2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
11 -- The above warning supression flag is a temporary kludge.
12 -- While working on this module you are encouraged to remove it and fix
13 -- any warnings in the module. See
14 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
17 module DmdAnal ( dmdAnalPgm, dmdAnalTopRhs,
18 both {- needed by WwLib -}
21 #include "HsVersions.h"
23 import DynFlags ( DynFlags, DynFlag(..) )
24 import StaticFlags ( opt_MaxWorkerArgs )
25 import NewDemand -- All of it
28 import CoreUtils ( exprIsHNF, exprIsTrivial, exprArity )
29 import DataCon ( dataConTyCon )
30 import TyCon ( isProductTyCon, isRecursiveTyCon )
31 import Id ( Id, idType, idInlinePragma,
32 isDataConWorkId, isGlobalId, idArity,
34 idDemandInfo, idStrictness, idCprInfo, idName,
36 idNewStrictness, idNewStrictness_maybe,
37 setIdNewStrictness, idNewDemandInfo,
38 idNewDemandInfo_maybe,
42 import IdInfo ( newStrictnessFromOld, newDemand )
46 import TysWiredIn ( unboxedPairDataCon )
47 import TysPrim ( realWorldStatePrimTy )
48 import LazyUniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly,
49 keysUFM, minusUFM, ufmToList, filterUFM )
50 import Type ( isUnLiftedType, coreEqType, splitTyConApp_maybe )
51 import Coercion ( coercionKind )
52 import CoreLint ( showPass, endPass )
53 import Util ( mapAndUnzip, lengthIs )
54 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive,
56 import Maybes ( orElse, expectJust )
64 * set a noinline pragma on bottoming Ids
66 * Consider f x = x+1 `fatbar` error (show x)
67 We'd like to unbox x, even if that means reboxing it in the error case.
70 %************************************************************************
72 \subsection{Top level stuff}
74 %************************************************************************
77 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
78 dmdAnalPgm dflags binds
80 showPass dflags "Demand analysis" ;
81 let { binds_plus_dmds = do_prog binds } ;
83 endPass dflags "Demand analysis"
84 Opt_D_dump_stranal binds_plus_dmds ;
86 -- Only if OLD_STRICTNESS is on, because only then is the old
87 -- strictness analyser run
88 let { dmd_changes = get_changes binds_plus_dmds } ;
89 printDump (text "Changes in demands" $$ dmd_changes) ;
91 return binds_plus_dmds
94 do_prog :: [CoreBind] -> [CoreBind]
95 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
97 dmdAnalTopBind :: SigEnv
100 dmdAnalTopBind sigs (NonRec id rhs)
102 ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs)
103 (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs1)
104 -- Do two passes to improve CPR information
105 -- See comments with ignore_cpr_info in mk_sig_ty
106 -- and with extendSigsWithLam
108 (sigs2, NonRec id2 rhs2)
110 dmdAnalTopBind sigs (Rec pairs)
112 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
113 -- We get two iterations automatically
114 -- c.f. the NonRec case above
120 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
121 -- Analyse the RHS and return
122 -- a) appropriate strictness info
123 -- b) the unfolding (decorated with stricntess info)
127 call_dmd = vanillaCall (exprArity rhs)
128 (_, rhs1) = dmdAnal emptySigEnv call_dmd rhs
129 (rhs_ty, rhs2) = dmdAnal emptySigEnv call_dmd rhs1
130 sig = mkTopSigTy rhs rhs_ty
131 -- Do two passes; see notes with extendSigsWithLam
132 -- Otherwise we get bogus CPR info for constructors like
133 -- newtype T a = MkT a
134 -- The constructor looks like (\x::T a -> x), modulo the coerce
135 -- extendSigsWithLam will optimistically give x a CPR tag the
136 -- first time, which is wrong in the end.
139 %************************************************************************
141 \subsection{The analyser itself}
143 %************************************************************************
146 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
148 dmdAnal sigs Abs e = (topDmdType, e)
151 | not (isStrictDmd dmd)
153 (res_ty, e') = dmdAnal sigs evalDmd e
155 (deferType res_ty, e')
156 -- It's important not to analyse e with a lazy demand because
157 -- a) When we encounter case s of (a,b) ->
158 -- we demand s with U(d1d2)... but if the overall demand is lazy
159 -- that is wrong, and we'd need to reduce the demand on s,
160 -- which is inconvenient
161 -- b) More important, consider
162 -- f (let x = R in x+x), where f is lazy
163 -- We still want to mark x as demanded, because it will be when we
164 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
165 -- just mark x as Lazy
166 -- c) The application rule wouldn't be right either
167 -- Evaluating (f x) in a L demand does *not* cause
168 -- evaluation of f in a C(L) demand!
