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, idDemandInfo, idInlinePragma,
24 isDataConId, isGlobalId, idArity,
25 idNewStrictness, idNewStrictness_maybe, getNewStrictness, setIdNewStrictness,
26 idNewDemandInfo, setIdNewDemandInfo, newStrictnessFromOld )
27 import IdInfo ( newDemand )
30 import UniqFM ( plusUFM_C, addToUFM_Directly, lookupUFM_Directly,
31 keysUFM, minusUFM, ufmToList, filterUFM )
32 import Type ( isUnLiftedType )
33 import CoreLint ( showPass, endPass )
34 import Util ( mapAndUnzip, mapAccumL, mapAccumR )
35 import BasicTypes ( Arity, TopLevelFlag(..), isTopLevel, isNeverActive )
36 import Maybes ( orElse, expectJust )
42 * set a noinline pragma on bottoming Ids
44 * Consider f x = x+1 `fatbar` error (show x)
45 We'd like to unbox x, even if that means reboxing it in the error case.
48 instance Outputable TopLevelFlag where
52 %************************************************************************
54 \subsection{Top level stuff}
56 %************************************************************************
59 dmdAnalPgm :: DynFlags -> [CoreBind] -> IO [CoreBind]
60 dmdAnalPgm dflags binds
62 showPass dflags "Demand analysis" ;
63 let { binds_plus_dmds = do_prog binds ;
64 dmd_changes = get_changes binds_plus_dmds } ;
65 endPass dflags "Demand analysis"
66 Opt_D_dump_stranal binds_plus_dmds ;
68 -- Only if DEBUG is on, because only then is the old strictness analyser run
69 printDump (text "Changes in demands" $$ dmd_changes) ;
71 return binds_plus_dmds
74 do_prog :: [CoreBind] -> [CoreBind]
75 do_prog binds = snd $ mapAccumL dmdAnalTopBind emptySigEnv binds
77 dmdAnalTopBind :: SigEnv
80 dmdAnalTopBind sigs (NonRec id rhs)
82 ( _, _, (_, rhs1)) = dmdAnalRhs TopLevel sigs (id, rhs)
83 (sigs2, _, (id2, rhs2)) = dmdAnalRhs TopLevel sigs (id, rhs1)
84 -- Do two passes to improve CPR information
85 -- See the comments with mkSigTy.ignore_cpr_info below
87 (sigs2, NonRec id2 rhs2)
89 dmdAnalTopBind sigs (Rec pairs)
91 (sigs', _, pairs') = dmdFix TopLevel sigs pairs
92 -- We get two iterations automatically
98 dmdAnalTopRhs :: CoreExpr -> (StrictSig, CoreExpr)
99 -- Analyse the RHS and return
100 -- a) appropriate strictness info
101 -- b) the unfolding (decorated with stricntess info)
105 arity = exprArity rhs
106 (rhs_ty, rhs') = dmdAnal emptySigEnv (vanillaCall arity) rhs
107 sig = mkTopSigTy rhs rhs_ty
110 %************************************************************************
112 \subsection{The analyser itself}
114 %************************************************************************
117 dmdAnal :: SigEnv -> Demand -> CoreExpr -> (DmdType, CoreExpr)
119 dmdAnal sigs Abs e = (topDmdType, e)
120 dmdAnal sigs Bot e = (botDmdType, e)
122 dmdAnal sigs Lazy e = let
123 (res_ty, e') = dmdAnal sigs Eval e
125 (deferType res_ty, e')
126 -- It's important not to analyse e with a lazy demand because
127 -- a) When we encounter case s of (a,b) ->
128 -- we demand s with U(d1d2)... but if the overall demand is lazy
129 -- that is wrong, and we'd need to reduce the demand on s,
130 -- which is inconvenient
131 -- b) More important, consider
132 -- f (let x = R in x+x), where f is lazy
133 -- We still want to mark x as demanded, because it will be when we
134 -- enter the let. If we analyse f's arg with a Lazy demand, we'll
135 -- just mark x as Lazy
136 -- c) The application rule wouldn't be right either
137 -- Evaluating (f x) in a L demand does *not* cause
138 -- evaluation of f in a C(L) demand!
