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 )
17 import StaticFlags ( opt_MaxWorkerArgs )
18 import Demand -- All of it
21 import CoreUtils ( exprIsHNF, exprIsTrivial )
22 import CoreArity ( exprArity )
23 import DataCon ( dataConTyCon, dataConRepStrictness )
24 import TyCon ( isProductTyCon, isRecursiveTyCon )
25 import Id ( Id, idType, idInlineActivation,
26 isDataConWorkId, isGlobalId, idArity,
28 setIdStrictness, idDemandInfo, idUnfolding,
29 idDemandInfo_maybe, setIdDemandInfo
33 import TysWiredIn ( unboxedPairDataCon )
34 import TysPrim ( realWorldStatePrimTy )
35 import UniqFM ( addToUFM_Directly, lookupUFM_Directly,
37 import Type ( isUnLiftedType, coreEqType, splitTyConApp_maybe )
38 import Coercion ( coercionKind )
39 import Util ( mapAndUnzip, lengthIs, zipEqual )
40 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive,
41 RecFlag(..), isRec, isMarkedStrict )
42 import Maybes ( orElse, expectJust )
50 * set a noinline pragma on bottoming Ids
52 * Consider f x = x+1 `fatbar` error (show x)
53 We'd like to unbox x, even if that means reboxing it in the error case.
56 %************************************************************************
58 \subsection{Top level stuff}
60 %************************************************************************
63 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
66 let { binds_plus_dmds = do_prog binds } ;
67 return binds_plus_dmds
70 do_prog :: [CoreBind] -> [CoreBind]
71 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
73 dmdAnalTopBind :: SigEnv
76 dmdAnalTopBind sigs (NonRec id rhs)
78 ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs)
79 (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs1)
80 -- Do two passes to improve CPR information
81 -- See comments with ignore_cpr_info in mk_sig_ty
82 -- and with extendSigsWithLam
84 (sigs2, NonRec id2 rhs2)
86 dmdAnalTopBind sigs (Rec pairs)
88 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
89 -- We get two iterations automatically
90 -- c.f. the NonRec case above
96 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
97 -- Analyse the RHS and return
98 -- a) appropriate strictness info
99 -- b) the unfolding (decorated with stricntess info)
103 call_dmd = vanillaCall (exprArity rhs)
104 (_, rhs1) = dmdAnal emptySigEnv call_dmd rhs
105 (rhs_ty, rhs2) = dmdAnal emptySigEnv call_dmd rhs1
106 sig = mkTopSigTy rhs rhs_ty
107 -- Do two passes; see notes with extendSigsWithLam
108 -- Otherwise we get bogus CPR info for constructors like
109 -- newtype T a = MkT a
110 -- The constructor looks like (\x::T a -> x), modulo the coerce
111 -- extendSigsWithLam will optimistically give x a CPR tag the
112 -- first time, which is wrong in the end.
115 %************************************************************************
117 \subsection{The analyser itself}
119 %************************************************************************
122 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
124 dmdAnal _ Abs e = (topDmdType, e)
127 | not (isStrictDmd dmd)
129 (res_ty, e') = dmdAnal sigs evalDmd e
131 (deferType res_ty, e')
132 -- It's important not to analyse e with a lazy demand because
133 -- a) When we encounter case s of (a,b) ->
134 -- we demand s with U(d1d2)... but if the overall demand is lazy
135 -- that is wrong, and we'd need to reduce the demand on s,
136 -- which is inconvenient
137 -- b) More important, consider
138 -- f (let x = R in x+x), where f is lazy
139 -- We still want to mark x as demanded, because it will be when we
140 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
141 -- just mark x as Lazy
142 -- c) The application rule wouldn't be right either
143 -- Evaluating (f x) in a L demand does *not* cause
144 -- evaluation of f in a C(L) demand!
147 dmdAnal _ _ (Lit lit) = (topDmdType, Lit lit)
148 dmdAnal _ _ (Type ty) = (topDmdType, Type ty) -- Doesn't happen, in fact
150 dmdAnal sigs dmd (Var var)
151 = (dmdTransform sigs var dmd, Var var)
153 dmdAnal sigs dmd (Cast e co)
154 = (dmd_ty, Cast e' co)
156 (dmd_ty, e') = dmdAnal sigs dmd' e
157 to_co = snd (coercionKind co)
159 | Just (tc, _) <- splitTyConApp_maybe to_co
160 , isRecursiveTyCon tc = evalDmd
162 -- This coerce usually arises from a recursive
163 -- newtype, and we don't want to look inside them
164 -- for exactly the same reason that we don't look
165 -- inside recursive products -- we might not reach
166 -- a fixpoint. So revert to a vanilla Eval demand
168 dmdAnal sigs dmd (Note n e)
169 = (dmd_ty, Note n e')
171 (dmd_ty, e') = dmdAnal sigs dmd e
173 dmdAnal sigs dmd (App fun (Type ty))
174 = (fun_ty, App fun' (Type ty))
176 (fun_ty, fun') = dmdAnal sigs dmd fun
178 -- Lots of the other code is there to make this
179 -- beautiful, compositional, application rule :-)
180 dmdAnal sigs dmd (App fun arg) -- Non-type arguments
181 = let -- [Type arg handled above]
182 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
183 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
184 (arg_dmd, res_ty) = splitDmdTy fun_ty
186 (res_ty `bothType` arg_ty, App fun' arg')
188 dmdAnal sigs dmd (Lam var body)
191 (body_ty, body') = dmdAnal sigs dmd body
193 (body_ty, Lam var body')
195 | Call body_dmd <- dmd -- A call demand: good!