171 dmdAnal sigs dmd (Lit lit)
172 = (topDmdType, Lit lit)
174 dmdAnal sigs dmd (Var var)
175 = (dmdTransform sigs var dmd, Var var)
177 dmdAnal sigs dmd (Cast e co)
178 = (dmd_ty, Cast e' co)
180 (dmd_ty, e') = dmdAnal sigs dmd' e
181 to_co = snd (coercionKind co)
183 | Just (tc, args) <- splitTyConApp_maybe to_co
184 , isRecursiveTyCon tc = evalDmd
186 -- This coerce usually arises from a recursive
187 -- newtype, and we don't want to look inside them
188 -- for exactly the same reason that we don't look
189 -- inside recursive products -- we might not reach
190 -- a fixpoint. So revert to a vanilla Eval demand
192 dmdAnal sigs dmd (Note n e)
193 = (dmd_ty, Note n e')
195 (dmd_ty, e') = dmdAnal sigs dmd e
197 dmdAnal sigs dmd (App fun (Type ty))
198 = (fun_ty, App fun' (Type ty))
200 (fun_ty, fun') = dmdAnal sigs dmd fun
202 -- Lots of the other code is there to make this
203 -- beautiful, compositional, application rule :-)
204 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
205 = let -- [Type arg handled above]
206 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
207 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
208 (arg_dmd, res_ty) = splitDmdTy fun_ty
210 (res_ty `bothType` arg_ty, App fun' arg')
212 dmdAnal sigs dmd (Lam var body)
215 (body_ty, body') = dmdAnal sigs dmd body
217 (body_ty, Lam var body')
219 | Call body_dmd <- dmd -- A call demand: good!
221 sigs' = extendSigsWithLam sigs var
222 (body_ty, body') = dmdAnal sigs' body_dmd body
223 (lam_ty, var') = annotateLamIdBndr body_ty var
225 (lam_ty, Lam var' body')
227 | otherwise -- Not enough demand on the lambda; but do the body
228 = let -- anyway to annotate it and gather free var info
229 (body_ty, body') = dmdAnal sigs evalDmd body
230 (lam_ty, var') = annotateLamIdBndr body_ty var
232 (deferType lam_ty, Lam var' body')
234 dmdAnal sigs dmd (Case scrut case_bndr ty [alt@(DataAlt dc,bndrs,rhs)])
235 | let tycon = dataConTyCon dc,
236 isProductTyCon tycon,
237 not (isRecursiveTyCon tycon)
239 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
240 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
241 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
242 (_, bndrs', _) = alt'
243 case_bndr_sig = cprSig
244 -- Inside the alternative, the case binder has the CPR property.
245 -- Meaning that a case on it will successfully cancel.
247 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
250 -- We want f to have the CPR property:
251 -- f b x = case fw b x of { r -> I# r }
252 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
255 -- Figure out whether the demand on the case binder is used, and use
256 -- that to set the scrut_dmd. This is utterly essential.
257 -- Consider f x = case x of y { (a,b) -> k y a }
258 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
259 -- worker, so the worker will rebuild
260 -- x = (a, absent-error)
261 -- and that'll crash.
262 -- So at one stage I had:
263 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
264 -- keepity | dead_case_bndr = Drop
265 -- | otherwise = Keep
268 -- case x of y { (a,b) -> h y + a }
269 -- where h : U(LL) -> T
270 -- The above code would compute a Keep for x, since y is not Abs, which is silly
271 -- The insight is, of course, that a demand on y is a demand on the
272 -- scrutinee, so we need to `both` it with the scrut demand
274 alt_dmd = Eval (Prod [idNewDemandInfo b | b <- bndrs', isId b])
275 scrut_dmd = alt_dmd `both`
276 idNewDemandInfo case_bndr'
278 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
280 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' ty [alt'])
282 dmdAnal sigs dmd (Case scrut case_bndr ty alts)
284 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
285 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
286 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
288 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
289 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' ty alts')
291 dmdAnal sigs dmd (Let (NonRec id rhs) body)
293 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
294 (body_ty, body') = dmdAnal sigs' dmd body
295 (body_ty1, id2) = annotateBndr body_ty id1
296 body_ty2 = addLazyFVs body_ty1 lazy_fv
298 -- If the actual demand is better than the vanilla call
299 -- demand, you might think that we might do better to re-analyse
300 -- the RHS with the stronger demand.
