2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
11 -- The above warning supression flag is a temporary kludge.
12 -- While working on this module you are encouraged to remove it and fix
13 -- any warnings in the module. See
14 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
17 module DmdAnal ( dmdAnalPgm, dmdAnalTopRhs,
18 both {- needed by WwLib -}
21 #include "HsVersions.h"
23 import DynFlags ( DynFlags, DynFlag(..) )
24 import StaticFlags ( opt_MaxWorkerArgs )
25 import NewDemand -- All of it
28 import CoreUtils ( exprIsHNF, exprIsTrivial, exprArity )
29 import DataCon ( dataConTyCon )
30 import TyCon ( isProductTyCon, isRecursiveTyCon )
31 import Id ( Id, idType, idInlinePragma,
32 isDataConWorkId, isGlobalId, idArity,
34 idDemandInfo, idStrictness, idCprInfo, idName,
36 idNewStrictness, idNewStrictness_maybe,
37 setIdNewStrictness, idNewDemandInfo,
38 idNewDemandInfo_maybe,
42 import IdInfo ( newStrictnessFromOld, newDemand )
46 import TysWiredIn ( unboxedPairDataCon )
47 import TysPrim ( realWorldStatePrimTy )
48 import UniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly,
49 keysUFM, minusUFM, ufmToList, filterUFM )
50 import Type ( isUnLiftedType, coreEqType, splitTyConApp_maybe )
51 import Coercion ( coercionKind )
52 import CoreLint ( showPass, endPass )
53 import Util ( mapAndUnzip, lengthIs )
54 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive,
56 import Maybes ( orElse, expectJust )
64 * set a noinline pragma on bottoming Ids
66 * Consider f x = x+1 `fatbar` error (show x)
67 We'd like to unbox x, even if that means reboxing it in the error case.
70 %************************************************************************
72 \subsection{Top level stuff}
74 %************************************************************************
77 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
78 dmdAnalPgm dflags binds
80 let { binds_plus_dmds = do_prog binds } ;
82 -- Only if OLD_STRICTNESS is on, because only then is the old
83 -- strictness analyser run
84 let { dmd_changes = get_changes binds_plus_dmds } ;
85 printDump (text "Changes in demands" $$ dmd_changes) ;
87 return binds_plus_dmds
90 do_prog :: [CoreBind] -> [CoreBind]
91 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
93 dmdAnalTopBind :: SigEnv
96 dmdAnalTopBind sigs (NonRec id rhs)
98 ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs)
99 (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs1)
100 -- Do two passes to improve CPR information
101 -- See comments with ignore_cpr_info in mk_sig_ty
102 -- and with extendSigsWithLam
104 (sigs2, NonRec id2 rhs2)
106 dmdAnalTopBind sigs (Rec pairs)
108 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
109 -- We get two iterations automatically
110 -- c.f. the NonRec case above
116 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
117 -- Analyse the RHS and return
118 -- a) appropriate strictness info
119 -- b) the unfolding (decorated with stricntess info)
123 call_dmd = vanillaCall (exprArity rhs)
124 (_, rhs1) = dmdAnal emptySigEnv call_dmd rhs
125 (rhs_ty, rhs2) = dmdAnal emptySigEnv call_dmd rhs1
126 sig = mkTopSigTy rhs rhs_ty
127 -- Do two passes; see notes with extendSigsWithLam
128 -- Otherwise we get bogus CPR info for constructors like
129 -- newtype T a = MkT a
130 -- The constructor looks like (\x::T a -> x), modulo the coerce
131 -- extendSigsWithLam will optimistically give x a CPR tag the
132 -- first time, which is wrong in the end.
135 %************************************************************************
137 \subsection{The analyser itself}
139 %************************************************************************
142 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
144 dmdAnal sigs Abs e = (topDmdType, e)
147 | not (isStrictDmd dmd)
149 (res_ty, e') = dmdAnal sigs evalDmd e
151 (deferType res_ty, e')
152 -- It's important not to analyse e with a lazy demand because
153 -- a) When we encounter case s of (a,b) ->
154 -- we demand s with U(d1d2)... but if the overall demand is lazy
155 -- that is wrong, and we'd need to reduce the demand on s,
156 -- which is inconvenient
157 -- b) More important, consider
158 -- f (let x = R in x+x), where f is lazy
159 -- We still want to mark x as demanded, because it will be when we
160 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
161 -- just mark x as Lazy
162 -- c) The application rule wouldn't be right either
163 -- Evaluating (f x) in a L demand does *not* cause
164 -- evaluation of f in a C(L) demand!
