2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
10 module DmdAnal ( dmdAnalPgm, dmdAnalTopRhs,
11 both {- needed by WwLib -}
14 #include "HsVersions.h"
16 import DynFlags ( DynFlags, DynFlag(..) )
17 import StaticFlags ( opt_MaxWorkerArgs )
18 import NewDemand -- All of it
21 import CoreUtils ( exprIsHNF, exprIsTrivial, exprArity )
22 import DataCon ( dataConTyCon )
23 import TyCon ( isProductTyCon, isRecursiveTyCon )
24 import Id ( Id, idType, idInlinePragma,
25 isDataConWorkId, isGlobalId, idArity,
27 idDemandInfo, idStrictness, idCprInfo, idName,
29 idNewStrictness, idNewStrictness_maybe,
30 setIdNewStrictness, idNewDemandInfo,
31 idNewDemandInfo_maybe,
35 import IdInfo ( newStrictnessFromOld, newDemand )
39 import TysWiredIn ( unboxedPairDataCon )
40 import TysPrim ( realWorldStatePrimTy )
41 import UniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly,
42 keysUFM, minusUFM, ufmToList, filterUFM )
43 import Type ( isUnLiftedType, coreEqType )
44 import CoreLint ( showPass, endPass )
45 import Util ( mapAndUnzip, mapAccumL, mapAccumR, lengthIs )
46 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive,
48 import Maybes ( orElse, expectJust )
54 * set a noinline pragma on bottoming Ids
56 * Consider f x = x+1 `fatbar` error (show x)
57 We'd like to unbox x, even if that means reboxing it in the error case.
60 %************************************************************************
62 \subsection{Top level stuff}
64 %************************************************************************
67 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
68 dmdAnalPgm dflags binds
70 showPass dflags "Demand analysis" ;
71 let { binds_plus_dmds = do_prog binds } ;
73 endPass dflags "Demand analysis"
74 Opt_D_dump_stranal binds_plus_dmds ;
76 -- Only if OLD_STRICTNESS is on, because only then is the old
77 -- strictness analyser run
78 let { dmd_changes = get_changes binds_plus_dmds } ;
79 printDump (text "Changes in demands" $$ dmd_changes) ;
81 return binds_plus_dmds
84 do_prog :: [CoreBind] -> [CoreBind]
85 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
87 dmdAnalTopBind :: SigEnv
90 dmdAnalTopBind sigs (NonRec id rhs)
92 ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs)
93 (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs1)
94 -- Do two passes to improve CPR information
95 -- See comments with ignore_cpr_info in mk_sig_ty
96 -- and with extendSigsWithLam
98 (sigs2, NonRec id2 rhs2)
100 dmdAnalTopBind sigs (Rec pairs)
102 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
103 -- We get two iterations automatically
104 -- c.f. the NonRec case above
110 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
111 -- Analyse the RHS and return
112 -- a) appropriate strictness info
113 -- b) the unfolding (decorated with stricntess info)
117 call_dmd = vanillaCall (exprArity rhs)
118 (_, rhs1) = dmdAnal emptySigEnv call_dmd rhs
119 (rhs_ty, rhs2) = dmdAnal emptySigEnv call_dmd rhs1
120 sig = mkTopSigTy rhs rhs_ty
121 -- Do two passes; see notes with extendSigsWithLam
122 -- Otherwise we get bogus CPR info for constructors like
123 -- newtype T a = MkT a
124 -- The constructor looks like (\x::T a -> x), modulo the coerce
125 -- extendSigsWithLam will optimistically give x a CPR tag the
126 -- first time, which is wrong in the end.
129 %************************************************************************
131 \subsection{The analyser itself}
133 %************************************************************************
136 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
138 dmdAnal sigs Abs e = (topDmdType, e)
141 | not (isStrictDmd dmd)
143 (res_ty, e') = dmdAnal sigs evalDmd e
145 (deferType res_ty, e')
146 -- It's important not to analyse e with a lazy demand because
147 -- a) When we encounter case s of (a,b) ->
148 -- we demand s with U(d1d2)... but if the overall demand is lazy
149 -- that is wrong, and we'd need to reduce the demand on s,
150 -- which is inconvenient
151 -- b) More important, consider
152 -- f (let x = R in x+x), where f is lazy
153 -- We still want to mark x as demanded, because it will be when we
154 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
155 -- just mark x as Lazy
156 -- c) The application rule wouldn't be right either
157 -- Evaluating (f x) in a L demand does *not* cause
158 -- evaluation of f in a C(L) demand!