301 -- But (a) That seldom happens, because it means that *every* path in
302 -- the body of the let has to use that stronger demand
303 -- (b) It often happens temporarily in when fixpointing, because
304 -- the recursive function at first seems to place a massive demand.
305 -- But we don't want to go to extra work when the function will
306 -- probably iterate to something less demanding.
307 -- In practice, all the times the actual demand on id2 is more than
308 -- the vanilla call demand seem to be due to (b). So we don't
309 -- bother to re-analyse the RHS.
310 (body_ty2, Let (NonRec id2 rhs') body')
312 dmdAnal sigs dmd (Let (Rec pairs) body)
314 bndrs = map fst pairs
315 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
316 (body_ty, body') = dmdAnal sigs' dmd body
317 body_ty1 = addLazyFVs body_ty lazy_fv
319 sigs' `seq` body_ty `seq`
321 (body_ty2, _) = annotateBndrs body_ty1 bndrs
322 -- Don't bother to add demand info to recursive
323 -- binders as annotateBndr does;
324 -- being recursive, we can't treat them strictly.
325 -- But we do need to remove the binders from the result demand env
327 (body_ty2, Let (Rec pairs') body')
330 dmdAnalAlt sigs dmd (con,bndrs,rhs)
332 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
333 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
334 final_alt_ty | io_hack_reqd = alt_ty `lubType` topDmdType
337 -- There's a hack here for I/O operations. Consider
338 -- case foo x s of { (# s, r #) -> y }
339 -- Is this strict in 'y'. Normally yes, but what if 'foo' is an I/O
340 -- operation that simply terminates the program (not in an erroneous way)?
341 -- In that case we should not evaluate y before the call to 'foo'.
342 -- Hackish solution: spot the IO-like situation and add a virtual branch,
346 -- other -> return ()
347 -- So the 'y' isn't necessarily going to be evaluated
349 -- A more complete example where this shows up is:
350 -- do { let len = <expensive> ;
351 -- ; when (...) (exitWith ExitSuccess)
354 io_hack_reqd = con == DataAlt unboxedPairDataCon &&
355 idType (head bndrs) `coreEqType` realWorldStatePrimTy
357 (final_alt_ty, (con, bndrs', rhs'))
360 %************************************************************************
362 \subsection{Bindings}
364 %************************************************************************
367 dmdFix :: TopLevelFlag
368 -> SigEnv -- Does not include bindings for this binding
371 [(Id,CoreExpr)]) -- Binders annotated with stricness info
373 dmdFix top_lvl sigs orig_pairs
374 = loop 1 initial_sigs orig_pairs
376 bndrs = map fst orig_pairs
377 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
380 -> SigEnv -- Already contains the current sigs
382 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
385 = (sigs', lazy_fv, pairs')
386 -- Note: use pairs', not pairs. pairs' is the result of
387 -- processing the RHSs with sigs (= sigs'), whereas pairs
388 -- is the result of processing the RHSs with the *previous*
389 -- iteration of sigs.
391 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
392 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
393 text "env:" <+> ppr (ufmToList sigs),
394 text "binds:" <+> pprCoreBinding (Rec pairs)]))
395 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
396 -- The lazy_fv part is really important! orig_pairs has no strictness
397 -- info, including nothing about free vars. But if we have
398 -- letrec f = ....y..... in ...f...
399 -- where 'y' is free in f, we must record that y is mentioned,
400 -- otherwise y will get recorded as absent altogether
402 | otherwise = loop (n+1) sigs' pairs'
404 found_fixpoint = all (same_sig sigs sigs') bndrs
405 -- Use the new signature to do the next pair
406 -- The occurrence analyser has arranged them in a good order
407 -- so this can significantly reduce the number of iterations needed
408 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
410 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
411 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
413 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
414 ((sigs', lazy_fv'), pair')
417 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
418 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
419 -- old_sig = lookup sigs id
420 -- new_sig = lookup sigs' id
422 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
423 lookup sigs var = case lookupVarEnv sigs var of
426 -- Get an initial strictness signature from the Id
427 -- itself. That way we make use of earlier iterations
428 -- of the fixpoint algorithm. (Cunning plan.)
429 -- Note that the cunning plan extends to the DmdEnv too,
430 -- since it is part of the strictness signature
431 initialSig id = idNewStrictness_maybe id `orElse` botSig
433 dmdAnalRhs :: TopLevelFlag -> RecFlag
434 -> SigEnv -> (Id, CoreExpr)
435 -> (SigEnv, DmdEnv, (Id, CoreExpr))
436 -- Process the RHS of the binding, add the strictness signature
437 -- to the Id, and augment the environment with the signature as well.