167 dmdAnal sigs dmd (Lit lit)
168 = (topDmdType, Lit lit)
170 dmdAnal sigs dmd (Var var)
171 = (dmdTransform sigs var dmd, Var var)
173 dmdAnal sigs dmd (Cast e co)
174 = (dmd_ty, Cast e' co)
176 (dmd_ty, e') = dmdAnal sigs dmd' e
177 to_co = snd (coercionKind co)
179 | Just (tc, args) <- splitTyConApp_maybe to_co
180 , isRecursiveTyCon tc = evalDmd
182 -- This coerce usually arises from a recursive
183 -- newtype, and we don't want to look inside them
184 -- for exactly the same reason that we don't look
185 -- inside recursive products -- we might not reach
186 -- a fixpoint. So revert to a vanilla Eval demand
188 dmdAnal sigs dmd (Note n e)
189 = (dmd_ty, Note n e')
191 (dmd_ty, e') = dmdAnal sigs dmd e
193 dmdAnal sigs dmd (App fun (Type ty))
194 = (fun_ty, App fun' (Type ty))
196 (fun_ty, fun') = dmdAnal sigs dmd fun
198 -- Lots of the other code is there to make this
199 -- beautiful, compositional, application rule :-)
200 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
201 = let -- [Type arg handled above]
202 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
203 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
204 (arg_dmd, res_ty) = splitDmdTy fun_ty
206 (res_ty `bothType` arg_ty, App fun' arg')
208 dmdAnal sigs dmd (Lam var body)
211 (body_ty, body') = dmdAnal sigs dmd body
213 (body_ty, Lam var body')
215 | Call body_dmd <- dmd -- A call demand: good!
217 sigs' = extendSigsWithLam sigs var
218 (body_ty, body') = dmdAnal sigs' body_dmd body
219 (lam_ty, var') = annotateLamIdBndr body_ty var
221 (lam_ty, Lam var' body')
223 | otherwise -- Not enough demand on the lambda; but do the body
224 = let -- anyway to annotate it and gather free var info
225 (body_ty, body') = dmdAnal sigs evalDmd body
226 (lam_ty, var') = annotateLamIdBndr body_ty var
228 (deferType lam_ty, Lam var' body')
230 dmdAnal sigs dmd (Case scrut case_bndr ty [alt@(DataAlt dc,bndrs,rhs)])
231 | let tycon = dataConTyCon dc,
232 isProductTyCon tycon,
233 not (isRecursiveTyCon tycon)
235 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
236 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
237 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
238 (_, bndrs', _) = alt'
239 case_bndr_sig = cprSig
240 -- Inside the alternative, the case binder has the CPR property.
241 -- Meaning that a case on it will successfully cancel.
243 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
246 -- We want f to have the CPR property:
247 -- f b x = case fw b x of { r -> I# r }
248 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
251 -- Figure out whether the demand on the case binder is used, and use
252 -- that to set the scrut_dmd. This is utterly essential.
253 -- Consider f x = case x of y { (a,b) -> k y a }
254 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
255 -- worker, so the worker will rebuild
256 -- x = (a, absent-error)
257 -- and that'll crash.
258 -- So at one stage I had:
259 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
260 -- keepity | dead_case_bndr = Drop
261 -- | otherwise = Keep
264 -- case x of y { (a,b) -> h y + a }
265 -- where h : U(LL) -> T
266 -- The above code would compute a Keep for x, since y is not Abs, which is silly
267 -- The insight is, of course, that a demand on y is a demand on the
268 -- scrutinee, so we need to `both` it with the scrut demand
270 alt_dmd = Eval (Prod [idNewDemandInfo b | b <- bndrs', isId b])
271 scrut_dmd = alt_dmd `both`
272 idNewDemandInfo case_bndr'
274 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
276 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' ty [alt'])
278 dmdAnal sigs dmd (Case scrut case_bndr ty alts)
280 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
281 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
282 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
284 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
285 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' ty alts')
287 dmdAnal sigs dmd (Let (NonRec id rhs) body)
289 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
290 (body_ty, body') = dmdAnal sigs' dmd body
291 (body_ty1, id2) = annotateBndr body_ty id1
292 body_ty2 = addLazyFVs body_ty1 lazy_fv
294 -- If the actual demand is better than the vanilla call
295 -- demand, you might think that we might do better to re-analyse
296 -- the RHS with the stronger demand.
297 -- But (a) That seldom happens, because it means that *every* path in
298 -- the body of the let has to use that stronger demand
299 -- (b) It often happens temporarily in when fixpointing, because
300 -- the recursive function at first seems to place a massive demand.