141 dmdAnal sigs dmd (Lit lit)
142 = (topDmdType, Lit lit)
144 dmdAnal sigs dmd (Var var)
145 = (dmdTransform sigs var dmd, Var var)
147 dmdAnal sigs dmd (Note n e)
148 = (dmd_ty, Note n e')
150 (dmd_ty, e') = dmdAnal sigs dmd' e
152 Coerce _ _ -> Eval -- This coerce usually arises from a recursive
153 other -> dmd -- newtype, and we don't want to look inside them
154 -- for exactly the same reason that we don't look
155 -- inside recursive products -- we might not reach
156 -- a fixpoint. So revert to a vanilla Eval demand
158 dmdAnal sigs dmd (App fun (Type ty))
159 = (fun_ty, App fun' (Type ty))
161 (fun_ty, fun') = dmdAnal sigs dmd fun
163 -- Lots of the other code is there to make this
164 -- beautiful, compositional, application rule :-)
165 dmdAnal sigs dmd e@(App fun arg) -- Non-type arguments
166 = let -- [Type arg handled above]
167 (fun_ty, fun') = dmdAnal sigs (Call dmd) fun
168 (arg_ty, arg') = dmdAnal sigs arg_dmd arg
169 (arg_dmd, res_ty) = splitDmdTy fun_ty
171 (res_ty `bothType` arg_ty, App fun' arg')
173 dmdAnal sigs dmd (Lam var body)
176 (body_ty, body') = dmdAnal sigs dmd body
178 (body_ty, Lam var body')
180 | Call body_dmd <- dmd -- A call demand: good!
182 (body_ty, body') = dmdAnal sigs body_dmd body
183 (lam_ty, var') = annotateLamIdBndr body_ty var
185 (lam_ty, Lam var' body')
187 | otherwise -- Not enough demand on the lambda; but do the body
188 = let -- anyway to annotate it and gather free var info
189 (body_ty, body') = dmdAnal sigs Eval body
190 (lam_ty, var') = annotateLamIdBndr body_ty var
192 (deferType lam_ty, Lam var' body')
194 dmdAnal sigs dmd (Case scrut case_bndr [alt@(DataAlt dc,bndrs,rhs)])
195 | let tycon = dataConTyCon dc,
196 isProductTyCon tycon,
197 not (isRecursiveTyCon tycon)
199 sigs_alt = extendSigEnv NotTopLevel sigs case_bndr case_bndr_sig
200 (alt_ty, alt') = dmdAnalAlt sigs_alt dmd alt
201 (alt_ty1, case_bndr') = annotateBndr alt_ty case_bndr
202 (_, bndrs', _) = alt'
203 case_bndr_sig = StrictSig (mkDmdType emptyVarEnv [] RetCPR)
204 -- Inside the alternative, the case binder has the CPR property.
205 -- Meaning that a case on it will successfully cancel.
207 -- f True x = case x of y { I# x' -> if x' ==# 3 then y else I# 8 }
210 -- We want f to have the CPR property:
211 -- f b x = case fw b x of { r -> I# r }
212 -- fw True x = case x of y { I# x' -> if x' ==# 3 then x' else 8 }
215 -- Figure out whether the demand on the case binder is used, and use
216 -- that to set the scrut_dmd. This is utterly essential.
217 -- Consider f x = case x of y { (a,b) -> k y a }
218 -- If we just take scrut_demand = U(L,A), then we won't pass x to the
219 -- worker, so the worker will rebuild
220 -- x = (a, absent-error)
221 -- and that'll crash.
222 -- So at one stage I had:
223 -- dead_case_bndr = isAbsentDmd (idNewDemandInfo case_bndr')
224 -- keepity | dead_case_bndr = Drop
225 -- | otherwise = Keep
228 -- case x of y { (a,b) -> h y + a }
229 -- where h : U(LL) -> T
230 -- The above code would compute a Keep for x, since y is not Abs, which is silly
231 -- The insight is, of course, that a demand on y is a demand on the
232 -- scrutinee, so we need to `both` it with the scrut demand
234 scrut_dmd = mkSeq Drop [idNewDemandInfo b | b <- bndrs', isId b]
236 idNewDemandInfo case_bndr'
238 (scrut_ty, scrut') = dmdAnal sigs scrut_dmd scrut
240 (alt_ty1 `bothType` scrut_ty, Case scrut' case_bndr' [alt'])
242 dmdAnal sigs dmd (Case scrut case_bndr alts)
244 (alt_tys, alts') = mapAndUnzip (dmdAnalAlt sigs dmd) alts
245 (scrut_ty, scrut') = dmdAnal sigs Eval scrut
246 (alt_ty, case_bndr') = annotateBndr (foldr1 lubType alt_tys) case_bndr
248 -- pprTrace "dmdAnal:Case" (ppr alts $$ ppr alt_tys)
249 (alt_ty `bothType` scrut_ty, Case scrut' case_bndr' alts')
251 dmdAnal sigs dmd (Let (NonRec id rhs) body)
253 (sigs', lazy_fv, (id1, rhs')) = dmdAnalRhs NotTopLevel sigs (id, rhs)
254 (body_ty, body') = dmdAnal sigs' dmd body
255 (body_ty1, id2) = annotateBndr body_ty id1
256 body_ty2 = addLazyFVs body_ty1 lazy_fv
258 -- pprTrace "dmdLet" (ppr id <+> ppr (sig,rhs_env))
259 (body_ty2, Let (NonRec id2 rhs') body')
261 dmdAnal sigs dmd (Let (Rec pairs) body)
263 bndrs = map fst pairs
264 (sigs', lazy_fv, pairs') = dmdFix NotTopLevel sigs pairs
265 (body_ty, body') = dmdAnal sigs' dmd body
266 body_ty1 = addLazyFVs body_ty lazy_fv
268 sigs' `seq` body_ty `seq`
270 (body_ty2, _) = annotateBndrs body_ty1 bndrs
271 -- Don't bother to add demand info to recursive
272 -- binders as annotateBndr does;
273 -- being recursive, we can't treat them strictly.
