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, exprIsTrivial, 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')
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
255 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' ty [alt'])
258 dmdAnal sigs dmd (Case scrut case_bndr ty alts)
260 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
261 (scrut_ty, scrut') = dmdAnal sigs evalDmd scrut
262 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
264 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
266 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' ty alts')
268 dmdAnal sigs dmd (Let (NonRec id rhs) body)
270 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel NonRecursive sigs (id, rhs)
271 (body_ty, body') = dmdAnal sigs' dmd body
272 (body_ty1, id2) = annotateBndr body_ty id1
273 body_ty2 = addLazyFVs body_ty1 lazy_fv
275 -- If the actual demand is better than the vanilla call
276 -- demand, you might think that we might do better to re-analyse
277 -- the RHS with the stronger demand.
278 -- But (a) That seldom happens, because it means that *every* path in
279 -- the body of the let has to use that stronger demand
280 -- (b) It often happens temporarily in when fixpointing, because
281 -- the recursive function at first seems to place a massive demand.
282 -- But we don't want to go to extra work when the function will
283 -- probably iterate to something less demanding.
284 -- In practice, all the times the actual demand on id2 is more than
285 -- the vanilla call demand seem to be due to (b). So we don't
286 -- bother to re-analyse the RHS.
287 (body_ty2, Let (NonRec id2 rhs') body')
289 dmdAnal sigs dmd (Let (Rec pairs) body)
291 bndrs = map fst pairs
292 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
293 (body_ty, body') = dmdAnal sigs' dmd body
294 body_ty1 = addLazyFVs body_ty lazy_fv
296 sigs' `seq` body_ty `seq`
298 (body_ty2, _) = annotateBndrs body_ty1 bndrs
299 -- Don't bother to add demand info to recursive
300 -- binders as annotateBndr does;
301 -- being recursive, we can't treat them strictly.
302 -- But we do need to remove the binders from the result demand env
304 (body_ty2, Let (Rec pairs') body')
307 dmdAnalAlt sigs dmd (con,bndrs,rhs)
309 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
310 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
311 final_alt_ty | io_hack_reqd = alt_ty `lubType` topDmdType
314 -- There's a hack here for I/O operations. Consider
315 -- case foo x s of { (# s, r #) -> y }
316 -- Is this strict in 'y'. Normally yes, but what if 'foo' is an I/O
317 -- operation that simply terminates the program (not in an erroneous way)?
318 -- In that case we should not evaluate y before the call to 'foo'.
319 -- Hackish solution: spot the IO-like situation and add a virtual branch,
323 -- other -> return ()
324 -- So the 'y' isn't necessarily going to be evaluated
326 -- A more complete example where this shows up is:
327 -- do { let len = <expensive> ;
328 -- ; when (...) (exitWith ExitSuccess)
331 io_hack_reqd = con == DataAlt unboxedPairDataCon &&
332 idType (head bndrs) `eqType` realWorldStatePrimTy
334 (final_alt_ty, (con, bndrs', rhs'))
337 %************************************************************************
339 \subsection{Bindings}
341 %************************************************************************
344 dmdFix :: TopLevelFlag
345 -> SigEnv -- Does not include bindings for this binding
348 [(Id,CoreExpr)]) -- Binders annotated with stricness info
350 dmdFix top_lvl sigs orig_pairs
351 = loop 1 initial_sigs orig_pairs
353 bndrs = map fst orig_pairs
354 initial_sigs = extendSigEnvList sigs [(id, (initialSig id, top_lvl)) | id <- bndrs]
357 -> SigEnv -- Already contains the current sigs
359 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
362 = (sigs', lazy_fv, pairs')
363 -- Note: use pairs', not pairs. pairs' is the result of
364 -- processing the RHSs with sigs (= sigs'), whereas pairs
365 -- is the result of processing the RHSs with the *previous*
366 -- iteration of sigs.
