2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[SpecConstr]{Specialise over constructors}
7 -- The above warning supression flag is a temporary kludge.
8 -- While working on this module you are encouraged to remove it and fix
9 -- any warnings in the module. See
10 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
14 specConstrProgram, SpecConstrAnnotation(..)
17 #include "HsVersions.h"
22 import CoreUnfold ( couldBeSmallEnoughToInline )
23 import CoreFVs ( exprsFreeVars )
25 import HscTypes ( ModGuts(..) )
26 import WwLib ( mkWorkerArgs )
27 import DataCon ( dataConTyCon, dataConRepArity, dataConUnivTyVars )
28 import TyCon ( TyCon )
29 import Literal ( literalType )
32 import Type hiding( substTy )
34 import MkId ( mkImpossibleExpr )
39 import DynFlags ( DynFlags(..) )
40 import StaticFlags ( opt_PprStyle_Debug )
41 import Maybes ( orElse, catMaybes, isJust, isNothing )
43 import DmdAnal ( both )
44 import Serialized ( deserializeWithData )
50 import qualified LazyUniqFM as L
52 import Control.Monad ( zipWithM )
54 #if __GLASGOW_HASKELL__ > 609
55 import Data.Data ( Data, Typeable )
57 import Data.Generics ( Data, Typeable )
61 -----------------------------------------------------
63 -----------------------------------------------------
68 drop n (x:xs) = drop (n-1) xs
70 After the first time round, we could pass n unboxed. This happens in
71 numerical code too. Here's what it looks like in Core:
73 drop n xs = case xs of
78 _ -> drop (I# (n# -# 1#)) xs
80 Notice that the recursive call has an explicit constructor as argument.
81 Noticing this, we can make a specialised version of drop
83 RULE: drop (I# n#) xs ==> drop' n# xs
85 drop' n# xs = let n = I# n# in ...orig RHS...
87 Now the simplifier will apply the specialisation in the rhs of drop', giving
89 drop' n# xs = case xs of
93 _ -> drop (n# -# 1#) xs
97 We'd also like to catch cases where a parameter is carried along unchanged,
98 but evaluated each time round the loop:
100 f i n = if i>0 || i>n then i else f (i*2) n
102 Here f isn't strict in n, but we'd like to avoid evaluating it each iteration.
103 In Core, by the time we've w/wd (f is strict in i) we get
105 f i# n = case i# ># 0 of
107 True -> case n of n' { I# n# ->
110 True -> f (i# *# 2#) n'
112 At the call to f, we see that the argument, n is know to be (I# n#),
113 and n is evaluated elsewhere in the body of f, so we can play the same
119 We must be careful not to allocate the same constructor twice. Consider
120 f p = (...(case p of (a,b) -> e)...p...,
121 ...let t = (r,s) in ...t...(f t)...)
122 At the recursive call to f, we can see that t is a pair. But we do NOT want
123 to make a specialised copy:
124 f' a b = let p = (a,b) in (..., ...)
125 because now t is allocated by the caller, then r and s are passed to the
126 recursive call, which allocates the (r,s) pair again.
129 (a) the argument p is used in other than a case-scrutinsation way.
130 (b) the argument to the call is not a 'fresh' tuple; you have to
131 look into its unfolding to see that it's a tuple
133 Hence the "OR" part of Note [Good arguments] below.
135 ALTERNATIVE 2: pass both boxed and unboxed versions. This no longer saves
136 allocation, but does perhaps save evals. In the RULE we'd have
139 f (I# x#) = f' (I# x#) x#
141 If at the call site the (I# x) was an unfolding, then we'd have to
142 rely on CSE to eliminate the duplicate allocation.... This alternative
143 doesn't look attractive enough to pursue.
145 ALTERNATIVE 3: ignore the reboxing problem. The trouble is that
146 the conservative reboxing story prevents many useful functions from being
147 specialised. Example:
148 foo :: Maybe Int -> Int -> Int
150 foo x@(Just m) n = foo x (n-m)
151 Here the use of 'x' will clearly not require boxing in the specialised function.
153 The strictness analyser has the same problem, in fact. Example:
155 If we pass just 'a' and 'b' to the worker, it might need to rebox the
156 pair to create (a,b). A more sophisticated analysis might figure out
157 precisely the cases in which this could happen, but the strictness
158 analyser does no such analysis; it just passes 'a' and 'b', and hopes
161 So my current choice is to make SpecConstr similarly aggressive, and
162 ignore the bad potential of reboxing.
165 Note [Good arguments]
166 ~~~~~~~~~~~~~~~~~~~~~
169 * A self-recursive function. Ignore mutual recursion for now,
170 because it's less common, and the code is simpler for self-recursion.
174 a) At a recursive call, one or more parameters is an explicit
175 constructor application
177 That same parameter is scrutinised by a case somewhere in
178 the RHS of the function
182 b) At a recursive call, one or more parameters has an unfolding
183 that is an explicit constructor application
185 That same parameter is scrutinised by a case somewhere in
186 the RHS of the function
188 Those are the only uses of the parameter (see Note [Reboxing])
191 What to abstract over
192 ~~~~~~~~~~~~~~~~~~~~~
193 There's a bit of a complication with type arguments. If the call
196 f p = ...f ((:) [a] x xs)...
198 then our specialised function look like
200 f_spec x xs = let p = (:) [a] x xs in ....as before....
202 This only makes sense if either
203 a) the type variable 'a' is in scope at the top of f, or
204 b) the type variable 'a' is an argument to f (and hence fs)
206 Actually, (a) may hold for value arguments too, in which case
207 we may not want to pass them. Supose 'x' is in scope at f's
208 defn, but xs is not. Then we'd like
210 f_spec xs = let p = (:) [a] x xs in ....as before....
212 Similarly (b) may hold too. If x is already an argument at the
213 call, no need to pass it again.
215 Finally, if 'a' is not in scope at the call site, we could abstract
216 it as we do the term variables:
218 f_spec a x xs = let p = (:) [a] x xs in ...as before...
220 So the grand plan is:
222 * abstract the call site to a constructor-only pattern
223 e.g. C x (D (f p) (g q)) ==> C s1 (D s2 s3)
225 * Find the free variables of the abstracted pattern
227 * Pass these variables, less any that are in scope at
228 the fn defn. But see Note [Shadowing] below.
231 NOTICE that we only abstract over variables that are not in scope,
232 so we're in no danger of shadowing variables used in "higher up"
238 In this pass we gather up usage information that may mention variables
239 that are bound between the usage site and the definition site; or (more
240 seriously) may be bound to something different at the definition site.
243 f x = letrec g y v = let x = ...
246 Since 'x' is in scope at the call site, we may make a rewrite rule that
248 RULE forall a,b. g (a,b) x = ...
249 But this rule will never match, because it's really a different 'x' at
250 the call site -- and that difference will be manifest by the time the
251 simplifier gets to it. [A worry: the simplifier doesn't *guarantee*
252 no-shadowing, so perhaps it may not be distinct?]
254 Anyway, the rule isn't actually wrong, it's just not useful. One possibility
255 is to run deShadowBinds before running SpecConstr, but instead we run the
256 simplifier. That gives the simplest possible program for SpecConstr to
257 chew on; and it virtually guarantees no shadowing.
259 Note [Specialising for constant parameters]
260 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
261 This one is about specialising on a *constant* (but not necessarily
262 constructor) argument
264 foo :: Int -> (Int -> Int) -> Int
266 foo m f = foo (f m) (+1)
270 lvl_rmV :: GHC.Base.Int -> GHC.Base.Int
272 \ (ds_dlk :: GHC.Base.Int) ->
273 case ds_dlk of wild_alH { GHC.Base.I# x_alG ->
274 GHC.Base.I# (GHC.Prim.+# x_alG 1)
276 T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
279 \ (ww_sme :: GHC.Prim.Int#) (w_smg :: GHC.Base.Int -> GHC.Base.Int) ->
280 case ww_sme of ds_Xlw {
282 case w_smg (GHC.Base.I# ds_Xlw) of w1_Xmo { GHC.Base.I# ww1_Xmz ->
283 T.$wfoo ww1_Xmz lvl_rmV
288 The recursive call has lvl_rmV as its argument, so we could create a specialised copy
289 with that argument baked in; that is, not passed at all. Now it can perhaps be inlined.
