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 CmdLineOpts ( DynFlags, DynFlag(..), opt_MaxWorkerArgs )
17 import NewDemand -- All of it
20 import CoreUtils ( exprIsValue, exprArity )
21 import DataCon ( dataConTyCon )
22 import TyCon ( isProductTyCon, isRecursiveTyCon )
23 import Id ( Id, idType, idInlinePragma,
24 isDataConWorkId, isGlobalId, idArity,
26 idDemandInfo, idStrictness, idCprInfo, idName,
28 idNewStrictness, idNewStrictness_maybe,
29 setIdNewStrictness, idNewDemandInfo,
30 idNewDemandInfo_maybe,
34 import IdInfo ( newStrictnessFromOld, newDemand )
38 import TysWiredIn ( unboxedPairDataCon )
39 import TysPrim ( realWorldStatePrimTy )
40 import UniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly,
41 keysUFM, minusUFM, ufmToList, filterUFM )
42 import Type ( isUnLiftedType, eqType )
43 import CoreLint ( showPass, endPass )
44 import Util ( mapAndUnzip, mapAccumL, mapAccumR, lengthIs )
45 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive,
47 import Maybes ( orElse, expectJust )
53 * set a noinline pragma on bottoming Ids
55 * Consider f x = x+1 `fatbar` error (show x)
56 We'd like to unbox x, even if that means reboxing it in the error case.
59 %************************************************************************
61 \subsection{Top level stuff}
63 %************************************************************************
66 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
67 dmdAnalPgm dflags binds
69 showPass dflags "Demand analysis" ;
70 let { binds_plus_dmds = do_prog binds } ;
72 endPass dflags "Demand analysis"
73 Opt_D_dump_stranal binds_plus_dmds ;
75 -- Only if OLD_STRICTNESS is on, because only then is the old
76 -- strictness analyser run
77 let { dmd_changes = get_changes binds_plus_dmds } ;
78 printDump (text "Changes in demands" $$ dmd_changes) ;
80 return binds_plus_dmds
83 do_prog :: [CoreBind] -> [CoreBind]
84 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
86 dmdAnalTopBind :: SigEnv
89 dmdAnalTopBind sigs (NonRec id rhs)
91 ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs)
92 (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel NonRecursive sigs (id, rhs1)
93 -- Do two passes to improve CPR information
94 -- See comments with ignore_cpr_info in mk_sig_ty
95 -- and with extendSigsWithLam
97 (sigs2, NonRec id2 rhs2)
99 dmdAnalTopBind sigs (Rec pairs)
101 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
102 -- We get two iterations automatically
103 -- c.f. the NonRec case above
109 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
110 -- Analyse the RHS and return
111 -- a) appropriate strictness info
112 -- b) the unfolding (decorated with stricntess info)
116 arity = exprArity rhs
117 (rhs_ty, rhs') = dmdAnal emptySigEnv (vanillaCall arity) rhs
118 sig = mkTopSigTy rhs rhs_ty
121 %************************************************************************
123 \subsection{The analyser itself}
125 %************************************************************************
128 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
130 dmdAnal sigs Abs e = (topDmdType, e)
133 | not (isStrictDmd dmd)
135 (res_ty, e') = dmdAnal sigs evalDmd e
137 (deferType res_ty, e')
138 -- It's important not to analyse e with a lazy demand because
139 -- a) When we encounter case s of (a,b) ->
140 -- we demand s with U(d1d2)... but if the overall demand is lazy
141 -- that is wrong, and we'd need to reduce the demand on s,
142 -- which is inconvenient
143 -- b) More important, consider
144 -- f (let x = R in x+x), where f is lazy
145 -- We still want to mark x as demanded, because it will be when we
146 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
147 -- just mark x as Lazy
148 -- c) The application rule wouldn't be right either
149 -- Evaluating (f x) in a L demand does *not* cause
150 -- evaluation of f in a C(L) demand!