274 -- But we do need to remove the binders from the result demand env
276 (body_ty2, Let (Rec pairs') body')
279 dmdAnalAlt sigs dmd (con,bndrs,rhs)
281 (rhs_ty, rhs') = dmdAnal sigs dmd rhs
282 (alt_ty, bndrs') = annotateBndrs rhs_ty bndrs
284 (alt_ty, (con, bndrs', rhs'))
287 %************************************************************************
289 \subsection{Bindings}
291 %************************************************************************
294 dmdFix :: TopLevelFlag
295 -> SigEnv -- Does not include bindings for this binding
298 [(Id,CoreExpr)]) -- Binders annotated with stricness info
300 dmdFix top_lvl sigs orig_pairs
301 = loop 1 initial_sigs orig_pairs
303 bndrs = map fst orig_pairs
304 initial_sigs = extendSigEnvList sigs [(id, (initial_sig id, top_lvl)) | id <- bndrs]
307 -> SigEnv -- Already contains the current sigs
309 -> (SigEnv, DmdEnv, [(Id,CoreExpr)])
311 | all (same_sig sigs sigs') bndrs
312 = (sigs', lazy_fv, pairs')
313 -- Note: use pairs', not pairs. pairs' is the result of
314 -- processing the RHSs with sigs (= sigs'), whereas pairs
315 -- is the result of processing the RHSs with the *previous*
316 -- iteration of sigs.
317 | n >= 10 = pprTrace "dmdFix loop" (ppr n <+> (vcat
318 [ text "Sigs:" <+> ppr [(id,lookup sigs id, lookup sigs' id) | (id,_) <- pairs],
319 text "env:" <+> ppr (ufmToList sigs),
320 text "binds:" <+> pprCoreBinding (Rec pairs)]))
321 (emptySigEnv, emptyDmdEnv, orig_pairs) -- Safe output
322 | otherwise = loop (n+1) sigs' pairs'
324 -- Use the new signature to do the next pair
325 -- The occurrence analyser has arranged them in a good order
326 -- so this can significantly reduce the number of iterations needed
327 ((sigs',lazy_fv), pairs') = mapAccumL (my_downRhs top_lvl) (sigs, emptyDmdEnv) pairs
329 my_downRhs top_lvl (sigs,lazy_fv) (id,rhs)
330 = -- pprTrace "downRhs {" (ppr id <+> (ppr old_sig))
332 -- pprTrace "downRhsEnd" (ppr id <+> ppr new_sig <+> char '}' )
333 ((sigs', lazy_fv'), pair')
336 (sigs', lazy_fv1, pair') = dmdAnalRhs top_lvl sigs (id,rhs)
337 lazy_fv' = plusUFM_C both lazy_fv lazy_fv1
338 -- old_sig = lookup sigs id
339 -- new_sig = lookup sigs' id
341 -- Get an initial strictness signature from the Id
342 -- itself. That way we make use of earlier iterations
343 -- of the fixpoint algorithm. (Cunning plan.)
344 -- Note that the cunning plan extends to the DmdEnv too,
345 -- since it is part of the strictness signature
346 initial_sig id = idNewStrictness_maybe id `orElse` botSig
348 same_sig sigs sigs' var = lookup sigs var == lookup sigs' var
349 lookup sigs var = case lookupVarEnv sigs var of
352 dmdAnalRhs :: TopLevelFlag
353 -> SigEnv -> (Id, CoreExpr)
354 -> (SigEnv, DmdEnv, (Id, CoreExpr))
355 -- Process the RHS of the binding, add the strictness signature
356 -- to the Id, and augment the environment with the signature as well.
