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 Demand -- All of it
28 import CoreUtils ( exprIsHNF, exprIsTrivial )
29 import CoreArity ( exprArity )
30 import DataCon ( dataConTyCon )
31 import TyCon ( isProductTyCon, isRecursiveTyCon )
32 import Id ( Id, idType, idInlineActivation,
33 isDataConWorkId, isGlobalId, idArity,
34 idStrictness, idStrictness_maybe,
35 setIdStrictness, idDemandInfo,
41 import TysWiredIn ( unboxedPairDataCon )
42 import TysPrim ( realWorldStatePrimTy )
43 import UniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly,
44 keysUFM, minusUFM, ufmToList, filterUFM )
45 import Type ( isUnLiftedType, coreEqType, splitTyConApp_maybe )
46 import Coercion ( coercionKind )
47 import Util ( mapAndUnzip, lengthIs )
48 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive,
50 import Maybes ( orElse, expectJust )
51 import ErrUtils ( showPass )
59 * set a noinline pragma on bottoming Ids
61 * Consider f x = x+1 `fatbar` error (show x)
62 We'd like to unbox x, even if that means reboxing it in the error case.
65 %************************************************************************
67 \subsection{Top level stuff}
69 %************************************************************************
72 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
73 dmdAnalPgm dflags binds
75 let { binds_plus_dmds = do_prog binds } ;
76 return binds_plus_dmds
79 do_prog :: [CoreBind] -> [CoreBind]
80 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
82 dmdAnalTopBind :: SigEnv
85 dmdAnalTopBind sigs (NonRec id rhs)
87 ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs)
88 (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs1)
89 -- Do two passes to improve CPR information
90 -- See comments with ignore_cpr_info in mk_sig_ty
91 -- and with extendSigsWithLam
93 (sigs2, NonRec id2 rhs2)
95 dmdAnalTopBind sigs (Rec pairs)
97 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
98 -- We get two iterations automatically
99 -- c.f. the NonRec case above
105 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
106 -- Analyse the RHS and return
107 -- a) appropriate strictness info
108 -- b) the unfolding (decorated with stricntess info)
112 call_dmd = vanillaCall (exprArity rhs)
113 (_, rhs1) = dmdAnal emptySigEnv call_dmd rhs
114 (rhs_ty, rhs2) = dmdAnal emptySigEnv call_dmd rhs1
115 sig = mkTopSigTy rhs rhs_ty
116 -- Do two passes; see notes with extendSigsWithLam
117 -- Otherwise we get bogus CPR info for constructors like
118 -- newtype T a = MkT a
119 -- The constructor looks like (\x::T a -> x), modulo the coerce
120 -- extendSigsWithLam will optimistically give x a CPR tag the
121 -- first time, which is wrong in the end.
124 %************************************************************************
126 \subsection{The analyser itself}
128 %************************************************************************
131 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
133 dmdAnal sigs Abs e = (topDmdType, e)
136 | not (isStrictDmd dmd)
138 (res_ty, e') = dmdAnal sigs evalDmd e
140 (deferType res_ty, e')
141 -- It's important not to analyse e with a lazy demand because
142 -- a) When we encounter case s of (a,b) ->
143 -- we demand s with U(d1d2)... but if the overall demand is lazy
144 -- that is wrong, and we'd need to reduce the demand on s,
145 -- which is inconvenient
146 -- b) More important, consider
147 -- f (let x = R in x+x), where f is lazy
148 -- We still want to mark x as demanded, because it will be when we
149 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
150 -- just mark x as Lazy
151 -- c) The application rule wouldn't be right either
152 -- Evaluating (f x) in a L demand does *not* cause
153 -- evaluation of f in a C(L) demand!