301 -- But we don't want to go to extra work when the function will
302 -- probably iterate to something less demanding.
303 -- In practice, all the times the actual demand on id2 is more than
304 -- the vanilla call demand seem to be due to (b). So we don't
305 -- bother to re-analyse the RHS.
306 (body_ty2, Let (NonRec id2 rhs') body')
308 dmdAnal sigs dmd (Let (Rec pairs) body)
310 bndrs = map fst pairs
311 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
312 (body_ty, body') = dmdAnal sigs' dmd body
313 body_ty1 = addLazyFVs body_ty lazy_fv
315 sigs' `seq` body_ty `seq`
317 (body_ty2, _) = annotateBndrs body_ty1 bndrs
318 -- Don't bother to add demand info to recursive
319 -- binders as annotateBndr does;
320 -- being recursive, we can't treat them strictly.
321 -- But we do need to remove the binders from the result demand env
323 (body_ty2, Let (Rec pairs') body')
326 dmdAnalAlt sigs dmd (con,bndrs,rhs)
328 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
329 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
330 final_alt_ty | io_hack_reqd = alt_ty `lubType` topDmdType
333 -- There's a hack here for I/O operations. Consider
334 -- case foo x s of { (# s, r #) -> y }
335 -- Is this strict in 'y'. Normally yes, but what if 'foo' is an I/O
336 -- operation that simply terminates the program (not in an erroneous way)?
337 -- In that case we should not evaluate y before the call to 'foo'.
338 -- Hackish solution: spot the IO-like situation and add a virtual branch,
342 -- other -> return ()
343 -- So the 'y' isn't necessarily going to be evaluated
345 -- A more complete example where this shows up is:
346 -- do { let len = <expensive> ;
347 -- ; when (...) (exitWith ExitSuccess)
350 io_hack_reqd = con == DataAlt unboxedPairDataCon &&
351 idType (head bndrs) `coreEqType` realWorldStatePrimTy
353 (final_alt_ty, (con, bndrs', rhs'))
356 %************************************************************************
358 \subsection{Bindings}
360 %************************************************************************
363 dmdFix :: TopLevelFlag
364 -> SigEnv -- Does not include bindings for this binding
367 [(Id,CoreExpr)]) -- Binders annotated with stricness info
369 dmdFix top_lvl sigs orig_pairs
370 = loop 1 initial_sigs orig_pairs
372 bndrs = map fst orig_pairs
373 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
376 -> SigEnv -- Already contains the current sigs
378 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
381 = (sigs', lazy_fv, pairs')
382 -- Note: use pairs', not pairs. pairs' is the result of
383 -- processing the RHSs with sigs (= sigs'), whereas pairs
384 -- is the result of processing the RHSs with the *previous*
385 -- iteration of sigs.
387 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
388 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
389 text "env:" <+> ppr (ufmToList sigs),
390 text "binds:" <+> pprCoreBinding (Rec pairs)]))
391 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
392 -- The lazy_fv part is really important! orig_pairs has no strictness
393 -- info, including nothing about free vars. But if we have
394 -- letrec f = ....y..... in ...f...
395 -- where 'y' is free in f, we must record that y is mentioned,
396 -- otherwise y will get recorded as absent altogether
398 | otherwise = loop (n+1) sigs' pairs'
400 found_fixpoint = all (same_sig sigs sigs') bndrs
401 -- Use the new signature to do the next pair
402 -- The occurrence analyser has arranged them in a good order
403 -- so this can significantly reduce the number of iterations needed
404 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
406 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
407 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
409 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
410 ((sigs', lazy_fv'), pair')
413 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
414 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
415 -- old_sig = lookup sigs id
416 -- new_sig = lookup sigs' id
418 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
419 lookup sigs var = case lookupVarEnv sigs var of
422 -- Get an initial strictness signature from the Id
423 -- itself. That way we make use of earlier iterations
424 -- of the fixpoint algorithm. (Cunning plan.)
425 -- Note that the cunning plan extends to the DmdEnv too,
426 -- since it is part of the strictness signature
427 initialSig id = idNewStrictness_maybe id `orElse` botSig
429 dmdAnalRhs :: TopLevelFlag -> RecFlag
430 -> SigEnv -> (Id, CoreExpr)
431 -> (SigEnv, DmdEnv, (Id, CoreExpr))
432 -- Process the RHS of the binding, add the strictness signature
433 -- to the Id, and augment the environment with the signature as well.