291 When is this worth it? Call the constant 'lvl'
292 - If 'lvl' has an unfolding that is a constructor, see if the corresponding
293 parameter is scrutinised anywhere in the body.
295 - If 'lvl' has an unfolding that is a inlinable function, see if the corresponding
296 parameter is applied (...to enough arguments...?)
298 Also do this is if the function has RULES?
302 Note [Specialising for lambda parameters]
303 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
304 foo :: Int -> (Int -> Int) -> Int
306 foo m f = foo (f m) (\n -> n-m)
308 This is subtly different from the previous one in that we get an
309 explicit lambda as the argument:
311 T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
314 \ (ww_sm8 :: GHC.Prim.Int#) (w_sma :: GHC.Base.Int -> GHC.Base.Int) ->
315 case ww_sm8 of ds_Xlr {
317 case w_sma (GHC.Base.I# ds_Xlr) of w1_Xmf { GHC.Base.I# ww1_Xmq ->
320 (\ (n_ad3 :: GHC.Base.Int) ->
321 case n_ad3 of wild_alB { GHC.Base.I# x_alA ->
322 GHC.Base.I# (GHC.Prim.-# x_alA ds_Xlr)
328 I wonder if SpecConstr couldn't be extended to handle this? After all,
329 lambda is a sort of constructor for functions and perhaps it already
330 has most of the necessary machinery?
332 Furthermore, there's an immediate win, because you don't need to allocate the lamda
333 at the call site; and if perchance it's called in the recursive call, then you
334 may avoid allocating it altogether. Just like for constructors.
336 Looks cool, but probably rare...but it might be easy to implement.
339 Note [SpecConstr for casts]
340 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
343 data instance T Int = T Int
348 go (T n) = go (T (n-1))
350 The recursive call ends up looking like
351 go (T (I# ...) `cast` g)
352 So we want to spot the construtor application inside the cast.
353 That's why we have the Cast case in argToPat
355 Note [Local recursive groups]
356 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
357 For a *local* recursive group, we can see all the calls to the
358 function, so we seed the specialisation loop from the calls in the
359 body, not from the calls in the RHS. Consider:
361 bar m n = foo n (n,n) (n,n) (n,n) (n,n)
365 | n > 3000 = case p of { (p1,p2) -> foo (n-1) (p2,p1) q r s }
366 | n > 2000 = case q of { (q1,q2) -> foo (n-1) p (q2,q1) r s }
367 | n > 1000 = case r of { (r1,r2) -> foo (n-1) p q (r2,r1) s }
368 | otherwise = case s of { (s1,s2) -> foo (n-1) p q r (s2,s1) }
370 If we start with the RHSs of 'foo', we get lots and lots of specialisations,
371 most of which are not needed. But if we start with the (single) call
372 in the rhs of 'bar' we get exactly one fully-specialised copy, and all
373 the recursive calls go to this fully-specialised copy. Indeed, the original
374 function is later collected as dead code. This is very important in
375 specialising the loops arising from stream fusion, for example in NDP where
376 we were getting literally hundreds of (mostly unused) specialisations of
379 Note [Do not specialise diverging functions]
380 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
381 Specialising a function that just diverges is a waste of code.
382 Furthermore, it broke GHC (simpl014) thus:
384 f = \x. case x of (a,b) -> f x
385 If we specialise f we get
386 f = \x. case x of (a,b) -> fspec a b
387 But fspec doesn't have decent strictnes info. As it happened,
388 (f x) :: IO t, so the state hack applied and we eta expanded fspec,
389 and hence f. But now f's strictness is less than its arity, which
392 -----------------------------------------------------
393 Stuff not yet handled
394 -----------------------------------------------------
396 Here are notes arising from Roman's work that I don't want to lose.
402 foo :: Int -> T Int -> Int
404 foo x t | even x = case t of { T n -> foo (x-n) t }
405 | otherwise = foo (x-1) t
407 SpecConstr does no specialisation, because the second recursive call
408 looks like a boxed use of the argument. A pity.
410 $wfoo_sFw :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
412 \ (ww_sFo [Just L] :: GHC.Prim.Int#) (w_sFq [Just L] :: T.T GHC.Base.Int) ->
413 case ww_sFo of ds_Xw6 [Just L] {
415 case GHC.Prim.remInt# ds_Xw6 2 of wild1_aEF [Dead Just A] {
416 __DEFAULT -> $wfoo_sFw (GHC.Prim.-# ds_Xw6 1) w_sFq;
418 case w_sFq of wild_Xy [Just L] { T.T n_ad5 [Just U(L)] ->
419 case n_ad5 of wild1_aET [Just A] { GHC.Base.I# y_aES [Just L] ->
420 $wfoo_sFw (GHC.Prim.-# ds_Xw6 y_aES) wild_Xy
426 data a :*: b = !a :*: !b
429 foo :: (Int :*: T Int) -> Int
431 foo (x :*: t) | even x = case t of { T n -> foo ((x-n) :*: t) }
432 | otherwise = foo ((x-1) :*: t)
434 Very similar to the previous one, except that the parameters are now in
435 a strict tuple. Before SpecConstr, we have
437 $wfoo_sG3 :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
439 \ (ww_sFU [Just L] :: GHC.Prim.Int#) (ww_sFW [Just L] :: T.T
441 case ww_sFU of ds_Xws [Just L] {
443 case GHC.Prim.remInt# ds_Xws 2 of wild1_aEZ [Dead Just A] {
445 case ww_sFW of tpl_B2 [Just L] { T.T a_sFo [Just A] ->
446 $wfoo_sG3 (GHC.Prim.-# ds_Xws 1) tpl_B2 -- $wfoo1
449 case ww_sFW of wild_XB [Just A] { T.T n_ad7 [Just S(L)] ->
450 case n_ad7 of wild1_aFd [Just L] { GHC.Base.I# y_aFc [Just L] ->
451 $wfoo_sG3 (GHC.Prim.-# ds_Xws y_aFc) wild_XB -- $wfoo2
455 We get two specialisations:
456 "SC:$wfoo1" [0] __forall {a_sFB :: GHC.Base.Int sc_sGC :: GHC.Prim.Int#}
457 Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int a_sFB)
458 = Foo.$s$wfoo1 a_sFB sc_sGC ;
459 "SC:$wfoo2" [0] __forall {y_aFp :: GHC.Prim.Int# sc_sGC :: GHC.Prim.Int#}
460 Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int (GHC.Base.I# y_aFp))
461 = Foo.$s$wfoo y_aFp sc_sGC ;
463 But perhaps the first one isn't good. After all, we know that tpl_B2 is
464 a T (I# x) really, because T is strict and Int has one constructor. (We can't
465 unbox the strict fields, becuase T is polymorphic!)
467 %************************************************************************
469 \subsection{Annotations}
471 %************************************************************************
473 Annotating a type with NoSpecConstr will make SpecConstr not specialise
474 for arguments of that type.