439 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
440 = (sigs', lazy_fv, (id', rhs'))
442 arity = idArity id -- The idArity should be up to date
443 -- The simplifier was run just beforehand
444 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
445 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs), ppr id )
446 -- The RHS can be eta-reduced to just a variable,
447 -- in which case we should not complain.
448 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
449 id' = id `setIdNewStrictness` sig_ty
450 sigs' = extendSigEnv top_lvl sigs id sig_ty
453 %************************************************************************
455 \subsection{Strictness signatures and types}
457 %************************************************************************
460 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
461 -- Take a DmdType and turn it into a StrictSig
462 -- NB: not used for never-inline things; hence False
463 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
465 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
466 mkSigTy top_lvl rec_flag id rhs dmd_ty
467 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
469 never_inline = isNeverActive (idInlinePragma id)
470 maybe_id_dmd = idNewDemandInfo_maybe id
471 -- Is Nothing the first time round
474 | isTopLevel top_lvl = False -- Top level things don't get
475 -- their demandInfo set at all
476 | isRec rec_flag = False -- Ditto recursive things
477 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
478 | otherwise = True -- Optimistic, first time round
482 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
483 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
484 If the rhs is a thunk, we usually forget the CPR info, because
485 it is presumably shared (else it would have been inlined, and
486 so we'd lose sharing if w/w'd it into a function). E.g.
488 let r = case expensive of
492 If we marked r as having the CPR property, then we'd w/w into
494 let $wr = \() -> case expensive of
500 But now r is a thunk, which won't be inlined, so we are no further ahead.
503 f x = let r = case expensive of (a,b) -> (b,a)
504 in if foo r then r else (x,x)
506 Does f have the CPR property? Well, no.
508 However, if the strictness analyser has figured out (in a previous
509 iteration) that it's strict, then we DON'T need to forget the CPR info.
510 Instead we can retain the CPR info and do the thunk-splitting transform
511 (see WorkWrap.splitThunk).
513 This made a big difference to PrelBase.modInt, which had something like
514 modInt = \ x -> let r = ... -> I# v in
515 ...body strict in r...
516 r's RHS isn't a value yet; but modInt returns r in various branches, so
517 if r doesn't have the CPR property then neither does modInt
518 Another case I found in practice (in Complex.magnitude), looks like this:
519 let k = if ... then I# a else I# b
520 in ... body strict in k ....
521 (For this example, it doesn't matter whether k is returned as part of
522 the overall result; but it does matter that k's RHS has the CPR property.)
523 Left to itself, the simplifier will make a join point thus:
524 let $j k = ...body strict in k...
525 if ... then $j (I# a) else $j (I# b)
526 With thunk-splitting, we get instead
527 let $j x = let k = I#x in ...body strict in k...
528 in if ... then $j a else $j b
529 This is much better; there's a good chance the I# won't get allocated.
531 The difficulty with this is that we need the strictness type to
532 look at the body... but we now need the body to calculate the demand
533 on the variable, so we can decide whether its strictness type should
534 have a CPR in it or not. Simple solution:
535 a) use strictness info from the previous iteration
536 b) make sure we do at least 2 iterations, by doing a second
537 round for top-level non-recs. Top level recs will get at
538 least 2 iterations except for totally-bottom functions
539 which aren't very interesting anyway.
541 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
543 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
544 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
545 Demand info now has a 'Nothing' state, just like strictness info.
546 The analysis works from 'dangerous' towards a 'safe' state; so we
547 start with botSig for 'Nothing' strictness infos, and we start with
548 "yes, it's demanded" for 'Nothing' in the demand info. The
549 fixpoint iteration will sort it all out.
551 We can't start with 'not-demanded' because then consider
555 if ... then t else I# y else f x'
557 In the first iteration we'd have no demand info for x, so assume
558 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
559 we'd see that t was demanded, and so give it the CPR property, but by
560 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
561 setting the first time round, we say 'yes t is demanded' the first
564 However, this does mean that for non-recursive bindings we must
565 iterate twice to be sure of not getting over-optimistic CPR info,
566 in the case where t turns out to be not-demanded. This is handled
570 Note [NOINLINE and strictness]
571 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
572 The strictness analyser used to have a HACK which ensured that NOINLNE
573 things were not strictness-analysed. The reason was unsafePerformIO.