368 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
369 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
370 text "env:" <+> ppr (ufmToList sigs),
371 text "binds:" <+> pprCoreBinding (Rec pairs)]))
372 (emptySigEnv, lazy_fv, orig_pairs) -- Safe output
373 -- The lazy_fv part is really important! orig_pairs has no strictness
374 -- info, including nothing about free vars. But if we have
375 -- letrec f = ....y..... in ...f...
376 -- where 'y' is free in f, we must record that y is mentioned,
377 -- otherwise y will get recorded as absent altogether
379 | otherwise = loop (n+1) sigs' pairs'
381 found_fixpoint = all (same_sig sigs sigs') bndrs
382 -- Use the new signature to do the next pair
383 -- The occurrence analyser has arranged them in a good order
384 -- so this can significantly reduce the number of iterations needed
385 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
387 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
388 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
390 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
391 ((sigs', lazy_fv'), pair')
394 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl Recursive sigs (id,rhs)
395 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
396 -- old_sig = lookup sigs id
397 -- new_sig = lookup sigs' id
399 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
400 lookup sigs var = case lookupVarEnv sigs var of
403 -- Get an initial strictness signature from the Id
404 -- itself. That way we make use of earlier iterations
405 -- of the fixpoint algorithm. (Cunning plan.)
406 -- Note that the cunning plan extends to the DmdEnv too,
407 -- since it is part of the strictness signature
408 initialSig id = idNewStrictness_maybe id `orElse` botSig
410 dmdAnalRhs :: TopLevelFlag -> RecFlag
411 -> SigEnv -> (Id, CoreExpr)
412 -> (SigEnv, DmdEnv, (Id, CoreExpr))
413 -- Process the RHS of the binding, add the strictness signature
414 -- to the Id, and augment the environment with the signature as well.
416 dmdAnalRhs top_lvl rec_flag sigs (id, rhs)
417 = (sigs', lazy_fv, (id', rhs'))
419 arity = idArity id -- The idArity should be up to date
420 -- The simplifier was run just beforehand
421 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
422 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty && not (exprIsTrivial rhs), ppr id )
423 -- The RHS can be eta-reduced to just a variable,
424 -- in which case we should not complain.
425 mkSigTy top_lvl rec_flag id rhs rhs_dmd_ty
426 id' = id `setIdNewStrictness` sig_ty
427 sigs' = extendSigEnv top_lvl sigs id sig_ty
430 %************************************************************************
432 \subsection{Strictness signatures and types}
434 %************************************************************************
437 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
438 -- Take a DmdType and turn it into a StrictSig
439 -- NB: not used for never-inline things; hence False
440 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
442 mkSigTy :: TopLevelFlag -> RecFlag -> Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
443 mkSigTy top_lvl rec_flag id rhs dmd_ty
444 = mk_sig_ty never_inline thunk_cpr_ok rhs dmd_ty
446 never_inline = isNeverActive (idInlinePragma id)
447 maybe_id_dmd = idNewDemandInfo_maybe id
448 -- Is Nothing the first time round
451 | isTopLevel top_lvl = False -- Top level things don't get
452 -- their demandInfo set at all
453 | isRec rec_flag = False -- Ditto recursive things
454 | Just dmd <- maybe_id_dmd = isStrictDmd dmd
455 | otherwise = True -- Optimistic, first time round
459 The thunk_cpr_ok stuff [CPR-AND-STRICTNESS]
460 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
461 If the rhs is a thunk, we usually forget the CPR info, because
462 it is presumably shared (else it would have been inlined, and
463 so we'd lose sharing if w/w'd it into a function.
465 However, if the strictness analyser has figured out (in a previous
466 iteration) that it's strict, then we DON'T need to forget the CPR info.
467 Instead we can retain the CPR info and do the thunk-splitting transform
468 (see WorkWrap.splitThunk).
470 This made a big difference to PrelBase.modInt, which had something like
471 modInt = \ x -> let r = ... -> I# v in
472 ...body strict in r...
473 r's RHS isn't a value yet; but modInt returns r in various branches, so
474 if r doesn't have the CPR property then neither does modInt
475 Another case I found in practice (in Complex.magnitude), looks like this:
476 let k = if ... then I# a else I# b
477 in ... body strict in k ....