358 dmdAnalRhs top_lvl sigs (id, rhs)
359 = (sigs', lazy_fv, (id', rhs'))
361 arity = idArity id -- The idArity should be up to date
362 -- The simplifier was run just beforehand
363 (rhs_dmd_ty, rhs') = dmdAnal sigs (vanillaCall arity) rhs
364 (lazy_fv, sig_ty) = WARN( arity /= dmdTypeDepth rhs_dmd_ty, ppr id )
365 mkSigTy id rhs rhs_dmd_ty
366 id' = id `setIdNewStrictness` sig_ty
367 sigs' = extendSigEnv top_lvl sigs id sig_ty
370 %************************************************************************
372 \subsection{Strictness signatures and types}
374 %************************************************************************
377 mkTopSigTy :: CoreExpr -> DmdType -> StrictSig
378 -- Take a DmdType and turn it into a StrictSig
379 -- NB: not used for never-inline things; hence False
380 mkTopSigTy rhs dmd_ty = snd (mk_sig_ty False False rhs dmd_ty)
382 mkSigTy :: Id -> CoreExpr -> DmdType -> (DmdEnv, StrictSig)
383 mkSigTy id rhs dmd_ty = mk_sig_ty (isNeverActive (idInlinePragma id))
384 (isStrictDmd (idNewDemandInfo id))
387 mk_sig_ty never_inline strictly_demanded rhs (DmdType fv dmds res)
388 | never_inline && not (isBotRes res)
390 -- Don't strictness-analyse NOINLINE things. Why not? Because
391 -- the NOINLINE says "don't expose any of the inner workings at the call
392 -- site" and the strictness is certainly an inner working.
394 -- More concretely, the demand analyser discovers the following strictness
395 -- for unsafePerformIO: C(U(AV))
397 -- unsafePerformIO (\s -> let r = f x in
398 -- case writeIORef v r s of (# s1, _ #) ->
400 -- The strictness analyser will find that the binding for r is strict,
401 -- (becuase of uPIO's strictness sig), and so it'll evaluate it before
402 -- doing the writeIORef. This actually makes tests/lib/should_run/memo002
405 -- Solution: don't expose the strictness of unsafePerformIO.
407 -- But we do want to expose the strictness of error functions,
408 -- which are also often marked NOINLINE
409 -- {-# NOINLINE foo #-}
410 -- foo x = error ("wubble buggle" ++ x)
411 -- So (hack, hack) we only drop the strictness for non-bottom things
412 -- This is all very unsatisfactory.
413 = (deferEnv fv, topSig)
416 = (lazy_fv, mkStrictSig dmd_ty)
418 dmd_ty = DmdType strict_fv final_dmds res'
420 lazy_fv = filterUFM (not . isStrictDmd) fv
421 strict_fv = filterUFM isStrictDmd fv
422 -- We put the strict FVs in the DmdType of the Id, so
423 -- that at its call sites we unleash demands on its strict fvs.
424 -- An example is 'roll' in imaginary/wheel-sieve2
425 -- Something like this:
427 -- go y = if ... then roll (x-1) else x+1
430 -- We want to see that roll is strict in x, which is because
431 -- go is called. So we put the DmdEnv for x in go's DmdType.
434 -- f :: Int -> Int -> Int
435 -- f x y = let t = x+1
436 -- h z = if z==0 then t else
437 -- if z==1 then x+1 else
441 -- Calling h does indeed evaluate x, but we can only see
442 -- that if we unleash a demand on x at the call site for t.
444 -- Incidentally, here's a place where lambda-lifting h would
445 -- lose the cigar --- we couldn't see the joint strictness in t/x
448 -- We don't want to put *all* the fv's from the RHS into the
449 -- DmdType, because that makes fixpointing very slow --- the
450 -- DmdType gets full of lazy demands that are slow to converge.
452 lazified_dmds = map funArgDemand dmds
453 -- Get rid of defers in the arguments
454 final_dmds = setUnpackStrategy lazified_dmds
455 -- Set the unpacking strategy
458 RetCPR | ignore_cpr_info -> TopRes
460 ignore_cpr_info = is_thunk && not strictly_demanded
461 is_thunk = not (exprIsValue rhs)
462 -- If the rhs is a thunk, we forget the CPR info, because
463 -- it is presumably shared (else it would have been inlined, and
464 -- so we'd lose sharing if w/w'd it into a function.
466 -- Also, if the strictness analyser has figured out (in a previous iteration)
467 -- that it's strict, the let-to-case transformation will happen, so again
469 -- This made a big difference to PrelBase.modInt, which had something like
470 -- modInt = \ x -> let r = ... -> I# v in
471 -- ...body strict in r...
472 -- r's RHS isn't a value yet; but modInt returns r in various branches, so
473 -- if r doesn't have the CPR property then neither does modInt
474 -- Another case I found in practice (in Complex.magnitude), looks like this:
475 -- let k = if ... then I# a else I# b
476 -- in ... body strict in k ....
477 -- (For this example, it doesn't matter whether k is returned as part of
478 -- the overall result.) Left to itself, the simplifier will make a join
480 -- let $j k = ...body strict in k...
481 -- if ... then $j (I# a) else $j (I# b)
484 -- The difficulty with this is that we need the strictness type to
485 -- look at the body... but we now need the body to calculate the demand
486 -- on the variable, so we can decide whether its strictness type should
487 -- have a CPR in it or not. Simple solution:
488 -- a) use strictness info from the previous iteration
489 -- b) make sure we do at least 2 iterations, by doing a second
490 -- round for top-level non-recs. Top level recs will get at
491 -- least 2 iterations except for totally-bottom functions
492 -- which aren't very interesting anyway.