156 dmdAnal sigs dmd (Lit lit)
157 = (topDmdType, Lit lit)
159 dmdAnal sigs dmd (Var var)
160 = (dmdTransform sigs var dmd, Var var)
162 dmdAnal sigs dmd (Cast e co)
163 = (dmd_ty, Cast e' co)
165 (dmd_ty, e') = dmdAnal sigs dmd' e
166 to_co = snd (coercionKind co)
168 | Just (tc, args) <- splitTyConApp_maybe to_co
169 , isRecursiveTyCon tc = evalDmd
171 -- This coerce usually arises from a recursive
172 -- newtype, and we don't want to look inside them
173 -- for exactly the same reason that we don't look
174 -- inside recursive products -- we might not reach
175 -- a fixpoint. So revert to a vanilla Eval demand
177 dmdAnal sigs dmd (Note n e)
178 = (dmd_ty, Note n e')
180 (dmd_ty, e') = dmdAnal sigs dmd e
182 dmdAnal sigs dmd (App fun (Type ty))
183 = (fun_ty, App fun' (Type ty))
185 (fun_ty, fun') = dmdAnal sigs dmd fun
187 -- Lots of the other code is there to make this
188 -- beautiful, compositional, application rule :-)
189 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
190 = let -- [Type arg handled above]
191 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
192 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
193 (arg_dmd, res_ty) = splitDmdTy fun_ty
195 (res_ty `bothType` arg_ty, App fun' arg')
197 dmdAnal sigs dmd (Lam var body)
200 (body_ty, body') = dmdAnal sigs dmd body
202 (body_ty, Lam var body')
204 | Call body_dmd <- dmd -- A call demand: good!
206 sigs' = extendSigsWithLam sigs var
207 (body_ty, body') = dmdAnal sigs' body_dmd body
208 (lam_ty, var') = annotateLamIdBndr body_ty var
210 (lam_ty, Lam var' body')
212 | otherwise -- Not enough demand on the lambda; but do the body
213 = let -- anyway to annotate it and gather free var info
214 (body_ty, body') = dmdAnal sigs evalDmd body
215 (lam_ty, var') = annotateLamIdBndr body_ty var
217 (deferType lam_ty, Lam var' body')
219 dmdAnal sigs dmd (Case scrut case_bndr ty [alt@(DataAlt dc,bndrs,rhs)])
220 | let tycon = dataConTyCon dc
221 , isProductTyCon tycon
222 , not (isRecursiveTyCon tycon)
224 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
225 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
226 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
227 (_, bndrs', _) = alt'
228 case_bndr_sig = cprSig
229 -- Inside the alternative, the case binder has the CPR property.
230 -- Meaning that a case on it will successfully cancel.
232 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
235 -- We want f to have the CPR property:
236 -- f b x = case fw b x of { r -> I# r }
237 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
240 -- Figure out whether the demand on the case binder is used, and use
241 -- that to set the scrut_dmd. This is utterly essential.
242 -- Consider f x = case x of y { (a,b) -> k y a }
243 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
244 -- worker, so the worker will rebuild
245 -- x = (a, absent-error)
246 -- and that'll crash.
247 -- So at one stage I had:
248 -- dead_case_bndr = isAbsentDmd (idDemandInfo case_bndr')
249 -- keepity | dead_case_bndr = Drop
250 -- | otherwise = Keep
253 -- case x of y { (a,b) -> h y + a }
254 -- where h : U(LL) -> T
255 -- The above code would compute a Keep for x, since y is not Abs, which is silly
256 -- The insight is, of course, that a demand on y is a demand on the
257 -- scrutinee, so we need to `both` it with the scrut demand
259 alt_dmd = Eval (Prod [idDemandInfo b | b <- bndrs', isId b])
260 scrut_dmd = alt_dmd `both`
261 idDemandInfo case_bndr'
263 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
265 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' ty [alt'])
267 dmdAnal sigs dmd (Case scrut case_bndr ty alts)
269 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
270 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
271 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
273 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
274 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' ty alts')
276 dmdAnal sigs dmd (Let (NonRec id rhs) body)
278 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
279 (body_ty, body') = dmdAnal sigs' dmd body
280 (body_ty1, id2) = annotateBndr body_ty id1
281 body_ty2 = addLazyFVs body_ty1 lazy_fv
283 -- If the actual demand is better than the vanilla call
284 -- demand, you might think that we might do better to re-analyse
285 -- the RHS with the stronger demand.
286 -- But (a) That seldom happens, because it means that *every* path in
287 -- the body of the let has to use that stronger demand
288 -- (b) It often happens temporarily in when fixpointing, because
289 -- the recursive function at first seems to place a massive demand.
290 -- But we don't want to go to extra work when the function will
291 -- probably iterate to something less demanding.
292 -- In practice, all the times the actual demand on id2 is more than
293 -- the vanilla call demand seem to be due to (b). So we don't
294 -- bother to re-analyse the RHS.