153 dmdAnal sigs dmd (Lit lit)
154 = (topDmdType, Lit lit)
156 dmdAnal sigs dmd (Var var)
157 = (dmdTransform sigs var dmd, Var var)
159 dmdAnal sigs dmd (Note n e)
160 = (dmd_ty, Note n e')
162 (dmd_ty, e') = dmdAnal sigs dmd' e
164 Coerce _ _ -> evalDmd -- This coerce usually arises from a recursive
165 other -> dmd -- newtype, and we don't want to look inside them
166 -- for exactly the same reason that we don't look
167 -- inside recursive products -- we might not reach
168 -- a fixpoint. So revert to a vanilla Eval demand
170 dmdAnal sigs dmd (App fun (Type ty))
171 = (fun_ty, App fun' (Type ty))
173 (fun_ty, fun') = dmdAnal sigs dmd fun
175 -- Lots of the other code is there to make this
176 -- beautiful, compositional, application rule :-)
177 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
178 = let -- [Type arg handled above]
179 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
180 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
181 (arg_dmd, res_ty) = splitDmdTy fun_ty
183 (res_ty `bothType` arg_ty, App fun' arg')
185 dmdAnal sigs dmd (Lam var body)
188 (body_ty, body') = dmdAnal sigs dmd body
190 (body_ty, Lam var body')
192 | Call body_dmd <- dmd -- A call demand: good!
194 sigs' = extendSigsWithLam sigs var
195 (body_ty, body') = dmdAnal sigs' body_dmd body
196 (lam_ty, var') = annotateLamIdBndr body_ty var
198 (lam_ty, Lam var' body')
200 | otherwise -- Not enough demand on the lambda; but do the body
201 = let -- anyway to annotate it and gather free var info
202 (body_ty, body') = dmdAnal sigs evalDmd body
203 (lam_ty, var') = annotateLamIdBndr body_ty var
205 (deferType lam_ty, Lam var' body')
207 dmdAnal sigs dmd (Case scrut case_bndr [alt@(DataAlt dc,bndrs,rhs)])
208 | let tycon = dataConTyCon dc,
209 isProductTyCon tycon,
210 not (isRecursiveTyCon tycon)
212 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
213 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
214 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
215 (_, bndrs', _) = alt'
216 case_bndr_sig = cprSig
217 -- Inside the alternative, the case binder has the CPR property.
218 -- Meaning that a case on it will successfully cancel.
220 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
223 -- We want f to have the CPR property:
224 -- f b x = case fw b x of { r -> I# r }
225 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
228 -- Figure out whether the demand on the case binder is used, and use
229 -- that to set the scrut_dmd. This is utterly essential.
230 -- Consider f x = case x of y { (a,b) -> k y a }
231 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
232 -- worker, so the worker will rebuild
233 -- x = (a, absent-error)
234 -- and that'll crash.
235 -- So at one stage I had:
236 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
237 -- keepity | dead_case_bndr = Drop
238 -- | otherwise = Keep
241 -- case x of y { (a,b) -> h y + a }
242 -- where h : U(LL) -> T
243 -- The above code would compute a Keep for x, since y is not Abs, which is silly
244 -- The insight is, of course, that a demand on y is a demand on the
245 -- scrutinee, so we need to `both` it with the scrut demand
247 scrut_dmd = Eval (Prod [idNewDemandInfo b | b <- bndrs', isId b])
249 idNewDemandInfo case_bndr'
251 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
253 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' [alt'])
255 dmdAnal sigs dmd (Case scrut case_bndr alts)
257 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
258 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
259 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
261 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
262 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' alts')
264 dmdAnal sigs dmd (Let (NonRec id rhs) body)
266 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
267 (body_ty, body') = dmdAnal sigs' dmd body
268 (body_ty1, id2) = annotateBndr body_ty id1
269 body_ty2 = addLazyFVs body_ty1 lazy_fv
271 -- If the actual demand is better than the vanilla call
272 -- demand, you might think that we might do better to re-analyse
273 -- the RHS with the stronger demand.
274 -- But (a) That seldom happens, because it means that *every* path in
275 -- the body of the let has to use that stronger demand
276 -- (b) It often happens temporarily in when fixpointing, because
277 -- the recursive function at first seems to place a massive demand.
278 -- But we don't want to go to extra work when the function will
279 -- probably iterate to something less demanding.
280 -- In practice, all the times the actual demand on id2 is more than
281 -- the vanilla call demand seem to be due to (b). So we don't
282 -- bother to re-analyse the RHS.
