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 arity = exprArity rhs
118 (rhs_ty, rhs') = dmdAnal emptySigEnv (vanillaCall arity) rhs
119 sig = mkTopSigTy rhs rhs_ty
122 %************************************************************************
124 \subsection{The analyser itself}
126 %************************************************************************
129 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
131 dmdAnal sigs Abs e = (topDmdType, e)
134 | not (isStrictDmd dmd)
136 (res_ty, e') = dmdAnal sigs evalDmd e
138 (deferType res_ty, e')
139 -- It's important not to analyse e with a lazy demand because
140 -- a) When we encounter case s of (a,b) ->
141 -- we demand s with U(d1d2)... but if the overall demand is lazy
142 -- that is wrong, and we'd need to reduce the demand on s,
143 -- which is inconvenient
144 -- b) More important, consider
145 -- f (let x = R in x+x), where f is lazy
146 -- We still want to mark x as demanded, because it will be when we
147 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
148 -- just mark x as Lazy
149 -- c) The application rule wouldn't be right either
150 -- Evaluating (f x) in a L demand does *not* cause
151 -- evaluation of f in a C(L) demand!
154 dmdAnal sigs dmd (Lit lit)
155 = (topDmdType, Lit lit)
157 dmdAnal sigs dmd (Var var)
158 = (dmdTransform sigs var dmd, Var var)
160 dmdAnal sigs dmd (Note n e)
161 = (dmd_ty, Note n e')
163 (dmd_ty, e') = dmdAnal sigs dmd' e
165 Coerce _ _ -> evalDmd -- This coerce usually arises from a recursive
166 other -> dmd -- newtype, and we don't want to look inside them
167 -- for exactly the same reason that we don't look
168 -- inside recursive products -- we might not reach
169 -- a fixpoint. So revert to a vanilla Eval demand
171 dmdAnal sigs dmd (App fun (Type ty))
172 = (fun_ty, App fun' (Type ty))
174 (fun_ty, fun') = dmdAnal sigs dmd fun
176 -- Lots of the other code is there to make this
177 -- beautiful, compositional, application rule :-)
178 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
179 = let -- [Type arg handled above]
180 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
181 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
182 (arg_dmd, res_ty) = splitDmdTy fun_ty
184 (res_ty `bothType` arg_ty, App fun' arg')
186 dmdAnal sigs dmd (Lam var body)
189 (body_ty, body') = dmdAnal sigs dmd body
191 (body_ty, Lam var body')
193 | Call body_dmd <- dmd -- A call demand: good!
195 sigs' = extendSigsWithLam sigs var
196 (body_ty, body') = dmdAnal sigs' body_dmd body
197 (lam_ty, var') = annotateLamIdBndr body_ty var
199 (lam_ty, Lam var' body')
201 | otherwise -- Not enough demand on the lambda; but do the body
202 = let -- anyway to annotate it and gather free var info
203 (body_ty, body') = dmdAnal sigs evalDmd body
204 (lam_ty, var') = annotateLamIdBndr body_ty var
206 (deferType lam_ty, Lam var' body')
208 dmdAnal sigs dmd (Case scrut case_bndr ty [alt@(DataAlt dc,bndrs,rhs)])
209 | let tycon = dataConTyCon dc,
210 isProductTyCon tycon,
211 not (isRecursiveTyCon tycon)
213 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
214 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
215 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
216 (_, bndrs', _) = alt'
217 case_bndr_sig = cprSig
218 -- Inside the alternative, the case binder has the CPR property.
219 -- Meaning that a case on it will successfully cancel.
221 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
224 -- We want f to have the CPR property:
225 -- f b x = case fw b x of { r -> I# r }
226 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
229 -- Figure out whether the demand on the case binder is used, and use
230 -- that to set the scrut_dmd. This is utterly essential.
231 -- Consider f x = case x of y { (a,b) -> k y a }
232 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
233 -- worker, so the worker will rebuild
234 -- x = (a, absent-error)
235 -- and that'll crash.