197 sigs' = extendSigsWithLam sigs var
198 (body_ty, body') = dmdAnal sigs' body_dmd body
199 (lam_ty, var') = annotateLamIdBndr sigs body_ty var
201 (lam_ty, Lam var' body')
203 | otherwise -- Not enough demand on the lambda; but do the body
204 = let -- anyway to annotate it and gather free var info
205 (body_ty, body') = dmdAnal sigs evalDmd body
206 (lam_ty, var') = annotateLamIdBndr sigs body_ty var
208 (deferType lam_ty, Lam var' body')
210 dmdAnal sigs dmd (Case scrut case_bndr ty [alt@(DataAlt dc, _, _)])
211 | let tycon = dataConTyCon dc
212 , isProductTyCon tycon
213 , not (isRecursiveTyCon tycon)
215 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
216 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
217 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
218 (_, bndrs', _) = alt'
219 case_bndr_sig = cprSig
220 -- Inside the alternative, the case binder has the CPR property.
221 -- Meaning that a case on it will successfully cancel.
223 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
226 -- We want f to have the CPR property:
227 -- f b x = case fw b x of { r -> I# r }
228 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
231 -- Figure out whether the demand on the case binder is used, and use
232 -- that to set the scrut_dmd. This is utterly essential.
233 -- Consider f x = case x of y { (a,b) -> k y a }
234 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
235 -- worker, so the worker will rebuild
236 -- x = (a, absent-error)
237 -- and that'll crash.
238 -- So at one stage I had:
239 -- dead_case_bndr = isAbsentDmd (idDemandInfo case_bndr')
240 -- keepity | dead_case_bndr = Drop
241 -- | otherwise = Keep
244 -- case x of y { (a,b) -> h y + a }
245 -- where h : U(LL) -> T
246 -- The above code would compute a Keep for x, since y is not Abs, which is silly
247 -- The insight is, of course, that a demand on y is a demand on the
248 -- scrutinee, so we need to `both` it with the scrut demand
250 alt_dmd = Eval (Prod [idDemandInfo b | b <- bndrs', isId b])
251 scrut_dmd = alt_dmd `both`
252 idDemandInfo case_bndr'
254 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
256 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' ty [alt'])
258 dmdAnal sigs dmd (Case scrut case_bndr ty alts)
260 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
261 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
262 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
264 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
265 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' ty alts')
267 dmdAnal sigs dmd (Let (NonRec id rhs) body)
269 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
270 (body_ty, body') = dmdAnal sigs' dmd body
271 (body_ty1, id2) = annotateBndr body_ty id1
272 body_ty2 = addLazyFVs body_ty1 lazy_fv
274 -- If the actual demand is better than the vanilla call
275 -- demand, you might think that we might do better to re-analyse
276 -- the RHS with the stronger demand.
277 -- But (a) That seldom happens, because it means that *every* path in
278 -- the body of the let has to use that stronger demand
279 -- (b) It often happens temporarily in when fixpointing, because
280 -- the recursive function at first seems to place a massive demand.
281 -- But we don't want to go to extra work when the function will
282 -- probably iterate to something less demanding.
283 -- In practice, all the times the actual demand on id2 is more than
284 -- the vanilla call demand seem to be due to (b). So we don't
285 -- bother to re-analyse the RHS.
286 (body_ty2, Let (NonRec id2 rhs') body')
288 dmdAnal sigs dmd (Let (Rec pairs) body)
290 bndrs = map fst pairs
291 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
292 (body_ty, body') = dmdAnal sigs' dmd body
293 body_ty1 = addLazyFVs body_ty lazy_fv
295 sigs' `seq` body_ty `seq`
297 (body_ty2, _) = annotateBndrs body_ty1 bndrs
298 -- Don't bother to add demand info to recursive
299 -- binders as annotateBndr does;
300 -- being recursive, we can't treat them strictly.
301 -- But we do need to remove the binders from the result demand env
303 (body_ty2, Let (Rec pairs') body')
306 dmdAnalAlt :: SigEnv -> Demand -> Alt Var -> (DmdType, Alt Var)
307 dmdAnalAlt sigs dmd (con,bndrs,rhs)
309 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
310 rhs_ty' = addDataConPatDmds con bndrs rhs_ty
311 (alt_ty, bndrs') = annotateBndrs rhs_ty' bndrs
312 final_alt_ty | io_hack_reqd = alt_ty `lubType` topDmdType
315 -- There's a hack here for I/O operations. Consider
316 -- case foo x s of { (# s, r #) -> y }
317 -- Is this strict in 'y'. Normally yes, but what if 'foo' is an I/O
318 -- operation that simply terminates the program (not in an erroneous way)?
319 -- In that case we should not evaluate y before the call to 'foo'.