161 dmdAnal sigs dmd (Lit lit)
162 = (topDmdType, Lit lit)
164 dmdAnal sigs dmd (Var var)
165 = (dmdTransform sigs var dmd, Var var)
167 dmdAnal sigs dmd (Note n e)
168 = (dmd_ty, Note n e')
170 (dmd_ty, e') = dmdAnal sigs dmd' e
172 Coerce _ _ -> evalDmd -- This coerce usually arises from a recursive
173 other -> dmd -- newtype, and we don't want to look inside them
174 -- for exactly the same reason that we don't look
175 -- inside recursive products -- we might not reach
176 -- a fixpoint. So revert to a vanilla Eval demand
178 dmdAnal sigs dmd (App fun (Type ty))
179 = (fun_ty, App fun' (Type ty))
181 (fun_ty, fun') = dmdAnal sigs dmd fun
183 -- Lots of the other code is there to make this
184 -- beautiful, compositional, application rule :-)
185 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
186 = let -- [Type arg handled above]
187 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
188 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
189 (arg_dmd, res_ty) = splitDmdTy fun_ty
191 (res_ty `bothType` arg_ty, App fun' arg')
193 dmdAnal sigs dmd (Lam var body)
196 (body_ty, body') = dmdAnal sigs dmd body
198 (body_ty, Lam var body')
200 | Call body_dmd <- dmd -- A call demand: good!
202 sigs' = extendSigsWithLam sigs var
203 (body_ty, body') = dmdAnal sigs' body_dmd body
204 (lam_ty, var') = annotateLamIdBndr body_ty var
206 (lam_ty, Lam var' body')
208 | otherwise -- Not enough demand on the lambda; but do the body
209 = let -- anyway to annotate it and gather free var info
210 (body_ty, body') = dmdAnal sigs evalDmd body
211 (lam_ty, var') = annotateLamIdBndr body_ty var
213 (deferType lam_ty, Lam var' body')
215 dmdAnal sigs dmd (Case scrut case_bndr ty [alt@(DataAlt dc,bndrs,rhs)])
216 | let tycon = dataConTyCon dc,
217 isProductTyCon tycon,
218 not (isRecursiveTyCon tycon)
220 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
221 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
222 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
223 (_, bndrs', _) = alt'
224 case_bndr_sig = cprSig
225 -- Inside the alternative, the case binder has the CPR property.
226 -- Meaning that a case on it will successfully cancel.
228 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
231 -- We want f to have the CPR property:
232 -- f b x = case fw b x of { r -> I# r }
233 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
236 -- Figure out whether the demand on the case binder is used, and use
237 -- that to set the scrut_dmd. This is utterly essential.
238 -- Consider f x = case x of y { (a,b) -> k y a }
239 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
240 -- worker, so the worker will rebuild
241 -- x = (a, absent-error)
242 -- and that'll crash.
243 -- So at one stage I had:
244 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
245 -- keepity | dead_case_bndr = Drop
246 -- | otherwise = Keep
249 -- case x of y { (a,b) -> h y + a }
250 -- where h : U(LL) -> T
251 -- The above code would compute a Keep for x, since y is not Abs, which is silly
252 -- The insight is, of course, that a demand on y is a demand on the
253 -- scrutinee, so we need to `both` it with the scrut demand
255 scrut_dmd = Eval (Prod [idNewDemandInfo b | b <- bndrs', isId b])
257 idNewDemandInfo case_bndr'
259 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
261 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' ty [alt'])
263 dmdAnal sigs dmd (Case scrut case_bndr ty alts)
265 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
266 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
267 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
269 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
270 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' ty alts')
272 dmdAnal sigs dmd (Let (NonRec id rhs) body)
274 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
275 (body_ty, body') = dmdAnal sigs' dmd body
276 (body_ty1, id2) = annotateBndr body_ty id1
277 body_ty2 = addLazyFVs body_ty1 lazy_fv
279 -- If the actual demand is better than the vanilla call
280 -- demand, you might think that we might do better to re-analyse
281 -- the RHS with the stronger demand.