236 -- So at one stage I had:
237 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
238 -- keepity | dead_case_bndr = Drop
239 -- | otherwise = Keep
242 -- case x of y { (a,b) -> h y + a }
243 -- where h : U(LL) -> T
244 -- The above code would compute a Keep for x, since y is not Abs, which is silly
245 -- The insight is, of course, that a demand on y is a demand on the
246 -- scrutinee, so we need to `both` it with the scrut demand
248 scrut_dmd = Eval (Prod [idNewDemandInfo b | b <- bndrs', isId b])
250 idNewDemandInfo case_bndr'
252 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
254 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' ty [alt'])
256 dmdAnal sigs dmd (Case scrut case_bndr ty alts)
258 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
259 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
260 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
262 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
263 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' ty alts')
265 dmdAnal sigs dmd (Let (NonRec id rhs) body)
267 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
268 (body_ty, body') = dmdAnal sigs' dmd body
269 (body_ty1, id2) = annotateBndr body_ty id1
270 body_ty2 = addLazyFVs body_ty1 lazy_fv
272 -- If the actual demand is better than the vanilla call
273 -- demand, you might think that we might do better to re-analyse
274 -- the RHS with the stronger demand.
275 -- But (a) That seldom happens, because it means that *every* path in
276 -- the body of the let has to use that stronger demand
277 -- (b) It often happens temporarily in when fixpointing, because
278 -- the recursive function at first seems to place a massive demand.
279 -- But we don't want to go to extra work when the function will
280 -- probably iterate to something less demanding.
281 -- In practice, all the times the actual demand on id2 is more than
282 -- the vanilla call demand seem to be due to (b). So we don't
283 -- bother to re-analyse the RHS.
284 (body_ty2, Let (NonRec id2 rhs') body')
286 dmdAnal sigs dmd (Let (Rec pairs) body)
288 bndrs = map fst pairs
289 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
290 (body_ty, body') = dmdAnal sigs' dmd body
291 body_ty1 = addLazyFVs body_ty lazy_fv
293 sigs' `seq` body_ty `seq`
295 (body_ty2, _) = annotateBndrs body_ty1 bndrs
296 -- Don't bother to add demand info to recursive
297 -- binders as annotateBndr does;
298 -- being recursive, we can't treat them strictly.
299 -- But we do need to remove the binders from the result demand env
301 (body_ty2, Let (Rec pairs') body')
304 dmdAnalAlt sigs dmd (con,bndrs,rhs)
306 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
307 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
308 final_alt_ty | io_hack_reqd = alt_ty `lubType` topDmdType
311 -- There's a hack here for I/O operations. Consider
312 -- case foo x s of { (# s, r #) -> y }
313 -- Is this strict in 'y'. Normally yes, but what if 'foo' is an I/O
314 -- operation that simply terminates the program (not in an erroneous way)?
315 -- In that case we should not evaluate y before the call to 'foo'.
316 -- Hackish solution: spot the IO-like situation and add a virtual branch,
320 -- other -> return ()
321 -- So the 'y' isn't necessarily going to be evaluated
323 -- A more complete example where this shows up is:
324 -- do { let len = <expensive> ;
325 -- ; when (...) (exitWith ExitSuccess)
328 io_hack_reqd = con == DataAlt unboxedPairDataCon &&
329 idType (head bndrs) `coreEqType` realWorldStatePrimTy
331 (final_alt_ty, (con, bndrs', rhs'))
334 %************************************************************************
336 \subsection{Bindings}
338 %************************************************************************
341 dmdFix :: TopLevelFlag
342 -> SigEnv -- Does not include bindings for this binding
345 [(Id,CoreExpr)]) -- Binders annotated with stricness info
347 dmdFix top_lvl sigs orig_pairs
348 = loop 1 initial_sigs orig_pairs
350 bndrs = map fst orig_pairs
351 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
354 -> SigEnv -- Already contains the current sigs
356 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
359 = (sigs', lazy_fv, pairs')
360 -- Note: use pairs', not pairs. pairs' is the result of
361 -- processing the RHSs with sigs (= sigs'), whereas pairs
362 -- is the result of processing the RHSs with the *previous*
363 -- iteration of sigs.