574 Left to itself, the strictness analyser would discover this strictness
576 unsafePerformIO: C(U(AV))
577 But then consider this sub-expression
578 unsafePerformIO (\s -> let r = f x in
579 case writeIORef v r s of (# s1, _ #) ->
581 The strictness analyser will now find that r is sure to be eval'd,
582 and may then hoist it out. This makes tests/lib/should_run/memo002
585 Solving this by making all NOINLINE things have no strictness info is overkill.
586 In particular, it's overkill for runST, which is perfectly respectable.
588 f x = runST (return x)
589 This should be strict in x.
591 So the new plan is to define unsafePerformIO using the 'lazy' combinator:
593 unsafePerformIO (IO m) = lazy (case m realWorld# of (# _, r #) -> r)
595 Remember, 'lazy' is a wired-in identity-function Id, of type a->a, which is
596 magically NON-STRICT, and is inlined after strictness analysis. So
597 unsafePerformIO will look non-strict, and that's what we want.
599 Now we don't need the hack in the strictness analyser. HOWEVER, this
600 decision does mean that even a NOINLINE function is not entirely
601 opaque: some aspect of its implementation leaks out, notably its
602 strictness. For example, if you have a function implemented by an
603 error stub, but which has RULES, you may want it not to be eliminated
608 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
609 = (lazy_fv, mkStrictSig dmd_ty)
610 -- Re unused never_inline, see Note [NOINLINE and strictness]
612 dmd_ty = DmdType strict_fv final_dmds res'
614 lazy_fv = filterUFM (not . isStrictDmd) fv
615 strict_fv = filterUFM isStrictDmd fv
616 -- We put the strict FVs in the DmdType of the Id, so
617 -- that at its call sites we unleash demands on its strict fvs.
618 -- An example is 'roll' in imaginary/wheel-sieve2
619 -- Something like this:
621 -- go y = if ... then roll (x-1) else x+1
624 -- We want to see that roll is strict in x, which is because
625 -- go is called. So we put the DmdEnv for x in go's DmdType.
628 -- f :: Int -> Int -> Int
629 -- f x y = let t = x+1
630 -- h z = if z==0 then t else
631 -- if z==1 then x+1 else
635 -- Calling h does indeed evaluate x, but we can only see
636 -- that if we unleash a demand on x at the call site for t.
638 -- Incidentally, here's a place where lambda-lifting h would
639 -- lose the cigar --- we couldn't see the joint strictness in t/x
642 -- We don't want to put *all* the fv's from the RHS into the
643 -- DmdType, because that makes fixpointing very slow --- the
644 -- DmdType gets full of lazy demands that are slow to converge.
646 final_dmds = setUnpackStrategy dmds
647 -- Set the unpacking strategy
650 RetCPR | ignore_cpr_info -> TopRes
652 ignore_cpr_info = not (exprIsHNF rhs || thunk_cpr_ok)
655 The unpack strategy determines whether we'll *really* unpack the argument,
656 or whether we'll just remember its strictness. If unpacking would give
657 rise to a *lot* of worker args, we may decide not to unpack after all.
660 setUnpackStrategy :: [Demand] -> [Demand]
662 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
664 go :: Int -- Max number of args available for sub-components of [Demand]
666 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
668 go n (Eval (Prod cs) : ds)
669 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
670 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
673 n' = n + 1 - non_abs_args
674 -- Add one to the budget 'cos we drop the top-level arg
675 non_abs_args = nonAbsentArgs cs
676 -- Delete # of non-absent args to which we'll now be committed
678 go n (d:ds) = d `cons` go n ds
681 cons d (n,ds) = (n, d:ds)
683 nonAbsentArgs :: [Demand] -> Int
685 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
686 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
690 %************************************************************************
692 \subsection{Strictness signatures and types}
694 %************************************************************************
697 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
699 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
700 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
701 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
703 addLazyFVs (DmdType fv ds res) lazy_fvs
704 = DmdType both_fv1 ds res
706 both_fv = (plusUFM_C both fv lazy_fvs)
707 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
708 -- This modifyEnv is vital. Consider
709 -- let f = \x -> (x,y)
711 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
712 -- demand with the bottom coming up from 'error'
714 -- I got a loop in the fixpointer without this, due to an interaction
715 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
717 -- = letrec g y = x `fatbar`
718 -- letrec h z = z + ...g...
721 -- In the initial iteration for f, f=Bot
722 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
723 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
724 -- places on its free variables. Suppose it places none. Then the
725 -- x `fatbar` ...call to h...
726 -- will give a x->V demand for x. That turns into a L demand for x,
727 -- which floats out of the defn for h. Without the modifyEnv, that
728 -- L demand doesn't get both'd with the Bot coming up from the inner
729 -- call to f. So we just get an L demand for x for g.