320 -- Hackish solution: spot the IO-like situation and add a virtual branch,
324 -- other -> return ()
325 -- So the 'y' isn't necessarily going to be evaluated
327 -- A more complete example where this shows up is:
328 -- do { let len = <expensive> ;
329 -- ; when (...) (exitWith ExitSuccess)
332 io_hack_reqd = con == DataAlt unboxedPairDataCon &&
333 idType (head bndrs) `coreEqType` realWorldStatePrimTy
335 (final_alt_ty, (con, bndrs', rhs'))
337 addDataConPatDmds :: AltCon -> [Var] -> DmdType -> DmdType
338 -- See Note [Add demands for strict constructors]
339 addDataConPatDmds DEFAULT _ dmd_ty = dmd_ty
340 addDataConPatDmds (LitAlt _) _ dmd_ty = dmd_ty
341 addDataConPatDmds (DataAlt con) bndrs dmd_ty
342 = foldr add dmd_ty str_bndrs
344 add bndr dmd_ty = addVarDmd dmd_ty bndr seqDmd
345 str_bndrs = [ b | (b,s) <- zipEqual "addDataConPatBndrs"
347 (dataConRepStrictness con)
351 Note [Add demands for strict constructors]
352 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
353 Consider this program (due to Roman):
357 foo :: X Int -> Int -> Int
360 go i | i < n = a + go (i+1)
363 We want the worker for 'foo' too look like this:
365 $wfoo :: Int# -> Int# -> Int#
367 with the first argument unboxed, so that it is not eval'd each time
368 around the loop (which would otherwise happen, since 'foo' is not
369 strict in 'a'. It is sound for the wrapper to pass an unboxed arg
370 because X is strict, so its argument must be evaluated. And if we
371 *don't* pass an unboxed argument, we can't even repair it by adding a
374 foo (X a) n = a `seq` go 0
376 because the seq is discarded (very early) since X is strict!
378 There is the usual danger of reboxing, which as usual we ignore. But
379 if X is monomorphic, and has an UNPACK pragma, then this optimisation
380 is even more important. We don't want the wrapper to rebox an unboxed
381 argument, and pass an Int to $wfoo!
384 %************************************************************************
388 %************************************************************************
391 dmdTransform :: SigEnv -- The strictness environment
392 -> Id -- The function
393 -> Demand -- The demand on the function
394 -> DmdType -- The demand type of the function in this context
395 -- Returned DmdEnv includes the demand on
396 -- this function plus demand on its free variables
398 dmdTransform sigs var dmd
400 ------ DATA CONSTRUCTOR
401 | isDataConWorkId var -- Data constructor
403 StrictSig dmd_ty = idStrictness var -- It must have a strictness sig
404 DmdType _ _ con_res = dmd_ty
407 if arity == call_depth then -- Saturated, so unleash the demand
409 -- Important! If we Keep the constructor application, then
410 -- we need the demands the constructor places (always lazy)
411 -- If not, we don't need to. For example:
412 -- f p@(x,y) = (p,y) -- S(AL)
414 -- It's vital that we don't calculate Absent for a!
415 dmd_ds = case res_dmd of
416 Box (Eval ds) -> mapDmds box ds
420 -- ds can be empty, when we are just seq'ing the thing
421 -- If so we must make up a suitable bunch of demands
422 arg_ds = case dmd_ds of
423 Poly d -> replicate arity d
424 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
427 mkDmdType emptyDmdEnv arg_ds con_res
428 -- Must remember whether it's a product, hence con_res, not TopRes
432 ------ IMPORTED FUNCTION
433 | isGlobalId var, -- Imported function
434 let StrictSig dmd_ty = idStrictness var
435 = -- pprTrace "strict-sig" (ppr var $$ ppr dmd_ty) $
436 if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
441 ------ LOCAL LET/REC BOUND THING
442 | Just (StrictSig dmd_ty, top_lvl) <- lookupSigEnv sigs var
444 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
445 | otherwise = deferType dmd_ty
446 -- NB: it's important to use deferType, and not just return topDmdType
447 -- Consider let { f x y = p + x } in f 1
448 -- The application isn't saturated, but we must nevertheless propagate
449 -- a lazy demand for p!
451 if isTopLevel top_lvl then fn_ty -- Don't record top level things
452 else addVarDmd fn_ty var dmd
454 ------ LOCAL NON-LET/REC BOUND THING
455 | otherwise -- Default case
459 (call_depth, res_dmd) = splitCallDmd dmd
462 %************************************************************************
464 \subsection{Bindings}
466 %************************************************************************
469 dmdFix :: TopLevelFlag
470 -> SigEnv -- Does not include bindings for this binding
473 [(Id,CoreExpr)]) -- Binders annotated with stricness info
475 dmdFix top_lvl sigs orig_pairs
476 = loop 1 initial_sigs orig_pairs
478 bndrs = map fst orig_pairs
479 initial_sigs = addInitialSigs top_lvl sigs bndrs
482 -> SigEnv -- Already contains the current sigs
484 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
487 = (sigs', lazy_fv, pairs')
488 -- Note: return pairs', not pairs. pairs' is the result of
489 -- processing the RHSs with sigs (= sigs'), whereas pairs
490 -- is the result of processing the RHSs with the *previous*
491 -- iteration of sigs.