365 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
366 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
367 text "env:" <+> ppr (ufmToList sigs),
368 text "binds:" <+> pprCoreBinding (Rec pairs)]))
369 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
370 -- The lazy_fv part is really important! orig_pairs has no strictness
371 -- info, including nothing about free vars. But if we have
372 -- letrec f = ....y..... in ...f...
373 -- where 'y' is free in f, we must record that y is mentioned,
374 -- otherwise y will get recorded as absent altogether
376 | otherwise = loop (n+1) sigs' pairs'
378 found_fixpoint = all (same_sig sigs sigs') bndrs
379 -- Use the new signature to do the next pair
380 -- The occurrence analyser has arranged them in a good order
381 -- so this can significantly reduce the number of iterations needed
382 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
384 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
385 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
387 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
388 ((sigs', lazy_fv'), pair')
391 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
392 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
393 -- old_sig = lookup sigs id
394 -- new_sig = lookup sigs' id
396 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
397 lookup sigs var = case lookupVarEnv sigs var of
400 -- Get an initial strictness signature from the Id
401 -- itself. That way we make use of earlier iterations
402 -- of the fixpoint algorithm. (Cunning plan.)
403 -- Note that the cunning plan extends to the DmdEnv too,
404 -- since it is part of the strictness signature
405 initialSig id = idNewStrictness_maybe id `orElse` botSig
407 dmdAnalRhs :: TopLevelFlag -> RecFlag
408 -> SigEnv -> (Id, CoreExpr)
409 -> (SigEnv, DmdEnv, (Id, CoreExpr))
410 -- Process the RHS of the binding, add the strictness signature
411 -- to the Id, and augment the environment with the signature as well.
413 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
414 = (sigs', lazy_fv, (id', rhs'))
416 arity = idArity id -- The idArity should be up to date
417 -- The simplifier was run just beforehand
418 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
419 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs), ppr id )
420 -- The RHS can be eta-reduced to just a variable,
421 -- in which case we should not complain.
422 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
423 id' = id `setIdNewStrictness` sig_ty
424 sigs' = extendSigEnv top_lvl sigs id sig_ty
427 %************************************************************************
429 \subsection{Strictness signatures and types}
431 %************************************************************************
434 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
435 -- Take a DmdType and turn it into a StrictSig
436 -- NB: not used for never-inline things; hence False
437 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
439 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
440 mkSigTy top_lvl rec_flag id rhs dmd_ty
441 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
443 never_inline = isNeverActive (idInlinePragma id)
444 maybe_id_dmd = idNewDemandInfo_maybe id
445 -- Is Nothing the first time round
448 | isTopLevel top_lvl = False -- Top level things don't get
449 -- their demandInfo set at all
450 | isRec rec_flag = False -- Ditto recursive things
451 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
452 | otherwise = True -- Optimistic, first time round
456 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
457 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
458 If the rhs is a thunk, we usually forget the CPR info, because
459 it is presumably shared (else it would have been inlined, and
460 so we'd lose sharing if w/w'd it into a function.
462 However, if the strictness analyser has figured out (in a previous
463 iteration) that it's strict, then we DON'T need to forget the CPR info.
464 Instead we can retain the CPR info and do the thunk-splitting transform
465 (see WorkWrap.splitThunk).
467 This made a big difference to PrelBase.modInt, which had something like
468 modInt = \ x -> let r = ... -> I# v in
469 ...body strict in r...
470 r's RHS isn't a value yet; but modInt returns r in various branches, so
471 if r doesn't have the CPR property then neither does modInt
472 Another case I found in practice (in Complex.magnitude), looks like this:
473 let k = if ... then I# a else I# b
474 in ... body strict in k ....