478 (For this example, it doesn't matter whether k is returned as part of
479 the overall result; but it does matter that k's RHS has the CPR property.)
480 Left to itself, the simplifier will make a join point thus:
481 let $j k = ...body strict in k...
482 if ... then $j (I# a) else $j (I# b)
483 With thunk-splitting, we get instead
484 let $j x = let k = I#x in ...body strict in k...
485 in if ... then $j a else $j b
486 This is much better; there's a good chance the I# won't get allocated.
488 The difficulty with this is that we need the strictness type to
489 look at the body... but we now need the body to calculate the demand
490 on the variable, so we can decide whether its strictness type should
491 have a CPR in it or not. Simple solution:
492 a) use strictness info from the previous iteration
493 b) make sure we do at least 2 iterations, by doing a second
494 round for top-level non-recs. Top level recs will get at
495 least 2 iterations except for totally-bottom functions
496 which aren't very interesting anyway.
498 NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
500 The Nothing case in thunk_cpr_ok [CPR-AND-STRICTNESS]
501 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
502 Demand info now has a 'Nothing' state, just like strictness info.
503 The analysis works from 'dangerous' towards a 'safe' state; so we
504 start with botSig for 'Nothing' strictness infos, and we start with
505 "yes, it's demanded" for 'Nothing' in the demand info. The
506 fixpoint iteration will sort it all out.
508 We can't start with 'not-demanded' because then consider
512 if ... then t else I# y else f x'
514 In the first iteration we'd have no demand info for x, so assume
515 not-demanded; then we'd get TopRes for f's CPR info. Next iteration
516 we'd see that t was demanded, and so give it the CPR property, but by
517 now f has TopRes, so it will stay TopRes. Instead, with the Nothing
518 setting the first time round, we say 'yes t is demanded' the first
521 However, this does mean that for non-recursive bindings we must
522 iterate twice to be sure of not getting over-optimistic CPR info,
523 in the case where t turns out to be not-demanded. This is handled
528 mk_sig_ty never_inline thunk_cpr_ok rhs (DmdType fv dmds res)
529 | never_inline && not (isBotRes res)
531 -- Don't strictness-analyse NOINLINE things. Why not? Because
532 -- the NOINLINE says "don't expose any of the inner workings at the call
533 -- site" and the strictness is certainly an inner working.
535 -- More concretely, the demand analyser discovers the following strictness
536 -- for unsafePerformIO: C(U(AV))
538 -- unsafePerformIO (\s -> let r = f x in
539 -- case writeIORef v r s of (# s1, _ #) ->
541 -- The strictness analyser will find that the binding for r is strict,
542 -- (becuase of uPIO's strictness sig), and so it'll evaluate it before
543 -- doing the writeIORef. This actually makes tests/lib/should_run/memo002
546 -- Solution: don't expose the strictness of unsafePerformIO.
548 -- But we do want to expose the strictness of error functions,
549 -- which are also often marked NOINLINE
550 -- {-# NOINLINE foo #-}
551 -- foo x = error ("wubble buggle" ++ x)
552 -- So (hack, hack) we only drop the strictness for non-bottom things
553 -- This is all very unsatisfactory.
554 = (deferEnv fv, topSig)
557 = (lazy_fv, mkStrictSig dmd_ty)
559 dmd_ty = DmdType strict_fv final_dmds res'
561 lazy_fv = filterUFM (not . isStrictDmd) fv
562 strict_fv = filterUFM isStrictDmd fv
563 -- We put the strict FVs in the DmdType of the Id, so
564 -- that at its call sites we unleash demands on its strict fvs.
565 -- An example is 'roll' in imaginary/wheel-sieve2
566 -- Something like this:
568 -- go y = if ... then roll (x-1) else x+1
571 -- We want to see that roll is strict in x, which is because
572 -- go is called. So we put the DmdEnv for x in go's DmdType.