282 -- But (a) That seldom happens, because it means that *every* path in
283 -- the body of the let has to use that stronger demand
284 -- (b) It often happens temporarily in when fixpointing, because
285 -- the recursive function at first seems to place a massive demand.
286 -- But we don't want to go to extra work when the function will
287 -- probably iterate to something less demanding.
288 -- In practice, all the times the actual demand on id2 is more than
289 -- the vanilla call demand seem to be due to (b). So we don't
290 -- bother to re-analyse the RHS.
291 (body_ty2, Let (NonRec id2 rhs') body')
293 dmdAnal sigs dmd (Let (Rec pairs) body)
295 bndrs = map fst pairs
296 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
297 (body_ty, body') = dmdAnal sigs' dmd body
298 body_ty1 = addLazyFVs body_ty lazy_fv
300 sigs' `seq` body_ty `seq`
302 (body_ty2, _) = annotateBndrs body_ty1 bndrs
303 -- Don't bother to add demand info to recursive
304 -- binders as annotateBndr does;
305 -- being recursive, we can't treat them strictly.
306 -- But we do need to remove the binders from the result demand env
308 (body_ty2, Let (Rec pairs') body')
311 dmdAnalAlt sigs dmd (con,bndrs,rhs)
313 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
314 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
315 final_alt_ty | io_hack_reqd = alt_ty `lubType` topDmdType
318 -- There's a hack here for I/O operations. Consider
319 -- case foo x s of { (# s, r #) -> y }
320 -- Is this strict in 'y'. Normally yes, but what if 'foo' is an I/O
321 -- operation that simply terminates the program (not in an erroneous way)?
322 -- In that case we should not evaluate y before the call to 'foo'.
323 -- Hackish solution: spot the IO-like situation and add a virtual branch,
327 -- other -> return ()
328 -- So the 'y' isn't necessarily going to be evaluated
330 -- A more complete example where this shows up is:
331 -- do { let len = <expensive> ;
332 -- ; when (...) (exitWith ExitSuccess)
335 io_hack_reqd = con == DataAlt unboxedPairDataCon &&
336 idType (head bndrs) `coreEqType` realWorldStatePrimTy
338 (final_alt_ty, (con, bndrs', rhs'))
341 %************************************************************************
343 \subsection{Bindings}
345 %************************************************************************
348 dmdFix :: TopLevelFlag
349 -> SigEnv -- Does not include bindings for this binding
352 [(Id,CoreExpr)]) -- Binders annotated with stricness info
354 dmdFix top_lvl sigs orig_pairs
355 = loop 1 initial_sigs orig_pairs
357 bndrs = map fst orig_pairs
358 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
361 -> SigEnv -- Already contains the current sigs
363 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
366 = (sigs', lazy_fv, pairs')
367 -- Note: use pairs', not pairs. pairs' is the result of
368 -- processing the RHSs with sigs (= sigs'), whereas pairs
369 -- is the result of processing the RHSs with the *previous*
370 -- iteration of sigs.
372 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
373 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
374 text "env:" <+> ppr (ufmToList sigs),
375 text "binds:" <+> pprCoreBinding (Rec pairs)]))
376 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
377 -- The lazy_fv part is really important! orig_pairs has no strictness
378 -- info, including nothing about free vars. But if we have
379 -- letrec f = ....y..... in ...f...
380 -- where 'y' is free in f, we must record that y is mentioned,
381 -- otherwise y will get recorded as absent altogether
383 | otherwise = loop (n+1) sigs' pairs'
385 found_fixpoint = all (same_sig sigs sigs') bndrs
386 -- Use the new signature to do the next pair
387 -- The occurrence analyser has arranged them in a good order
388 -- so this can significantly reduce the number of iterations needed
389 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
391 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
392 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
394 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
395 ((sigs', lazy_fv'), pair')
398 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
399 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
400 -- old_sig = lookup sigs id
401 -- new_sig = lookup sigs' id
403 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
404 lookup sigs var = case lookupVarEnv sigs var of
407 -- Get an initial strictness signature from the Id
408 -- itself. That way we make use of earlier iterations
409 -- of the fixpoint algorithm. (Cunning plan.)