475 (For this example, it doesn't matter whether k is returned as part of
476 the overall result; but it does matter that k's RHS has the CPR property.)
477 Left to itself, the simplifier will make a join point thus:
478 let $j k = ...body strict in k...
479 if ... then $j (I# a) else $j (I# b)
480 With thunk-splitting, we get instead
481 let $j x = let k = I#x in ...body strict in k...
482 in if ... then $j a else $j b
483 This is much better; there's a good chance the I# won't get allocated.
485 The difficulty with this is that we need the strictness type to
486 look at the body... but we now need the body to calculate the demand
487 on the variable, so we can decide whether its strictness type should
488 have a CPR in it or not. Simple solution:
489 a) use strictness info from the previous iteration
490 b) make sure we do at least 2 iterations, by doing a second
491 round for top-level non-recs. Top level recs will get at
492 least 2 iterations except for totally-bottom functions
493 which aren't very interesting anyway.
495 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
497 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
498 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
499 Demand info now has a 'Nothing' state, just like strictness info.
500 The analysis works from 'dangerous' towards a 'safe' state; so we
501 start with botSig for 'Nothing' strictness infos, and we start with
502 "yes, it's demanded" for 'Nothing' in the demand info. The
503 fixpoint iteration will sort it all out.
505 We can't start with 'not-demanded' because then consider
509 if ... then t else I# y else f x'
511 In the first iteration we'd have no demand info for x, so assume
512 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
513 we'd see that t was demanded, and so give it the CPR property, but by
514 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
515 setting the first time round, we say 'yes t is demanded' the first
518 However, this does mean that for non-recursive bindings we must
519 iterate twice to be sure of not getting over-optimistic CPR info,
520 in the case where t turns out to be not-demanded. This is handled
525 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
526 | never_inline && not (isBotRes res)
528 -- Don't strictness-analyse NOINLINE things. Why not? Because
529 -- the NOINLINE says "don't expose any of the inner workings at the call
530 -- site" and the strictness is certainly an inner working.
532 -- More concretely, the demand analyser discovers the following strictness
533 -- for unsafePerformIO: C(U(AV))
535 -- unsafePerformIO (\s -> let r = f x in
536 -- case writeIORef v r s of (# s1, _ #) ->
538 -- The strictness analyser will find that the binding for r is strict,
539 -- (becuase of uPIO's strictness sig), and so it'll evaluate it before
540 -- doing the writeIORef. This actually makes tests/lib/should_run/memo002
543 -- Solution: don't expose the strictness of unsafePerformIO.
545 -- But we do want to expose the strictness of error functions,
546 -- which are also often marked NOINLINE
547 -- {-# NOINLINE foo #-}
548 -- foo x = error ("wubble buggle" ++ x)
549 -- So (hack, hack) we only drop the strictness for non-bottom things
550 -- This is all very unsatisfactory.
551 = (deferEnv fv, topSig)
554 = (lazy_fv, mkStrictSig dmd_ty)
556 dmd_ty = DmdType strict_fv final_dmds res'
558 lazy_fv = filterUFM (not . isStrictDmd) fv
559 strict_fv = filterUFM isStrictDmd fv
560 -- We put the strict FVs in the DmdType of the Id, so
561 -- that at its call sites we unleash demands on its strict fvs.
562 -- An example is 'roll' in imaginary/wheel-sieve2
563 -- Something like this:
565 -- go y = if ... then roll (x-1) else x+1
568 -- We want to see that roll is strict in x, which is because
569 -- go is called. So we put the DmdEnv for x in go's DmdType.
572 -- f :: Int -> Int -> Int
573 -- f x y = let t = x+1
574 -- h z = if z==0 then t else
575 -- if z==1 then x+1 else
579 -- Calling h does indeed evaluate x, but we can only see
580 -- that if we unleash a demand on x at the call site for t.