575 -- f :: Int -> Int -> Int
576 -- f x y = let t = x+1
577 -- h z = if z==0 then t else
578 -- if z==1 then x+1 else
582 -- Calling h does indeed evaluate x, but we can only see
583 -- that if we unleash a demand on x at the call site for t.
585 -- Incidentally, here's a place where lambda-lifting h would
586 -- lose the cigar --- we couldn't see the joint strictness in t/x
589 -- We don't want to put *all* the fv's from the RHS into the
590 -- DmdType, because that makes fixpointing very slow --- the
591 -- DmdType gets full of lazy demands that are slow to converge.
593 final_dmds = setUnpackStrategy dmds
594 -- Set the unpacking strategy
597 RetCPR | ignore_cpr_info -> TopRes
599 ignore_cpr_info = not (exprIsValue rhs || thunk_cpr_ok)
602 The unpack strategy determines whether we'll *really* unpack the argument,
603 or whether we'll just remember its strictness. If unpacking would give
604 rise to a *lot* of worker args, we may decide not to unpack after all.
607 setUnpackStrategy :: [Demand] -> [Demand]
609 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
611 go :: Int -- Max number of args available for sub-components of [Demand]
613 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
615 go n (Eval (Prod cs) : ds)
616 | n' >= 0 = Eval (Prod cs') `cons` go n'' ds
617 | otherwise = Box (Eval (Prod cs)) `cons` go n ds
620 n' = n + 1 - non_abs_args
621 -- Add one to the budget 'cos we drop the top-level arg
622 non_abs_args = nonAbsentArgs cs
623 -- Delete # of non-absent args to which we'll now be committed
625 go n (d:ds) = d `cons` go n ds
628 cons d (n,ds) = (n, d:ds)
630 nonAbsentArgs :: [Demand] -> Int
632 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
633 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
637 %************************************************************************
639 \subsection{Strictness signatures and types}
641 %************************************************************************
644 splitDmdTy :: DmdType -> (Demand, DmdType)
645 -- Split off one function argument
646 -- We already have a suitable demand on all
647 -- free vars, so no need to add more!
648 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
649 splitDmdTy ty@(DmdType fv [] res_ty) = (resTypeArgDmd res_ty, ty)
653 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
655 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
656 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
657 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
659 addLazyFVs (DmdType fv ds res) lazy_fvs
660 = DmdType both_fv1 ds res
662 both_fv = (plusUFM_C both fv lazy_fvs)
663 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
664 -- This modifyEnv is vital. Consider
665 -- let f = \x -> (x,y)
667 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
668 -- demand with the bottom coming up from 'error'
670 -- I got a loop in the fixpointer without this, due to an interaction
671 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
673 -- = letrec g y = x `fatbar`
674 -- letrec h z = z + ...g...
677 -- In the initial iteration for f, f=Bot
678 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
679 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
680 -- places on its free variables. Suppose it places none. Then the
681 -- x `fatbar` ...call to h...
682 -- will give a x->V demand for x. That turns into a L demand for x,
683 -- which floats out of the defn for h. Without the modifyEnv, that
684 -- L demand doesn't get both'd with the Bot coming up from the inner
685 -- call to f. So we just get an L demand for x for g.
687 -- A better way to say this is that the lazy-fv filtering should give the
688 -- same answer as putting the lazy fv demands in the function's type.