283 (body_ty2, Let (NonRec id2 rhs') body')
285 dmdAnal sigs dmd (Let (Rec pairs) body)
287 bndrs = map fst pairs
288 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
289 (body_ty, body') = dmdAnal sigs' dmd body
290 body_ty1 = addLazyFVs body_ty lazy_fv
292 sigs' `seq` body_ty `seq`
294 (body_ty2, _) = annotateBndrs body_ty1 bndrs
295 -- Don't bother to add demand info to recursive
296 -- binders as annotateBndr does;
297 -- being recursive, we can't treat them strictly.
298 -- But we do need to remove the binders from the result demand env
300 (body_ty2, Let (Rec pairs') body')
303 dmdAnalAlt sigs dmd (con,bndrs,rhs)
305 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
306 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
307 final_alt_ty | io_hack_reqd = alt_ty `lubType` topDmdType
310 -- There's a hack here for I/O operations. Consider
311 -- case foo x s of { (# s, r #) -> y }
312 -- Is this strict in 'y'. Normally yes, but what if 'foo' is an I/O
313 -- operation that simply terminates the program (not in an erroneous way)?
314 -- In that case we should not evaluate y before the call to 'foo'.
315 -- Hackish solution: spot the IO-like situation and add a virtual branch,
319 -- other -> return ()
320 -- So the 'y' isn't necessarily going to be evaluated
322 -- A more complete example where this shows up is:
323 -- do { let len = <expensive> ;
324 -- ; when (...) (exitWith ExitSuccess)
327 io_hack_reqd = con == DataAlt unboxedPairDataCon &&
328 idType (head bndrs) `eqType` realWorldStatePrimTy
330 (final_alt_ty, (con, bndrs', rhs'))
333 %************************************************************************
335 \subsection{Bindings}
337 %************************************************************************
340 dmdFix :: TopLevelFlag
341 -> SigEnv -- Does not include bindings for this binding
344 [(Id,CoreExpr)]) -- Binders annotated with stricness info
346 dmdFix top_lvl sigs orig_pairs
347 = loop 1 initial_sigs orig_pairs
349 bndrs = map fst orig_pairs
350 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
353 -> SigEnv -- Already contains the current sigs
355 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
358 = (sigs', lazy_fv, pairs')
359 -- Note: use pairs', not pairs. pairs' is the result of
360 -- processing the RHSs with sigs (= sigs'), whereas pairs
361 -- is the result of processing the RHSs with the *previous*
362 -- iteration of sigs.
364 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
365 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
366 text "env:" <+> ppr (ufmToList sigs),
367 text "binds:" <+> pprCoreBinding (Rec pairs)]))
368 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
369 -- The lazy_fv part is really important! orig_pairs has no strictness
370 -- info, including nothing about free vars. But if we have
371 -- letrec f = ....y..... in ...f...
372 -- where 'y' is free in f, we must record that y is mentioned,
373 -- otherwise y will get recorded as absent altogether
375 | otherwise = loop (n+1) sigs' pairs'
377 found_fixpoint = all (same_sig sigs sigs') bndrs
378 -- Use the new signature to do the next pair
379 -- The occurrence analyser has arranged them in a good order
380 -- so this can significantly reduce the number of iterations needed
381 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
383 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
384 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
386 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
387 ((sigs', lazy_fv'), pair')
390 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
391 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
392 -- old_sig = lookup sigs id
393 -- new_sig = lookup sigs' id
395 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
396 lookup sigs var = case lookupVarEnv sigs var of
399 -- Get an initial strictness signature from the Id
400 -- itself. That way we make use of earlier iterations
401 -- of the fixpoint algorithm. (Cunning plan.)
402 -- Note that the cunning plan extends to the DmdEnv too,
403 -- since it is part of the strictness signature
404 initialSig id = idNewStrictness_maybe id `orElse` botSig
406 dmdAnalRhs :: TopLevelFlag -> RecFlag
407 -> SigEnv -> (Id, CoreExpr)
408 -> (SigEnv, DmdEnv, (Id, CoreExpr))
409 -- Process the RHS of the binding, add the strictness signature
410 -- to the Id, and augment the environment with the signature as well.