295 (body_ty2, Let (NonRec id2 rhs') body')
297 dmdAnal sigs dmd (Let (Rec pairs) body)
299 bndrs = map fst pairs
300 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
301 (body_ty, body') = dmdAnal sigs' dmd body
302 body_ty1 = addLazyFVs body_ty lazy_fv
304 sigs' `seq` body_ty `seq`
306 (body_ty2, _) = annotateBndrs body_ty1 bndrs
307 -- Don't bother to add demand info to recursive
308 -- binders as annotateBndr does;
309 -- being recursive, we can't treat them strictly.
310 -- But we do need to remove the binders from the result demand env
312 (body_ty2, Let (Rec pairs') body')
315 dmdAnalAlt sigs dmd (con,bndrs,rhs)
317 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
318 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
319 final_alt_ty | io_hack_reqd = alt_ty `lubType` topDmdType
322 -- There's a hack here for I/O operations. Consider
323 -- case foo x s of { (# s, r #) -> y }
324 -- Is this strict in 'y'. Normally yes, but what if 'foo' is an I/O
325 -- operation that simply terminates the program (not in an erroneous way)?
326 -- In that case we should not evaluate y before the call to 'foo'.
327 -- Hackish solution: spot the IO-like situation and add a virtual branch,
331 -- other -> return ()
332 -- So the 'y' isn't necessarily going to be evaluated
334 -- A more complete example where this shows up is:
335 -- do { let len = <expensive> ;
336 -- ; when (...) (exitWith ExitSuccess)
339 io_hack_reqd = con == DataAlt unboxedPairDataCon &&
340 idType (head bndrs) `coreEqType` realWorldStatePrimTy
342 (final_alt_ty, (con, bndrs', rhs'))
345 %************************************************************************
347 \subsection{Bindings}
349 %************************************************************************
352 dmdFix :: TopLevelFlag
353 -> SigEnv -- Does not include bindings for this binding
356 [(Id,CoreExpr)]) -- Binders annotated with stricness info
358 dmdFix top_lvl sigs orig_pairs
359 = loop 1 initial_sigs orig_pairs
361 bndrs = map fst orig_pairs
362 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
365 -> SigEnv -- Already contains the current sigs
367 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
370 = (sigs', lazy_fv, pairs')
371 -- Note: use pairs', not pairs. pairs' is the result of
372 -- processing the RHSs with sigs (= sigs'), whereas pairs
373 -- is the result of processing the RHSs with the *previous*
374 -- iteration of sigs.
376 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
377 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
378 text "env:" <+> ppr (ufmToList sigs),
379 text "binds:" <+> pprCoreBinding (Rec pairs)]))
380 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
381 -- The lazy_fv part is really important! orig_pairs has no strictness
382 -- info, including nothing about free vars. But if we have
383 -- letrec f = ....y..... in ...f...
384 -- where 'y' is free in f, we must record that y is mentioned,
385 -- otherwise y will get recorded as absent altogether
387 | otherwise = loop (n+1) sigs' pairs'
389 found_fixpoint = all (same_sig sigs sigs') bndrs
390 -- Use the new signature to do the next pair
391 -- The occurrence analyser has arranged them in a good order
392 -- so this can significantly reduce the number of iterations needed
393 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
395 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
396 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
398 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
399 ((sigs', lazy_fv'), pair')
402 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
403 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
404 -- old_sig = lookup sigs id
405 -- new_sig = lookup sigs' id
407 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
408 lookup sigs var = case lookupVarEnv sigs var of
411 -- Get an initial strictness signature from the Id
412 -- itself. That way we make use of earlier iterations
413 -- of the fixpoint algorithm. (Cunning plan.)
414 -- Note that the cunning plan extends to the DmdEnv too,
415 -- since it is part of the strictness signature
416 initialSig id = idStrictness_maybe id `orElse` botSig
418 dmdAnalRhs :: TopLevelFlag -> RecFlag
419 -> SigEnv -> (Id, CoreExpr)
420 -> (SigEnv, DmdEnv, (Id, CoreExpr))
421 -- Process the RHS of the binding, add the strictness signature
422 -- to the Id, and augment the environment with the signature as well.