477 data SpecConstrAnnotation = NoSpecConstr | ForceSpecConstr
478 deriving( Data, Typeable, Eq )
481 %************************************************************************
483 \subsection{Top level wrapper stuff}
485 %************************************************************************
488 specConstrProgram :: ModGuts -> CoreM ModGuts
489 specConstrProgram guts
491 dflags <- getDynFlags
492 us <- getUniqueSupplyM
493 annos <- getFirstAnnotations deserializeWithData guts
494 let binds' = fst $ initUs us (go (initScEnv dflags annos) (mg_binds guts))
495 return (guts { mg_binds = binds' })
498 go env (bind:binds) = do (env', bind') <- scTopBind env bind
499 binds' <- go env' binds
500 return (bind' : binds')
504 %************************************************************************
506 \subsection{Environment: goes downwards}
508 %************************************************************************
511 data ScEnv = SCE { sc_size :: Maybe Int, -- Size threshold
512 sc_count :: Maybe Int, -- Max # of specialisations for any one fn
513 -- See Note [Avoiding exponential blowup]
515 sc_subst :: Subst, -- Current substitution
516 -- Maps InIds to OutExprs
518 sc_how_bound :: HowBoundEnv,
519 -- Binds interesting non-top-level variables
520 -- Domain is OutVars (*after* applying the substitution)
523 -- Domain is OutIds (*after* applying the substitution)
524 -- Used even for top-level bindings (but not imported ones)
526 sc_annotations :: L.UniqFM SpecConstrAnnotation
529 ---------------------
530 -- As we go, we apply a substitution (sc_subst) to the current term
531 type InExpr = CoreExpr -- _Before_ applying the subst
534 type OutExpr = CoreExpr -- _After_ applying the subst
538 ---------------------
539 type HowBoundEnv = VarEnv HowBound -- Domain is OutVars
541 ---------------------
542 type ValueEnv = IdEnv Value -- Domain is OutIds
543 data Value = ConVal AltCon [CoreArg] -- _Saturated_ constructors
544 | LambdaVal -- Inlinable lambdas or PAPs
546 instance Outputable Value where
547 ppr (ConVal con args) = ppr con <+> interpp'SP args
548 ppr LambdaVal = ptext (sLit "<Lambda>")
550 ---------------------
551 initScEnv :: DynFlags -> L.UniqFM SpecConstrAnnotation -> ScEnv
552 initScEnv dflags anns
553 = SCE { sc_size = specConstrThreshold dflags,
554 sc_count = specConstrCount dflags,
555 sc_subst = emptySubst,
556 sc_how_bound = emptyVarEnv,
557 sc_vals = emptyVarEnv,
558 sc_annotations = anns }
560 data HowBound = RecFun -- These are the recursive functions for which
561 -- we seek interesting call patterns
563 | RecArg -- These are those functions' arguments, or their sub-components;
564 -- we gather occurrence information for these
566 instance Outputable HowBound where
567 ppr RecFun = text "RecFun"
568 ppr RecArg = text "RecArg"
570 lookupHowBound :: ScEnv -> Id -> Maybe HowBound
571 lookupHowBound env id = lookupVarEnv (sc_how_bound env) id
573 scSubstId :: ScEnv -> Id -> CoreExpr
574 scSubstId env v = lookupIdSubst (text "scSubstId") (sc_subst env) v
576 scSubstTy :: ScEnv -> Type -> Type
577 scSubstTy env ty = substTy (sc_subst env) ty
579 zapScSubst :: ScEnv -> ScEnv
580 zapScSubst env = env { sc_subst = zapSubstEnv (sc_subst env) }
582 extendScInScope :: ScEnv -> [Var] -> ScEnv
583 -- Bring the quantified variables into scope
584 extendScInScope env qvars = env { sc_subst = extendInScopeList (sc_subst env) qvars }
586 -- Extend the substitution
587 extendScSubst :: ScEnv -> Var -> OutExpr -> ScEnv
588 extendScSubst env var expr = env { sc_subst = extendSubst (sc_subst env) var expr }
590 extendScSubstList :: ScEnv -> [(Var,OutExpr)] -> ScEnv
591 extendScSubstList env prs = env { sc_subst = extendSubstList (sc_subst env) prs }
593 extendHowBound :: ScEnv -> [Var] -> HowBound -> ScEnv
594 extendHowBound env bndrs how_bound
595 = env { sc_how_bound = extendVarEnvList (sc_how_bound env)
596 [(bndr,how_bound) | bndr <- bndrs] }
598 extendBndrsWith :: HowBound -> ScEnv -> [Var] -> (ScEnv, [Var])
599 extendBndrsWith how_bound env bndrs
600 = (env { sc_subst = subst', sc_how_bound = hb_env' }, bndrs')
602 (subst', bndrs') = substBndrs (sc_subst env) bndrs
603 hb_env' = sc_how_bound env `extendVarEnvList`
604 [(bndr,how_bound) | bndr <- bndrs']
606 extendBndrWith :: HowBound -> ScEnv -> Var -> (ScEnv, Var)
607 extendBndrWith how_bound env bndr
608 = (env { sc_subst = subst', sc_how_bound = hb_env' }, bndr')
610 (subst', bndr') = substBndr (sc_subst env) bndr
611 hb_env' = extendVarEnv (sc_how_bound env) bndr' how_bound
613 extendRecBndrs :: ScEnv -> [Var] -> (ScEnv, [Var])
614 extendRecBndrs env bndrs = (env { sc_subst = subst' }, bndrs')
616 (subst', bndrs') = substRecBndrs (sc_subst env) bndrs
618 extendBndr :: ScEnv -> Var -> (ScEnv, Var)
619 extendBndr env bndr = (env { sc_subst = subst' }, bndr')
621 (subst', bndr') = substBndr (sc_subst env) bndr
623 extendValEnv :: ScEnv -> Id -> Maybe Value -> ScEnv
624 extendValEnv env _ Nothing = env
625 extendValEnv env id (Just cv) = env { sc_vals = extendVarEnv (sc_vals env) id cv }
627 extendCaseBndrs :: ScEnv -> Id -> AltCon -> [Var] -> (ScEnv, [Var])
631 -- we want to bind b, to (C x y)
632 -- NB1: Extends only the sc_vals part of the envt
633 -- NB2: Kill the dead-ness info on the pattern binders x,y, since
634 -- they are potentially made alive by the [b -> C x y] binding
635 extendCaseBndrs env case_bndr con alt_bndrs
636 | isDeadBinder case_bndr
639 = (env1, map zap alt_bndrs)
640 -- NB: We used to bind v too, if scrut = (Var v); but
641 -- the simplifer has already done this so it seems
642 -- redundant to do so here
644 -- Var v -> extendValEnv env1 v cval
647 zap v | isTyVar v = v -- See NB2 above
648 | otherwise = zapIdOccInfo v
649 env1 = extendValEnv env case_bndr cval
652 LitAlt {} -> Just (ConVal con [])
653 DataAlt {} -> Just (ConVal con vanilla_args)
655 vanilla_args = map Type (tyConAppArgs (idType case_bndr)) ++
656 varsToCoreExprs alt_bndrs
658 ignoreTyCon :: ScEnv -> TyCon -> Bool
659 ignoreTyCon env tycon
660 = L.lookupUFM (sc_annotations env) tycon == Just NoSpecConstr
662 ignoreType :: ScEnv -> Type -> Bool
664 = case splitTyConApp_maybe ty of
665 Just (tycon, _) -> ignoreTyCon env tycon
668 ignoreAltCon :: ScEnv -> AltCon -> Bool
669 ignoreAltCon env (DataAlt dc) = ignoreTyCon env (dataConTyCon dc)
670 ignoreAltCon env (LitAlt lit) = ignoreType env (literalType lit)
671 ignoreAltCon _ DEFAULT = True
673 forceSpecBndr :: ScEnv -> Var -> Bool
674 forceSpecBndr env var = forceSpecFunTy env . snd . splitForAllTys . varType $ var
676 forceSpecFunTy :: ScEnv -> Type -> Bool
677 forceSpecFunTy env = any (forceSpecArgTy env) . fst . splitFunTys
679 forceSpecArgTy :: ScEnv -> Type -> Bool
680 forceSpecArgTy env ty
681 | Just ty' <- coreView ty = forceSpecArgTy env ty'
683 forceSpecArgTy env ty
684 | Just (tycon, tys) <- splitTyConApp_maybe ty
686 = L.lookupUFM (sc_annotations env) tycon == Just ForceSpecConstr
687 || any (forceSpecArgTy env) tys
689 forceSpecArgTy _ _ = False
691 decreaseSpecCount :: ScEnv -> Int -> ScEnv
692 -- See Note [Avoiding exponential blowup]
693 decreaseSpecCount env n_specs
694 = env { sc_count = case sc_count env of
696 Just n -> Just (n `div` (n_specs + 1)) }
697 -- The "+1" takes account of the original function;
698 -- See Note [Avoiding exponential blowup]
701 Note [Avoiding exponential blowup]
702 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
703 The sc_count field of the ScEnv says how many times we are prepared to
704 duplicate a single function. But we must take care with recursive
705 specialiations. Consider
707 let $j1 = let $j2 = let $j3 = ...