731 -- A better way to say this is that the lazy-fv filtering should give the
732 -- same answer as putting the lazy fv demands in the function's type.
734 annotateBndr :: DmdType -> Var -> (DmdType, Var)
735 -- The returned env has the var deleted
736 -- The returned var is annotated with demand info
737 -- No effect on the argument demands
738 annotateBndr dmd_ty@(DmdType fv ds res) var
739 | isTyVar var = (dmd_ty, var)
740 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
742 (fv', dmd) = removeFV fv var res
744 annotateBndrs = mapAccumR annotateBndr
746 annotateLamIdBndr :: DmdType -- Demand type of body
747 -> Id -- Lambda binder
748 -> (DmdType, -- Demand type of lambda
749 Id) -- and binder annotated with demand
751 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
752 -- For lambdas we add the demand to the argument demands
753 -- Only called for Ids
755 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
757 (fv', dmd) = removeFV fv id res
758 hacked_dmd = argDemand dmd
759 -- This call to argDemand is vital, because otherwise we label
760 -- a lambda binder with demand 'B'. But in terms of calling
761 -- conventions that's Abs, because we don't pass it. But
762 -- when we do a w/w split we get
763 -- fw x = (\x y:B -> ...) x (error "oops")
764 -- And then the simplifier things the 'B' is a strict demand
765 -- and evaluates the (error "oops"). Sigh
767 removeFV fv id res = (fv', zapUnlifted id dmd)
769 fv' = fv `delVarEnv` id
770 dmd = lookupVarEnv fv id `orElse` deflt
771 deflt | isBotRes res = Bot
774 -- For unlifted-type variables, we are only
775 -- interested in Bot/Abs/Box Abs
776 zapUnlifted is Bot = Bot
777 zapUnlifted id Abs = Abs
778 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
782 %************************************************************************
784 \subsection{Strictness signatures}
786 %************************************************************************
789 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
790 -- We use the SigEnv to tell us whether to
791 -- record info about a variable in the DmdEnv
792 -- We do so if it's a LocalId, but not top-level
794 -- The DmdEnv gives the demand on the free vars of the function
795 -- when it is given enough args to satisfy the strictness signature
797 emptySigEnv = emptyVarEnv
799 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
800 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
802 extendSigEnvList = extendVarEnvList
804 extendSigsWithLam :: SigEnv -> Id -> SigEnv
805 -- Extend the SigEnv when we meet a lambda binder
806 -- If the binder is marked demanded with a product demand, then give it a CPR
807 -- signature, because in the likely event that this is a lambda on a fn defn
808 -- [we only use this when the lambda is being consumed with a call demand],
809 -- it'll be w/w'd and so it will be CPR-ish. E.g.
810 -- f = \x::(Int,Int). if ...strict in x... then
814 -- We want f to have the CPR property because x does, by the time f has been w/w'd
816 -- Also note that we only want to do this for something that
817 -- definitely has product type, else we may get over-optimistic
818 -- CPR results (e.g. from \x -> x!).
820 extendSigsWithLam sigs id
821 = case idNewDemandInfo_maybe id of
822 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
823 -- Optimistic in the Nothing case;
824 -- See notes [CPR-AND-STRICTNESS]
825 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
829 dmdTransform :: SigEnv -- The strictness environment
830 -> Id -- The function
831 -> Demand -- The demand on the function
832 -> DmdType -- The demand type of the function in this context
833 -- Returned DmdEnv includes the demand on
834 -- this function plus demand on its free variables
836 dmdTransform sigs var dmd
838 ------ DATA CONSTRUCTOR
839 | isDataConWorkId var -- Data constructor
841 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
842 DmdType _ _ con_res = dmd_ty
845 if arity == call_depth then -- Saturated, so unleash the demand
847 -- Important! If we Keep the constructor application, then
848 -- we need the demands the constructor places (always lazy)
849 -- If not, we don't need to. For example:
850 -- f p@(x,y) = (p,y) -- S(AL)
852 -- It's vital that we don't calculate Absent for a!
853 dmd_ds = case res_dmd of
854 Box (Eval ds) -> mapDmds box ds
858 -- ds can be empty, when we are just seq'ing the thing
859 -- If so we must make up a suitable bunch of demands
860 arg_ds = case dmd_ds of
861 Poly d -> replicate arity d
862 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
865 mkDmdType emptyDmdEnv arg_ds con_res
866 -- Must remember whether it's a product, hence con_res, not TopRes
870 ------ IMPORTED FUNCTION
871 | isGlobalId var, -- Imported function
872 let StrictSig dmd_ty = idNewStrictness var
873 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
878 ------ LOCAL LET/REC BOUND THING
879 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
881 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
882 | otherwise = deferType dmd_ty
883 -- NB: it's important to use deferType, and not just return topDmdType
884 -- Consider let { f x y = p + x } in f 1
885 -- The application isn't saturated, but we must nevertheless propagate
886 -- a lazy demand for p!