494 -- NB: strictly_demanded is never true of a top-level Id, or of a recursive Id.
497 The unpack strategy determines whether we'll *really* unpack the argument,
498 or whether we'll just remember its strictness. If unpacking would give
499 rise to a *lot* of worker args, we may decide not to unpack after all.
502 setUnpackStrategy :: [Demand] -> [Demand]
504 = snd (go (opt_MaxWorkerArgs - nonAbsentArgs ds) ds)
506 go :: Int -- Max number of args available for sub-components of [Demand]
508 -> (Int, [Demand]) -- Args remaining after subcomponents of [Demand] are unpacked
510 go n (Seq keep cs : ds)
511 | n' >= 0 = Seq keep cs' `cons` go n'' ds
512 | otherwise = Eval `cons` go n ds
515 n' = n + box - non_abs_args
518 Drop -> 1 -- Add one to the budget if we drop the top-level arg
519 non_abs_args = nonAbsentArgs cs
520 -- Delete # of non-absent args to which we'll now be committed
522 go n (d:ds) = d `cons` go n ds
525 cons d (n,ds) = (n, d:ds)
527 nonAbsentArgs :: [Demand] -> Int
529 nonAbsentArgs (Abs : ds) = nonAbsentArgs ds
530 nonAbsentArgs (d : ds) = 1 + nonAbsentArgs ds
534 %************************************************************************
536 \subsection{Strictness signatures and types}
538 %************************************************************************
541 splitDmdTy :: DmdType -> (Demand, DmdType)
542 -- Split off one function argument
543 -- We already have a suitable demand on all
544 -- free vars, so no need to add more!
545 splitDmdTy (DmdType fv (dmd:dmds) res_ty) = (dmd, DmdType fv dmds res_ty)
546 splitDmdTy ty@(DmdType fv [] TopRes) = (Lazy, ty)
547 splitDmdTy ty@(DmdType fv [] BotRes) = (Bot, ty)
549 splitDmdTy ty@(DmdType fv [] RetCPR) = panic "splitDmdTy"
550 -- We should not be applying a product as a function!
554 unitVarDmd var dmd = DmdType (unitVarEnv var dmd) [] TopRes
556 addVarDmd top_lvl dmd_ty@(DmdType fv ds res) var dmd
557 | isTopLevel top_lvl = dmd_ty -- Don't record top level things
558 | otherwise = DmdType (extendVarEnv fv var dmd) ds res
560 addLazyFVs (DmdType fv ds res) lazy_fvs
561 = DmdType both_fv1 ds res
563 both_fv = (plusUFM_C both fv lazy_fvs)
564 both_fv1 = modifyEnv (isBotRes res) (`both` Bot) lazy_fvs fv both_fv
565 -- This modifyEnv is vital. Consider
566 -- let f = \x -> (x,y)
568 -- Here, y is treated as a lazy-fv of f, but we must `both` that L
569 -- demand with the bottom coming up from 'error'
571 -- I got a loop in the fixpointer without this, due to an interaction
572 -- with the lazy_fv filtering in mkSigTy. Roughly, it was
574 -- = letrec g y = x `fatbar`
575 -- letrec h z = z + ...g...
578 -- In the initial iteration for f, f=Bot
579 -- Suppose h is found to be strict in z, but the occurrence of g in its RHS
580 -- is lazy. Now consider the fixpoint iteration for g, esp the demands it
581 -- places on its free variables. Suppose it places none. Then the
582 -- x `fatbar` ...call to h...
583 -- will give a x->V demand for x. That turns into a L demand for x,
584 -- which floats out of the defn for h. Without the modifyEnv, that
585 -- L demand doesn't get both'd with the Bot coming up from the inner
586 -- call to f. So we just get an L demand for x for g.
588 -- A better way to say this is that the lazy-fv filtering should give the
589 -- same answer as putting the lazy fv demands in the function's type.