412 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
413 = (sigs', lazy_fv, (id', rhs'))
415 arity = idArity id -- The idArity should be up to date
416 -- The simplifier was run just beforehand
417 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
418 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty, ppr id )
419 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
420 id' = id `setIdNewStrictness` sig_ty
421 sigs' = extendSigEnv top_lvl sigs id sig_ty
424 %************************************************************************
426 \subsection{Strictness signatures and types}
428 %************************************************************************
431 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
432 -- Take a DmdType and turn it into a StrictSig
433 -- NB: not used for never-inline things; hence False
434 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
436 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
437 mkSigTy top_lvl rec_flag id rhs dmd_ty
438 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
440 never_inline = isNeverActive (idInlinePragma id)
441 maybe_id_dmd = idNewDemandInfo_maybe id
442 -- Is Nothing the first time round
445 | isTopLevel top_lvl = False -- Top level things don't get
446 -- their demandInfo set at all
447 | isRec rec_flag = False -- Ditto recursive things
448 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
449 | otherwise = True -- Optimistic, first time round
453 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
454 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
455 If the rhs is a thunk, we usually forget the CPR info, because
456 it is presumably shared (else it would have been inlined, and
457 so we'd lose sharing if w/w'd it into a function.
459 However, if the strictness analyser has figured out (in a previous
460 iteration) that it's strict, then we DON'T need to forget the CPR info.
461 Instead we can retain the CPR info and do the thunk-splitting transform
462 (see WorkWrap.splitThunk).
464 This made a big difference to PrelBase.modInt, which had something like
465 modInt = \ x -> let r = ... -> I# v in
466 ...body strict in r...
467 r's RHS isn't a value yet; but modInt returns r in various branches, so
468 if r doesn't have the CPR property then neither does modInt
469 Another case I found in practice (in Complex.magnitude), looks like this:
470 let k = if ... then I# a else I# b
471 in ... body strict in k ....
472 (For this example, it doesn't matter whether k is returned as part of
473 the overall result; but it does matter that k's RHS has the CPR property.)
474 Left to itself, the simplifier will make a join point thus:
475 let $j k = ...body strict in k...
476 if ... then $j (I# a) else $j (I# b)
477 With thunk-splitting, we get instead
478 let $j x = let k = I#x in ...body strict in k...
479 in if ... then $j a else $j b
480 This is much better; there's a good chance the I# won't get allocated.
482 The difficulty with this is that we need the strictness type to
483 look at the body... but we now need the body to calculate the demand
484 on the variable, so we can decide whether its strictness type should
485 have a CPR in it or not. Simple solution:
486 a) use strictness info from the previous iteration
487 b) make sure we do at least 2 iterations, by doing a second
488 round for top-level non-recs. Top level recs will get at
489 least 2 iterations except for totally-bottom functions
490 which aren't very interesting anyway.
492 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
494 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
495 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
496 Demand info now has a 'Nothing' state, just like strictness info.
497 The analysis works from 'dangerous' towards a 'safe' state; so we
498 start with botSig for 'Nothing' strictness infos, and we start with
499 "yes, it's demanded" for 'Nothing' in the demand info. The
500 fixpoint iteration will sort it all out.
502 We can't start with 'not-demanded' because then consider
506 if ... then t else I# y else f x'
508 In the first iteration we'd have no demand info for x, so assume
509 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
510 we'd see that t was demanded, and so give it the CPR property, but by
511 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
512 setting the first time round, we say 'yes t is demanded' the first
515 However, this does mean that for non-recursive bindings we must
516 iterate twice to be sure of not getting over-optimistic CPR info,
517 in the case where t turns out to be not-demanded. This is handled
522 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
523 | never_inline && not (isBotRes res)
525 -- Don't strictness-analyse NOINLINE things. Why not? Because
526 -- the NOINLINE says "don't expose any of the inner workings at the call
527 -- site" and the strictness is certainly an inner working.
529 -- More concretely, the demand analyser discovers the following strictness
530 -- for unsafePerformIO: C(U(AV))
532 -- unsafePerformIO (\s -> let r = f x in
533 -- case writeIORef v r s of (# s1, _ #) ->
535 -- The strictness analyser will find that the binding for r is strict,
536 -- (becuase of uPIO's strictness sig), and so it'll evaluate it before
537 -- doing the writeIORef. This actually makes tests/lib/should_run/memo002
540 -- Solution: don't expose the strictness of unsafePerformIO.