715 If we specialise $j1 then in each specialisation (as well as the original)
716 we can specialise $j2, and similarly $j3. Even if we make just *one*
717 specialisation of each, becuase we also have the original we'll get 2^n
718 copies of $j3, which is not good.
720 So when recursively specialising we divide the sc_count by the number of
721 copies we are making at this level, including the original.
724 %************************************************************************
726 \subsection{Usage information: flows upwards}
728 %************************************************************************
733 scu_calls :: CallEnv, -- Calls
734 -- The functions are a subset of the
735 -- RecFuns in the ScEnv
737 scu_occs :: !(IdEnv ArgOcc) -- Information on argument occurrences
738 } -- The domain is OutIds
740 type CallEnv = IdEnv [Call]
741 type Call = (ValueEnv, [CoreArg])
742 -- The arguments of the call, together with the
743 -- env giving the constructor bindings at the call site
746 nullUsage = SCU { scu_calls = emptyVarEnv, scu_occs = emptyVarEnv }
748 combineCalls :: CallEnv -> CallEnv -> CallEnv
749 combineCalls = plusVarEnv_C (++)
751 combineUsage :: ScUsage -> ScUsage -> ScUsage
752 combineUsage u1 u2 = SCU { scu_calls = combineCalls (scu_calls u1) (scu_calls u2),
753 scu_occs = plusVarEnv_C combineOcc (scu_occs u1) (scu_occs u2) }
755 combineUsages :: [ScUsage] -> ScUsage
756 combineUsages [] = nullUsage
757 combineUsages us = foldr1 combineUsage us
759 lookupOcc :: ScUsage -> OutVar -> (ScUsage, ArgOcc)
760 lookupOcc (SCU { scu_calls = sc_calls, scu_occs = sc_occs }) bndr
761 = (SCU {scu_calls = sc_calls, scu_occs = delVarEnv sc_occs bndr},
762 lookupVarEnv sc_occs bndr `orElse` NoOcc)
764 lookupOccs :: ScUsage -> [OutVar] -> (ScUsage, [ArgOcc])
765 lookupOccs (SCU { scu_calls = sc_calls, scu_occs = sc_occs }) bndrs
766 = (SCU {scu_calls = sc_calls, scu_occs = delVarEnvList sc_occs bndrs},
767 [lookupVarEnv sc_occs b `orElse` NoOcc | b <- bndrs])
769 data ArgOcc = NoOcc -- Doesn't occur at all; or a type argument
770 | UnkOcc -- Used in some unknown way
772 | ScrutOcc (UniqFM [ArgOcc]) -- See Note [ScrutOcc]
774 | BothOcc -- Definitely taken apart, *and* perhaps used in some other way
778 An occurrence of ScrutOcc indicates that the thing, or a `cast` version of the thing,
779 is *only* taken apart or applied.
781 Functions, literal: ScrutOcc emptyUFM
782 Data constructors: ScrutOcc subs,
784 where (subs :: UniqFM [ArgOcc]) gives usage of the *pattern-bound* components,
785 The domain of the UniqFM is the Unique of the data constructor
787 The [ArgOcc] is the occurrences of the *pattern-bound* components
788 of the data structure. E.g.
789 data T a = forall b. MkT a b (b->a)
790 A pattern binds b, x::a, y::b, z::b->a, but not 'a'!
794 instance Outputable ArgOcc where
795 ppr (ScrutOcc xs) = ptext (sLit "scrut-occ") <> ppr xs
796 ppr UnkOcc = ptext (sLit "unk-occ")
797 ppr BothOcc = ptext (sLit "both-occ")
798 ppr NoOcc = ptext (sLit "no-occ")
800 -- Experimentally, this vesion of combineOcc makes ScrutOcc "win", so
801 -- that if the thing is scrutinised anywhere then we get to see that
802 -- in the overall result, even if it's also used in a boxed way
803 -- This might be too agressive; see Note [Reboxing] Alternative 3
804 combineOcc :: ArgOcc -> ArgOcc -> ArgOcc
805 combineOcc NoOcc occ = occ
806 combineOcc occ NoOcc = occ
807 combineOcc (ScrutOcc xs) (ScrutOcc ys) = ScrutOcc (plusUFM_C combineOccs xs ys)
808 combineOcc _occ (ScrutOcc ys) = ScrutOcc ys
809 combineOcc (ScrutOcc xs) _occ = ScrutOcc xs
810 combineOcc UnkOcc UnkOcc = UnkOcc
811 combineOcc _ _ = BothOcc
813 combineOccs :: [ArgOcc] -> [ArgOcc] -> [ArgOcc]
814 combineOccs xs ys = zipWithEqual "combineOccs" combineOcc xs ys
816 setScrutOcc :: ScEnv -> ScUsage -> OutExpr -> ArgOcc -> ScUsage
817 -- _Overwrite_ the occurrence info for the scrutinee, if the scrutinee
818 -- is a variable, and an interesting variable
819 setScrutOcc env usg (Cast e _) occ = setScrutOcc env usg e occ
820 setScrutOcc env usg (Note _ e) occ = setScrutOcc env usg e occ
821 setScrutOcc env usg (Var v) occ
822 | Just RecArg <- lookupHowBound env v = usg { scu_occs = extendVarEnv (scu_occs usg) v occ }
824 setScrutOcc _env usg _other _occ -- Catch-all
827 conArgOccs :: ArgOcc -> AltCon -> [ArgOcc]
828 -- Find usage of components of data con; returns [UnkOcc...] if unknown
829 -- See Note [ScrutOcc] for the extra UnkOccs in the vanilla datacon case
831 conArgOccs (ScrutOcc fm) (DataAlt dc)
832 | Just pat_arg_occs <- lookupUFM fm dc
833 = [UnkOcc | _ <- dataConUnivTyVars dc] ++ pat_arg_occs
835 conArgOccs _other _con = repeat UnkOcc
838 %************************************************************************
840 \subsection{The main recursive function}
842 %************************************************************************
844 The main recursive function gathers up usage information, and
845 creates specialised versions of functions.