424 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
425 = (sigs', lazy_fv, (id', rhs'))
427 arity = idArity id -- The idArity should be up to date
428 -- The simplifier was run just beforehand
429 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
430 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs), ppr id )
431 -- The RHS can be eta-reduced to just a variable,
432 -- in which case we should not complain.
433 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
434 id' = id `setIdStrictness` sig_ty
435 sigs' = extendSigEnv top_lvl sigs id sig_ty
438 %************************************************************************
440 \subsection{Strictness signatures and types}
442 %************************************************************************
445 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
446 -- Take a DmdType and turn it into a StrictSig
447 -- NB: not used for never-inline things; hence False
448 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
450 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
451 mkSigTy top_lvl rec_flag id rhs dmd_ty
452 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
454 never_inline = isNeverActive (idInlineActivation id)
455 maybe_id_dmd = idDemandInfo_maybe id
456 -- Is Nothing the first time round
459 | isTopLevel top_lvl = False -- Top level things don't get
460 -- their demandInfo set at all
461 | isRec rec_flag = False -- Ditto recursive things
462 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
463 | otherwise = True -- Optimistic, first time round
467 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
468 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
469 If the rhs is a thunk, we usually forget the CPR info, because
470 it is presumably shared (else it would have been inlined, and
471 so we'd lose sharing if w/w'd it into a function). E.g.
473 let r = case expensive of
477 If we marked r as having the CPR property, then we'd w/w into
479 let $wr = \() -> case expensive of
485 But now r is a thunk, which won't be inlined, so we are no further ahead.
488 f x = let r = case expensive of (a,b) -> (b,a)
489 in if foo r then r else (x,x)
491 Does f have the CPR property? Well, no.
493 However, if the strictness analyser has figured out (in a previous
494 iteration) that it's strict, then we DON'T need to forget the CPR info.
495 Instead we can retain the CPR info and do the thunk-splitting transform
496 (see WorkWrap.splitThunk).
498 This made a big difference to PrelBase.modInt, which had something like
499 modInt = \ x -> let r = ... -> I# v in
500 ...body strict in r...
501 r's RHS isn't a value yet; but modInt returns r in various branches, so
502 if r doesn't have the CPR property then neither does modInt
503 Another case I found in practice (in Complex.magnitude), looks like this:
504 let k = if ... then I# a else I# b
505 in ... body strict in k ....
506 (For this example, it doesn't matter whether k is returned as part of
507 the overall result; but it does matter that k's RHS has the CPR property.)
508 Left to itself, the simplifier will make a join point thus:
509 let $j k = ...body strict in k...
510 if ... then $j (I# a) else $j (I# b)
511 With thunk-splitting, we get instead
512 let $j x = let k = I#x in ...body strict in k...
513 in if ... then $j a else $j b
514 This is much better; there's a good chance the I# won't get allocated.
516 The difficulty with this is that we need the strictness type to
517 look at the body... but we now need the body to calculate the demand
518 on the variable, so we can decide whether its strictness type should
519 have a CPR in it or not. Simple solution:
520 a) use strictness info from the previous iteration
521 b) make sure we do at least 2 iterations, by doing a second
522 round for top-level non-recs. Top level recs will get at
523 least 2 iterations except for totally-bottom functions
524 which aren't very interesting anyway.
526 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
528 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
529 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
530 Demand info now has a 'Nothing' state, just like strictness info.
531 The analysis works from 'dangerous' towards a 'safe' state; so we
532 start with botSig for 'Nothing' strictness infos, and we start with
533 "yes, it's demanded" for 'Nothing' in the demand info. The
534 fixpoint iteration will sort it all out.
536 We can't start with 'not-demanded' because then consider
540 if ... then t else I# y else f x'
542 In the first iteration we'd have no demand info for x, so assume
543 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
544 we'd see that t was demanded, and so give it the CPR property, but by
545 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
546 setting the first time round, we say 'yes t is demanded' the first
549 However, this does mean that for non-recursive bindings we must
550 iterate twice to be sure of not getting over-optimistic CPR info,
551 in the case where t turns out to be not-demanded. This is handled
555 Note [NOINLINE and strictness]
556 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
557 The strictness analyser used to have a HACK which ensured that NOINLNE
558 things were not strictness-analysed. The reason was unsafePerformIO.