888 addVarDmd top_lvl fn_ty var dmd
890 ------ LOCAL NON-LET/REC BOUND THING
891 | otherwise -- Default case
895 (call_depth, res_dmd) = splitCallDmd dmd
899 %************************************************************************
903 %************************************************************************
906 splitDmdTy :: DmdType -> (Demand, DmdType)
907 -- Split off one function argument
908 -- We already have a suitable demand on all
909 -- free vars, so no need to add more!
910 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
911 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
913 splitCallDmd :: Demand -> (Int, Demand)
914 splitCallDmd (Call d) = case splitCallDmd d of
916 splitCallDmd d = (0, d)
918 vanillaCall :: Arity -> Demand
919 vanillaCall 0 = evalDmd
920 vanillaCall n = Call (vanillaCall (n-1))
922 deferType :: DmdType -> DmdType
923 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
924 -- Notice that we throw away info about both arguments and results
925 -- For example, f = let ... in \x -> x
926 -- We don't want to get a stricness type V->T for f.
928 deferEnv :: DmdEnv -> DmdEnv
929 deferEnv fv = mapVarEnv defer fv
933 argDemand :: Demand -> Demand
934 -- The 'Defer' demands are just Lazy at function boundaries
935 -- Ugly! Ask John how to improve it.
936 argDemand Top = lazyDmd
937 argDemand (Defer d) = lazyDmd
938 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
939 argDemand (Box Bot) = evalDmd
940 argDemand (Box d) = box (argDemand d)
941 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
946 -------------------------
947 -- Consider (if x then y else []) with demand V
948 -- Then the first branch gives {y->V} and the second
949 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
950 -- in the result env.
951 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
952 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
954 lub_fv = plusUFM_C lub fv1 fv2
955 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
956 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
957 -- lub is the identity for Bot
959 -- Extend the shorter argument list to match the longer
960 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
962 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
963 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
965 -----------------------------------
966 -- (t1 `bothType` t2) takes the argument/result info from t1,
967 -- using t2 just for its free-var info
968 -- NB: Don't forget about r2! It might be BotRes, which is
969 -- a bottom demand on all the in-scope variables.
970 -- Peter: can this be done more neatly?
971 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
972 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
974 both_fv = plusUFM_C both fv1 fv2
975 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
976 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
977 -- both is the identity for Abs
984 lubRes RetCPR RetCPR = RetCPR
985 lubRes r1 r2 = TopRes
987 -- If either diverges, the whole thing does
988 -- Otherwise take CPR info from the first
989 bothRes r1 BotRes = BotRes
994 modifyEnv :: Bool -- No-op if False
995 -> (Demand -> Demand) -- The zapper
996 -> DmdEnv -> DmdEnv -- Env1 and Env2
997 -> DmdEnv -> DmdEnv -- Transform this env
998 -- Zap anything in Env1 but not in Env2
999 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
1001 modifyEnv need_to_modify zapper env1 env2 env
1002 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
1005 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
1007 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
1011 %************************************************************************
1013 \subsection{LUB and BOTH}
1015 %************************************************************************
1018 lub :: Demand -> Demand -> Demand
1021 lub Abs d2 = absLub d2
1023 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
1025 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
1026 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
1027 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
1028 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
1030 -- For the Eval case, we use these approximation rules
1031 -- Box Bot <= Eval (Box Bot ...)
1032 -- Box Top <= Defer (Box Bot ...)
1033 -- Box (Eval ds) <= Eval (map Box ds)
1034 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
1035 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
1036 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
1037 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
1038 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
1040 lub (Box d1) (Box d2) = box (d1 `lub` d2)
1041 lub d1@(Box _) d2 = d2 `lub` d1
1043 lubs ds1 ds2 = zipWithDmds lub ds1 ds2
1045 ---------------------
1046 -- box is the smart constructor for Box
1047 -- It computes <B,bot> & d
1048 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
1049 -- Seems to be no point in allowing (Box (Call d))
1050 box (Call d) = Call d -- The odd man out. Why?