848 scExpr, scExpr' :: ScEnv -> CoreExpr -> UniqSM (ScUsage, CoreExpr)
849 -- The unique supply is needed when we invent
850 -- a new name for the specialised function and its args
852 scExpr env e = scExpr' env e
855 scExpr' env (Var v) = case scSubstId env v of
856 Var v' -> return (varUsage env v' UnkOcc, Var v')
857 e' -> scExpr (zapScSubst env) e'
859 scExpr' env (Type t) = return (nullUsage, Type (scSubstTy env t))
860 scExpr' _ e@(Lit {}) = return (nullUsage, e)
861 scExpr' env (Note n e) = do (usg,e') <- scExpr env e
862 return (usg, Note n e')
863 scExpr' env (Cast e co) = do (usg, e') <- scExpr env e
864 return (usg, Cast e' (scSubstTy env co))
865 scExpr' env e@(App _ _) = scApp env (collectArgs e)
866 scExpr' env (Lam b e) = do let (env', b') = extendBndr env b
867 (usg, e') <- scExpr env' e
868 return (usg, Lam b' e')
870 scExpr' env (Case scrut b ty alts)
871 = do { (scrut_usg, scrut') <- scExpr env scrut
872 ; case isValue (sc_vals env) scrut' of
873 Just (ConVal con args) -> sc_con_app con args scrut'
874 _other -> sc_vanilla scrut_usg scrut'
877 sc_con_app con args scrut' -- Known constructor; simplify
878 = do { let (_, bs, rhs) = findAlt con alts
879 `orElse` (DEFAULT, [], mkImpossibleExpr (coreAltsType alts))
880 alt_env' = extendScSubstList env ((b,scrut') : bs `zip` trimConArgs con args)
881 ; scExpr alt_env' rhs }
883 sc_vanilla scrut_usg scrut' -- Normal case
884 = do { let (alt_env,b') = extendBndrWith RecArg env b
885 -- Record RecArg for the components
887 ; (alt_usgs, alt_occs, alts')
888 <- mapAndUnzip3M (sc_alt alt_env scrut' b') alts
890 ; let (alt_usg, b_occ) = lookupOcc (combineUsages alt_usgs) b'
891 scrut_occ = foldr combineOcc b_occ alt_occs
892 scrut_usg' = setScrutOcc env scrut_usg scrut' scrut_occ
893 -- The combined usage of the scrutinee is given
894 -- by scrut_occ, which is passed to scScrut, which
895 -- in turn treats a bare-variable scrutinee specially
897 ; return (alt_usg `combineUsage` scrut_usg',
898 Case scrut' b' (scSubstTy env ty) alts') }
900 sc_alt env _scrut' b' (con,bs,rhs)
901 = do { let (env1, bs1) = extendBndrsWith RecArg env bs
902 (env2, bs2) = extendCaseBndrs env1 b' con bs1
903 ; (usg,rhs') <- scExpr env2 rhs
904 ; let (usg', arg_occs) = lookupOccs usg bs2
905 scrut_occ = case con of
906 DataAlt dc -> ScrutOcc (unitUFM dc arg_occs)
907 _ -> ScrutOcc emptyUFM
908 ; return (usg', scrut_occ, (con, bs2, rhs')) }
910 scExpr' env (Let (NonRec bndr rhs) body)
911 | isTyVar bndr -- Type-lets may be created by doBeta
912 = scExpr' (extendScSubst env bndr rhs) body
914 | otherwise -- Note [Local let bindings]
915 = do { let (body_env, bndr') = extendBndr env bndr
916 body_env2 = extendHowBound body_env [bndr'] RecFun
917 ; (body_usg, body') <- scExpr body_env2 body
919 ; (rhs_usg, rhs_info) <- scRecRhs env (bndr',rhs)
921 ; let force_spec = False
922 ; (spec_usg, specs) <- specialise env force_spec
925 (SI [] 0 (Just rhs_usg))
927 ; return (body_usg { scu_calls = scu_calls body_usg `delVarEnv` bndr' }
928 `combineUsage` spec_usg,
929 mkLets [NonRec b r | (b,r) <- specInfoBinds rhs_info specs] body')
933 -- A *local* recursive group: see Note [Local recursive groups]
934 scExpr' env (Let (Rec prs) body)
935 = do { let (bndrs,rhss) = unzip prs
936 (rhs_env1,bndrs') = extendRecBndrs env bndrs
937 rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun
938 force_spec = any (forceSpecBndr env) bndrs'
940 ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss)
941 ; (body_usg, body') <- scExpr rhs_env2 body
943 -- NB: start specLoop from body_usg
944 ; (spec_usg, specs) <- specLoop rhs_env2 force_spec
945 (scu_calls body_usg) rhs_infos nullUsage
946 [SI [] 0 (Just usg) | usg <- rhs_usgs]
947 -- Do not unconditionally use rhs_usgs.
948 -- Instead use them only if we find an unspecialised call
949 -- See Note [Local recursive groups]
951 ; let all_usg = spec_usg `combineUsage` body_usg
952 bind' = Rec (concat (zipWith specInfoBinds rhs_infos specs))
954 ; return (all_usg { scu_calls = scu_calls all_usg `delVarEnvList` bndrs' },
958 Note [Local let bindings]
959 ~~~~~~~~~~~~~~~~~~~~~~~~~
960 It is not uncommon to find this
962 let $j = \x. <blah> in ...$j True...$j True...
964 Here $j is an arbitrary let-bound function, but it often comes up for
965 join points. We might like to specialise $j for its call patterns.
966 Notice the difference from a letrec, where we look for call patterns
967 in the *RHS* of the function. Here we look for call patterns in the
970 At one point I predicated this on the RHS mentioning the outer
971 recursive function, but that's not essential and might even be
972 harmful. I'm not sure.
976 scApp :: ScEnv -> (InExpr, [InExpr]) -> UniqSM (ScUsage, CoreExpr)
978 scApp env (Var fn, args) -- Function is a variable
979 = ASSERT( not (null args) )
980 do { args_w_usgs <- mapM (scExpr env) args
981 ; let (arg_usgs, args') = unzip args_w_usgs
982 arg_usg = combineUsages arg_usgs
983 ; case scSubstId env fn of
984 fn'@(Lam {}) -> scExpr (zapScSubst env) (doBeta fn' args')
985 -- Do beta-reduction and try again
987 Var fn' -> return (arg_usg `combineUsage` fn_usg, mkApps (Var fn') args')
989 fn_usg = case lookupHowBound env fn' of
990 Just RecFun -> SCU { scu_calls = unitVarEnv fn' [(sc_vals env, args')],
991 scu_occs = emptyVarEnv }
992 Just RecArg -> SCU { scu_calls = emptyVarEnv,
993 scu_occs = unitVarEnv fn' (ScrutOcc emptyUFM) }
997 other_fn' -> return (arg_usg, mkApps other_fn' args') }
998 -- NB: doing this ignores any usage info from the substituted
999 -- function, but I don't think that matters. If it does
1002 doBeta :: OutExpr -> [OutExpr] -> OutExpr
1003 -- ToDo: adjust for System IF
1004 doBeta (Lam bndr body) (arg : args) = Let (NonRec bndr arg) (doBeta body args)
1005 doBeta fn args = mkApps fn args
1007 -- The function is almost always a variable, but not always.