494 = pprTrace "dmdFix loop" (ppr n <+> (vcat
495 [ text "Sigs:" <+> ppr [ (id,lookupSigEnv sigs id, lookupSigEnv sigs' id)
497 text "env:" <+> ppr sigs,
498 text "binds:" <+> pprCoreBinding (Rec pairs)]))
499 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
500 -- The lazy_fv part is really important! orig_pairs has no strictness
501 -- info, including nothing about free vars. But if we have
502 -- letrec f = ....y..... in ...f...
503 -- where 'y' is free in f, we must record that y is mentioned,
504 -- otherwise y will get recorded as absent altogether
507 = loop (n+1) (setNonVirgin sigs') pairs'
509 found_fixpoint = all (same_sig sigs sigs') bndrs
510 -- Use the new signature to do the next pair
511 -- The occurrence analyser has arranged them in a good order
512 -- so this can significantly reduce the number of iterations needed
513 ((sigs',lazy_fv), pairs') = mapAccumL my_downRhs (sigs, emptyDmdEnv) pairs
515 my_downRhs (sigs,lazy_fv) (id,rhs) = ((sigs', lazy_fv'), pair')
517 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
518 lazy_fv' = plusVarEnv_C both lazy_fv lazy_fv1
520 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
521 lookup sigs var = case lookupSigEnv sigs var of
523 Nothing -> pprPanic "dmdFix" (ppr var)
525 dmdAnalRhs :: TopLevelFlag -> RecFlag
526 -> SigEnv -> (Id, CoreExpr)
527 -> (SigEnv, DmdEnv, (Id, CoreExpr))
528 -- Process the RHS of the binding, add the strictness signature
529 -- to the Id, and augment the environment with the signature as well.
531 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
532 = (sigs', lazy_fv, (id', rhs'))
534 arity = idArity id -- The idArity should be up to date
535 -- The simplifier was run just beforehand
536 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
537 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs), ppr id )
538 -- The RHS can be eta-reduced to just a variable,
539 -- in which case we should not complain.
540 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
541 id' = id `setIdStrictness` sig_ty
542 sigs' = extendSigEnv top_lvl sigs id sig_ty
546 %************************************************************************
548 \subsection{Strictness signatures and types}
550 %************************************************************************
553 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
554 -- Take a DmdType and turn it into a StrictSig
555 -- NB: not used for never-inline things; hence False
556 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
558 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
559 mkSigTy top_lvl rec_flag id rhs dmd_ty
560 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
562 never_inline = isNeverActive (idInlineActivation id)
563 maybe_id_dmd = idDemandInfo_maybe id
564 -- Is Nothing the first time round
567 | isTopLevel top_lvl = False -- Top level things don't get
568 -- their demandInfo set at all
569 | isRec rec_flag = False -- Ditto recursive things
570 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
571 | otherwise = True -- Optimistic, first time round
575 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
576 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
577 If the rhs is a thunk, we usually forget the CPR info, because
578 it is presumably shared (else it would have been inlined, and
579 so we'd lose sharing if w/w'd it into a function). E.g.
581 let r = case expensive of
585 If we marked r as having the CPR property, then we'd w/w into
587 let $wr = \() -> case expensive of
593 But now r is a thunk, which won't be inlined, so we are no further ahead.
596 f x = let r = case expensive of (a,b) -> (b,a)
597 in if foo r then r else (x,x)
599 Does f have the CPR property? Well, no.
601 However, if the strictness analyser has figured out (in a previous
602 iteration) that it's strict, then we DON'T need to forget the CPR info.
603 Instead we can retain the CPR info and do the thunk-splitting transform
604 (see WorkWrap.splitThunk).
606 This made a big difference to PrelBase.modInt, which had something like
607 modInt = \ x -> let r = ... -> I# v in
608 ...body strict in r...
609 r's RHS isn't a value yet; but modInt returns r in various branches, so
610 if r doesn't have the CPR property then neither does modInt
611 Another case I found in practice (in Complex.magnitude), looks like this:
612 let k = if ... then I# a else I# b
613 in ... body strict in k ....
614 (For this example, it doesn't matter whether k is returned as part of
615 the overall result; but it does matter that k's RHS has the CPR property.)
616 Left to itself, the simplifier will make a join point thus:
617 let $j k = ...body strict in k...
618 if ... then $j (I# a) else $j (I# b)
619 With thunk-splitting, we get instead
620 let $j x = let k = I#x in ...body strict in k...
621 in if ... then $j a else $j b
622 This is much better; there's a good chance the I# won't get allocated.
624 The difficulty with this is that we need the strictness type to
625 look at the body... but we now need the body to calculate the demand
626 on the variable, so we can decide whether its strictness type should
627 have a CPR in it or not. Simple solution:
628 a) use strictness info from the previous iteration
629 b) make sure we do at least 2 iterations, by doing a second
630 round for top-level non-recs. Top level recs will get at
631 least 2 iterations except for totally-bottom functions
632 which aren't very interesting anyway.
634 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
636 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
637 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
638 Demand info now has a 'Nothing' state, just like strictness info.
639 The analysis works from 'dangerous' towards a 'safe' state; so we
640 start with botSig for 'Nothing' strictness infos, and we start with
641 "yes, it's demanded" for 'Nothing' in the demand info. The
642 fixpoint iteration will sort it all out.