435 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
436 = (sigs', lazy_fv, (id', rhs'))
438 arity = idArity id -- The idArity should be up to date
439 -- The simplifier was run just beforehand
440 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
441 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs), ppr id )
442 -- The RHS can be eta-reduced to just a variable,
443 -- in which case we should not complain.
444 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
445 id' = id `setIdNewStrictness` sig_ty
446 sigs' = extendSigEnv top_lvl sigs id sig_ty
449 %************************************************************************
451 \subsection{Strictness signatures and types}
453 %************************************************************************
456 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
457 -- Take a DmdType and turn it into a StrictSig
458 -- NB: not used for never-inline things; hence False
459 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
461 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
462 mkSigTy top_lvl rec_flag id rhs dmd_ty
463 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
465 never_inline = isNeverActive (idInlinePragma id)
466 maybe_id_dmd = idNewDemandInfo_maybe id
467 -- Is Nothing the first time round
470 | isTopLevel top_lvl = False -- Top level things don't get
471 -- their demandInfo set at all
472 | isRec rec_flag = False -- Ditto recursive things
473 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
474 | otherwise = True -- Optimistic, first time round
478 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
479 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
480 If the rhs is a thunk, we usually forget the CPR info, because
481 it is presumably shared (else it would have been inlined, and
482 so we'd lose sharing if w/w'd it into a function). E.g.
484 let r = case expensive of
488 If we marked r as having the CPR property, then we'd w/w into
490 let $wr = \() -> case expensive of
496 But now r is a thunk, which won't be inlined, so we are no further ahead.
499 f x = let r = case expensive of (a,b) -> (b,a)
500 in if foo r then r else (x,x)
502 Does f have the CPR property? Well, no.
504 However, if the strictness analyser has figured out (in a previous
505 iteration) that it's strict, then we DON'T need to forget the CPR info.
506 Instead we can retain the CPR info and do the thunk-splitting transform
507 (see WorkWrap.splitThunk).
509 This made a big difference to PrelBase.modInt, which had something like
510 modInt = \ x -> let r = ... -> I# v in
511 ...body strict in r...
512 r's RHS isn't a value yet; but modInt returns r in various branches, so
513 if r doesn't have the CPR property then neither does modInt
514 Another case I found in practice (in Complex.magnitude), looks like this:
515 let k = if ... then I# a else I# b
516 in ... body strict in k ....
517 (For this example, it doesn't matter whether k is returned as part of
518 the overall result; but it does matter that k's RHS has the CPR property.)
519 Left to itself, the simplifier will make a join point thus:
520 let $j k = ...body strict in k...
521 if ... then $j (I# a) else $j (I# b)
522 With thunk-splitting, we get instead
523 let $j x = let k = I#x in ...body strict in k...
524 in if ... then $j a else $j b
525 This is much better; there's a good chance the I# won't get allocated.
527 The difficulty with this is that we need the strictness type to
528 look at the body... but we now need the body to calculate the demand
529 on the variable, so we can decide whether its strictness type should
530 have a CPR in it or not. Simple solution:
531 a) use strictness info from the previous iteration
532 b) make sure we do at least 2 iterations, by doing a second
533 round for top-level non-recs. Top level recs will get at
534 least 2 iterations except for totally-bottom functions
535 which aren't very interesting anyway.
537 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
539 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
540 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
541 Demand info now has a 'Nothing' state, just like strictness info.
542 The analysis works from 'dangerous' towards a 'safe' state; so we
543 start with botSig for 'Nothing' strictness infos, and we start with
544 "yes, it's demanded" for 'Nothing' in the demand info. The
545 fixpoint iteration will sort it all out.
547 We can't start with 'not-demanded' because then consider
551 if ... then t else I# y else f x'
553 In the first iteration we'd have no demand info for x, so assume
554 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
555 we'd see that t was demanded, and so give it the CPR property, but by
556 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
557 setting the first time round, we say 'yes t is demanded' the first
560 However, this does mean that for non-recursive bindings we must
561 iterate twice to be sure of not getting over-optimistic CPR info,
562 in the case where t turns out to be not-demanded. This is handled
566 Note [NOINLINE and strictness]
567 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
568 The strictness analyser used to have a HACK which ensured that NOINLNE
569 things were not strictness-analysed. The reason was unsafePerformIO.
