict FVs in the DmdType of the Id, so
570 -- that at its call sites we unleash demands on its strict fvs.
571 -- An example is 'roll' in imaginary/wheel-sieve2
572 -- Something like this:
574 -- go y = if ... then roll (x-1) else x+1
577 -- We want to see that roll is strict in x, which is because
578 -- go is called. So we put the DmdEnv for x in go's DmdType.
581 -- f :: Int -> Int -> Int
582 -- f x y = let t = x+1
583 -- h z = if z==0 then t else
584 -- if z==1 then x+1 else
588 -- Calling h does indeed evaluate x, but we can only see
589 -- that if we unleash a demand on x at the call site for t.
591 -- Incidentally, here's a place where lambda-lifting h would
592 -- lose the cigar --- we couldn't see the joint strictness in t/x
595 -- We don't want to put *all* the fv's from the RHS into the
596 -- DmdType, because that makes fixpointing very slow --- the
597 -- DmdType gets full of lazy demands that are slow to converge.
599 final_dmds = setUnpackStrategy dmds
600 -- Set the unpacking strategy
603 RetCPR | ignore_cpr_info -> TopRes
605 ignore_cpr_info = not (exprIsHNF rhs || thunk_cpr_ok)
608 The unpack strategy determines whether we'll *really* unpack the argument,
609 or whether we'll just remember its strictness. If unpacking would give
610 rise to a *lot* of worker args, we may decide not to unpack after all.
613 setUnpackStrategy :: [Demand] -> [Demand]
615 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
617 go :: Int -- Max number of args available for sub-components of [Demand]
619 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
621 go n (Eval (Prod cs) : ds)
622 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
623 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
626 n' = n + 1 - non_abs_args
627 -- Add one to the budget 'cos we drop the top-level arg
628 non_abs_args = nonAbsentArgs cs
629 -- Delete # of non-absent args to which we'll now be committed
631 go n (d:ds) = d `cons` go n ds
634 cons d (n,ds) = (n, d:ds)
636 nonAbsentArgs :: [Demand] -> Int
638 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
639 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
643 %************************************************************************
645 \subsection{Strictness signatures and types}
647 %************************************************************************
650 splitDmdTy :: DmdType -> (Demand, DmdType)
651 -- Split off one function argument
652 -- We already have a suitable demand on all
653 -- free vars, so no need to add more!
654 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
655 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
659 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
661 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
662 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
663 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
665 addLazyFVs (DmdType fv ds res) lazy_fvs
666 = DmdType both_fv1 ds res
668 both_fv = (plusUFM_C both fv lazy_fvs)
669 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
670 -- This modifyEnv is vital. Consider
671 -- let f = \x -> (x,y)
673 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
674 -- demand with the bottom coming up from 'error'
676 -- I got a loop in the fixpointer without this, due to an interaction
677 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
679 -- = letrec g y = x `fatbar`
680 -- letrec h z = z + ...g...
683 -- In the initial iteration for f, f=Bot
684 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
685 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
686 -- places on its free variables. Suppose it places none. Then the
687 -- x `fatbar` ...call to h...
688 -- will give a x->V demand for x. That turns into a L demand for x,
689 -- which floats out of the defn for h. Without the modifyEnv, that
690 -- L demand doesn't get both'd with the Bot coming up from the inner
691 -- call to f. So we just get an L demand for x for g.
693 -- A better way to say this is that the lazy-fv filtering should give the
694 -- same answer as putting the lazy fv demands in the function's type.