410 -- Note that the cunning plan extends to the DmdEnv too,
411 -- since it is part of the strictness signature
412 initialSig id = idNewStrictness_maybe id `orElse` botSig
414 dmdAnalRhs :: TopLevelFlag -> RecFlag
415 -> SigEnv -> (Id, CoreExpr)
416 -> (SigEnv, DmdEnv, (Id, CoreExpr))
417 -- Process the RHS of the binding, add the strictness signature
418 -- to the Id, and augment the environment with the signature as well.
420 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
421 = (sigs', lazy_fv, (id', rhs'))
423 arity = idArity id -- The idArity should be up to date
424 -- The simplifier was run just beforehand
425 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
426 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs), ppr id )
427 -- The RHS can be eta-reduced to just a variable,
428 -- in which case we should not complain.
429 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
430 id' = id `setIdNewStrictness` sig_ty
431 sigs' = extendSigEnv top_lvl sigs id sig_ty
434 %************************************************************************
436 \subsection{Strictness signatures and types}
438 %************************************************************************
441 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
442 -- Take a DmdType and turn it into a StrictSig
443 -- NB: not used for never-inline things; hence False
444 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
446 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
447 mkSigTy top_lvl rec_flag id rhs dmd_ty
448 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
450 never_inline = isNeverActive (idInlinePragma id)
451 maybe_id_dmd = idNewDemandInfo_maybe id
452 -- Is Nothing the first time round
455 | isTopLevel top_lvl = False -- Top level things don't get
456 -- their demandInfo set at all
457 | isRec rec_flag = False -- Ditto recursive things
458 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
459 | otherwise = True -- Optimistic, first time round
463 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
464 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
465 If the rhs is a thunk, we usually forget the CPR info, because
466 it is presumably shared (else it would have been inlined, and
467 so we'd lose sharing if w/w'd it into a function.
469 However, if the strictness analyser has figured out (in a previous
470 iteration) that it's strict, then we DON'T need to forget the CPR info.
471 Instead we can retain the CPR info and do the thunk-splitting transform
472 (see WorkWrap.splitThunk).
474 This made a big difference to PrelBase.modInt, which had something like
475 modInt = \ x -> let r = ... -> I# v in
476 ...body strict in r...
477 r's RHS isn't a value yet; but modInt returns r in various branches, so
478 if r doesn't have the CPR property then neither does modInt
479 Another case I found in practice (in Complex.magnitude), looks like this:
480 let k = if ... then I# a else I# b
481 in ... body strict in k ....
482 (For this example, it doesn't matter whether k is returned as part of
483 the overall result; but it does matter that k's RHS has the CPR property.)
484 Left to itself, the simplifier will make a join point thus:
485 let $j k = ...body strict in k...
486 if ... then $j (I# a) else $j (I# b)
487 With thunk-splitting, we get instead
488 let $j x = let k = I#x in ...body strict in k...
489 in if ... then $j a else $j b
490 This is much better; there's a good chance the I# won't get allocated.
492 The difficulty with this is that we need the strictness type to
493 look at the body... but we now need the body to calculate the demand
494 on the variable, so we can decide whether its strictness type should
495 have a CPR in it or not. Simple solution:
496 a) use strictness info from the previous iteration
497 b) make sure we do at least 2 iterations, by doing a second
498 round for top-level non-recs. Top level recs will get at
499 least 2 iterations except for totally-bottom functions
500 which aren't very interesting anyway.
502 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
504 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
505 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
506 Demand info now has a 'Nothing' state, just like strictness info.
507 The analysis works from 'dangerous' towards a 'safe' state; so we
508 start with botSig for 'Nothing' strictness infos, and we start with
509 "yes, it's demanded" for 'Nothing' in the demand info. The
510 fixpoint iteration will sort it all out.