1008 -- In particular, if this pass follows float-in,
1009 -- which it may, we can get
1010 -- (let f = ...f... in f) arg1 arg2
1011 scApp env (other_fn, args)
1012 = do { (fn_usg, fn') <- scExpr env other_fn
1013 ; (arg_usgs, args') <- mapAndUnzipM (scExpr env) args
1014 ; return (combineUsages arg_usgs `combineUsage` fn_usg, mkApps fn' args') }
1016 ----------------------
1017 scTopBind :: ScEnv -> CoreBind -> UniqSM (ScEnv, CoreBind)
1018 scTopBind env (Rec prs)
1019 | Just threshold <- sc_size env
1021 , not (all (couldBeSmallEnoughToInline threshold) rhss)
1022 -- No specialisation
1023 = do { let (rhs_env,bndrs') = extendRecBndrs env bndrs
1024 ; (_, rhss') <- mapAndUnzipM (scExpr rhs_env) rhss
1025 ; return (rhs_env, Rec (bndrs' `zip` rhss')) }
1026 | otherwise -- Do specialisation
1027 = do { let (rhs_env1,bndrs') = extendRecBndrs env bndrs
1028 rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun
1030 ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss)
1031 ; let rhs_usg = combineUsages rhs_usgs
1033 ; (_, specs) <- specLoop rhs_env2 force_spec
1034 (scu_calls rhs_usg) rhs_infos nullUsage
1035 [SI [] 0 Nothing | _ <- bndrs]
1037 ; return (rhs_env1, -- For the body of the letrec, delete the RecFun business
1038 Rec (concat (zipWith specInfoBinds rhs_infos specs))) }
1040 (bndrs,rhss) = unzip prs
1041 force_spec = any (forceSpecBndr env) bndrs
1043 scTopBind env (NonRec bndr rhs)
1044 = do { (_, rhs') <- scExpr env rhs
1045 ; let (env1, bndr') = extendBndr env bndr
1046 env2 = extendValEnv env1 bndr' (isValue (sc_vals env) rhs')
1047 ; return (env2, NonRec bndr' rhs') }
1049 ----------------------
1050 scRecRhs :: ScEnv -> (OutId, InExpr) -> UniqSM (ScUsage, RhsInfo)
1051 scRecRhs env (bndr,rhs)
1052 = do { let (arg_bndrs,body) = collectBinders rhs
1053 (body_env, arg_bndrs') = extendBndrsWith RecArg env arg_bndrs
1054 ; (body_usg, body') <- scExpr body_env body
1055 ; let (rhs_usg, arg_occs) = lookupOccs body_usg arg_bndrs'
1056 ; return (rhs_usg, RI bndr (mkLams arg_bndrs' body')
1057 arg_bndrs body arg_occs) }
1058 -- The arg_occs says how the visible,
1059 -- lambda-bound binders of the RHS are used
1060 -- (including the TyVar binders)
1061 -- Two pats are the same if they match both ways
1063 ----------------------
1064 specInfoBinds :: RhsInfo -> SpecInfo -> [(Id,CoreExpr)]
1065 specInfoBinds (RI fn new_rhs _ _ _) (SI specs _ _)
1066 = [(id,rhs) | OS _ _ id rhs <- specs] ++
1067 [(fn `addIdSpecialisations` rules, new_rhs)]
1069 rules = [r | OS _ r _ _ <- specs]
1071 ----------------------
1072 varUsage :: ScEnv -> OutVar -> ArgOcc -> ScUsage
1074 | Just RecArg <- lookupHowBound env v = SCU { scu_calls = emptyVarEnv
1075 , scu_occs = unitVarEnv v use }
1076 | otherwise = nullUsage
1080 %************************************************************************
1082 The specialiser itself
1084 %************************************************************************
1087 data RhsInfo = RI OutId -- The binder
1088 OutExpr -- The new RHS
1089 [InVar] InExpr -- The *original* RHS (\xs.body)
1090 -- Note [Specialise original body]
1091 [ArgOcc] -- Info on how the xs occur in body
1093 data SpecInfo = SI [OneSpec] -- The specialisations we have generated
1095 Int -- Length of specs; used for numbering them
1097 (Maybe ScUsage) -- Nothing => we have generated specialisations
1098 -- from calls in the *original* RHS
1099 -- Just cs => we haven't, and this is the usage
1100 -- of the original RHS
1101 -- See Note [Local recursive groups]
1103 -- One specialisation: Rule plus definition
1104 data OneSpec = OS CallPat -- Call pattern that generated this specialisation
1105 CoreRule -- Rule connecting original id with the specialisation
1106 OutId OutExpr -- Spec id + its rhs
1110 -> Bool -- force specialisation?
1113 -> ScUsage -> [SpecInfo] -- One per binder; acccumulating parameter
1114 -> UniqSM (ScUsage, [SpecInfo]) -- ...ditto...
1115 specLoop env force_spec all_calls rhs_infos usg_so_far specs_so_far
1116 = do { specs_w_usg <- zipWithM (specialise env force_spec all_calls) rhs_infos specs_so_far
1117 ; let (new_usg_s, all_specs) = unzip specs_w_usg
1118 new_usg = combineUsages new_usg_s
1119 new_calls = scu_calls new_usg
1120 all_usg = usg_so_far `combineUsage` new_usg
1121 ; if isEmptyVarEnv new_calls then
1122 return (all_usg, all_specs)
1124 specLoop env force_spec new_calls rhs_infos all_usg all_specs }
1128 -> Bool -- force specialisation?
1129 -> CallEnv -- Info on calls
1131 -> SpecInfo -- Original RHS plus patterns dealt with
1132 -> UniqSM (ScUsage, SpecInfo) -- New specialised versions and their usage
1134 -- Note: the rhs here is the optimised version of the original rhs
1135 -- So when we make a specialised copy of the RHS, we're starting
1136 -- from an RHS whose nested functions have been optimised already.
1138 specialise env force_spec bind_calls (RI fn _ arg_bndrs body arg_occs)
1139 spec_info@(SI specs spec_count mb_unspec)
1140 | not (isBottomingId fn) -- Note [Do not specialise diverging functions]
1141 , notNull arg_bndrs -- Only specialise functions
1142 , Just all_calls <- lookupVarEnv bind_calls fn
1143 = do { (boring_call, pats) <- callsToPats env specs arg_occs all_calls
1144 -- ; pprTrace "specialise" (vcat [ ppr fn <+> text "with" <+> int (length pats) <+> text "good patterns"
1145 -- , text "arg_occs" <+> ppr arg_occs
1146 -- , text "calls" <+> ppr all_calls
1147 -- , text "good pats" <+> ppr pats]) $
1150 -- Bale out if too many specialisations
1151 ; let n_pats = length pats
1152 spec_count' = n_pats + spec_count
1153 ; case sc_count env of
1154 Just max | not force_spec && spec_count' > max
1155 -> pprTrace "SpecConstr" msg $
1156 return (nullUsage, spec_info)
1158 msg = vcat [ sep [ ptext (sLit "Function") <+> quotes (ppr fn)
1159 , nest 2 (ptext (sLit "has") <+>
1160 speakNOf spec_count' (ptext (sLit "call pattern")) <> comma <+>
1161 ptext (sLit "but the limit is") <+> int max) ]
1162 , ptext (sLit "Use -fspec-constr-count=n to set the bound")
1164 extra | not opt_PprStyle_Debug = ptext (sLit "Use -dppr-debug to see specialisations")
1165 | otherwise = ptext (sLit "Specialisations:") <+> ppr (pats ++ [p | OS p _ _ _ <- specs])
1167 _normal_case -> do {
1169 let spec_env = decreaseSpecCount env n_pats
1170 ; (spec_usgs, new_specs) <- mapAndUnzipM (spec_one spec_env fn arg_bndrs body)
1171 (pats `zip` [spec_count..])
1172 -- See Note [Specialise original body]
1174 ; let spec_usg = combineUsages spec_usgs
1175 (new_usg, mb_unspec')
1177 Just rhs_usg | boring_call -> (spec_usg `combineUsage` rhs_usg, Nothing)
1178 _ -> (spec_usg, mb_unspec)
1180 ; return (new_usg, SI (new_specs ++ specs) spec_count' mb_unspec') } }
1182 = return (nullUsage, spec_info) -- The boring case
1185 ---------------------
1187 -> OutId -- Function
1188 -> [InVar] -- Lambda-binders of RHS; should match patterns
1189 -> InExpr -- Body of the original function
1191 -> UniqSM (ScUsage, OneSpec) -- Rule and binding
1193 -- spec_one creates a specialised copy of the function, together
1194 -- with a rule for using it. I'm very proud of how short this
1195 -- function is, considering what it does :-).
1201 f = /\b \y::[(a,b)] -> ....f (b,c) ((:) (a,(b,c)) (x,v) (h w))...