582 -- Incidentally, here's a place where lambda-lifting h would
583 -- lose the cigar --- we couldn't see the joint strictness in t/x
586 -- We don't want to put *all* the fv's from the RHS into the
587 -- DmdType, because that makes fixpointing very slow --- the
588 -- DmdType gets full of lazy demands that are slow to converge.
590 final_dmds = setUnpackStrategy dmds
591 -- Set the unpacking strategy
594 RetCPR | ignore_cpr_info -> TopRes
596 ignore_cpr_info = not (exprIsHNF rhs || thunk_cpr_ok)
599 The unpack strategy determines whether we'll *really* unpack the argument,
600 or whether we'll just remember its strictness. If unpacking would give
601 rise to a *lot* of worker args, we may decide not to unpack after all.
604 setUnpackStrategy :: [Demand] -> [Demand]
606 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
608 go :: Int -- Max number of args available for sub-components of [Demand]
610 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
612 go n (Eval (Prod cs) : ds)
613 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
614 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
617 n' = n + 1 - non_abs_args
618 -- Add one to the budget 'cos we drop the top-level arg
619 non_abs_args = nonAbsentArgs cs
620 -- Delete # of non-absent args to which we'll now be committed
622 go n (d:ds) = d `cons` go n ds
625 cons d (n,ds) = (n, d:ds)
627 nonAbsentArgs :: [Demand] -> Int
629 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
630 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
634 %************************************************************************
636 \subsection{Strictness signatures and types}
638 %************************************************************************
641 splitDmdTy :: DmdType -> (Demand, DmdType)
642 -- Split off one function argument
643 -- We already have a suitable demand on all
644 -- free vars, so no need to add more!
645 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
646 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
650 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
652 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
653 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
654 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
656 addLazyFVs (DmdType fv ds res) lazy_fvs
657 = DmdType both_fv1 ds res
659 both_fv = (plusUFM_C both fv lazy_fvs)
660 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
661 -- This modifyEnv is vital. Consider
662 -- let f = \x -> (x,y)
664 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
665 -- demand with the bottom coming up from 'error'
667 -- I got a loop in the fixpointer without this, due to an interaction
668 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
670 -- = letrec g y = x `fatbar`
671 -- letrec h z = z + ...g...
674 -- In the initial iteration for f, f=Bot
675 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
676 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
677 -- places on its free variables. Suppose it places none. Then the
678 -- x `fatbar` ...call to h...
679 -- will give a x->V demand for x. That turns into a L demand for x,
680 -- which floats out of the defn for h. Without the modifyEnv, that
681 -- L demand doesn't get both'd with the Bot coming up from the inner
682 -- call to f. So we just get an L demand for x for g.
684 -- A better way to say this is that the lazy-fv filtering should give the
685 -- same answer as putting the lazy fv demands in the function's type.
687 annotateBndr :: DmdType -> Var -> (DmdType, Var)
688 -- The returned env has the var deleted
689 -- The returned var is annotated with demand info
690 -- No effect on the argument demands
691 annotateBndr dmd_ty@(DmdType fv ds res) var
692 | isTyVar var = (dmd_ty, var)
693 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
695 (fv', dmd) = removeFV fv var res
697 annotateBndrs = mapAccumR annotateBndr
699 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
700 -- For lambdas we add the demand to the argument demands
701 -- Only called for Ids
703 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
705 (fv', dmd) = removeFV fv id res
706 hacked_dmd = argDemand dmd
707 -- This call to argDemand is vital, because otherwise we label
708 -- a lambda binder with demand 'B'. But in terms of calling
709 -- conventions that's Abs, because we don't pass it. But
710 -- when we do a w/w split we get
711 -- fw x = (\x y:B -> ...) x (error "oops")
712 -- And then the simplifier things the 'B' is a strict demand
713 -- and evaluates the (error "oops"). Sigh
715 removeFV fv id res = (fv', zapUnlifted id dmd)
717 fv' = fv `delVarEnv` id
718 dmd = lookupVarEnv fv id `orElse` deflt
719 deflt | isBotRes res = Bot
722 -- For unlifted-type variables, we are only
723 -- interested in Bot/Abs/Box Abs
724 zapUnlifted is Bot = Bot
725 zapUnlifted id Abs = Abs
726 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
730 %************************************************************************
732 \subsection{Strictness signatures}
734 %************************************************************************
737 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
738 -- We use the SigEnv to tell us whether to
739 -- record info about a variable in the DmdEnv
740 -- We do so if it's a LocalId, but not top-level
742 -- The DmdEnv gives the demand on the free vars of the function
743 -- when it is given enough args to satisfy the strictness signature
745 emptySigEnv = emptyVarEnv
747 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
748 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
750 extendSigEnvList = extendVarEnvList
752 extendSigsWithLam :: SigEnv -> Id -> SigEnv
753 -- Extend the SigEnv when we meet a lambda binder
754 -- If the binder is marked demanded with a product demand, then give it a CPR
755 -- signature, because in the likely event that this is a lambda on a fn defn
756 -- [we only use this when the lambda is being consumed with a call demand],
757 -- it'll be w/w'd and so it will be CPR-ish. E.g.