690 annotateBndr :: DmdType -> Var -> (DmdType, Var)
691 -- The returned env has the var deleted
692 -- The returned var is annotated with demand info
693 -- No effect on the argument demands
694 annotateBndr dmd_ty@(DmdType fv ds res) var
695 | isTyVar var = (dmd_ty, var)
696 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var dmd)
698 (fv', dmd) = removeFV fv var res
700 annotateBndrs = mapAccumR annotateBndr
702 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
703 -- For lambdas we add the demand to the argument demands
704 -- Only called for Ids
706 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
708 (fv', dmd) = removeFV fv id res
709 hacked_dmd = argDemand dmd
710 -- This call to argDemand is vital, because otherwise we label
711 -- a lambda binder with demand 'B'. But in terms of calling
712 -- conventions that's Abs, because we don't pass it. But
713 -- when we do a w/w split we get
714 -- fw x = (\x y:B -> ...) x (error "oops")
715 -- And then the simplifier things the 'B' is a strict demand
716 -- and evaluates the (error "oops"). Sigh
718 removeFV fv id res = (fv', zapUnlifted id dmd)
720 fv' = fv `delVarEnv` id
721 dmd = lookupVarEnv fv id `orElse` deflt
722 deflt | isBotRes res = Bot
725 -- For unlifted-type variables, we are only
726 -- interested in Bot/Abs/Box Abs
727 zapUnlifted is Bot = Bot
728 zapUnlifted id Abs = Abs
729 zapUnlifted id dmd | isUnLiftedType (idType id) = lazyDmd
733 %************************************************************************
735 \subsection{Strictness signatures}
737 %************************************************************************
740 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
741 -- We use the SigEnv to tell us whether to
742 -- record info about a variable in the DmdEnv
743 -- We do so if it's a LocalId, but not top-level
745 -- The DmdEnv gives the demand on the free vars of the function
746 -- when it is given enough args to satisfy the strictness signature
748 emptySigEnv = emptyVarEnv
750 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
751 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
753 extendSigEnvList = extendVarEnvList
755 extendSigsWithLam :: SigEnv -> Id -> SigEnv
756 -- Extend the SigEnv when we meet a lambda binder
757 -- If the binder is marked demanded with a product demand, then give it a CPR
758 -- signature, because in the likely event that this is a lambda on a fn defn
759 -- [we only use this when the lambda is being consumed with a call demand],
760 -- it'll be w/w'd and so it will be CPR-ish. E.g.
761 -- f = \x::(Int,Int). if ...strict in x... then
765 -- We want f to have the CPR property because x does, by the time f has been w/w'd
767 -- NOTE: see notes [CPR-AND-STRICTNESS]
769 -- Also note that we only want to do this for something that
770 -- definitely has product type, else we may get over-optimistic
771 -- CPR results (e.g. from \x -> x!).
773 extendSigsWithLam sigs id
774 = case idNewDemandInfo_maybe id of
775 Nothing -> extendVarEnv sigs id (cprSig, NotTopLevel)
776 Just (Eval (Prod ds)) -> extendVarEnv sigs id (cprSig, NotTopLevel)
780 dmdTransform :: SigEnv -- The strictness environment
781 -> Id -- The function
782 -> Demand -- The demand on the function
783 -> DmdType -- The demand type of the function in this context
784 -- Returned DmdEnv includes the demand on
785 -- this function plus demand on its free variables
787 dmdTransform sigs var dmd
789 ------ DATA CONSTRUCTOR
790 | isDataConWorkId var -- Data constructor
792 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
793 DmdType _ _ con_res = dmd_ty
796 if arity == call_depth then -- Saturated, so unleash the demand
798 -- Important! If we Keep the constructor application, then
799 -- we need the demands the constructor places (always lazy)
800 -- If not, we don't need to. For example:
801 -- f p@(x,y) = (p,y) -- S(AL)
803 -- It's vital that we don't calculate Absent for a!
804 dmd_ds = case res_dmd of
805 Box (Eval ds) -> mapDmds box ds
809 -- ds can be empty, when we are just seq'ing the thing
810 -- If so we must make up a suitable bunch of demands
811 arg_ds = case dmd_ds of
812 Poly d -> replicate arity d
813 Prod ds -> ASSERT( ds `lengthIs` arity ) ds
816 mkDmdType emptyDmdEnv arg_ds con_res
817 -- Must remember whether it's a product, hence con_res, not TopRes
821 ------ IMPORTED FUNCTION
822 | isGlobalId var, -- Imported function
823 let StrictSig dmd_ty = idNewStrictness var
824 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
829 ------ LOCAL LET/REC BOUND THING
830 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
832 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
833 | otherwise = deferType dmd_ty
834 -- NB: it's important to use deferType, and not just return topDmdType
835 -- Consider let { f x y = p + x } in f 1
836 -- The application isn't saturated, but we must nevertheless propagate
837 -- a lazy demand for p!