512 We can't start with 'not-demanded' because then consider
516 if ... then t else I# y else f x'
518 In the first iteration we'd have no demand info for x, so assume
519 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
520 we'd see that t was demanded, and so give it the CPR property, but by
521 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
522 setting the first time round, we say 'yes t is demanded' the first
525 However, this does mean that for non-recursive bindings we must
526 iterate twice to be sure of not getting over-optimistic CPR info,
527 in the case where t turns out to be not-demanded. This is handled
532 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
533 = (lazy_fv, mkStrictSig dmd_ty)
535 dmd_ty = DmdType strict_fv final_dmds res'
537 lazy_fv = filterUFM (not . isStrictDmd) fv
538 strict_fv = filterUFM isStrictDmd fv
539 -- We put the strict FVs in the DmdType of the Id, so
540 -- that at its call sites we unleash demands on its strict fvs.
541 -- An example is 'roll' in imaginary/wheel-sieve2
542 -- Something like this:
544 -- go y = if ... then roll (x-1) else x+1
547 -- We want to see that roll is strict in x, which is because
548 -- go is called. So we put the DmdEnv for x in go's DmdType.
551 -- f :: Int -> Int -> Int
552 -- f x y = let t = x+1
553 -- h z = if z==0 then t else
554 -- if z==1 then x+1 else
558 -- Calling h does indeed evaluate x, but we can only see
559 -- that if we unleash a demand on x at the call site for t.
561 -- Incidentally, here's a place where lambda-lifting h would
562 -- lose the cigar --- we couldn't see the joint strictness in t/x
565 -- We don't want to put *all* the fv's from the RHS into the
566 -- DmdType, because that makes fixpointing very slow --- the
567 -- DmdType gets full of lazy demands that are slow to converge.
569 final_dmds = setUnpackStrategy dmds
570 -- Set the unpacking strategy
573 RetCPR | ignore_cpr_info -> TopRes
575 ignore_cpr_info = not (exprIsHNF rhs || thunk_cpr_ok)
578 The unpack strategy determines whether we'll *really* unpack the argument,
579 or whether we'll just remember its strictness. If unpacking would give
580 rise to a *lot* of worker args, we may decide not to unpack after all.
583 setUnpackStrategy :: [Demand] -> [Demand]
585 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
587 go :: Int -- Max number of args available for sub-components of [Demand]
589 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
591 go n (Eval (Prod cs) : ds)
592 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
593 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
596 n' = n + 1 - non_abs_args
597 -- Add one to the budget 'cos we drop the top-level arg
598 non_abs_args = nonAbsentArgs cs
599 -- Delete # of non-absent args to which we'll now be committed
601 go n (d:ds) = d `cons` go n ds
604 cons d (n,ds) = (n, d:ds)
606 nonAbsentArgs :: [Demand] -> Int
608 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
609 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
613 %************************************************************************
615 \subsection{Strictness signatures and types}
617 %************************************************************************
620 splitDmdTy :: DmdType -> (Demand, DmdType)
621 -- Split off one function argument
622 -- We already have a suitable demand on all
623 -- free vars, so no need to add more!
624 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
625 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
629 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
631 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
632 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
633 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
635 addLazyFVs (DmdType fv ds res) lazy_fvs
636 = DmdType both_fv1 ds res
638 both_fv = (plusUFM_C both fv lazy_fvs)
639 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
640 -- This modifyEnv is vital. Consider
641 -- let f = \x -> (x,y)
643 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
644 -- demand with the bottom coming up from 'error'
646 -- I got a loop in the fixpointer without this, due to an interaction
647 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
649 -- = letrec g y = x `fatbar`
650 -- letrec h z = z + ...g...
653 -- In the initial iteration for f, f=Bot
654 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
655 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
656 -- places on its free variables. Suppose it places none. Then the
657 -- x `fatbar` ...call to h...
658 -- will give a x->V demand for x. That turns into a L demand for x,
659 -- which floats out of the defn for h. Without the modifyEnv, that
660 -- L demand doesn't get both'd with the Bot coming up from the inner
661 -- call to f. So we just get an L demand for x for g.
663 -- A better way to say this is that the lazy-fv filtering should give the
664 -- same answer as putting the lazy fv demands in the function's type.