559 Left to itself, the strictness analyser would discover this strictness
561 unsafePerformIO: C(U(AV))
562 But then consider this sub-expression
563 unsafePerformIO (\s -> let r = f x in
564 case writeIORef v r s of (# s1, _ #) ->
566 The strictness analyser will now find that r is sure to be eval'd,
567 and may then hoist it out. This makes tests/lib/should_run/memo002
570 Solving this by making all NOINLINE things have no strictness info is overkill.
571 In particular, it's overkill for runST, which is perfectly respectable.
573 f x = runST (return x)
574 This should be strict in x.
576 So the new plan is to define unsafePerformIO using the 'lazy' combinator:
578 unsafePerformIO (IO m) = lazy (case m realWorld# of (# _, r #) -> r)
580 Remember, 'lazy' is a wired-in identity-function Id, of type a->a, which is
581 magically NON-STRICT, and is inlined after strictness analysis. So
582 unsafePerformIO will look non-strict, and that's what we want.
584 Now we don't need the hack in the strictness analyser. HOWEVER, this
585 decision does mean that even a NOINLINE function is not entirely
586 opaque: some aspect of its implementation leaks out, notably its
587 strictness. For example, if you have a function implemented by an
588 error stub, but which has RULES, you may want it not to be eliminated
593 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
594 = (lazy_fv, mkStrictSig dmd_ty)
595 -- Re unused never_inline, see Note [NOINLINE and strictness]
597 dmd_ty = DmdType strict_fv final_dmds res'
599 lazy_fv = filterUFM (not . isStrictDmd) fv
600 strict_fv = filterUFM isStrictDmd fv
601 -- We put the strict FVs in the DmdType of the Id, so
602 -- that at its call sites we unleash demands on its strict fvs.
603 -- An example is 'roll' in imaginary/wheel-sieve2
604 -- Something like this:
606 -- go y = if ... then roll (x-1) else x+1
609 -- We want to see that roll is strict in x, which is because
610 -- go is called. So we put the DmdEnv for x in go's DmdType.
613 -- f :: Int -> Int -> Int
614 -- f x y = let t = x+1
615 -- h z = if z==0 then t else
616 -- if z==1 then x+1 else
620 -- Calling h does indeed evaluate x, but we can only see
621 -- that if we unleash a demand on x at the call site for t.
623 -- Incidentally, here's a place where lambda-lifting h would
624 -- lose the cigar --- we couldn't see the joint strictness in t/x
627 -- We don't want to put *all* the fv's from the RHS into the
628 -- DmdType, because that makes fixpointing very slow --- the
629 -- DmdType gets full of lazy demands that are slow to converge.
631 final_dmds = setUnpackStrategy dmds
632 -- Set the unpacking strategy
635 RetCPR | ignore_cpr_info -> TopRes
637 ignore_cpr_info = not (exprIsHNF rhs || thunk_cpr_ok)
640 The unpack strategy determines whether we'll *really* unpack the argument,
641 or whether we'll just remember its strictness. If unpacking would give
642 rise to a *lot* of worker args, we may decide not to unpack after all.
645 setUnpackStrategy :: [Demand] -> [Demand]
647 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
649 go :: Int -- Max number of args available for sub-components of [Demand]
651 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
653 go n (Eval (Prod cs) : ds)
654 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
655 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
658 n' = n + 1 - non_abs_args
659 -- Add one to the budget 'cos we drop the top-level arg
660 non_abs_args = nonAbsentArgs cs
661 -- Delete # of non-absent args to which we'll now be committed
663 go n (d:ds) = d `cons` go n ds
666 cons d (n,ds) = (n, d:ds)
668 nonAbsentArgs :: [Demand] -> Int
670 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
671 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
675 %************************************************************************
677 \subsection{Strictness signatures and types}
679 %************************************************************************
682 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
684 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
685 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
686 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
688 addLazyFVs (DmdType fv ds res) lazy_fvs
689 = DmdType both_fv1 ds res
691 both_fv = (plusUFM_C both fv lazy_fvs)
692 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
693 -- This modifyEnv is vital. Consider
694 -- let f = \x -> (x,y)
696 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
697 -- demand with the bottom coming up from 'error'
699 -- I got a loop in the fixpointer without this, due to an interaction
700 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
702 -- = letrec g y = x `fatbar`
703 -- letrec h z = z + ...g...