591 annotateBndr :: DmdType -> Var -> (DmdType, Var)
592 -- The returned env has the var deleted
593 -- The returned var is annotated with demand info
594 -- No effect on the argument demands
595 annotateBndr dmd_ty@(DmdType fv ds res) var
596 | isTyVar var = (dmd_ty, var)
597 | otherwise = (DmdType fv' ds res, setIdNewDemandInfo var hacked_dmd)
599 (fv', dmd) = removeFV fv var res
600 hacked_dmd | isUnLiftedType (idType var) = unliftedDemand dmd
603 annotateBndrs = mapAccumR annotateBndr
605 annotateLamIdBndr dmd_ty@(DmdType fv ds res) id
606 -- For lambdas we add the demand to the argument demands
607 -- Only called for Ids
609 (DmdType fv' (hacked_dmd:ds) res, setIdNewDemandInfo id hacked_dmd)
611 (fv', dmd) = removeFV fv id res
612 hacked_dmd | isUnLiftedType (idType id) = unliftedDemand dmd
613 | otherwise = funArgDemand dmd
614 -- This call to funArgDemand is vital, because otherwise we label
615 -- a lambda binder with demand 'B'. But in terms of calling
616 -- conventions that's Abs, because we don't pass it. But
617 -- when we do a w/w split we get
618 -- fw x = (\x y:B -> ...) x (error "oops")
619 -- And then the simplifier things the 'B' is a strict demand
620 -- and evaluates the (error "oops"). Sigh
622 removeFV fv var res = (fv', dmd)
624 fv' = fv `delVarEnv` var
625 dmd = lookupVarEnv fv var `orElse` deflt
626 deflt | isBotRes res = Bot
630 %************************************************************************
632 \subsection{Strictness signatures}
634 %************************************************************************
637 type SigEnv = VarEnv (StrictSig, TopLevelFlag)
638 -- We use the SigEnv to tell us whether to
639 -- record info about a variable in the DmdEnv
640 -- We do so if it's a LocalId, but not top-level
642 -- The DmdEnv gives the demand on the free vars of the function
643 -- when it is given enough args to satisfy the strictness signature
645 emptySigEnv = emptyVarEnv
647 extendSigEnv :: TopLevelFlag -> SigEnv -> Id -> StrictSig -> SigEnv
648 extendSigEnv top_lvl env var sig = extendVarEnv env var (sig, top_lvl)
650 extendSigEnvList = extendVarEnvList
652 dmdTransform :: SigEnv -- The strictness environment
653 -> Id -- The function
654 -> Demand -- The demand on the function
655 -> DmdType -- The demand type of the function in this context
656 -- Returned DmdEnv includes the demand on
657 -- this function plus demand on its free variables
659 dmdTransform sigs var dmd
661 ------ DATA CONSTRUCTOR
662 | isDataConId var, -- Data constructor
663 Seq k ds <- res_dmd -- and the demand looks inside its fields
665 StrictSig dmd_ty = idNewStrictness var -- It must have a strictness sig
666 DmdType _ _ con_res = dmd_ty
669 if arity == call_depth then -- Saturated, so unleash the demand
671 -- ds can be empty, when we are just seq'ing the thing
672 -- If so we must make up a suitable bunch of demands
673 dmd_ds | null ds = replicate arity Abs
674 | otherwise = ASSERT( length ds == arity ) ds
677 Keep -> bothLazy_s dmd_ds
679 Defer -> pprTrace "dmdTransform: surprising!" (ppr var)
680 -- I don't think this can happen
682 -- Important! If we Keep the constructor application, then
683 -- we need the demands the constructor places (always lazy)
684 -- If not, we don't need to. For example:
685 -- f p@(x,y) = (p,y) -- S(AL)
687 -- It's vital that we don't calculate Absent for a!
689 mkDmdType emptyDmdEnv arg_ds con_res
690 -- Must remember whether it's a product, hence con_res, not TopRes
694 ------ IMPORTED FUNCTION
695 | isGlobalId var, -- Imported function
696 let StrictSig dmd_ty = getNewStrictness var
697 = if dmdTypeDepth dmd_ty <= call_depth then -- Saturated, so unleash the demand
702 ------ LOCAL LET/REC BOUND THING
703 | Just (StrictSig dmd_ty, top_lvl) <- lookupVarEnv sigs var
705 fn_ty | dmdTypeDepth dmd_ty <= call_depth = dmd_ty
706 | otherwise = deferType dmd_ty
707 -- NB: it's important to use deferType, and not just return topDmdType
708 -- Consider let { f x y = p + x } in f 1
709 -- The application isn't saturated, but we must nevertheless propagate
710 -- a lazy demand for p!
712 addVarDmd top_lvl fn_ty var dmd
714 ------ LOCAL NON-LET/REC BOUND THING
715 | otherwise -- Default case
719 (call_depth, res_dmd) = splitCallDmd dmd
723 %************************************************************************
727 %************************************************************************
730 splitCallDmd :: Demand -> (Int, Demand)
731 splitCallDmd (Call d) = case splitCallDmd d of
733 splitCallDmd d = (0, d)
735 vanillaCall :: Arity -> Demand
737 vanillaCall n = Call (vanillaCall (n-1))
739 deferType :: DmdType -> DmdType
740 deferType (DmdType fv _ _) = DmdType (deferEnv fv) [] TopRes
741 -- Notice that we throw away info about both arguments and results
742 -- For example, f = let ... in \x -> x
743 -- We don't want to get a stricness type V->T for f.
746 deferEnv :: DmdEnv -> DmdEnv
747 deferEnv fv = mapVarEnv defer fv
750 bothLazy :: Demand -> Demand
752 bothLazy_s :: [Demand] -> [Demand]
753 bothLazy_s = map bothLazy
755 funArgDemand :: Demand -> Demand
756 -- The 'Defer' demands are just Lazy at function boundaries
757 -- Ugly! Ask John how to improve it.