839 addVarDmd top_lvl fn_ty var dmd
841 ------ LOCAL NON-LET/REC BOUND THING
842 | otherwise -- Default case
846 (call_depth, res_dmd) = splitCallDmd dmd
850 %************************************************************************
854 %************************************************************************
857 splitCallDmd :: Demand -> (Int, Demand)
858 splitCallDmd (Call d) = case splitCallDmd d of
860 splitCallDmd d = (0, d)
862 vanillaCall :: Arity -> Demand
863 vanillaCall 0 = evalDmd
864 vanillaCall n = Call (vanillaCall (n-1))
866 deferType :: DmdType -> DmdType
867 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
868 -- Notice that we throw away info about both arguments and results
869 -- For example, f = let ... in \x -> x
870 -- We don't want to get a stricness type V->T for f.
873 deferEnv :: DmdEnv -> DmdEnv
874 deferEnv fv = mapVarEnv defer fv
878 argDemand :: Demand -> Demand
879 -- The 'Defer' demands are just Lazy at function boundaries
880 -- Ugly! Ask John how to improve it.
881 argDemand Top = lazyDmd
882 argDemand (Defer d) = lazyDmd
883 argDemand (Eval ds) = Eval (mapDmds argDemand ds)
884 argDemand (Box Bot) = evalDmd
885 argDemand (Box d) = box (argDemand d)
886 argDemand Bot = Abs -- Don't pass args that are consumed (only) by bottom
891 -------------------------
892 -- Consider (if x then y else []) with demand V
893 -- Then the first branch gives {y->V} and the second
894 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
895 -- in the result env.
896 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
897 = DmdType lub_fv2 (lub_ds ds1 ds2) (r1 `lubRes` r2)
899 lub_fv = plusUFM_C lub fv1 fv2
900 lub_fv1 = modifyEnv (not (isBotRes r1)) absLub fv2 fv1 lub_fv
901 lub_fv2 = modifyEnv (not (isBotRes r2)) absLub fv1 fv2 lub_fv1
902 -- lub is the identity for Bot
904 -- Extend the shorter argument list to match the longer
905 lub_ds (d1:ds1) (d2:ds2) = lub d1 d2 : lub_ds ds1 ds2
907 lub_ds ds1 [] = map (`lub` resTypeArgDmd r2) ds1
908 lub_ds [] ds2 = map (resTypeArgDmd r1 `lub`) ds2
910 -----------------------------------
911 -- (t1 `bothType` t2) takes the argument/result info from t1,
912 -- using t2 just for its free-var info
913 -- NB: Don't forget about r2! It might be BotRes, which is
914 -- a bottom demand on all the in-scope variables.
915 -- Peter: can this be done more neatly?
916 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
917 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
919 both_fv = plusUFM_C both fv1 fv2
920 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
921 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
922 -- both is the identity for Abs
929 lubRes RetCPR RetCPR = RetCPR
930 lubRes r1 r2 = TopRes
932 -- If either diverges, the whole thing does
933 -- Otherwise take CPR info from the first
934 bothRes r1 BotRes = BotRes
939 modifyEnv :: Bool -- No-op if False
940 -> (Demand -> Demand) -- The zapper
941 -> DmdEnv -> DmdEnv -- Env1 and Env2
942 -> DmdEnv -> DmdEnv -- Transform this env
943 -- Zap anything in Env1 but not in Env2
944 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
946 modifyEnv need_to_modify zapper env1 env2 env
947 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
950 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
952 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
956 %************************************************************************
958 \subsection{LUB and BOTH}
960 %************************************************************************
963 lub :: Demand -> Demand -> Demand
966 lub Abs d2 = absLub d2
968 lub (Defer ds1) d2 = defer (Eval ds1 `lub` d2)
970 lub (Call d1) (Call d2) = Call (d1 `lub` d2)
971 lub d1@(Call _) (Box d2) = d1 `lub` d2 -- Just strip the box
972 lub d1@(Call _) d2@(Eval _) = d2 -- Presumably seq or vanilla eval
973 lub d1@(Call _) d2 = d2 `lub` d1 -- Bot, Abs, Top
975 -- For the Eval case, we use these approximation rules
976 -- Box Bot <= Eval (Box Bot ...)