570 Left to itself, the strictness analyser would discover this strictness
572 unsafePerformIO: C(U(AV))
573 But then consider this sub-expression
574 unsafePerformIO (\s -> let r = f x in
575 case writeIORef v r s of (# s1, _ #) ->
577 The strictness analyser will now find that r is sure to be eval'd,
578 and may then hoist it out. This makes tests/lib/should_run/memo002
581 Solving this by making all NOINLINE things have no strictness info is overkill.
582 In particular, it's overkill for runST, which is perfectly respectable.
584 f x = runST (return x)
585 This should be strict in x.
587 So the new plan is to define unsafePerformIO using the 'lazy' combinator:
589 unsafePerformIO (IO m) = lazy (case m realWorld# of (# _, r #) -> r)
591 Remember, 'lazy' is a wired-in identity-function Id, of type a->a, which is
592 magically NON-STRICT, and is inlined after strictness analysis. So
593 unsafePerformIO will look non-strict, and that's what we want.
595 Now we don't need the hack in the strictness analyser. HOWEVER, this
596 decision does mean that even a NOINLINE function is not entirely
597 opaque: some aspect of its implementation leaks out, notably its
598 strictness. For example, if you have a function implemented by an
599 error stub, but which has RULES, you may want it not to be eliminated
604 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
605 = (lazy_fv, mkStrictSig dmd_ty)
606 -- Re unused never_inline, see Note [NOINLINE and strictness]
608 dmd_ty = DmdType strict_fv final_dmds res'
610 lazy_fv = filterUFM (not . isStrictDmd) fv
611 strict_fv = filterUFM isStrictDmd fv
612 -- We put the strict FVs in the DmdType of the Id, so
613 -- that at its call sites we unleash demands on its strict fvs.
614 -- An example is 'roll' in imaginary/wheel-sieve2
615 -- Something like this:
617 -- go y = if ... then roll (x-1) else x+1
620 -- We want to see that roll is strict in x, which is because
621 -- go is called. So we put the DmdEnv for x in go's DmdType.
624 -- f :: Int -> Int -> Int
625 -- f x y = let t = x+1
626 -- h z = if z==0 then t else
627 -- if z==1 then x+1 else
631 -- Calling h does indeed evaluate x, but we can only see
632 -- that if we unleash a demand on x at the call site for t.
634 -- Incidentally, here's a place where lambda-lifting h would
635 -- lose the cigar --- we couldn't see the joint strictness in t/x
638 -- We don't want to put *all* the fv's from the RHS into the
639 -- DmdType, because that makes fixpointing very slow --- the
640 -- DmdType gets full of lazy demands that are slow to converge.
642 final_dmds = setUnpackStrategy dmds
643 -- Set the unpacking strategy
646 RetCPR | ignore_cpr_info -> TopRes
648 ignore_cpr_info = not (exprIsHNF rhs || thunk_cpr_ok)
651 The unpack strategy determines whether we'll *really* unpack the argument,
652 or whether we'll just remember its strictness. If unpacking would give
653 rise to a *lot* of worker args, we may decide not to unpack after all.
656 setUnpackStrategy :: [Demand] -> [Demand]
658 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
660 go :: Int -- Max number of args available for sub-components of [Demand]
662 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
664 go n (Eval (Prod cs) : ds)
665 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
666 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
669 n' = n + 1 - non_abs_args
670 -- Add one to the budget 'cos we drop the top-level arg
671 non_abs_args = nonAbsentArgs cs
672 -- Delete # of non-absent args to which we'll now be committed
674 go n (d:ds) = d `cons` go n ds
677 cons d (n,ds) = (n, d:ds)
679 nonAbsentArgs :: [Demand] -> Int
681 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
682 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
686 %************************************************************************
688 \subsection{Strictness signatures and types}
690 %************************************************************************
693 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
695 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
696 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
697 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
699 addLazyFVs (DmdType fv ds res) lazy_fvs
700 = DmdType both_fv1 ds res
702 both_fv = (plusUFM_C both fv lazy_fvs)
703 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
704 -- This modifyEnv is vital. Consider
705 -- let f = \x -> (x,y)
707 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
708 -- demand with the bottom coming up from 'error'
710 -- I got a loop in the fixpointer without this, due to an interaction
711 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
713 -- = letrec g y = x `fatbar`
714 -- letrec h z = z + ...g...
717 -- In the initial iteration for f, f=Bot
718 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
719 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
720 -- places on its free variables. Suppose it places none. Then the
721 -- x `fatbar` ...call to h...
722 -- will give a x->V demand for x. That turns into a L demand for x,
723 -- which floats out of the defn for h. Without the modifyEnv, that
724 -- L demand doesn't get both'd with the Bot coming up from the inner
725 -- call to f. So we just get an L demand for x for g.
727 -- A better way to say this is that the lazy-fv filtering should give the
728 -- same answer as putting the lazy fv demands in the function's type.