758 funArgDemand (Seq Defer ds) = Lazy
759 funArgDemand (Seq k ds) = Seq k (map funArgDemand ds)
760 funArgDemand Err = Eval -- Args passed to a bottoming function
761 funArgDemand Bot = Abs -- Don't pass args that are consumed by bottom/err
764 unliftedDemand :: Demand -> Demand
765 -- Same idea, but for unlifted types the domain is much simpler:
766 -- Either we use it (Lazy) or we don't (Abs)
767 unliftedDemand Bot = Abs
768 unliftedDemand Abs = Abs
769 unliftedDemand other = Lazy
773 betterStrictness :: StrictSig -> StrictSig -> Bool
774 betterStrictness (StrictSig t1) (StrictSig t2) = betterDmdType t1 t2
776 betterDmdType t1 t2 = (t1 `lubType` t2) == t2
778 betterDemand :: Demand -> Demand -> Bool
779 -- If d1 `better` d2, and d2 `better` d2, then d1==d2
780 betterDemand d1 d2 = (d1 `lub` d2) == d2
782 squashDmdEnv (StrictSig (DmdType fv ds res)) = StrictSig (DmdType emptyDmdEnv ds res)
786 -------------------------
787 -- Consider (if x then y else []) with demand V
788 -- Then the first branch gives {y->V} and the second
789 -- *implicitly* has {y->A}. So we must put {y->(V `lub` A)}
790 -- in the result env.
791 lubType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
792 = DmdType lub_fv2 (zipWith lub ds1 ds2) (r1 `lubRes` r2)
794 lub_fv = plusUFM_C lub fv1 fv2
795 lub_fv1 = modifyEnv (not (isBotRes r1)) defer fv2 fv1 lub_fv
796 lub_fv2 = modifyEnv (not (isBotRes r2)) defer fv1 fv2 lub_fv1
797 -- lub is the identity for Bot
799 -----------------------------------
800 -- (t1 `bothType` t2) takes the argument/result info from t1,
801 -- using t2 just for its free-var info
802 -- NB: Don't forget about r2! It might be BotRes, which is
803 -- a bottom demand on all the in-scope variables.
804 -- Peter: can this be done more neatly?
805 bothType (DmdType fv1 ds1 r1) (DmdType fv2 ds2 r2)
806 = DmdType both_fv2 ds1 (r1 `bothRes` r2)
808 both_fv = plusUFM_C both fv1 fv2
809 both_fv1 = modifyEnv (isBotRes r1) (`both` Bot) fv2 fv1 both_fv
810 both_fv2 = modifyEnv (isBotRes r2) (`both` Bot) fv1 fv2 both_fv1
811 -- both is the identity for Abs
818 lubRes RetCPR RetCPR = RetCPR
819 lubRes r1 r2 = TopRes
821 -- If either diverges, the whole thing does
822 -- Otherwise take CPR info from the first
823 bothRes r1 BotRes = BotRes
828 -- A Seq can have an empty list of demands, in the polymorphic case.
831 lubs ds1 ds2 = ASSERT( length ds1 == length ds2 ) zipWith lub ds1 ds2
833 -----------------------------------
834 -- A Seq can have an empty list of demands, in the polymorphic case.
837 boths ds1 ds2 = ASSERT( length ds1 == length ds2 ) zipWith both ds1 ds2
841 modifyEnv :: Bool -- No-op if False
842 -> (Demand -> Demand) -- The zapper
843 -> DmdEnv -> DmdEnv -- Env1 and Env2
844 -> DmdEnv -> DmdEnv -- Transform this env
845 -- Zap anything in Env1 but not in Env2
846 -- Assume: dom(env) includes dom(Env1) and dom(Env2)
848 modifyEnv need_to_modify zapper env1 env2 env
849 | need_to_modify = foldr zap env (keysUFM (env1 `minusUFM` env2))
852 zap uniq env = addToUFM_Directly env uniq (zapper current_val)
854 current_val = expectJust "modifyEnv" (lookupUFM_Directly env uniq)
858 %************************************************************************
860 \subsection{LUB and BOTH}
862 %************************************************************************
866 lub :: Demand -> Demand -> Demand
871 lub Err Abs = Lazy -- E.g. f x = if ... then True else error x
880 lub Eval (Seq Drop ds) | not (null ds) = Seq Drop [Lazy | d <- ds]
882 -- For the Seq case, consier
884 -- f n (x:xs) = f (n+x) xs
885 -- Here we want to do better than just V for n. It's
886 -- unboxed in the (x:xs) case, and we might be prepared to
887 -- rebox it in the [] case.