666 annotateBndr :: DmdType -> Var -> (DmdType, Var)
667 -- The returned env has the var deleted
668 -- The returned var is annotated with demand info
669 -- No effect on the argument demands
670 annotateBndr dmd_ty@(DmdType fv ds res) var
671 | isTyVar var = (dmd_ty, var)
672 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
674 (fv', dmd) = removeFV fv var res
676 annotateBndrs = mapAccumR annotateBndr
678 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
679 -- For lambdas we add the demand to the argument demands
680 -- Only called for Ids
682 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
684 (fv', dmd) = removeFV fv id res
685 hacked_dmd = argDemand dmd
686 -- This call to argDemand is vital, because otherwise we label
687 -- a lambda binder with demand 'B'. But in terms of calling
688 -- conventions that's Abs, because we don't pass it. But
689 -- when we do a w/w split we get
690 -- fw x = (\x y:B -> ...) x (error "oops")
691 -- And then the simplifier things the 'B' is a strict demand
692 -- and evaluates the (error "oops"). Sigh
694 removeFV fv id res = (fv', zapUnlifted id dmd)
696 fv' = fv `delVarEnv` id
697 dmd = lookupVarEnv fv id `orElse` deflt
698 deflt | isBotRes res = Bot
701 -- For unlifted-type variables, we are only
702 -- interested in Bot/Abs/Box Abs
703 zapUnlifted is Bot = Bot
704 zapUnlifted id Abs = Abs
705 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
709 %************************************************************************
711 \subsection{Strictness signatures}
713 %************************************************************************
716 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
717 -- We use the SigEnv to tell us whether to
718 -- record info about a variable in the DmdEnv
719 -- We do so if it's a LocalId, but not top-level
721 -- The DmdEnv gives the demand on the free vars of the function
722 -- when it is given enough args to satisfy the strictness signature
724 emptySigEnv = emptyVarEnv
726 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
727 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
729 extendSigEnvList = extendVarEnvList
731 extendSigsWithLam :: SigEnv -> Id -> SigEnv
732 -- Extend the SigEnv when we meet a lambda binder
733 -- If the binder is marked demanded with a product demand, then give it a CPR
734 -- signature, because in the likely event that this is a lambda on a fn defn
735 -- [we only use this when the lambda is being consumed with a call demand],
736 -- it'll be w/w'd and so it will be CPR-ish. E.g.
737 -- f = \x::(Int,Int). if ...strict in x... then
741 -- We want f to have the CPR property because x does, by the time f has been w/w'd
743 -- Also note that we only want to do this for something that
744 -- definitely has product type, else we may get over-optimistic
745 -- CPR results (e.g. from \x -> x!).
747 extendSigsWithLam sigs id
748 = case idNewDemandInfo_maybe id of
749 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
750 -- Optimistic in the Nothing case;
751 -- See notes [CPR-AND-STRICTNESS]
752 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
756 dmdTransform :: SigEnv -- The strictness environment
757 -> Id -- The function
758 -> Demand -- The demand on the function
759 -> DmdType -- The demand type of the function in this context
760 -- Returned DmdEnv includes the demand on
761 -- this function plus demand on its free variables
763 dmdTransform sigs var dmd
765 ------ DATA CONSTRUCTOR
766 | isDataConWorkId var -- Data constructor
768 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
769 DmdType _ _ con_res = dmd_ty
772 if arity == call_depth then -- Saturated, so unleash the demand
774 -- Important! If we Keep the constructor application, then
775 -- we need the demands the constructor places (always lazy)
776 -- If not, we don't need to. For example:
777 -- f p@(x,y) = (p,y) -- S(AL)
779 -- It's vital that we don't calculate Absent for a!
780 dmd_ds = case res_dmd of
781 Box (Eval ds) -> mapDmds box ds
785 -- ds can be empty, when we are just seq'ing the thing
786 -- If so we must make up a suitable bunch of demands
787 arg_ds = case dmd_ds of
788 Poly d -> replicate arity d
789 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
792 mkDmdType emptyDmdEnv arg_ds con_res
793 -- Must remember whether it's a product, hence con_res, not TopRes
797 ------ IMPORTED FUNCTION
798 | isGlobalId var, -- Imported function
799 let StrictSig dmd_ty = idNewStrictness var
800 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
805 ------ LOCAL LET/REC BOUND THING
806 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
808 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
809 | otherwise = deferType dmd_ty
810 -- NB: it's important to use deferType, and not just return topDmdType
811 -- Consider let { f x y = p + x } in f 1
812 -- The application isn't saturated, but we must nevertheless propagate
813 -- a lazy demand for p!