542 -- But we do want to expose the strictness of error functions,
543 -- which are also often marked NOINLINE
544 -- {-# NOINLINE foo #-}
545 -- foo x = error ("wubble buggle" ++ x)
546 -- So (hack, hack) we only drop the strictness for non-bottom things
547 -- This is all very unsatisfactory.
548 = (deferEnv fv, topSig)
551 = (lazy_fv, mkStrictSig dmd_ty)
553 dmd_ty = DmdType strict_fv final_dmds res'
555 lazy_fv = filterUFM (not . isStrictDmd) fv
556 strict_fv = filterUFM isStrictDmd fv
557 -- We put the strict FVs in the DmdType of the Id, so
558 -- that at its call sites we unleash demands on its strict fvs.
559 -- An example is 'roll' in imaginary/wheel-sieve2
560 -- Something like this:
562 -- go y = if ... then roll (x-1) else x+1
565 -- We want to see that roll is strict in x, which is because
566 -- go is called. So we put the DmdEnv for x in go's DmdType.
569 -- f :: Int -> Int -> Int
570 -- f x y = let t = x+1
571 -- h z = if z==0 then t else
572 -- if z==1 then x+1 else
576 -- Calling h does indeed evaluate x, but we can only see
577 -- that if we unleash a demand on x at the call site for t.
579 -- Incidentally, here's a place where lambda-lifting h would
580 -- lose the cigar --- we couldn't see the joint strictness in t/x
583 -- We don't want to put *all* the fv's from the RHS into the
584 -- DmdType, because that makes fixpointing very slow --- the
585 -- DmdType gets full of lazy demands that are slow to converge.
587 final_dmds = setUnpackStrategy dmds
588 -- Set the unpacking strategy
591 RetCPR | ignore_cpr_info -> TopRes
593 ignore_cpr_info = not (exprIsValue rhs || thunk_cpr_ok)
596 The unpack strategy determines whether we'll *really* unpack the argument,
597 or whether we'll just remember its strictness. If unpacking would give
598 rise to a *lot* of worker args, we may decide not to unpack after all.
601 setUnpackStrategy :: [Demand] -> [Demand]
603 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
605 go :: Int -- Max number of args available for sub-components of [Demand]
607 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
609 go n (Eval (Prod cs) : ds)
610 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
611 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
614 n' = n + 1 - non_abs_args
615 -- Add one to the budget 'cos we drop the top-level arg
616 non_abs_args = nonAbsentArgs cs
617 -- Delete # of non-absent args to which we'll now be committed
619 go n (d:ds) = d `cons` go n ds
622 cons d (n,ds) = (n, d:ds)
624 nonAbsentArgs :: [Demand] -> Int
626 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
627 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
631 %************************************************************************
633 \subsection{Strictness signatures and types}
635 %************************************************************************
638 splitDmdTy :: DmdType -> (Demand, DmdType)
639 -- Split off one function argument
640 -- We already have a suitable demand on all
641 -- free vars, so no need to add more!
642 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
643 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
647 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
649 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
650 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
651 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
653 addLazyFVs (DmdType fv ds res) lazy_fvs
654 = DmdType both_fv1 ds res
656 both_fv = (plusUFM_C both fv lazy_fvs)
657 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
658 -- This modifyEnv is vital. Consider
659 -- let f = \x -> (x,y)
661 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
662 -- demand with the bottom coming up from 'error'
664 -- I got a loop in the fixpointer without this, due to an interaction
665 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
667 -- = letrec g y = x `fatbar`
668 -- letrec h z = z + ...g...
671 -- In the initial iteration for f, f=Bot
672 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
673 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
674 -- places on its free variables. Suppose it places none. Then the
675 -- x `fatbar` ...call to h...
676 -- will give a x->V demand for x. That turns into a L demand for x,
677 -- which floats out of the defn for h. Without the modifyEnv, that
678 -- L demand doesn't get both'd with the Bot coming up from the inner
679 -- call to f. So we just get an L demand for x for g.
681 -- A better way to say this is that the lazy-fv filtering should give the
682 -- same answer as putting the lazy fv demands in the function's type.