730 annotateBndr :: DmdType -> Var -> (DmdType, Var)
731 -- The returned env has the var deleted
732 -- The returned var is annotated with demand info
733 -- No effect on the argument demands
734 annotateBndr dmd_ty@(DmdType fv ds res) var
735 | isTyVar var = (dmd_ty, var)
736 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
738 (fv', dmd) = removeFV fv var res
740 annotateBndrs = mapAccumR annotateBndr
742 annotateLamIdBndr :: DmdType -- Demand type of body
743 -> Id -- Lambda binder
744 -> (DmdType, -- Demand type of lambda
745 Id) -- and binder annotated with demand
747 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
748 -- For lambdas we add the demand to the argument demands
749 -- Only called for Ids
751 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
753 (fv', dmd) = removeFV fv id res
754 hacked_dmd = argDemand dmd
755 -- This call to argDemand is vital, because otherwise we label
756 -- a lambda binder with demand 'B'. But in terms of calling
757 -- conventions that's Abs, because we don't pass it. But
758 -- when we do a w/w split we get
759 -- fw x = (\x y:B -> ...) x (error "oops")
760 -- And then the simplifier things the 'B' is a strict demand
761 -- and evaluates the (error "oops"). Sigh
763 removeFV fv id res = (fv', zapUnlifted id dmd)
765 fv' = fv `delVarEnv` id
766 dmd = lookupVarEnv fv id `orElse` deflt
767 deflt | isBotRes res = Bot
770 -- For unlifted-type variables, we are only
771 -- interested in Bot/Abs/Box Abs
772 zapUnlifted is Bot = Bot
773 zapUnlifted id Abs = Abs
774 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
778 %************************************************************************
780 \subsection{Strictness signatures}
782 %************************************************************************
785 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
786 -- We use the SigEnv to tell us whether to
787 -- record info about a variable in the DmdEnv
788 -- We do so if it's a LocalId, but not top-level
790 -- The DmdEnv gives the demand on the free vars of the function
791 -- when it is given enough args to satisfy the strictness signature
793 emptySigEnv = emptyVarEnv
795 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
796 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
798 extendSigEnvList = extendVarEnvList
800 extendSigsWithLam :: SigEnv -> Id -> SigEnv
801 -- Extend the SigEnv when we meet a lambda binder
802 -- If the binder is marked demanded with a product demand, then give it a CPR
803 -- signature, because in the likely event that this is a lambda on a fn defn
804 -- [we only use this when the lambda is being consumed with a call demand],
805 -- it'll be w/w'd and so it will be CPR-ish. E.g.
806 -- f = \x::(Int,Int). if ...strict in x... then
810 -- We want f to have the CPR property because x does, by the time f has been w/w'd
812 -- Also note that we only want to do this for something that
813 -- definitely has product type, else we may get over-optimistic
814 -- CPR results (e.g. from \x -> x!).
816 extendSigsWithLam sigs id
817 = case idNewDemandInfo_maybe id of
818 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
819 -- Optimistic in the Nothing case;
820 -- See notes [CPR-AND-STRICTNESS]
821 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
825 dmdTransform :: SigEnv -- The strictness environment
826 -> Id -- The function
827 -> Demand -- The demand on the function
828 -> DmdType -- The demand type of the function in this context
829 -- Returned DmdEnv includes the demand on
830 -- this function plus demand on its free variables
832 dmdTransform sigs var dmd
834 ------ DATA CONSTRUCTOR
835 | isDataConWorkId var -- Data constructor
837 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
838 DmdType _ _ con_res = dmd_ty
841 if arity == call_depth then -- Saturated, so unleash the demand
843 -- Important! If we Keep the constructor application, then
844 -- we need the demands the constructor places (always lazy)
845 -- If not, we don't need to. For example:
846 -- f p@(x,y) = (p,y) -- S(AL)
848 -- It's vital that we don't calculate Absent for a!
849 dmd_ds = case res_dmd of
850 Box (Eval ds) -> mapDmds box ds
854 -- ds can be empty, when we are just seq'ing the thing
855 -- If so we must make up a suitable bunch of demands
856 arg_ds = case dmd_ds of
857 Poly d -> replicate arity d
858 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
861 mkDmdType emptyDmdEnv arg_ds con_res
862 -- Must remember whether it's a product, hence con_res, not TopRes
866 ------ IMPORTED FUNCTION
867 | isGlobalId var, -- Imported function
868 let StrictSig dmd_ty = idNewStrictness var
869 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
874 ------ LOCAL LET/REC BOUND THING
875 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
877 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
878 | otherwise = deferType dmd_ty
879 -- NB: it's important to use deferType, and not just return topDmdType
880 -- Consider let { f x y = p + x } in f 1
881 -- The application isn't saturated, but we must nevertheless propagate
882 -- a lazy demand for p!