1052 box (Defer _) = lazyDmd
1053 box Top = lazyDmd -- Box Abs and Box Top
1054 box Abs = lazyDmd -- are the same <B,L>
1055 box d = Box d -- Bot, Eval
1058 defer :: Demand -> Demand
1060 -- defer is the smart constructor for Defer
1061 -- The idea is that (Defer ds) = <U(ds), L>
1063 -- It specifies what happens at a lazy function argument
1064 -- or a lambda; the L* operator
1065 -- Set the strictness part to L, but leave
1066 -- the boxity side unaffected
1067 -- It also ensures that Defer (Eval [LLLL]) = L
1072 defer (Call _) = lazyDmd -- Approximation here?
1073 defer (Box _) = lazyDmd
1074 defer (Defer ds) = Defer ds
1075 defer (Eval ds) = deferEval ds
1077 -- deferEval ds = defer (Eval ds)
1078 deferEval ds | allTop ds = Top
1079 | otherwise = Defer ds
1081 ---------------------
1082 absLub :: Demand -> Demand
1083 -- Computes (Abs `lub` d)
1084 -- For the Bot case consider
1085 -- f x y = if ... then x else error x
1086 -- Then for y we get Abs `lub` Bot, and we really
1091 absLub (Call _) = Top
1092 absLub (Box _) = Top
1093 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1094 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1096 absLubs = mapDmds absLub
1099 both :: Demand -> Demand -> Demand
1105 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1108 -- From 'error' itself we get demand Bot on x
1109 -- From the arg demand on x we get
1110 -- x :-> evalDmd = Box (Eval (Poly Abs))
1111 -- So we get Bot `both` Box (Eval (Poly Abs))
1112 -- = Seq Keep (Poly Bot)
1115 -- f x = if ... then error (fst x) else fst x
1116 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1118 -- which is what we want.
1121 both Top Bot = errDmd
1124 both Top (Box d) = Box d
1125 both Top (Call d) = Call d
1126 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1127 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1128 -- = defer (Eval (mapDmds (`both` Top) ds))
1129 = deferEval (mapDmds (`both` Top) ds)
1132 both (Box d1) (Box d2) = box (d1 `both` d2)
1133 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1134 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1135 both (Box d1) (Defer d2) = Box d1
1136 both d1@(Box _) d2 = d2 `both` d1
1138 both (Call d1) (Call d2) = Call (d1 `both` d2)
1139 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1140 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1141 both d1@(Call _) d2 = d1 `both` d1
1143 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1144 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1145 both d1@(Eval ds1) d2 = d2 `both` d1
1147 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1148 both d1@(Defer ds1) d2 = d2 `both` d1
1150 boths ds1 ds2 = zipWithDmds both ds1 ds2
1155 %************************************************************************
1157 \subsection{Miscellaneous
1159 %************************************************************************
1163 #ifdef OLD_STRICTNESS
1164 get_changes binds = vcat (map get_changes_bind binds)
1166 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1167 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1169 get_changes_pr (id,rhs)
1170 = get_changes_var id $$ get_changes_expr rhs
1173 | isId var = get_changes_str var $$ get_changes_dmd var
1176 get_changes_expr (Type t) = empty
1177 get_changes_expr (Var v) = empty
1178 get_changes_expr (Lit l) = empty
1179 get_changes_expr (Note n e) = get_changes_expr e
1180 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1181 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1182 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1183 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1185 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1188 | new_better && old_better = empty
1189 | new_better = message "BETTER"
1190 | old_better = message "WORSE"
1191 | otherwise = message "INCOMPARABLE"
1193 message word = text word <+> text "strictness for" <+> ppr id <+> info
1194 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1195 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1196 -- strictness analyser can't track
1197 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1198 old_better = old `betterStrictness` new
1199 new_better = new `betterStrictness` old
1202 | isUnLiftedType (idType id) = empty -- Not useful
1203 | new_better && old_better = empty
1204 | new_better = message "BETTER"
1205 | old_better = message "WORSE"
1206 | otherwise = message "INCOMPARABLE"
1208 message word = text word <+> text "demand for" <+> ppr id <+> info
1209 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1210 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1212 old = newDemand (idDemandInfo id)
1213 new_better = new `betterDemand` old
1214 old_better = old `betterDemand` new
1216 betterStrictness :: StrictSig -> StrictSig -> Bool
1217 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
1219 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
1221 betterDemand :: Demand -> Demand -> Bool
1222 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
1223 betterDemand d1 d2 = (d1 `lub` d2) == d2
1225 squashSig (StrictSig (DmdType fv ds res))
1226 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1228 -- squash just gets rid of call demands
1229 -- which the old analyser doesn't track
1230 squashDmd (Call d) = evalDmd
1231 squashDmd (Box d) = Box (squashDmd d)
1232 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1233 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)