888 -- But if we don't use *any* of the components, give up
891 lub (Call d1) (Call d2) = Call (lub d1 d2)
892 lub d1@(Call _) d2 = d2 `lub` d1
894 lub (Seq k1 ds1) (Seq k2 ds2)
895 = Seq (k1 `lub_keep` k2) (lub_ds k1 ds1 k2 ds2)
898 lub_ds Keep ds1 Keep ds2 = ds1 `lubs` ds2
899 lub_ds Keep ds1 non_keep ds2 | null ds1 = [Lazy | d <- ds2]
900 | otherwise = bothLazy_s ds1 `lubs` ds2
902 lub_ds non_keep ds1 Keep ds2 | null ds2 = [Lazy | d <- ds1]
903 | otherwise = ds1 `lubs` bothLazy_s ds2
905 lub_ds k1 ds1 k2 ds2 = ds1 `lubs` ds2
910 lub_keep Drop Defer = Defer
911 lub_keep Drop k = Drop
913 lub_keep Defer k = Defer
915 lub d1@(Seq _ _) d2 = d2 `lub` d1
918 both :: Demand -> Demand -> Demand
930 both Lazy Eval = Eval
931 both Lazy (Call d) = Call d
932 both Lazy (Seq Defer ds) = Lazy
933 both Lazy (Seq k ds) = Seq Keep ds
936 -- For the (Eval `both` Bot) case, consider
938 -- From 'error' itself we get demand Bot on x
939 -- From the arg demand on x we get Eval
940 -- So we want Eval `both` Bot to be Err.
941 -- That's what Err is *for*
944 both Eval (Seq k ds) = Seq Keep ds
947 both (Call d1) (Call d2) = Call (d1 `both` d2)
948 both d1@(Call _) d2 = d2 `both` d1
950 both (Seq k1 ds1) (Seq k2 ds2)
951 = Seq (k1 `both_keep` k2) (both_ds k1 ds1 k2 ds2)
954 both_keep Keep k2 = Keep
956 both_keep Drop Keep = Keep
957 both_keep Drop k2 = Drop
959 both_keep Defer k2 = k2
962 both_ds Defer ds1 Defer ds2 = ds1 `boths` ds2
963 both_ds Defer ds1 non_defer ds2 = map defer ds1 `boths` ds2
965 both_ds non_defer ds1 Defer ds2 = ds1 `boths` map defer ds2
967 both_ds k1 ds1 k2 ds2 = ds1 `boths` ds2
969 both d1@(Seq _ _) d2 = d2 `both` d1
973 %************************************************************************
975 \subsection{Miscellaneous
977 %************************************************************************
981 get_changes binds = vcat (map get_changes_bind binds)
983 get_changes_bind (Rec pairs) = vcat (map get_changes_pr pairs)
984 get_changes_bind (NonRec id rhs) = get_changes_pr (id,rhs)
986 get_changes_pr (id,rhs)
987 = get_changes_var id $$ get_changes_expr rhs
990 | isId var = get_changes_str var $$ get_changes_dmd var
993 get_changes_expr (Type t) = empty
994 get_changes_expr (Var v) = empty
995 get_changes_expr (Lit l) = empty
996 get_changes_expr (Note n e) = get_changes_expr e
997 get_changes_expr (App e1 e2) = get_changes_expr e1 $$ get_changes_expr e2
998 get_changes_expr (Lam b e) = {- get_changes_var b $$ -} get_changes_expr e
999 get_changes_expr (Let b e) = get_changes_bind b $$ get_changes_expr e
1000 get_changes_expr (Case e b a) = get_changes_expr e $$ {- get_changes_var b $$ -} vcat (map get_changes_alt a)
1002 get_changes_alt (con,bs,rhs) = {- vcat (map get_changes_var bs) $$ -} get_changes_expr rhs
1005 | new_better && old_better = empty
1006 | new_better = message "BETTER"
1007 | old_better = message "WORSE"
1008 | otherwise = message "INCOMPARABLE"
1010 message word = text word <+> text "strictness for" <+> ppr id <+> info
1011 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1012 new = squashDmdEnv (idNewStrictness id) -- Don't report diffs in the env
1013 old = newStrictnessFromOld id
1014 old_better = old `betterStrictness` new
1015 new_better = new `betterStrictness` old
1018 | isUnLiftedType (idType id) = empty -- Not useful
1019 | new_better && old_better = empty
1020 | new_better = message "BETTER"
1021 | old_better = message "WORSE"
1022 | otherwise = message "INCOMPARABLE"
1024 message word = text word <+> text "demand for" <+> ppr id <+> info
1025 info = (text "Old" <+> ppr old) $$ (text "New" <+> ppr new)
1026 new = funArgDemand (idNewDemandInfo id) -- FunArgDemand to avoid spurious improvements
1027 old = newDemand (idDemandInfo id)
1028 new_better = new `betterDemand` old
1029 old_better = old `betterDemand` new