758 -- f = \x::(Int,Int). if ...strict in x... then
762 -- We want f to have the CPR property because x does, by the time f has been w/w'd
764 -- NOTE: see notes [CPR-AND-STRICTNESS]
766 -- Also note that we only want to do this for something that
767 -- definitely has product type, else we may get over-optimistic
768 -- CPR results (e.g. from \x -> x!).
770 extendSigsWithLam sigs id
771 = case idNewDemandInfo_maybe id of
772 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
773 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
777 dmdTransform :: SigEnv -- The strictness environment
778 -> Id -- The function
779 -> Demand -- The demand on the function
780 -> DmdType -- The demand type of the function in this context
781 -- Returned DmdEnv includes the demand on
782 -- this function plus demand on its free variables
784 dmdTransform sigs var dmd
786 ------ DATA CONSTRUCTOR
787 | isDataConWorkId var -- Data constructor
789 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
790 DmdType _ _ con_res = dmd_ty
793 if arity == call_depth then -- Saturated, so unleash the demand
795 -- Important! If we Keep the constructor application, then
796 -- we need the demands the constructor places (always lazy)
797 -- If not, we don't need to. For example:
798 -- f p@(x,y) = (p,y) -- S(AL)
800 -- It's vital that we don't calculate Absent for a!
801 dmd_ds = case res_dmd of
802 Box (Eval ds) -> mapDmds box ds
806 -- ds can be empty, when we are just seq'ing the thing
807 -- If so we must make up a suitable bunch of demands
808 arg_ds = case dmd_ds of
809 Poly d -> replicate arity d
810 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
813 mkDmdType emptyDmdEnv arg_ds con_res
814 -- Must remember whether it's a product, hence con_res, not TopRes
818 ------ IMPORTED FUNCTION
819 | isGlobalId var, -- Imported function
820 let StrictSig dmd_ty = idNewStrictness var
821 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
826 ------ LOCAL LET/REC BOUND THING
827 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
829 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
830 | otherwise = deferType dmd_ty
831 -- NB: it's important to use deferType, and not just return topDmdType
832 -- Consider let { f x y = p + x } in f 1
833 -- The application isn't saturated, but we must nevertheless propagate
834 -- a lazy demand for p!
836 addVarDmd top_lvl fn_ty var dmd
838 ------ LOCAL NON-LET/REC BOUND THING
839 | otherwise -- Default case
843 (call_depth, res_dmd) = splitCallDmd dmd
847 %************************************************************************
851 %************************************************************************
854 splitCallDmd :: Demand -> (Int, Demand)
855 splitCallDmd (Call d) = case splitCallDmd d of
857 splitCallDmd d = (0, d)
859 vanillaCall :: Arity -> Demand
860 vanillaCall 0 = evalDmd
861 vanillaCall n = Call (vanillaCall (n-1))
863 deferType :: DmdType -> DmdType
864 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
865 -- Notice that we throw away info about both arguments and results
866 -- For example, f = let ... in \x -> x
867 -- We don't want to get a stricness type V->T for f.