884 addVarDmd top_lvl fn_ty var dmd
886 ------ LOCAL NON-LET/REC BOUND THING
887 | otherwise -- Default case
891 (call_depth, res_dmd) = splitCallDmd dmd
895 %************************************************************************
899 %************************************************************************
902 splitDmdTy :: DmdType -> (Demand, DmdType)
903 -- Split off one function argument
904 -- We already have a suitable demand on all
905 -- free vars, so no need to add more!
906 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
907 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
909 splitCallDmd :: Demand -> (Int, Demand)
910 splitCallDmd (Call d) = case splitCallDmd d of
912 splitCallDmd d = (0, d)
914 vanillaCall :: Arity -> Demand
915 vanillaCall 0 = evalDmd
916 vanillaCall n = Call (vanillaCall (n-1))
918 deferType :: DmdType -> DmdType
919 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
920 -- Notice that we throw away info about both arguments and results
921 -- For example, f = let ... in \x -> x
922 -- We don't want to get a stricness type V->T for f.
924 deferEnv :: DmdEnv -> DmdEnv
925 deferEnv fv = mapVarEnv defer fv
929 argDemand :: Demand -> Demand
930 -- The 'Defer' demands are just Lazy at function boundaries
931 -- Ugly! Ask John how to improve it.
932 argDemand Top = lazyDmd
933 argDemand (Defer d) = lazyDmd
934 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
935 argDemand (Box Bot) = evalDmd
936 argDemand (Box d) = box (argDemand d)
937 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
942 -------------------------
943 -- Consider (if x then y else []) with demand V
944 -- Then the first branch gives {y->V} and the second
945 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
946 -- in the result env.
947 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
948 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
950 lub_fv = plusUFM_C lub fv1 fv2
951 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
952 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
953 -- lub is the identity for Bot
955 -- Extend the shorter argument list to match the longer
956 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
958 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
959 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
961 -----------------------------------
962 -- (t1 `bothType` t2) takes the argument/result info from t1,
963 -- using t2 just for its free-var info
964 -- NB: Don't forget about r2! It might be BotRes, which is
965 -- a bottom demand on all the in-scope variables.
966 -- Peter: can this be done more neatly?
967 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
968 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
970 both_fv = plusUFM_C both fv1 fv2
971 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
972 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
973 -- both is the identity for Abs
980 lubRes RetCPR RetCPR = RetCPR
981 lubRes r1 r2 = TopRes
983 -- If either diverges, the whole thing does
984 -- Otherwise take CPR info from the first
985 bothRes r1 BotRes = BotRes
990 modifyEnv :: Bool -- No-op if False
991 -> (Demand -> Demand) -- The zapper
992 -> DmdEnv -> DmdEnv -- Env1 and Env2
993 -> DmdEnv -> DmdEnv -- Transform this env
994 -- Zap anything in Env1 but not in Env2
995 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
997 modifyEnv need_to_modify zapper env1 env2 env
998 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
1001 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
1003 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
1007 %************************************************************************
1009 \subsection{LUB and BOTH}
1011 %************************************************************************
1014 lub :: Demand -> Demand -> Demand
1017 lub Abs d2 = absLub d2
1019 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
1021 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
1022 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
1023 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
1024 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
1026 -- For the Eval case, we use these approximation rules
1027 -- Box Bot <= Eval (Box Bot ...)
1028 -- Box Top <= Defer (Box Bot ...)
1029 -- Box (Eval ds) <= Eval (map Box ds)
1030 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
1031 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
1032 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
1033 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
1034 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
1036 lub (Box d1) (Box d2) = box (d1 `lub` d2)
1037 lub d1@(Box _) d2 = d2 `lub` d1
1039 lubs ds1 ds2 = zipWithDmds lub ds1 ds2
1041 ---------------------
1042 -- box is the smart constructor for Box
1043 -- It computes <B,bot> & d
1044 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
1045 -- Seems to be no point in allowing (Box (Call d))
1046 box (Call d) = Call d -- The odd man out. Why?