644 We can't start with 'not-demanded' because then consider
648 if ... then t else I# y else f x'
650 In the first iteration we'd have no demand info for x, so assume
651 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
652 we'd see that t was demanded, and so give it the CPR property, but by
653 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
654 setting the first time round, we say 'yes t is demanded' the first
657 However, this does mean that for non-recursive bindings we must
658 iterate twice to be sure of not getting over-optimistic CPR info,
659 in the case where t turns out to be not-demanded. This is handled
663 Note [NOINLINE and strictness]
664 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
665 The strictness analyser used to have a HACK which ensured that NOINLNE
666 things were not strictness-analysed. The reason was unsafePerformIO.
667 Left to itself, the strictness analyser would discover this strictness
669 unsafePerformIO: C(U(AV))
670 But then consider this sub-expression
671 unsafePerformIO (\s -> let r = f x in
672 case writeIORef v r s of (# s1, _ #) ->
674 The strictness analyser will now find that r is sure to be eval'd,
675 and may then hoist it out. This makes tests/lib/should_run/memo002
678 Solving this by making all NOINLINE things have no strictness info is overkill.
679 In particular, it's overkill for runST, which is perfectly respectable.
681 f x = runST (return x)
682 This should be strict in x.
684 So the new plan is to define unsafePerformIO using the 'lazy' combinator:
686 unsafePerformIO (IO m) = lazy (case m realWorld# of (# _, r #) -> r)
688 Remember, 'lazy' is a wired-in identity-function Id, of type a->a, which is
689 magically NON-STRICT, and is inlined after strictness analysis. So
690 unsafePerformIO will look non-strict, and that's what we want.
692 Now we don't need the hack in the strictness analyser. HOWEVER, this
693 decision does mean that even a NOINLINE function is not entirely
694 opaque: some aspect of its implementation leaks out, notably its
695 strictness. For example, if you have a function implemented by an
696 error stub, but which has RULES, you may want it not to be eliminated
701 mk_sig_ty :: Bool -> Bool -> CoreExpr
702 -> DmdType -> (DmdEnv, StrictSig)
703 mk_sig_ty _never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
704 = (lazy_fv, mkStrictSig dmd_ty)
705 -- Re unused never_inline, see Note [NOINLINE and strictness]
707 dmd_ty = DmdType strict_fv final_dmds res'
709 lazy_fv = filterUFM (not . isStrictDmd) fv
710 strict_fv = filterUFM isStrictDmd fv
711 -- We put the strict FVs in the DmdType of the Id, so
712 -- that at its call sites we unleash demands on its strict fvs.
713 -- An example is 'roll' in imaginary/wheel-sieve2
714 -- Something like this:
716 -- go y = if ... then roll (x-1) else x+1
719 -- We want to see that roll is strict in x, which is because
720 -- go is called. So we put the DmdEnv for x in go's DmdType.
723 -- f :: Int -> Int -> Int
724 -- f x y = let t = x+1
725 -- h z = if z==0 then t else
726 -- if z==1 then x+1 else
730 -- Calling h does indeed evaluate x, but we can only see
731 -- that if we unleash a demand on x at the call site for t.
733 -- Incidentally, here's a place where lambda-lifting h would
734 -- lose the cigar --- we couldn't see the joint strictness in t/x
737 -- We don't want to put *all* the fv's from the RHS into the
738 -- DmdType, because that makes fixpointing very slow --- the
739 -- DmdType gets full of lazy demands that are slow to converge.
741 final_dmds = setUnpackStrategy dmds
742 -- Set the unpacking strategy
745 RetCPR | ignore_cpr_info -> TopRes
747 ignore_cpr_info = not (exprIsHNF rhs || thunk_cpr_ok)
750 The unpack strategy determines whether we'll *really* unpack the argument,
751 or whether we'll just remember its strictness. If unpacking would give
752 rise to a *lot* of worker args, we may decide not to unpack after all.
755 setUnpackStrategy :: [Demand] -> [Demand]
757 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
759 go :: Int -- Max number of args available for sub-components of [Demand]
761 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
763 go n (Eval (Prod cs) : ds)
764 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
765 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
768 n' = n + 1 - non_abs_args
769 -- Add one to the budget 'cos we drop the top-level arg
770 non_abs_args = nonAbsentArgs cs
771 -- Delete # of non-absent args to which we'll now be committed
773 go n (d:ds) = d `cons` go n ds
776 cons d (n,ds) = (n, d:ds)
778 nonAbsentArgs :: [Demand] -> Int
780 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
781 nonAbsentArgs (_ : ds) = 1 + nonAbsentArgs ds
785 %************************************************************************
787 \subsection{Strictness signatures and types}
789 %************************************************************************
792 unitVarDmd :: Var -> Demand -> DmdType
793 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
795 addVarDmd :: DmdType -> Var -> Demand -> DmdType
796 addVarDmd (DmdType fv ds res) var dmd
797 = DmdType (extendVarEnv_C both fv var dmd) ds res
799 addLazyFVs :: DmdType -> DmdEnv -> DmdType
800 addLazyFVs (DmdType fv ds res) lazy_fvs
801 = DmdType both_fv1 ds res
803 both_fv = plusVarEnv_C both fv lazy_fvs
804 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
805 -- This modifyEnv is vital. Consider
806 -- let f = \x -> (x,y)
808 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
809 -- demand with the bottom coming up from 'error'
811 -- I got a loop in the fixpointer without this, due to an interaction
812 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
814 -- = letrec g y = x `fatbar`
815 -- letrec h z = z + ...g...