706 -- In the initial iteration for f, f=Bot
707 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
708 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
709 -- places on its free variables. Suppose it places none. Then the
710 -- x `fatbar` ...call to h...
711 -- will give a x->V demand for x. That turns into a L demand for x,
712 -- which floats out of the defn for h. Without the modifyEnv, that
713 -- L demand doesn't get both'd with the Bot coming up from the inner
714 -- call to f. So we just get an L demand for x for g.
716 -- A better way to say this is that the lazy-fv filtering should give the
717 -- same answer as putting the lazy fv demands in the function's type.
719 annotateBndr :: DmdType -> Var -> (DmdType, Var)
720 -- The returned env has the var deleted
721 -- The returned var is annotated with demand info
722 -- No effect on the argument demands
723 annotateBndr dmd_ty@(DmdType fv ds res) var
724 | isTyVar var = (dmd_ty, var)
725 | otherwise = (DmdType fv' ds res, setIdDemandInfo var dmd)
727 (fv', dmd) = removeFV fv var res
729 annotateBndrs = mapAccumR annotateBndr
731 annotateLamIdBndr :: DmdType -- Demand type of body
732 -> Id -- Lambda binder
733 -> (DmdType, -- Demand type of lambda
734 Id) -- and binder annotated with demand
736 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
737 -- For lambdas we add the demand to the argument demands
738 -- Only called for Ids
740 (DmdType fv' (hacked_dmd:ds) res, setIdDemandInfo id hacked_dmd)
742 (fv', dmd) = removeFV fv id res
743 hacked_dmd = argDemand dmd
744 -- This call to argDemand is vital, because otherwise we label
745 -- a lambda binder with demand 'B'. But in terms of calling
746 -- conventions that's Abs, because we don't pass it. But
747 -- when we do a w/w split we get
748 -- fw x = (\x y:B -> ...) x (error "oops")
749 -- And then the simplifier things the 'B' is a strict demand
750 -- and evaluates the (error "oops"). Sigh
752 removeFV fv id res = (fv', zapUnlifted id dmd)
754 fv' = fv `delVarEnv` id
755 dmd = lookupVarEnv fv id `orElse` deflt
756 deflt | isBotRes res = Bot
759 -- For unlifted-type variables, we are only
760 -- interested in Bot/Abs/Box Abs
761 zapUnlifted is Bot = Bot
762 zapUnlifted id Abs = Abs
763 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
767 %************************************************************************
769 \subsection{Strictness signatures}
771 %************************************************************************
774 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
775 -- We use the SigEnv to tell us whether to
776 -- record info about a variable in the DmdEnv
777 -- We do so if it's a LocalId, but not top-level
779 -- The DmdEnv gives the demand on the free vars of the function
780 -- when it is given enough args to satisfy the strictness signature
782 emptySigEnv = emptyVarEnv
784 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
785 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
787 extendSigEnvList = extendVarEnvList
789 extendSigsWithLam :: SigEnv -> Id -> SigEnv
790 -- Extend the SigEnv when we meet a lambda binder
791 -- If the binder is marked demanded with a product demand, then give it a CPR
792 -- signature, because in the likely event that this is a lambda on a fn defn
793 -- [we only use this when the lambda is being consumed with a call demand],
794 -- it'll be w/w'd and so it will be CPR-ish. E.g.
795 -- f = \x::(Int,Int). if ...strict in x... then
799 -- We want f to have the CPR property because x does, by the time f has been w/w'd
801 -- Also note that we only want to do this for something that
802 -- definitely has product type, else we may get over-optimistic
803 -- CPR results (e.g. from \x -> x!).
805 extendSigsWithLam sigs id
806 = case idDemandInfo_maybe id of
807 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
808 -- Optimistic in the Nothing case;
809 -- See notes [CPR-AND-STRICTNESS]
810 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
814 dmdTransform :: SigEnv -- The strictness environment
815 -> Id -- The function
816 -> Demand -- The demand on the function
817 -> DmdType -- The demand type of the function in this context
818 -- Returned DmdEnv includes the demand on
819 -- this function plus demand on its free variables
821 dmdTransform sigs var dmd
823 ------ DATA CONSTRUCTOR
824 | isDataConWorkId var -- Data constructor
826 StrictSig dmd_ty = idStrictness var -- It must have a strictness sig
827 DmdType _ _ con_res = dmd_ty
830 if arity == call_depth then -- Saturated, so unleash the demand
832 -- Important! If we Keep the constructor application, then
833 -- we need the demands the constructor places (always lazy)
834 -- If not, we don't need to. For example:
835 -- f p@(x,y) = (p,y) -- S(AL)
837 -- It's vital that we don't calculate Absent for a!