696 annotateBndr :: DmdType -> Var -> (DmdType, Var)
697 -- The returned env has the var deleted
698 -- The returned var is annotated with demand info
699 -- No effect on the argument demands
700 annotateBndr dmd_ty@(DmdType fv ds res) var
701 | isTyVar var = (dmd_ty, var)
702 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
704 (fv', dmd) = removeFV fv var res
706 annotateBndrs = mapAccumR annotateBndr
708 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
709 -- For lambdas we add the demand to the argument demands
710 -- Only called for Ids
712 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
714 (fv', dmd) = removeFV fv id res
715 hacked_dmd = argDemand dmd
716 -- This call to argDemand is vital, because otherwise we label
717 -- a lambda binder with demand 'B'. But in terms of calling
718 -- conventions that's Abs, because we don't pass it. But
719 -- when we do a w/w split we get
720 -- fw x = (\x y:B -> ...) x (error "oops")
721 -- And then the simplifier things the 'B' is a strict demand
722 -- and evaluates the (error "oops"). Sigh
724 removeFV fv id res = (fv', zapUnlifted id dmd)
726 fv' = fv `delVarEnv` id
727 dmd = lookupVarEnv fv id `orElse` deflt
728 deflt | isBotRes res = Bot
731 -- For unlifted-type variables, we are only
732 -- interested in Bot/Abs/Box Abs
733 zapUnlifted is Bot = Bot
734 zapUnlifted id Abs = Abs
735 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
739 %************************************************************************
741 \subsection{Strictness signatures}
743 %************************************************************************
746 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
747 -- We use the SigEnv to tell us whether to
748 -- record info about a variable in the DmdEnv
749 -- We do so if it's a LocalId, but not top-level
751 -- The DmdEnv gives the demand on the free vars of the function
752 -- when it is given enough args to satisfy the strictness signature
754 emptySigEnv = emptyVarEnv
756 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
757 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
759 extendSigEnvList = extendVarEnvList
761 extendSigsWithLam :: SigEnv -> Id -> SigEnv
762 -- Extend the SigEnv when we meet a lambda binder
763 -- If the binder is marked demanded with a product demand, then give it a CPR
764 -- signature, because in the likely event that this is a lambda on a fn defn
765 -- [we only use this when the lambda is being consumed with a call demand],
766 -- it'll be w/w'd and so it will be CPR-ish. E.g.
767 -- f = \x::(Int,Int). if ...strict in x... then
771 -- We want f to have the CPR property because x does, by the time f has been w/w'd
773 -- Also note that we only want to do this for something that
774 -- definitely has product type, else we may get over-optimistic
775 -- CPR results (e.g. from \x -> x!).
777 extendSigsWithLam sigs id
778 = case idNewDemandInfo_maybe id of
779 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
780 -- Optimistic in the Nothing case;
781 -- See notes [CPR-AND-STRICTNESS]
782 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
786 dmdTransform :: SigEnv -- The strictness environment
787 -> Id -- The function
788 -> Demand -- The demand on the function
789 -> DmdType -- The demand type of the function in this context
790 -- Returned DmdEnv includes the demand on
791 -- this function plus demand on its free variables
793 dmdTransform sigs var dmd
795 ------ DATA CONSTRUCTOR
796 | isDataConWorkId var -- Data constructor
798 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
799 DmdType _ _ con_res = dmd_ty
802 if arity == call_depth then -- Saturated, so unleash the demand
804 -- Important! If we Keep the constructor application, then
805 -- we need the demands the constructor places (always lazy)
806 -- If not, we don't need to. For example:
807 -- f p@(x,y) = (p,y) -- S(AL)
809 -- It's vital that we don't calculate Absent for a!
810 dmd_ds = case res_dmd of
811 Box (Eval ds) -> mapDmds box ds
815 -- ds can be empty, when we are just seq'ing the thing
816 -- If so we must make up a suitable bunch of demands
817 arg_ds = case dmd_ds of
818 Poly d -> replicate arity d
819 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
822 mkDmdType emptyDmdEnv arg_ds con_res
823 -- Must remember whether it's a product, hence con_res, not TopRes
827 ------ IMPORTED FUNCTION
828 | isGlobalId var, -- Imported function
829 let StrictSig dmd_ty = idNewStrictness var
830 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
835 ------ LOCAL LET/REC BOUND THING
836 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
838 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
839 | otherwise = deferType dmd_ty
840 -- NB: it's important to use deferType, and not just return topDmdType
841 -- Consider let { f x y = p + x } in f 1
842 -- The application isn't saturated, but we must nevertheless propagate
843 -- a lazy demand for p!