1048 box (Defer _) = lazyDmd
1049 box Top = lazyDmd -- Box Abs and Box Top
1050 box Abs = lazyDmd -- are the same <B,L>
1051 box d = Box d -- Bot, Eval
1054 defer :: Demand -> Demand
1056 -- defer is the smart constructor for Defer
1057 -- The idea is that (Defer ds) = <U(ds), L>
1059 -- It specifies what happens at a lazy function argument
1060 -- or a lambda; the L* operator
1061 -- Set the strictness part to L, but leave
1062 -- the boxity side unaffected
1063 -- It also ensures that Defer (Eval [LLLL]) = L
1068 defer (Call _) = lazyDmd -- Approximation here?
1069 defer (Box _) = lazyDmd
1070 defer (Defer ds) = Defer ds
1071 defer (Eval ds) = deferEval ds
1073 -- deferEval ds = defer (Eval ds)
1074 deferEval ds | allTop ds = Top
1075 | otherwise = Defer ds
1077 ---------------------
1078 absLub :: Demand -> Demand
1079 -- Computes (Abs `lub` d)
1080 -- For the Bot case consider
1081 -- f x y = if ... then x else error x
1082 -- Then for y we get Abs `lub` Bot, and we really
1087 absLub (Call _) = Top
1088 absLub (Box _) = Top
1089 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1090 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1092 absLubs = mapDmds absLub
1095 both :: Demand -> Demand -> Demand
1101 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1104 -- From 'error' itself we get demand Bot on x
1105 -- From the arg demand on x we get
1106 -- x :-> evalDmd = Box (Eval (Poly Abs))
1107 -- So we get Bot `both` Box (Eval (Poly Abs))
1108 -- = Seq Keep (Poly Bot)
1111 -- f x = if ... then error (fst x) else fst x
1112 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1114 -- which is what we want.
1117 both Top Bot = errDmd
1120 both Top (Box d) = Box d
1121 both Top (Call d) = Call d
1122 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1123 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1124 -- = defer (Eval (mapDmds (`both` Top) ds))
1125 = deferEval (mapDmds (`both` Top) ds)
1128 both (Box d1) (Box d2) = box (d1 `both` d2)
1129 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1130 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1131 both (Box d1) (Defer d2) = Box d1
1132 both d1@(Box _) d2 = d2 `both` d1
1134 both (Call d1) (Call d2) = Call (d1 `both` d2)
1135 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1136 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1137 both d1@(Call _) d2 = d1 `both` d1
1139 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1140 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1141 both d1@(Eval ds1) d2 = d2 `both` d1
1143 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1144 both d1@(Defer ds1) d2 = d2 `both` d1
1146 boths ds1 ds2 = zipWithDmds both ds1 ds2
1151 %************************************************************************
1153 \subsection{Miscellaneous
1155 %************************************************************************
1159 #ifdef OLD_STRICTNESS
1160 get_changes binds = vcat (map get_changes_bind binds)
1162 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1163 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1165 get_changes_pr (id,rhs)
1166 = get_changes_var id $$ get_changes_expr rhs
1169 | isId var = get_changes_str var $$ get_changes_dmd var
1172 get_changes_expr (Type t) = empty
1173 get_changes_expr (Var v) = empty
1174 get_changes_expr (Lit l) = empty
1175 get_changes_expr (Note n e) = get_changes_expr e
1176 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1177 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1178 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1179 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1181 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1184 | new_better && old_better = empty
1185 | new_better = message "BETTER"
1186 | old_better = message "WORSE"
1187 | otherwise = message "INCOMPARABLE"
1189 message word = text word <+> text "strictness for" <+> ppr id <+> info
1190 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1191 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1192 -- strictness analyser can't track
1193 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1194 old_better = old `betterStrictness` new
1195 new_better = new `betterStrictness` old
1198 | isUnLiftedType (idType id) = empty -- Not useful
1199 | new_better && old_better = empty
1200 | new_better = message "BETTER"
1201 | old_better = message "WORSE"
1202 | otherwise = message "INCOMPARABLE"
1204 message word = text word <+> text "demand for" <+> ppr id <+> info
1205 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1206 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1208 old = newDemand (idDemandInfo id)
1209 new_better = new `betterDemand` old
1210 old_better = old `betterDemand` new
1212 betterStrictness :: StrictSig -> StrictSig -> Bool
1213 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
1215 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
1217 betterDemand :: Demand -> Demand -> Bool
1218 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
1219 betterDemand d1 d2 = (d1 `lub` d2) == d2
1221 squashSig (StrictSig (DmdType fv ds res))
1222 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1224 -- squash just gets rid of call demands
1225 -- which the old analyser doesn't track
1226 squashDmd (Call d) = evalDmd
1227 squashDmd (Box d) = Box (squashDmd d)
1228 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1229 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)