684 annotateBndr :: DmdType -> Var -> (DmdType, Var)
685 -- The returned env has the var deleted
686 -- The returned var is annotated with demand info
687 -- No effect on the argument demands
688 annotateBndr dmd_ty@(DmdType fv ds res) var
689 | isTyVar var = (dmd_ty, var)
690 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
692 (fv', dmd) = removeFV fv var res
694 annotateBndrs = mapAccumR annotateBndr
696 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
697 -- For lambdas we add the demand to the argument demands
698 -- Only called for Ids
700 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
702 (fv', dmd) = removeFV fv id res
703 hacked_dmd = argDemand dmd
704 -- This call to argDemand is vital, because otherwise we label
705 -- a lambda binder with demand 'B'. But in terms of calling
706 -- conventions that's Abs, because we don't pass it. But
707 -- when we do a w/w split we get
708 -- fw x = (\x y:B -> ...) x (error "oops")
709 -- And then the simplifier things the 'B' is a strict demand
710 -- and evaluates the (error "oops"). Sigh
712 removeFV fv id res = (fv', zapUnlifted id dmd)
714 fv' = fv `delVarEnv` id
715 dmd = lookupVarEnv fv id `orElse` deflt
716 deflt | isBotRes res = Bot
719 -- For unlifted-type variables, we are only
720 -- interested in Bot/Abs/Box Abs
721 zapUnlifted is Bot = Bot
722 zapUnlifted id Abs = Abs
723 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
727 %************************************************************************
729 \subsection{Strictness signatures}
731 %************************************************************************
734 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
735 -- We use the SigEnv to tell us whether to
736 -- record info about a variable in the DmdEnv
737 -- We do so if it's a LocalId, but not top-level
739 -- The DmdEnv gives the demand on the free vars of the function
740 -- when it is given enough args to satisfy the strictness signature
742 emptySigEnv = emptyVarEnv
744 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
745 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
747 extendSigEnvList = extendVarEnvList
749 extendSigsWithLam :: SigEnv -> Id -> SigEnv
750 -- Extend the SigEnv when we meet a lambda binder
751 -- If the binder is marked demanded with a product demand, then give it a CPR
752 -- signature, because in the likely event that this is a lambda on a fn defn
753 -- [we only use this when the lambda is being consumed with a call demand],
754 -- it'll be w/w'd and so it will be CPR-ish. E.g.
755 -- f = \x::(Int,Int). if ...strict in x... then
759 -- We want f to have the CPR property because x does, by the time f has been w/w'd
761 -- NOTE: see notes [CPR-AND-STRICTNESS]
763 -- Also note that we only want to do this for something that
764 -- definitely has product type, else we may get over-optimistic
765 -- CPR results (e.g. from \x -> x!).
767 extendSigsWithLam sigs id
768 = case idNewDemandInfo_maybe id of
769 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
770 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
774 dmdTransform :: SigEnv -- The strictness environment
775 -> Id -- The function
776 -> Demand -- The demand on the function
777 -> DmdType -- The demand type of the function in this context
778 -- Returned DmdEnv includes the demand on
779 -- this function plus demand on its free variables
781 dmdTransform sigs var dmd
783 ------ DATA CONSTRUCTOR
784 | isDataConWorkId var -- Data constructor
786 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
787 DmdType _ _ con_res = dmd_ty
790 if arity == call_depth then -- Saturated, so unleash the demand
792 -- Important! If we Keep the constructor application, then
793 -- we need the demands the constructor places (always lazy)
794 -- If not, we don't need to. For example:
795 -- f p@(x,y) = (p,y) -- S(AL)
797 -- It's vital that we don't calculate Absent for a!
798 dmd_ds = case res_dmd of
799 Box (Eval ds) -> mapDmds box ds
803 -- ds can be empty, when we are just seq'ing the thing
804 -- If so we must make up a suitable bunch of demands
805 arg_ds = case dmd_ds of
806 Poly d -> replicate arity d
807 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
810 mkDmdType emptyDmdEnv arg_ds con_res
811 -- Must remember whether it's a product, hence con_res, not TopRes
815 ------ IMPORTED FUNCTION
816 | isGlobalId var, -- Imported function
817 let StrictSig dmd_ty = idNewStrictness var
818 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
823 ------ LOCAL LET/REC BOUND THING
824 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
826 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
827 | otherwise = deferType dmd_ty
828 -- NB: it's important to use deferType, and not just return topDmdType
829 -- Consider let { f x y = p + x } in f 1
830 -- The application isn't saturated, but we must nevertheless propagate
831 -- a lazy demand for p!