838 dmd_ds = case res_dmd of
839 Box (Eval ds) -> mapDmds box ds
843 -- ds can be empty, when we are just seq'ing the thing
844 -- If so we must make up a suitable bunch of demands
845 arg_ds = case dmd_ds of
846 Poly d -> replicate arity d
847 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
850 mkDmdType emptyDmdEnv arg_ds con_res
851 -- Must remember whether it's a product, hence con_res, not TopRes
855 ------ IMPORTED FUNCTION
856 | isGlobalId var, -- Imported function
857 let StrictSig dmd_ty = idStrictness var
858 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
863 ------ LOCAL LET/REC BOUND THING
864 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
866 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
867 | otherwise = deferType dmd_ty
868 -- NB: it's important to use deferType, and not just return topDmdType
869 -- Consider let { f x y = p + x } in f 1
870 -- The application isn't saturated, but we must nevertheless propagate
871 -- a lazy demand for p!
873 addVarDmd top_lvl fn_ty var dmd
875 ------ LOCAL NON-LET/REC BOUND THING
876 | otherwise -- Default case
880 (call_depth, res_dmd) = splitCallDmd dmd
884 %************************************************************************
888 %************************************************************************
891 splitDmdTy :: DmdType -> (Demand, DmdType)
892 -- Split off one function argument
893 -- We already have a suitable demand on all
894 -- free vars, so no need to add more!
895 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
896 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
898 splitCallDmd :: Demand -> (Int, Demand)
899 splitCallDmd (Call d) = case splitCallDmd d of
901 splitCallDmd d = (0, d)
903 vanillaCall :: Arity -> Demand
904 vanillaCall 0 = evalDmd
905 vanillaCall n = Call (vanillaCall (n-1))
907 deferType :: DmdType -> DmdType
908 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
909 -- Notice that we throw away info about both arguments and results
910 -- For example, f = let ... in \x -> x
911 -- We don't want to get a stricness type V->T for f.
913 deferEnv :: DmdEnv -> DmdEnv
914 deferEnv fv = mapVarEnv defer fv
918 argDemand :: Demand -> Demand
919 -- The 'Defer' demands are just Lazy at function boundaries
920 -- Ugly! Ask John how to improve it.
921 argDemand Top = lazyDmd
922 argDemand (Defer d) = lazyDmd
923 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
924 argDemand (Box Bot) = evalDmd
925 argDemand (Box d) = box (argDemand d)
926 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
931 -------------------------
932 -- Consider (if x then y else []) with demand V
933 -- Then the first branch gives {y->V} and the second
934 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
935 -- in the result env.
936 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
937 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
939 lub_fv = plusUFM_C lub fv1 fv2
940 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
941 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
942 -- lub is the identity for Bot
944 -- Extend the shorter argument list to match the longer
945 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
947 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
948 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
950 -----------------------------------
951 -- (t1 `bothType` t2) takes the argument/result info from t1,
952 -- using t2 just for its free-var info
953 -- NB: Don't forget about r2! It might be BotRes, which is
954 -- a bottom demand on all the in-scope variables.
955 -- Peter: can this be done more neatly?
956 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
957 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
959 both_fv = plusUFM_C both fv1 fv2
960 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
961 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
962 -- both is the identity for Abs
969 lubRes RetCPR RetCPR = RetCPR
970 lubRes r1 r2 = TopRes
972 -- If either diverges, the whole thing does
973 -- Otherwise take CPR info from the first
974 bothRes r1 BotRes = BotRes
979 modifyEnv :: Bool -- No-op if False
980 -> (Demand -> Demand) -- The zapper
981 -> DmdEnv -> DmdEnv -- Env1 and Env2
982 -> DmdEnv -> DmdEnv -- Transform this env
983 -- Zap anything in Env1 but not in Env2
984 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
986 modifyEnv need_to_modify zapper env1 env2 env
987 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
990 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
992 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
996 %************************************************************************
998 \subsection{LUB and BOTH}
1000 %************************************************************************
1003 lub :: Demand -> Demand -> Demand
1006 lub Abs d2 = absLub d2
1008 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
1010 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
1011 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
1012 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
1013 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
1015 -- For the Eval case, we use these approximation rules
1016 -- Box Bot <= Eval (Box Bot ...)