818 -- In the initial iteration for f, f=Bot
819 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
820 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
821 -- places on its free variables. Suppose it places none. Then the
822 -- x `fatbar` ...call to h...
823 -- will give a x->V demand for x. That turns into a L demand for x,
824 -- which floats out of the defn for h. Without the modifyEnv, that
825 -- L demand doesn't get both'd with the Bot coming up from the inner
826 -- call to f. So we just get an L demand for x for g.
828 -- A better way to say this is that the lazy-fv filtering should give the
829 -- same answer as putting the lazy fv demands in the function's type.
831 annotateBndr :: DmdType -> Var -> (DmdType, Var)
832 -- The returned env has the var deleted
833 -- The returned var is annotated with demand info
834 -- No effect on the argument demands
835 annotateBndr dmd_ty@(DmdType fv ds res) var
836 | isTyCoVar var = (dmd_ty, var)
837 | otherwise = (DmdType fv' ds res, setIdDemandInfo var dmd)
839 (fv', dmd) = removeFV fv var res
841 annotateBndrs :: DmdType -> [Var] -> (DmdType, [Var])
842 annotateBndrs = mapAccumR annotateBndr
844 annotateLamIdBndr :: SigEnv
845 -> DmdType -- Demand type of body
846 -> Id -- Lambda binder
847 -> (DmdType, -- Demand type of lambda
848 Id) -- and binder annotated with demand
850 annotateLamIdBndr sigs (DmdType fv ds res) id
851 -- For lambdas we add the demand to the argument demands
852 -- Only called for Ids
854 (final_ty, setIdDemandInfo id hacked_dmd)
856 -- Watch out! See note [Lambda-bound unfoldings]
857 final_ty = case maybeUnfoldingTemplate (idUnfolding id) of
859 Just unf -> main_ty `bothType` unf_ty
861 (unf_ty, _) = dmdAnal sigs dmd unf
863 main_ty = DmdType fv' (hacked_dmd:ds) res
865 (fv', dmd) = removeFV fv id res
866 hacked_dmd = argDemand dmd
867 -- This call to argDemand is vital, because otherwise we label
868 -- a lambda binder with demand 'B'. But in terms of calling
869 -- conventions that's Abs, because we don't pass it. But
870 -- when we do a w/w split we get
871 -- fw x = (\x y:B -> ...) x (error "oops")
872 -- And then the simplifier things the 'B' is a strict demand
873 -- and evaluates the (error "oops"). Sigh
875 removeFV :: DmdEnv -> Var -> DmdResult -> (DmdEnv, Demand)
876 removeFV fv id res = (fv', zapUnlifted id dmd)
878 fv' = fv `delVarEnv` id
879 dmd = lookupVarEnv fv id `orElse` deflt
880 deflt | isBotRes res = Bot
883 zapUnlifted :: Id -> Demand -> Demand
884 -- For unlifted-type variables, we are only
885 -- interested in Bot/Abs/Box Abs
886 zapUnlifted _ Bot = Bot
887 zapUnlifted _ Abs = Abs
888 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
892 Note [Lamba-bound unfoldings]
893 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
894 We allow a lambda-bound variable to carry an unfolding, a facility that is used
895 exclusively for join points; see Note [Case binders and join points]. If so,
896 we must be careful to demand-analyse the RHS of the unfolding! Example
897 \x. \y{=Just x}. <body>
898 Then if <body> uses 'y', then transitively it uses 'x', and we must not
899 forget that fact, otherwise we might make 'x' absent when it isn't.
902 %************************************************************************
904 \subsection{Strictness signatures}
906 %************************************************************************
910 = SE { se_env :: VarEnv (StrictSig, TopLevelFlag)
911 , se_virgin :: Bool } -- True on first iteration only
912 -- See Note [Initialising strictness]
913 -- We use the se_env to tell us whether to
914 -- record info about a variable in the DmdEnv
915 -- We do so if it's a LocalId, but not top-level
917 -- The DmdEnv gives the demand on the free vars of the function
918 -- when it is given enough args to satisfy the strictness signature
920 instance Outputable SigEnv where
921 ppr (SE { se_env = env, se_virgin = virgin })
922 = ptext (sLit "SE") <+> braces (vcat
923 [ ptext (sLit "se_virgin =") <+> ppr virgin
924 , ptext (sLit "se_env =") <+> ppr env ])
926 emptySigEnv :: SigEnv
927 emptySigEnv = SE { se_env = emptyVarEnv, se_virgin = True }
929 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
930 extendSigEnv top_lvl sigs var sig
931 = sigs { se_env = extendVarEnv (se_env sigs) var (sig, top_lvl) }
933 lookupSigEnv :: SigEnv -> Id -> Maybe (StrictSig, TopLevelFlag)
934 lookupSigEnv sigs id = lookupVarEnv (se_env sigs) id
936 addInitialSigs :: TopLevelFlag -> SigEnv -> [Id] -> SigEnv
937 -- See Note [Initialising strictness]
938 addInitialSigs top_lvl sigs@(SE { se_env = env, se_virgin = virgin }) ids
939 = sigs { se_env = extendVarEnvList env [ (id, (init_sig id, top_lvl))
942 init_sig | virgin = \_ -> botSig
943 | otherwise = idStrictness
945 setNonVirgin :: SigEnv -> SigEnv
946 setNonVirgin sigs = sigs { se_virgin = False }
948 extendSigsWithLam :: SigEnv -> Id -> SigEnv
949 -- Extend the SigEnv when we meet a lambda binder
950 -- If the binder is marked demanded with a product demand, then give it a CPR
951 -- signature, because in the likely event that this is a lambda on a fn defn
952 -- [we only use this when the lambda is being consumed with a call demand],
953 -- it'll be w/w'd and so it will be CPR-ish. E.g.