977 -- Box Top <= Defer (Box Bot ...)
978 -- Box (Eval ds) <= Eval (map Box ds)
979 lub (Eval ds1) (Eval ds2) = Eval (ds1 `lubs` ds2)
980 lub (Eval ds1) (Box Bot) = Eval (mapDmds (`lub` Box Bot) ds1)
981 lub (Eval ds1) (Box (Eval ds2)) = Eval (ds1 `lubs` mapDmds box ds2)
982 lub (Eval ds1) (Box Abs) = deferEval (mapDmds (`lub` Box Bot) ds1)
983 lub d1@(Eval _) d2 = d2 `lub` d1 -- Bot,Abs,Top,Call,Defer
985 lub (Box d1) (Box d2) = box (d1 `lub` d2)
986 lub d1@(Box _) d2 = d2 `lub` d1
988 lubs = zipWithDmds lub
990 ---------------------
991 -- box is the smart constructor for Box
992 -- It computes <B,bot> & d
993 -- INVARIANT: (Box d) => d = Bot, Abs, Eval
994 -- Seems to be no point in allowing (Box (Call d))
995 box (Call d) = Call d -- The odd man out. Why?
997 box (Defer _) = lazyDmd
998 box Top = lazyDmd -- Box Abs and Box Top
999 box Abs = lazyDmd -- are the same <B,L>
1000 box d = Box d -- Bot, Eval
1003 defer :: Demand -> Demand
1005 -- defer is the smart constructor for Defer
1006 -- The idea is that (Defer ds) = <U(ds), L>
1008 -- It specifies what happens at a lazy function argument
1009 -- or a lambda; the L* operator
1010 -- Set the strictness part to L, but leave
1011 -- the boxity side unaffected
1012 -- It also ensures that Defer (Eval [LLLL]) = L
1017 defer (Call _) = lazyDmd -- Approximation here?
1018 defer (Box _) = lazyDmd
1019 defer (Defer ds) = Defer ds
1020 defer (Eval ds) = deferEval ds
1022 -- deferEval ds = defer (Eval ds)
1023 deferEval ds | allTop ds = Top
1024 | otherwise = Defer ds
1026 ---------------------
1027 absLub :: Demand -> Demand
1028 -- Computes (Abs `lub` d)
1029 -- For the Bot case consider
1030 -- f x y = if ... then x else error x
1031 -- Then for y we get Abs `lub` Bot, and we really
1036 absLub (Call _) = Top
1037 absLub (Box _) = Top
1038 absLub (Eval ds) = Defer (absLubs ds) -- Or (Defer ds)?
1039 absLub (Defer ds) = Defer (absLubs ds) -- Or (Defer ds)?
1041 absLubs = mapDmds absLub
1044 both :: Demand -> Demand -> Demand
1050 both Bot (Eval ds) = Eval (mapDmds (`both` Bot) ds)
1053 -- From 'error' itself we get demand Bot on x
1054 -- From the arg demand on x we get
1055 -- x :-> evalDmd = Box (Eval (Poly Abs))
1056 -- So we get Bot `both` Box (Eval (Poly Abs))
1057 -- = Seq Keep (Poly Bot)
1060 -- f x = if ... then error (fst x) else fst x
1061 -- Then we get (Eval (Box Bot, Bot) `lub` Eval (SA))
1063 -- which is what we want.