870 deferEnv :: DmdEnv -> DmdEnv
871 deferEnv fv = mapVarEnv defer fv
875 argDemand :: Demand -> Demand
876 -- The 'Defer' demands are just Lazy at function boundaries
877 -- Ugly! Ask John how to improve it.
878 argDemand Top = lazyDmd
879 argDemand (Defer d) = lazyDmd
880 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
881 argDemand (Box Bot) = evalDmd
882 argDemand (Box d) = box (argDemand d)
883 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
888 -------------------------
889 -- Consider (if x then y else []) with demand V
890 -- Then the first branch gives {y->V} and the second
891 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
892 -- in the result env.
893 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
894 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
896 lub_fv = plusUFM_C lub fv1 fv2
897 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
898 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
899 -- lub is the identity for Bot
901 -- Extend the shorter argument list to match the longer
902 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
904 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
905 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
907 -----------------------------------
908 -- (t1 `bothType` t2) takes the argument/result info from t1,
909 -- using t2 just for its free-var info
910 -- NB: Don't forget about r2! It might be BotRes, which is
911 -- a bottom demand on all the in-scope variables.
912 -- Peter: can this be done more neatly?
913 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
914 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
916 both_fv = plusUFM_C both fv1 fv2
917 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
918 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
919 -- both is the identity for Abs
926 lubRes RetCPR RetCPR = RetCPR
927 lubRes r1 r2 = TopRes
929 -- If either diverges, the whole thing does
930 -- Otherwise take CPR info from the first
931 bothRes r1 BotRes = BotRes
936 modifyEnv :: Bool -- No-op if False
937 -> (Demand -> Demand) -- The zapper
938 -> DmdEnv -> DmdEnv -- Env1 and Env2
939 -> DmdEnv -> DmdEnv -- Transform this env
940 -- Zap anything in Env1 but not in Env2
941 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
943 modifyEnv need_to_modify zapper env1 env2 env
944 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
947 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
949 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
953 %************************************************************************
955 \subsection{LUB and BOTH}
957 %************************************************************************
960 lub :: Demand -> Demand -> Demand
963 lub Abs d2 = absLub d2
965 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
967 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
968 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
969 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
970 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
972 -- For the Eval case, we use these approximation rules
973 -- Box Bot <= Eval (Box Bot ...)
974 -- Box Top <= Defer (Box Bot ...)
975 -- Box (Eval ds) <= Eval (map Box ds)
976 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
977 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
978 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
979 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
980 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
982 lub (Box d1) (Box d2) = box (d1 `lub` d2)
983 lub d1@(Box _) d2 = d2 `lub` d1
985 lubs = zipWithDmds lub
987 ---------------------
988 -- box is the smart constructor for Box
989 -- It computes <B,bot> & d
990 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
991 -- Seems to be no point in allowing (Box (Call d))
992 box (Call d) = Call d -- The odd man out. Why?
994 box (Defer _) = lazyDmd
995 box Top = lazyDmd -- Box Abs and Box Top
996 box Abs = lazyDmd -- are the same <B,L>
997 box d = Box d -- Bot, Eval
1000 defer :: Demand -> Demand
1002 -- defer is the smart constructor for Defer
1003 -- The idea is that (Defer ds) = <U(ds), L>
1005 -- It specifies what happens at a lazy function argument
1006 -- or a lambda; the L* operator
1007 -- Set the strictness part to L, but leave
1008 -- the boxity side unaffected
1009 -- It also ensures that Defer (Eval [LLLL]) = L
1014 defer (Call _) = lazyDmd -- Approximation here?