954 -- f = \x::(Int,Int). if ...strict in x... then
958 -- We want f to have the CPR property because x does, by the time f has been w/w'd
960 -- Also note that we only want to do this for something that
961 -- definitely has product type, else we may get over-optimistic
962 -- CPR results (e.g. from \x -> x!).
964 extendSigsWithLam sigs id
965 = case idDemandInfo_maybe id of
966 Nothing -> extendSigEnv NotTopLevel sigs id cprSig
967 -- Optimistic in the Nothing case;
968 -- See notes [CPR-AND-STRICTNESS]
969 Just (Eval (Prod _)) -> extendSigEnv NotTopLevel sigs id cprSig
973 Note [Initialising strictness]
974 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
975 Our basic plan is to initialise the strictness of each Id in
976 a recursive group to "bottom", and find a fixpoint from there.
977 However, this group A might be inside an *enclosing* recursive
978 group B, in which case we'll do the entire fixpoint shebang on A
979 for each iteration of B.
981 To speed things up, we initialise each iteration of B from the result
982 of the last one, which is neatly recorded in each binder. That way we
983 make use of earlier iterations of the fixpoint algorithm. (Cunning
986 But on the *first* iteration we want to *ignore* the current strictness
987 of the Id, and start from "bottom". Nowadays the Id can have a current
988 strictness, because interface files record strictness for nested bindings.
989 To know when we are in the first iteration, we look at the se_virgin
993 %************************************************************************
997 %************************************************************************
1000 splitDmdTy :: DmdType -> (Demand, DmdType)
1001 -- Split off one function argument
1002 -- We already have a suitable demand on all
1003 -- free vars, so no need to add more!
1004 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
1005 splitDmdTy ty@(DmdType _ [] res_ty) = (resTypeArgDmd res_ty, ty)
1007 splitCallDmd :: Demand -> (Int, Demand)
1008 splitCallDmd (Call d) = case splitCallDmd d of
1010 splitCallDmd d = (0, d)
1012 vanillaCall :: Arity -> Demand
1013 vanillaCall 0 = evalDmd
1014 vanillaCall n = Call (vanillaCall (n-1))
1016 deferType :: DmdType -> DmdType
1017 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
1018 -- Notice that we throw away info about both arguments and results
1019 -- For example, f = let ... in \x -> x
1020 -- We don't want to get a stricness type V->T for f.
1022 deferEnv :: DmdEnv -> DmdEnv
1023 deferEnv fv = mapVarEnv defer fv
1027 argDemand :: Demand -> Demand
1028 -- The 'Defer' demands are just Lazy at function boundaries
1029 -- Ugly! Ask John how to improve it.
1030 argDemand Top = lazyDmd
1031 argDemand (Defer _) = lazyDmd
1032 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
1033 argDemand (Box Bot) = evalDmd
1034 argDemand (Box d) = box (argDemand d)
1035 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
1040 -------------------------
1041 lubType :: DmdType -> DmdType -> DmdType
1042 -- Consider (if x then y else []) with demand V
1043 -- Then the first branch gives {y->V} and the second
1044 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
1045 -- in the result env.
1046 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
1047 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
1049 lub_fv = plusVarEnv_C lub fv1 fv2
1050 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
1051 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
1052 -- lub is the identity for Bot
1054 -- Extend the shorter argument list to match the longer
1055 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
1057 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
1058 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
1060 -----------------------------------
1061 bothType :: DmdType -> DmdType -> DmdType
1062 -- (t1 `bothType` t2) takes the argument/result info from t1,
1063 -- using t2 just for its free-var info
1064 -- NB: Don't forget about r2! It might be BotRes, which is
1065 -- a bottom demand on all the in-scope variables.
1066 -- Peter: can this be done more neatly?