833 addVarDmd top_lvl fn_ty var dmd
835 ------ LOCAL NON-LET/REC BOUND THING
836 | otherwise -- Default case
840 (call_depth, res_dmd) = splitCallDmd dmd
844 %************************************************************************
848 %************************************************************************
851 splitCallDmd :: Demand -> (Int, Demand)
852 splitCallDmd (Call d) = case splitCallDmd d of
854 splitCallDmd d = (0, d)
856 vanillaCall :: Arity -> Demand
857 vanillaCall 0 = evalDmd
858 vanillaCall n = Call (vanillaCall (n-1))
860 deferType :: DmdType -> DmdType
861 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
862 -- Notice that we throw away info about both arguments and results
863 -- For example, f = let ... in \x -> x
864 -- We don't want to get a stricness type V->T for f.
867 deferEnv :: DmdEnv -> DmdEnv
868 deferEnv fv = mapVarEnv defer fv
872 argDemand :: Demand -> Demand
873 -- The 'Defer' demands are just Lazy at function boundaries
874 -- Ugly! Ask John how to improve it.
875 argDemand Top = lazyDmd
876 argDemand (Defer d) = lazyDmd
877 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
878 argDemand (Box Bot) = evalDmd
879 argDemand (Box d) = box (argDemand d)
880 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
885 -------------------------
886 -- Consider (if x then y else []) with demand V
887 -- Then the first branch gives {y->V} and the second
888 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
889 -- in the result env.
890 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
891 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
893 lub_fv = plusUFM_C lub fv1 fv2
894 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
895 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
896 -- lub is the identity for Bot
898 -- Extend the shorter argument list to match the longer
899 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
901 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
902 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
904 -----------------------------------
905 -- (t1 `bothType` t2) takes the argument/result info from t1,
906 -- using t2 just for its free-var info
907 -- NB: Don't forget about r2! It might be BotRes, which is
908 -- a bottom demand on all the in-scope variables.
909 -- Peter: can this be done more neatly?
910 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
911 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
913 both_fv = plusUFM_C both fv1 fv2
914 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
915 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
916 -- both is the identity for Abs
923 lubRes RetCPR RetCPR = RetCPR
924 lubRes r1 r2 = TopRes
926 -- If either diverges, the whole thing does
927 -- Otherwise take CPR info from the first
928 bothRes r1 BotRes = BotRes
933 modifyEnv :: Bool -- No-op if False
934 -> (Demand -> Demand) -- The zapper
935 -> DmdEnv -> DmdEnv -- Env1 and Env2
936 -> DmdEnv -> DmdEnv -- Transform this env
937 -- Zap anything in Env1 but not in Env2
938 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
940 modifyEnv need_to_modify zapper env1 env2 env
941 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
944 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
946 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
950 %************************************************************************
952 \subsection{LUB and BOTH}
954 %************************************************************************
957 lub :: Demand -> Demand -> Demand
960 lub Abs d2 = absLub d2
962 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
964 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
965 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
966 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
967 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
969 -- For the Eval case, we use these approximation rules
970 -- Box Bot <= Eval (Box Bot ...)
971 -- Box Top <= Defer (Box Bot ...)
972 -- Box (Eval ds) <= Eval (map Box ds)
973 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
974 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
975 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
976 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
977 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
979 lub (Box d1) (Box d2) = box (d1 `lub` d2)
980 lub d1@(Box _) d2 = d2 `lub` d1
982 lubs = zipWithDmds lub
984 ---------------------
985 -- box is the smart constructor for Box
986 -- It computes <B,bot> & d
987 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
988 -- Seems to be no point in allowing (Box (Call d))
989 box (Call d) = Call d -- The odd man out. Why?
991 box (Defer _) = lazyDmd
992 box Top = lazyDmd -- Box Abs and Box Top
993 box Abs = lazyDmd -- are the same <B,L>
994 box d = Box d -- Bot, Eval
997 defer :: Demand -> Demand
999 -- defer is the smart constructor for Defer
1000 -- The idea is that (Defer ds) = <U(ds), L>
1002 -- It specifies what happens at a lazy function argument
1003 -- or a lambda; the L* operator
1004 -- Set the strictness part to L, but leave
1005 -- the boxity side unaffected
1006 -- It also ensures that Defer (Eval [LLLL]) = L
1011 defer (Call _) = lazyDmd -- Approximation here?