1202 [c::*, v::(b,c) are presumably bound by the (...) part]
1204 f_spec = /\ b c \ v::(b,c) hw::[(a,(b,c))] ->
1205 (...entire body of f...) [b -> (b,c),
1206 y -> ((:) (a,(b,c)) (x,v) hw)]
1208 RULE: forall b::* c::*, -- Note, *not* forall a, x
1212 f (b,c) ((:) (a,(b,c)) (x,v) hw) = f_spec b c v hw
1215 spec_one env fn arg_bndrs body (call_pat@(qvars, pats), rule_number)
1216 = do { spec_uniq <- getUniqueUs
1217 ; let spec_env = extendScSubstList (extendScInScope env qvars)
1218 (arg_bndrs `zip` pats)
1220 fn_loc = nameSrcSpan fn_name
1221 spec_occ = mkSpecOcc (nameOccName fn_name)
1222 rule_name = mkFastString ("SC:" ++ showSDoc (ppr fn <> int rule_number))
1223 spec_name = mkInternalName spec_uniq spec_occ fn_loc
1224 -- ; pprTrace "{spec_one" (ppr (sc_count env) <+> ppr fn <+> ppr pats <+> text "-->" <+> ppr spec_name) $
1227 -- Specialise the body
1228 ; (spec_usg, spec_body) <- scExpr spec_env body
1230 -- ; pprTrace "done spec_one}" (ppr fn) $
1233 -- And build the results
1234 ; let spec_id = mkLocalId spec_name (mkPiTypes spec_lam_args body_ty)
1235 `setIdStrictness` spec_str -- See Note [Transfer strictness]
1236 `setIdArity` count isId spec_lam_args
1237 spec_str = calcSpecStrictness fn spec_lam_args pats
1238 (spec_lam_args, spec_call_args) = mkWorkerArgs qvars body_ty
1239 -- Usual w/w hack to avoid generating
1240 -- a spec_rhs of unlifted type and no args
1242 spec_rhs = mkLams spec_lam_args spec_body
1243 body_ty = exprType spec_body
1244 rule_rhs = mkVarApps (Var spec_id) spec_call_args
1245 inline_act = idInlineActivation fn
1246 rule = mkLocalRule rule_name inline_act fn_name qvars pats rule_rhs
1247 ; return (spec_usg, OS call_pat rule spec_id spec_rhs) }
1249 calcSpecStrictness :: Id -- The original function
1250 -> [Var] -> [CoreExpr] -- Call pattern
1251 -> StrictSig -- Strictness of specialised thing
1252 -- See Note [Transfer strictness]
1253 calcSpecStrictness fn qvars pats
1254 = StrictSig (mkTopDmdType spec_dmds TopRes)
1256 spec_dmds = [ lookupVarEnv dmd_env qv `orElse` lazyDmd | qv <- qvars, isId qv ]
1257 StrictSig (DmdType _ dmds _) = idStrictness fn
1259 dmd_env = go emptyVarEnv dmds pats
1261 go env ds (Type {} : pats) = go env ds pats
1262 go env (d:ds) (pat : pats) = go (go_one env d pat) ds pats
1265 go_one env d (Var v) = extendVarEnv_C both env v d
1266 go_one env (Box d) e = go_one env d e
1267 go_one env (Eval (Prod ds)) e
1268 | (Var _, args) <- collectArgs e = go env ds args
1269 go_one env _ _ = env
1273 Note [Specialise original body]
1274 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1275 The RhsInfo for a binding keeps the *original* body of the binding. We
1276 must specialise that, *not* the result of applying specExpr to the RHS
1277 (which is also kept in RhsInfo). Otherwise we end up specialising a
1278 specialised RHS, and that can lead directly to exponential behaviour.
1280 Note [Transfer activation]
1281 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1282 In which phase should the specialise-constructor rules be active?
1283 Originally I made them always-active, but Manuel found that this
1284 defeated some clever user-written rules. Then I made them active only
1285 in Phase 0; after all, currently, the specConstr transformation is
1286 only run after the simplifier has reached Phase 0, but that meant
1287 that specialisations didn't fire inside wrappers; see test
1288 simplCore/should_compile/spec-inline.
1290 So now I just use the inline-activation of the parent Id, as the
1291 activation for the specialiation RULE, just like the main specialiser;
1292 see Note [Auto-specialisation and RULES] in Specialise.
1295 Note [Transfer strictness]
1296 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1297 We must transfer strictness information from the original function to
1298 the specialised one. Suppose, for example
1301 and a RULE f (a:as) b = f_spec a as b
1303 Now we want f_spec to have strictess LLS, otherwise we'll use call-by-need
1304 when calling f_spec instead of call-by-value. And that can result in
1305 unbounded worsening in space (cf the classic foldl vs foldl')
1307 See Trac #3437 for a good example.
1309 The function calcSpecStrictness performs the calculation.
1312 %************************************************************************
1314 \subsection{Argument analysis}
1316 %************************************************************************
1318 This code deals with analysing call-site arguments to see whether
1319 they are constructor applications.
1323 type CallPat = ([Var], [CoreExpr]) -- Quantified variables and arguments
1326 callsToPats :: ScEnv -> [OneSpec] -> [ArgOcc] -> [Call] -> UniqSM (Bool, [CallPat])
1327 -- Result has no duplicate patterns,
1328 -- nor ones mentioned in done_pats
1329 -- Bool indicates that there was at least one boring pattern
1330 callsToPats env done_specs bndr_occs calls
1331 = do { mb_pats <- mapM (callToPats env bndr_occs) calls
1333 ; let good_pats :: [([Var], [CoreArg])]
1334 good_pats = catMaybes mb_pats
1335 done_pats = [p | OS p _ _ _ <- done_specs]
1336 is_done p = any (samePat p) done_pats
1338 ; return (any isNothing mb_pats,
1339 filterOut is_done (nubBy samePat good_pats)) }
1341 callToPats :: ScEnv -> [ArgOcc] -> Call -> UniqSM (Maybe CallPat)
1342 -- The [Var] is the variables to quantify over in the rule
1343 -- Type variables come first, since they may scope
1344 -- over the following term variables
1345 -- The [CoreExpr] are the argument patterns for the rule
1346 callToPats env bndr_occs (con_env, args)
1347 | length args < length bndr_occs -- Check saturated
1350 = do { let in_scope = substInScope (sc_subst env)
1351 ; prs <- argsToPats env in_scope con_env (args `zip` bndr_occs)
1352 ; let (interesting_s, pats) = unzip prs
1353 pat_fvs = varSetElems (exprsFreeVars pats)
1354 qvars = filterOut (`elemInScopeSet` in_scope) pat_fvs
1355 -- Quantify over variables that are not in sccpe
1357 -- See Note [Shadowing] at the top
1359 (tvs, ids) = partition isTyVar qvars
1361 -- Put the type variables first; the type of a term
1362 -- variable may mention a type variable
1364 ; -- pprTrace "callToPats" (ppr args $$ ppr prs $$ ppr bndr_occs) $
1366 then return (Just (qvars', pats))
1367 else return Nothing }
1369 -- argToPat takes an actual argument, and returns an abstracted
1370 -- version, consisting of just the "constructor skeleton" of the
1371 -- argument, with non-constructor sub-expression replaced by new
1372 -- placeholder variables. For example:
1373 -- C a (D (f x) (g y)) ==> C p1 (D p2 p3)
1376 -> InScopeSet -- What's in scope at the fn defn site
1377 -> ValueEnv -- ValueEnv at the call site
1378 -> CoreArg -- A call arg (or component thereof)
1380 -> UniqSM (Bool, CoreArg)
1381 -- Returns (interesting, pat),
1382 -- where pat is the pattern derived from the argument
1383 -- intersting=True if the pattern is non-trivial (not a variable or type)
1384 -- E.g. x:xs --> (True, x:xs)
1385 -- f xs --> (False, w) where w is a fresh wildcard
1386 -- (f xs, 'c') --> (True, (w, 'c')) where w is a fresh wildcard
1387 -- \x. x+y --> (True, \x. x+y)
1388 -- lvl7 --> (True, lvl7) if lvl7 is bound
1389 -- somewhere further out
1391 argToPat _env _in_scope _val_env arg@(Type {}) _arg_occ
1392 = return (False, arg)
1394 argToPat env in_scope val_env (Note _ arg) arg_occ
1395 = argToPat env in_scope val_env arg arg_occ
1396 -- Note [Notes in call patterns]
1397 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1398 -- Ignore Notes. In particular, we want to ignore any InlineMe notes
1399 -- Perhaps we should not ignore profiling notes, but I'm going to
1400 -- ride roughshod over them all for now.