1017 -- Box Top <= Defer (Box Bot ...)
1018 -- Box (Eval ds) <= Eval (map Box ds)
1019 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
1020 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
1021 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
1022 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
1023 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
1025 lub (Box d1) (Box d2) = box (d1 `lub` d2)
1026 lub d1@(Box _) d2 = d2 `lub` d1
1028 lubs ds1 ds2 = zipWithDmds lub ds1 ds2
1030 ---------------------
1031 -- box is the smart constructor for Box
1032 -- It computes <B,bot> & d
1033 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
1034 -- Seems to be no point in allowing (Box (Call d))
1035 box (Call d) = Call d -- The odd man out. Why?
1037 box (Defer _) = lazyDmd
1038 box Top = lazyDmd -- Box Abs and Box Top
1039 box Abs = lazyDmd -- are the same <B,L>
1040 box d = Box d -- Bot, Eval
1043 defer :: Demand -> Demand
1045 -- defer is the smart constructor for Defer
1046 -- The idea is that (Defer ds) = <U(ds), L>
1048 -- It specifies what happens at a lazy function argument
1049 -- or a lambda; the L* operator
1050 -- Set the strictness part to L, but leave
1051 -- the boxity side unaffected
1052 -- It also ensures that Defer (Eval [LLLL]) = L
1057 defer (Call _) = lazyDmd -- Approximation here?
1058 defer (Box _) = lazyDmd
1059 defer (Defer ds) = Defer ds
1060 defer (Eval ds) = deferEval ds
1062 -- deferEval ds = defer (Eval ds)
1063 deferEval ds | allTop ds = Top
1064 | otherwise = Defer ds
1066 ---------------------
1067 absLub :: Demand -> Demand
1068 -- Computes (Abs `lub` d)
1069 -- For the Bot case consider
1070 -- f x y = if ... then x else error x
1071 -- Then for y we get Abs `lub` Bot, and we really
1076 absLub (Call _) = Top
1077 absLub (Box _) = Top
1078 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1079 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1081 absLubs = mapDmds absLub
1084 both :: Demand -> Demand -> Demand
1088 -- Note [Bottom demands]
1091 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1092 both Bot (Defer ds) = Eval (mapDmds (`both` Bot) ds)
1095 both Top Bot = errDmd
1098 both Top (Box d) = Box d
1099 both Top (Call d) = Call d
1100 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1101 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1102 -- = defer (Eval (mapDmds (`both` Top) ds))
1103 = deferEval (mapDmds (`both` Top) ds)
1106 both (Box d1) (Box d2) = box (d1 `both` d2)
1107 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1108 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1109 both (Box d1) (Defer d2) = Box d1
1110 both d1@(Box _) d2 = d2 `both` d1
1112 both (Call d1) (Call d2) = Call (d1 `both` d2)
1113 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1114 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1115 both d1@(Call _) d2 = d1 `both` d1
1117 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1118 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1119 both d1@(Eval ds1) d2 = d2 `both` d1
1121 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1122 both d1@(Defer ds1) d2 = d2 `both` d1
1124 boths ds1 ds2 = zipWithDmds both ds1 ds2
1127 Note [Bottom demands]
1128 ~~~~~~~~~~~~~~~~~~~~~
1131 From 'error' itself we get demand Bot on x
1132 From the arg demand on x we get
1133 x :-> evalDmd = Box (Eval (Poly Abs))
1134 So we get Bot `both` Box (Eval (Poly Abs))
1135 = Seq Keep (Poly Bot)
1138 f x = if ... then error (fst x) else fst x
1139 Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1141 which is what we want.
1146 x :-> Bot `both` Defer [SA]
1147 and we want the Bot demand to cancel out the Defer
1148 so that we get Eval [SA]. Otherwise we'd have the odd
1150 f x = error (fst x) -- Strictness U(SA)b
1151 g x = error ('y':fst x) -- Strictness Tb