815 addVarDmd top_lvl fn_ty var dmd
817 ------ LOCAL NON-LET/REC BOUND THING
818 | otherwise -- Default case
822 (call_depth, res_dmd) = splitCallDmd dmd
826 %************************************************************************
830 %************************************************************************
833 splitCallDmd :: Demand -> (Int, Demand)
834 splitCallDmd (Call d) = case splitCallDmd d of
836 splitCallDmd d = (0, d)
838 vanillaCall :: Arity -> Demand
839 vanillaCall 0 = evalDmd
840 vanillaCall n = Call (vanillaCall (n-1))
842 deferType :: DmdType -> DmdType
843 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
844 -- Notice that we throw away info about both arguments and results
845 -- For example, f = let ... in \x -> x
846 -- We don't want to get a stricness type V->T for f.
849 deferEnv :: DmdEnv -> DmdEnv
850 deferEnv fv = mapVarEnv defer fv
854 argDemand :: Demand -> Demand
855 -- The 'Defer' demands are just Lazy at function boundaries
856 -- Ugly! Ask John how to improve it.
857 argDemand Top = lazyDmd
858 argDemand (Defer d) = lazyDmd
859 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
860 argDemand (Box Bot) = evalDmd
861 argDemand (Box d) = box (argDemand d)
862 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
867 -------------------------
868 -- Consider (if x then y else []) with demand V
869 -- Then the first branch gives {y->V} and the second
870 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
871 -- in the result env.
872 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
873 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
875 lub_fv = plusUFM_C lub fv1 fv2
876 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
877 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
878 -- lub is the identity for Bot
880 -- Extend the shorter argument list to match the longer
881 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
883 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
884 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
886 -----------------------------------
887 -- (t1 `bothType` t2) takes the argument/result info from t1,
888 -- using t2 just for its free-var info
889 -- NB: Don't forget about r2! It might be BotRes, which is
890 -- a bottom demand on all the in-scope variables.
891 -- Peter: can this be done more neatly?
892 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
893 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
895 both_fv = plusUFM_C both fv1 fv2
896 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
897 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
898 -- both is the identity for Abs
905 lubRes RetCPR RetCPR = RetCPR
906 lubRes r1 r2 = TopRes
908 -- If either diverges, the whole thing does
909 -- Otherwise take CPR info from the first
910 bothRes r1 BotRes = BotRes
915 modifyEnv :: Bool -- No-op if False
916 -> (Demand -> Demand) -- The zapper
917 -> DmdEnv -> DmdEnv -- Env1 and Env2
918 -> DmdEnv -> DmdEnv -- Transform this env
919 -- Zap anything in Env1 but not in Env2
920 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
922 modifyEnv need_to_modify zapper env1 env2 env
923 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
926 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
928 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
932 %************************************************************************
934 \subsection{LUB and BOTH}
936 %************************************************************************
939 lub :: Demand -> Demand -> Demand
942 lub Abs d2 = absLub d2
944 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
946 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
947 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
948 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
949 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
951 -- For the Eval case, we use these approximation rules
952 -- Box Bot <= Eval (Box Bot ...)
953 -- Box Top <= Defer (Box Bot ...)
954 -- Box (Eval ds) <= Eval (map Box ds)
955 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
956 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
957 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
958 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
959 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
961 lub (Box d1) (Box d2) = box (d1 `lub` d2)
962 lub d1@(Box _) d2 = d2 `lub` d1
964 lubs = zipWithDmds lub
966 ---------------------
967 -- box is the smart constructor for Box
968 -- It computes <B,bot> & d
969 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
970 -- Seems to be no point in allowing (Box (Call d))
971 box (Call d) = Call d -- The odd man out. Why?
973 box (Defer _) = lazyDmd
974 box Top = lazyDmd -- Box Abs and Box Top
975 box Abs = lazyDmd -- are the same <B,L>
976 box d = Box d -- Bot, Eval
979 defer :: Demand -> Demand
981 -- defer is the smart constructor for Defer
982 -- The idea is that (Defer ds) = <U(ds), L>
984 -- It specifies what happens at a lazy function argument
985 -- or a lambda; the L* operator
986 -- Set the strictness part to L, but leave
987 -- the boxity side unaffected
988 -- It also ensures that Defer (Eval [LLLL]) = L
993 defer (Call _) = lazyDmd -- Approximation here?