1401 --- See Note [Notes in RULE matching] in Rules
1403 argToPat env in_scope val_env (Let _ arg) arg_occ
1404 = argToPat env in_scope val_env arg arg_occ
1405 -- Look through let expressions
1406 -- e.g. f (let v = rhs in \y -> ...v...)
1407 -- Here we can specialise for f (\y -> ...)
1408 -- because the rule-matcher will look through the let.
1410 argToPat env in_scope val_env (Cast arg co) arg_occ
1411 | not (ignoreType env ty2)
1412 = do { (interesting, arg') <- argToPat env in_scope val_env arg arg_occ
1413 ; if not interesting then
1416 { -- Make a wild-card pattern for the coercion
1418 ; let co_name = mkSysTvName uniq (fsLit "sg")
1419 co_var = mkCoVar co_name (mkCoKind ty1 ty2)
1420 ; return (interesting, Cast arg' (mkTyVarTy co_var)) } }
1422 (ty1, ty2) = coercionKind co
1426 {- Disabling lambda specialisation for now
1427 It's fragile, and the spec_loop can be infinite
1428 argToPat in_scope val_env arg arg_occ
1430 = return (True, arg)
1432 is_value_lam (Lam v e) -- Spot a value lambda, even if
1433 | isId v = True -- it is inside a type lambda
1434 | otherwise = is_value_lam e
1435 is_value_lam other = False
1438 -- Check for a constructor application
1439 -- NB: this *precedes* the Var case, so that we catch nullary constrs
1440 argToPat env in_scope val_env arg arg_occ
1441 | Just (ConVal dc args) <- isValue val_env arg
1442 , not (ignoreAltCon env dc)
1444 ScrutOcc _ -> True -- Used only by case scrutinee
1445 BothOcc -> case arg of -- Used elsewhere
1446 App {} -> True -- see Note [Reboxing]
1448 _other -> False -- No point; the arg is not decomposed
1449 = do { args' <- argsToPats env in_scope val_env (args `zip` conArgOccs arg_occ dc)
1450 ; return (True, mk_con_app dc (map snd args')) }
1452 -- Check if the argument is a variable that
1453 -- is in scope at the function definition site
1454 -- It's worth specialising on this if
1455 -- (a) it's used in an interesting way in the body
1456 -- (b) we know what its value is
1457 argToPat env in_scope val_env (Var v) arg_occ
1458 | case arg_occ of { UnkOcc -> False; _other -> True }, -- (a)
1460 not (ignoreType env (varType v))
1461 = return (True, Var v)
1464 | isLocalId v = v `elemInScopeSet` in_scope
1465 && isJust (lookupVarEnv val_env v)
1466 -- Local variables have values in val_env
1467 | otherwise = isValueUnfolding (idUnfolding v)
1468 -- Imports have unfoldings
1470 -- I'm really not sure what this comment means
1471 -- And by not wild-carding we tend to get forall'd
1472 -- variables that are in soope, which in turn can
1473 -- expose the weakness in let-matching
1474 -- See Note [Matching lets] in Rules
1476 -- Check for a variable bound inside the function.
1477 -- Don't make a wild-card, because we may usefully share
1478 -- e.g. f a = let x = ... in f (x,x)
1479 -- NB: this case follows the lambda and con-app cases!!
1480 -- argToPat _in_scope _val_env (Var v) _arg_occ
1481 -- = return (False, Var v)
1482 -- SLPJ : disabling this to avoid proliferation of versions
1483 -- also works badly when thinking about seeding the loop
1484 -- from the body of the let
1485 -- f x y = letrec g z = ... in g (x,y)
1486 -- We don't want to specialise for that *particular* x,y
1488 -- The default case: make a wild-card
1489 argToPat _env _in_scope _val_env arg _arg_occ
1490 = wildCardPat (exprType arg)
1492 wildCardPat :: Type -> UniqSM (Bool, CoreArg)
1493 wildCardPat ty = do { uniq <- getUniqueUs
1494 ; let id = mkSysLocal (fsLit "sc") uniq ty
1495 ; return (False, Var id) }
1497 argsToPats :: ScEnv -> InScopeSet -> ValueEnv
1498 -> [(CoreArg, ArgOcc)]
1499 -> UniqSM [(Bool, CoreArg)]
1500 argsToPats env in_scope val_env args
1503 do_one (arg,occ) = argToPat env in_scope val_env arg occ
1508 isValue :: ValueEnv -> CoreExpr -> Maybe Value
1509 isValue _env (Lit lit)
1510 = Just (ConVal (LitAlt lit) [])
1513 | Just stuff <- lookupVarEnv env v
1514 = Just stuff -- You might think we could look in the idUnfolding here
1515 -- but that doesn't take account of which branch of a
1516 -- case we are in, which is the whole point
1518 | not (isLocalId v) && isCheapUnfolding unf
1519 = isValue env (unfoldingTemplate unf)
1522 -- However we do want to consult the unfolding
1523 -- as well, for let-bound constructors!
1525 isValue env (Lam b e)
1526 | isTyVar b = case isValue env e of
1527 Just _ -> Just LambdaVal
1529 | otherwise = Just LambdaVal
1531 isValue _env expr -- Maybe it's a constructor application
1532 | (Var fun, args) <- collectArgs expr
1533 = case isDataConWorkId_maybe fun of
1535 Just con | args `lengthAtLeast` dataConRepArity con
1536 -- Check saturated; might be > because the
1537 -- arity excludes type args
1538 -> Just (ConVal (DataAlt con) args)
1540 _other | valArgCount args < idArity fun
1541 -- Under-applied function
1542 -> Just LambdaVal -- Partial application
1546 isValue _env _expr = Nothing
1548 mk_con_app :: AltCon -> [CoreArg] -> CoreExpr
1549 mk_con_app (LitAlt lit) [] = Lit lit
1550 mk_con_app (DataAlt con) args = mkConApp con args
1551 mk_con_app _other _args = panic "SpecConstr.mk_con_app"
1553 samePat :: CallPat -> CallPat -> Bool
1554 samePat (vs1, as1) (vs2, as2)
1557 same (Var v1) (Var v2)
1558 | v1 `elem` vs1 = v2 `elem` vs2
1559 | v2 `elem` vs2 = False
1560 | otherwise = v1 == v2
1562 same (Lit l1) (Lit l2) = l1==l2
1563 same (App f1 a1) (App f2 a2) = same f1 f2 && same a1 a2
1565 same (Type {}) (Type {}) = True -- Note [Ignore type differences]
1566 same (Note _ e1) e2 = same e1 e2 -- Ignore casts and notes
1567 same (Cast e1 _) e2 = same e1 e2
1568 same e1 (Note _ e2) = same e1 e2
1569 same e1 (Cast e2 _) = same e1 e2
1571 same e1 e2 = WARN( bad e1 || bad e2, ppr e1 $$ ppr e2)
1572 False -- Let, lambda, case should not occur
1573 bad (Case {}) = True
1579 Note [Ignore type differences]
1580 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1581 We do not want to generate specialisations where the call patterns
1582 differ only in their type arguments! Not only is it utterly useless,
1583 but it also means that (with polymorphic recursion) we can generate
1584 an infinite number of specialisations. Example is Data.Sequence.adjustTree,