845 addVarDmd top_lvl fn_ty var dmd
847 ------ LOCAL NON-LET/REC BOUND THING
848 | otherwise -- Default case
852 (call_depth, res_dmd) = splitCallDmd dmd
856 %************************************************************************
860 %************************************************************************
863 splitCallDmd :: Demand -> (Int, Demand)
864 splitCallDmd (Call d) = case splitCallDmd d of
866 splitCallDmd d = (0, d)
868 vanillaCall :: Arity -> Demand
869 vanillaCall 0 = evalDmd
870 vanillaCall n = Call (vanillaCall (n-1))
872 deferType :: DmdType -> DmdType
873 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
874 -- Notice that we throw away info about both arguments and results
875 -- For example, f = let ... in \x -> x
876 -- We don't want to get a stricness type V->T for f.
879 deferEnv :: DmdEnv -> DmdEnv
880 deferEnv fv = mapVarEnv defer fv
884 argDemand :: Demand -> Demand
885 -- The 'Defer' demands are just Lazy at function boundaries
886 -- Ugly! Ask John how to improve it.
887 argDemand Top = lazyDmd
888 argDemand (Defer d) = lazyDmd
889 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
890 argDemand (Box Bot) = evalDmd
891 argDemand (Box d) = box (argDemand d)
892 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
897 -------------------------
898 -- Consider (if x then y else []) with demand V
899 -- Then the first branch gives {y->V} and the second
900 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
901 -- in the result env.
902 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
903 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
905 lub_fv = plusUFM_C lub fv1 fv2
906 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
907 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
908 -- lub is the identity for Bot
910 -- Extend the shorter argument list to match the longer
911 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
913 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
914 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
916 -----------------------------------
917 -- (t1 `bothType` t2) takes the argument/result info from t1,
918 -- using t2 just for its free-var info
919 -- NB: Don't forget about r2! It might be BotRes, which is
920 -- a bottom demand on all the in-scope variables.
921 -- Peter: can this be done more neatly?
922 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
923 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
925 both_fv = plusUFM_C both fv1 fv2
926 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
927 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
928 -- both is the identity for Abs
935 lubRes RetCPR RetCPR = RetCPR
936 lubRes r1 r2 = TopRes
938 -- If either diverges, the whole thing does
939 -- Otherwise take CPR info from the first
940 bothRes r1 BotRes = BotRes
945 modifyEnv :: Bool -- No-op if False
946 -> (Demand -> Demand) -- The zapper
947 -> DmdEnv -> DmdEnv -- Env1 and Env2
948 -> DmdEnv -> DmdEnv -- Transform this env
949 -- Zap anything in Env1 but not in Env2
950 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
952 modifyEnv need_to_modify zapper env1 env2 env
953 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
956 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
958 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
962 %************************************************************************
964 \subsection{LUB and BOTH}
966 %************************************************************************
969 lub :: Demand -> Demand -> Demand
972 lub Abs d2 = absLub d2
974 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
976 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
977 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
978 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
979 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
981 -- For the Eval case, we use these approximation rules
982 -- Box Bot <= Eval (Box Bot ...)
983 -- Box Top <= Defer (Box Bot ...)
984 -- Box (Eval ds) <= Eval (map Box ds)
985 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
986 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
987 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
988 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
989 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
991 lub (Box d1) (Box d2) = box (d1 `lub` d2)
992 lub d1@(Box _) d2 = d2 `lub` d1
994 lubs ds1 ds2 = zipWithDmds lub ds1 ds2
996 ---------------------
997 -- box is the smart constructor for Box
998 -- It computes <B,bot> & d
999 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
1000 -- Seems to be no point in allowing (Box (Call d))
1001 box (Call d) = Call d -- The odd man out. Why?
1003 box (Defer _) = lazyDmd
1004 box Top = lazyDmd -- Box Abs and Box Top
1005 box Abs = lazyDmd -- are the same <B,L>
1006 box d = Box d -- Bot, Eval
1009 defer :: Demand -> Demand
1011 -- defer is the smart constructor for Defer
1012 -- The idea is that (Defer ds) = <U(ds), L>
1014 -- It specifies what happens at a lazy function argument
1015 -- or a lambda; the L* operator
1016 -- Set the strictness part to L, but leave
1017 -- the boxity side unaffected
1018 -- It also ensures that Defer (Eval [LLLL]) = L
1023 defer (Call _) = lazyDmd -- Approximation here?