1012 defer (Box _) = lazyDmd
1013 defer (Defer ds) = Defer ds
1014 defer (Eval ds) = deferEval ds
1016 -- deferEval ds = defer (Eval ds)
1017 deferEval ds | allTop ds = Top
1018 | otherwise = Defer ds
1020 ---------------------
1021 absLub :: Demand -> Demand
1022 -- Computes (Abs `lub` d)
1023 -- For the Bot case consider
1024 -- f x y = if ... then x else error x
1025 -- Then for y we get Abs `lub` Bot, and we really
1030 absLub (Call _) = Top
1031 absLub (Box _) = Top
1032 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1033 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1035 absLubs = mapDmds absLub
1038 both :: Demand -> Demand -> Demand
1044 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1047 -- From 'error' itself we get demand Bot on x
1048 -- From the arg demand on x we get
1049 -- x :-> evalDmd = Box (Eval (Poly Abs))
1050 -- So we get Bot `both` Box (Eval (Poly Abs))
1051 -- = Seq Keep (Poly Bot)
1054 -- f x = if ... then error (fst x) else fst x
1055 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1057 -- which is what we want.
1060 both Top Bot = errDmd
1063 both Top (Box d) = Box d
1064 both Top (Call d) = Call d
1065 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1066 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1067 -- = defer (Eval (mapDmds (`both` Top) ds))
1068 = deferEval (mapDmds (`both` Top) ds)
1071 both (Box d1) (Box d2) = box (d1 `both` d2)
1072 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1073 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1074 both (Box d1) (Defer d2) = Box d1
1075 both d1@(Box _) d2 = d2 `both` d1
1077 both (Call d1) (Call d2) = Call (d1 `both` d2)
1078 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1079 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1080 both d1@(Call _) d2 = d1 `both` d1
1082 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1083 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1084 both d1@(Eval ds1) d2 = d2 `both` d1
1086 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1087 both d1@(Defer ds1) d2 = d2 `both` d1
1089 boths = zipWithDmds both
1094 %************************************************************************
1096 \subsection{Miscellaneous
1098 %************************************************************************
1102 #ifdef OLD_STRICTNESS
1103 get_changes binds = vcat (map get_changes_bind binds)
1105 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1106 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1108 get_changes_pr (id,rhs)
1109 = get_changes_var id $$ get_changes_expr rhs
1112 | isId var = get_changes_str var $$ get_changes_dmd var
1115 get_changes_expr (Type t) = empty
1116 get_changes_expr (Var v) = empty
1117 get_changes_expr (Lit l) = empty
1118 get_changes_expr (Note n e) = get_changes_expr e
1119 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1120 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1121 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1122 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1124 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1127 | new_better && old_better = empty
1128 | new_better = message "BETTER"
1129 | old_better = message "WORSE"
1130 | otherwise = message "INCOMPARABLE"
1132 message word = text word <+> text "strictness for" <+> ppr id <+> info
1133 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1134 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1135 -- strictness analyser can't track
1136 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1137 old_better = old `betterStrictness` new
1138 new_better = new `betterStrictness` old
1141 | isUnLiftedType (idType id) = empty -- Not useful
1142 | new_better && old_better = empty
1143 | new_better = message "BETTER"
1144 | old_better = message "WORSE"
1145 | otherwise = message "INCOMPARABLE"
1147 message word = text word <+> text "demand for" <+> ppr id <+> info
1148 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1149 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1151 old = newDemand (idDemandInfo id)
1152 new_better = new `betterDemand` old
1153 old_better = old `betterDemand` new
1155 betterStrictness :: StrictSig -> StrictSig -> Bool
1156 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
1158 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
1160 betterDemand :: Demand -> Demand -> Bool
1161 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
1162 betterDemand d1 d2 = (d1 `lub` d2) == d2
1164 squashSig (StrictSig (DmdType fv ds res))
1165 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1167 -- squash just gets rid of call demands
1168 -- which the old analyser doesn't track
1169 squashDmd (Call d) = evalDmd
1170 squashDmd (Box d) = Box (squashDmd d)
1171 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1172 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)