1067 bothType (DmdType fv1 ds1 r1) (DmdType fv2 _ r2)
1068 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
1070 both_fv = plusVarEnv_C both fv1 fv2
1071 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
1072 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
1073 -- both is the identity for Abs
1078 lubRes :: DmdResult -> DmdResult -> DmdResult
1081 lubRes RetCPR RetCPR = RetCPR
1084 bothRes :: DmdResult -> DmdResult -> DmdResult
1085 -- If either diverges, the whole thing does
1086 -- Otherwise take CPR info from the first
1087 bothRes _ BotRes = BotRes
1092 modifyEnv :: Bool -- No-op if False
1093 -> (Demand -> Demand) -- The zapper
1094 -> DmdEnv -> DmdEnv -- Env1 and Env2
1095 -> DmdEnv -> DmdEnv -- Transform this env
1096 -- Zap anything in Env1 but not in Env2
1097 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
1099 modifyEnv need_to_modify zapper env1 env2 env
1100 | need_to_modify = foldr zap env (varEnvKeys (env1 `minusUFM` env2))
1103 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
1105 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
1109 %************************************************************************
1111 \subsection{LUB and BOTH}
1113 %************************************************************************
1116 lub :: Demand -> Demand -> Demand
1119 lub Abs d2 = absLub d2
1121 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
1123 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
1124 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
1125 lub (Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
1126 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
1128 -- For the Eval case, we use these approximation rules
1129 -- Box Bot <= Eval (Box Bot ...)
1130 -- Box Top <= Defer (Box Bot ...)
1131 -- Box (Eval ds) <= Eval (map Box ds)
1132 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
1133 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
1134 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
1135 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
1136 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
1138 lub (Box d1) (Box d2) = box (d1 `lub` d2)
1139 lub d1@(Box _) d2 = d2 `lub` d1
1141 lubs :: Demands -> Demands -> Demands
1142 lubs ds1 ds2 = zipWithDmds lub ds1 ds2
1144 ---------------------
1145 box :: Demand -> Demand
1146 -- box is the smart constructor for Box
1147 -- It computes <B,bot> & d
1148 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
1149 -- Seems to be no point in allowing (Box (Call d))
1150 box (Call d) = Call d -- The odd man out. Why?
1152 box (Defer _) = lazyDmd
1153 box Top = lazyDmd -- Box Abs and Box Top
1154 box Abs = lazyDmd -- are the same <B,L>
1155 box d = Box d -- Bot, Eval
1158 defer :: Demand -> Demand
1160 -- defer is the smart constructor for Defer
1161 -- The idea is that (Defer ds) = <U(ds), L>
1163 -- It specifies what happens at a lazy function argument
1164 -- or a lambda; the L* operator
1165 -- Set the strictness part to L, but leave
1166 -- the boxity side unaffected
1167 -- It also ensures that Defer (Eval [LLLL]) = L
1172 defer (Call _) = lazyDmd -- Approximation here?
1173 defer (Box _) = lazyDmd
1174 defer (Defer ds) = Defer ds
1175 defer (Eval ds) = deferEval ds
1177 deferEval :: Demands -> Demand
1178 -- deferEval ds = defer (Eval ds)
1179 deferEval ds | allTop ds = Top
1180 | otherwise = Defer ds
1182 ---------------------
1183 absLub :: Demand -> Demand
1184 -- Computes (Abs `lub` d)
1185 -- For the Bot case consider
1186 -- f x y = if ... then x else error x
1187 -- Then for y we get Abs `lub` Bot, and we really
1192 absLub (Call _) = Top
1193 absLub (Box _) = Top
1194 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1195 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1197 absLubs :: Demands -> Demands
1198 absLubs = mapDmds absLub
1201 both :: Demand -> Demand -> Demand
1205 -- Note [Bottom demands]
1208 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1209 both Bot (Defer ds) = Eval (mapDmds (`both` Bot) ds)
1212 both Top Bot = errDmd
1215 both Top (Box d) = Box d
1216 both Top (Call d) = Call d
1217 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1218 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1219 -- = defer (Eval (mapDmds (`both` Top) ds))
1220 = deferEval (mapDmds (`both` Top) ds)
1223 both (Box d1) (Box d2) = box (d1 `both` d2)
1224 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1225 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1226 both (Box d1) (Defer _) = Box d1
1227 both d1@(Box _) d2 = d2 `both` d1
1229 both (Call d1) (Call d2) = Call (d1 `both` d2)
1230 both (Call d1) (Eval _) = Call d1 -- Could do better for (Poly Bot)?
1231 both (Call d1) (Defer _) = Call d1 -- Ditto
1232 both d1@(Call _) d2 = d2 `both` d1
1234 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1235 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1236 both d1@(Eval _) d2 = d2 `both` d1
1238 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1239 both d1@(Defer _) d2 = d2 `both` d1
1241 boths :: Demands -> Demands -> Demands
1242 boths ds1 ds2 = zipWithDmds both ds1 ds2
1245 Note [Bottom demands]
1246 ~~~~~~~~~~~~~~~~~~~~~
1249 From 'error' itself we get demand Bot on x
1250 From the arg demand on x we get
1251 x :-> evalDmd = Box (Eval (Poly Abs))
1252 So we get Bot `both` Box (Eval (Poly Abs))
1253 = Seq Keep (Poly Bot)
1256 f x = if ... then error (fst x) else fst x
1257 Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1259 which is what we want.
1264 x :-> Bot `both` Defer [SA]
1265 and we want the Bot demand to cancel out the Defer
1266 so that we get Eval [SA]. Otherwise we'd have the odd
1268 f x = error (fst x) -- Strictness U(SA)b
1269 g x = error ('y':fst x) -- Strictness Tb