1066 both Top Bot = errDmd
1069 both Top (Box d) = Box d
1070 both Top (Call d) = Call d
1071 both Top (Eval ds) = Eval (mapDmds (`both` Top) ds)
1072 both Top (Defer ds) -- = defer (Top `both` Eval ds)
1073 -- = defer (Eval (mapDmds (`both` Top) ds))
1074 = deferEval (mapDmds (`both` Top) ds)
1077 both (Box d1) (Box d2) = box (d1 `both` d2)
1078 both (Box d1) d2@(Call _) = box (d1 `both` d2)
1079 both (Box d1) d2@(Eval _) = box (d1 `both` d2)
1080 both (Box d1) (Defer d2) = Box d1
1081 both d1@(Box _) d2 = d2 `both` d1
1083 both (Call d1) (Call d2) = Call (d1 `both` d2)
1084 both (Call d1) (Eval ds2) = Call d1 -- Could do better for (Poly Bot)?
1085 both (Call d1) (Defer ds2) = Call d1 -- Ditto
1086 both d1@(Call _) d2 = d1 `both` d1
1088 both (Eval ds1) (Eval ds2) = Eval (ds1 `boths` ds2)
1089 both (Eval ds1) (Defer ds2) = Eval (ds1 `boths` mapDmds defer ds2)
1090 both d1@(Eval ds1) d2 = d2 `both` d1
1092 both (Defer ds1) (Defer ds2) = deferEval (ds1 `boths` ds2)
1093 both d1@(Defer ds1) d2 = d2 `both` d1
1095 boths = zipWithDmds both
1100 %************************************************************************
1102 \subsection{Miscellaneous
1104 %************************************************************************
1108 #ifdef OLD_STRICTNESS
1109 get_changes binds = vcat (map get_changes_bind binds)
1111 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
1112 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
1114 get_changes_pr (id,rhs)
1115 = get_changes_var id $$ get_changes_expr rhs
1118 | isId var = get_changes_str var $$ get_changes_dmd var
1121 get_changes_expr (Type t) = empty
1122 get_changes_expr (Var v) = empty
1123 get_changes_expr (Lit l) = empty
1124 get_changes_expr (Note n e) = get_changes_expr e
1125 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
1126 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
1127 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1128 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1130 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1133 | new_better && old_better = empty
1134 | new_better = message "BETTER"
1135 | old_better = message "WORSE"
1136 | otherwise = message "INCOMPARABLE"
1138 message word = text word <+> text "strictness for" <+> ppr id <+> info
1139 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1140 new = squashSig (idNewStrictness id) -- Don't report spurious diffs that the old
1141 -- strictness analyser can't track
1142 old = newStrictnessFromOld (idName id) (idArity id) (idStrictness id) (idCprInfo id)
1143 old_better = old `betterStrictness` new
1144 new_better = new `betterStrictness` old
1147 | isUnLiftedType (idType id) = empty -- Not useful
1148 | new_better && old_better = empty
1149 | new_better = message "BETTER"
1150 | old_better = message "WORSE"
1151 | otherwise = message "INCOMPARABLE"
1153 message word = text word <+> text "demand for" <+> ppr id <+> info
1154 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1155 new = squashDmd (argDemand (idNewDemandInfo id)) -- To avoid spurious improvements
1157 old = newDemand (idDemandInfo id)
1158 new_better = new `betterDemand` old
1159 old_better = old `betterDemand` new
1161 betterStrictness :: StrictSig -> StrictSig -> Bool
1162 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
1164 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
1166 betterDemand :: Demand -> Demand -> Bool
1167 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
1168 betterDemand d1 d2 = (d1 `lub` d2) == d2
1170 squashSig (StrictSig (DmdType fv ds res))
1171 = StrictSig (DmdType emptyDmdEnv (map squashDmd ds) res)
1173 -- squash just gets rid of call demands
1174 -- which the old analyser doesn't track
1175 squashDmd (Call d) = evalDmd
1176 squashDmd (Box d) = Box (squashDmd d)
1177 squashDmd (Eval ds) = Eval (mapDmds squashDmd ds)
1178 squashDmd (Defer ds) = Defer (mapDmds squashDmd ds)