994 defer (Box _) = lazyDmd
995 defer (Defer ds) = Defer ds
996 defer (Eval ds) = deferEval ds
998 -- deferEval ds = defer (Eval ds)
999 deferEval ds | allTop ds = Top
1000 | otherwise = Defer ds
1002 ---------------------
1003 absLub :: Demand -> Demand
1004 -- Computes (Abs `lub` d)
1005 -- For the Bot case consider
1006 -- f x y = if ... then x else error x
1007 -- Then for y we get Abs `lub` Bot, and we really
1012 absLub (Call _) = Top
1013 absLub (Box _) = Top
1014 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1015 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1017 absLubs = mapDmds absLub
1020 both :: Demand -> Demand -> Demand
1026 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1029 -- From 'error' itself we get demand Bot on x
1030 -- From the arg demand on x we get
1031 -- x :-> evalDmd = Box (Eval (Poly Abs))
1032 -- So we get Bot `both` Box (Eval (Poly Abs))
1033 -- = Seq Keep (Poly Bot)
1036 -- f x = if ... then error (fst x) else fst x
1037 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1039 -- which is what we want.
1042 both Top Bot = errDmd
1045 both Top (Box d) = Box d
1046 both Top (Call d) = Call d
1047 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1048 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1049 -- = defer (Eval (mapDmds (`both` Top) ds))
1050 = deferEval (mapDmds (`both` Top) ds)
1053 both (Box d1) (Box d2) = box (d1 `both` d2)
1054 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1055 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1056 both (Box d1) (Defer d2) = Box d1
1057 both d1@(Box _) d2 = d2 `both` d1
1059 both (Call d1) (Call d2) = Call (d1 `both` d2)
1060 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1061 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1062 both d1@(Call _) d2 = d1 `both` d1
1064 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1065 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1066 both d1@(Eval ds1) d2 = d2 `both` d1
1068 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1069 both d1@(Defer ds1) d2 = d2 `both` d1
1071 boths = zipWithDmds both
1076 %************************************************************************
1078 \subsection{Miscellaneous
1080 %************************************************************************
1084 #ifdef OLD_STRICTNESS
1085 get_changes binds = vcat (map get_changes_bind binds)
1087 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1088 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1090 get_changes_pr (id,rhs)
1091 = get_changes_var id $$ get_changes_expr rhs
1094 | isId var = get_changes_str var $$ get_changes_dmd var
1097 get_changes_expr (Type t) = empty
1098 get_changes_expr (Var v) = empty
1099 get_changes_expr (Lit l) = empty
1100 get_changes_expr (Note n e) = get_changes_expr e
1101 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1102 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1103 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1104 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1106 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1109 | new_better && old_better = empty
1110 | new_better = message "BETTER"
1111 | old_better = message "WORSE"
1112 | otherwise = message "INCOMPARABLE"
1114 message word = text word <+> text "strictness for" <+> ppr id <+> info
1115 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1116 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1117 -- strictness analyser can't track
1118 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1119 old_better = old `betterStrictness` new
1120 new_better = new `betterStrictness` old
1123 | isUnLiftedType (idType id) = empty -- Not useful
1124 | new_better && old_better = empty
1125 | new_better = message "BETTER"
1126 | old_better = message "WORSE"
1127 | otherwise = message "INCOMPARABLE"
1129 message word = text word <+> text "demand for" <+> ppr id <+> info
1130 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1131 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1133 old = newDemand (idDemandInfo id)
1134 new_better = new `betterDemand` old
1135 old_better = old `betterDemand` new
1137 betterStrictness :: StrictSig -> StrictSig -> Bool
1138 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
1140 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
1142 betterDemand :: Demand -> Demand -> Bool
1143 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
1144 betterDemand d1 d2 = (d1 `lub` d2) == d2
1146 squashSig (StrictSig (DmdType fv ds res))
1147 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1149 -- squash just gets rid of call demands
1150 -- which the old analyser doesn't track
1151 squashDmd (Call d) = evalDmd
1152 squashDmd (Box d) = Box (squashDmd d)
1153 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1154 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)