1015 defer (Box _) = lazyDmd
1016 defer (Defer ds) = Defer ds
1017 defer (Eval ds) = deferEval ds
1019 -- deferEval ds = defer (Eval ds)
1020 deferEval ds | allTop ds = Top
1021 | otherwise = Defer ds
1023 ---------------------
1024 absLub :: Demand -> Demand
1025 -- Computes (Abs `lub` d)
1026 -- For the Bot case consider
1027 -- f x y = if ... then x else error x
1028 -- Then for y we get Abs `lub` Bot, and we really
1033 absLub (Call _) = Top
1034 absLub (Box _) = Top
1035 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1036 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1038 absLubs = mapDmds absLub
1041 both :: Demand -> Demand -> Demand
1047 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1050 -- From 'error' itself we get demand Bot on x
1051 -- From the arg demand on x we get
1052 -- x :-> evalDmd = Box (Eval (Poly Abs))
1053 -- So we get Bot `both` Box (Eval (Poly Abs))
1054 -- = Seq Keep (Poly Bot)
1057 -- f x = if ... then error (fst x) else fst x
1058 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1060 -- which is what we want.
1063 both Top Bot = errDmd
1066 both Top (Box d) = Box d
1067 both Top (Call d) = Call d
1068 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1069 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1070 -- = defer (Eval (mapDmds (`both` Top) ds))
1071 = deferEval (mapDmds (`both` Top) ds)
1074 both (Box d1) (Box d2) = box (d1 `both` d2)
1075 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1076 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1077 both (Box d1) (Defer d2) = Box d1
1078 both d1@(Box _) d2 = d2 `both` d1
1080 both (Call d1) (Call d2) = Call (d1 `both` d2)
1081 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1082 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1083 both d1@(Call _) d2 = d1 `both` d1
1085 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1086 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1087 both d1@(Eval ds1) d2 = d2 `both` d1
1089 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1090 both d1@(Defer ds1) d2 = d2 `both` d1
1092 boths = zipWithDmds both
1097 %************************************************************************
1099 \subsection{Miscellaneous
1101 %************************************************************************
1105 #ifdef OLD_STRICTNESS
1106 get_changes binds = vcat (map get_changes_bind binds)
1108 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1109 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1111 get_changes_pr (id,rhs)
1112 = get_changes_var id $$ get_changes_expr rhs
1115 | isId var = get_changes_str var $$ get_changes_dmd var
1118 get_changes_expr (Type t) = empty
1119 get_changes_expr (Var v) = empty
1120 get_changes_expr (Lit l) = empty
1121 get_changes_expr (Note n e) = get_changes_expr e
1122 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1123 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1124 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1125 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1127 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1130 | new_better && old_better = empty
1131 | new_better = message "BETTER"
1132 | old_better = message "WORSE"
1133 | otherwise = message "INCOMPARABLE"
1135 message word = text word <+> text "strictness for" <+> ppr id <+> info
1136 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1137 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1138 -- strictness analyser can't track
1139 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1140 old_better = old `betterStrictness` new
1141 new_better = new `betterStrictness` old
1144 | isUnLiftedType (idType id) = empty -- Not useful
1145 | new_better && old_better = empty
1146 | new_better = message "BETTER"
1147 | old_better = message "WORSE"
1148 | otherwise = message "INCOMPARABLE"
1150 message word = text word <+> text "demand for" <+> ppr id <+> info
1151 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1152 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1154 old = newDemand (idDemandInfo id)
1155 new_better = new `betterDemand` old
1156 old_better = old `betterDemand` new
1158 betterStrictness :: StrictSig -> StrictSig -> Bool
1159 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
1161 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
1163 betterDemand :: Demand -> Demand -> Bool
1164 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
1165 betterDemand d1 d2 = (d1 `lub` d2) == d2
1167 squashSig (StrictSig (DmdType fv ds res))
1168 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1170 -- squash just gets rid of call demands
1171 -- which the old analyser doesn't track
1172 squashDmd (Call d) = evalDmd
1173 squashDmd (Box d) = Box (squashDmd d)
1174 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1175 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)