1024 defer (Box _) = lazyDmd
1025 defer (Defer ds) = Defer ds
1026 defer (Eval ds) = deferEval ds
1028 -- deferEval ds = defer (Eval ds)
1029 deferEval ds | allTop ds = Top
1030 | otherwise = Defer ds
1032 ---------------------
1033 absLub :: Demand -> Demand
1034 -- Computes (Abs `lub` d)
1035 -- For the Bot case consider
1036 -- f x y = if ... then x else error x
1037 -- Then for y we get Abs `lub` Bot, and we really
1042 absLub (Call _) = Top
1043 absLub (Box _) = Top
1044 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1045 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1047 absLubs = mapDmds absLub
1050 both :: Demand -> Demand -> Demand
1056 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1059 -- From 'error' itself we get demand Bot on x
1060 -- From the arg demand on x we get
1061 -- x :-> evalDmd = Box (Eval (Poly Abs))
1062 -- So we get Bot `both` Box (Eval (Poly Abs))
1063 -- = Seq Keep (Poly Bot)
1066 -- f x = if ... then error (fst x) else fst x
1067 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1069 -- which is what we want.
1072 both Top Bot = errDmd
1075 both Top (Box d) = Box d
1076 both Top (Call d) = Call d
1077 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1078 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1079 -- = defer (Eval (mapDmds (`both` Top) ds))
1080 = deferEval (mapDmds (`both` Top) ds)
1083 both (Box d1) (Box d2) = box (d1 `both` d2)
1084 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1085 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1086 both (Box d1) (Defer d2) = Box d1
1087 both d1@(Box _) d2 = d2 `both` d1
1089 both (Call d1) (Call d2) = Call (d1 `both` d2)
1090 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1091 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1092 both d1@(Call _) d2 = d1 `both` d1
1094 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1095 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1096 both d1@(Eval ds1) d2 = d2 `both` d1
1098 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1099 both d1@(Defer ds1) d2 = d2 `both` d1
1101 boths ds1 ds2 = zipWithDmds both ds1 ds2
1106 %************************************************************************
1108 \subsection{Miscellaneous
1110 %************************************************************************
1114 #ifdef OLD_STRICTNESS
1115 get_changes binds = vcat (map get_changes_bind binds)
1117 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1118 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1120 get_changes_pr (id,rhs)
1121 = get_changes_var id $$ get_changes_expr rhs
1124 | isId var = get_changes_str var $$ get_changes_dmd var
1127 get_changes_expr (Type t) = empty
1128 get_changes_expr (Var v) = empty
1129 get_changes_expr (Lit l) = empty
1130 get_changes_expr (Note n e) = get_changes_expr e
1131 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1132 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1133 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1134 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1136 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1139 | new_better && old_better = empty
1140 | new_better = message "BETTER"
1141 | old_better = message "WORSE"
1142 | otherwise = message "INCOMPARABLE"
1144 message word = text word <+> text "strictness for" <+> ppr id <+> info
1145 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1146 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1147 -- strictness analyser can't track
1148 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1149 old_better = old `betterStrictness` new
1150 new_better = new `betterStrictness` old
1153 | isUnLiftedType (idType id) = empty -- Not useful
1154 | new_better && old_better = empty
1155 | new_better = message "BETTER"
1156 | old_better = message "WORSE"
1157 | otherwise = message "INCOMPARABLE"
1159 message word = text word <+> text "demand for" <+> ppr id <+> info
1160 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1161 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1163 old = newDemand (idDemandInfo id)
1164 new_better = new `betterDemand` old
1165 old_better = old `betterDemand` new
1167 betterStrictness :: StrictSig -> StrictSig -> Bool
1168 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
1170 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
1172 betterDemand :: Demand -> Demand -> Bool
1173 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
1174 betterDemand d1 d2 = (d1 `lub` d2) == d2
1176 squashSig (StrictSig (DmdType fv ds res))
1177 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1179 -- squash just gets rid of call demands
1180 -- which the old analyser doesn't track
1181 squashDmd (Call d) = evalDmd
1182 squashDmd (Box d) = Box (squashDmd d)
1183 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1184 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)