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 )
51 import Control.Monad ( zipWithM )
53 import Data.Data ( Data, Typeable )
56 -----------------------------------------------------
58 -----------------------------------------------------
63 drop n (x:xs) = drop (n-1) xs
65 After the first time round, we could pass n unboxed. This happens in
66 numerical code too. Here's what it looks like in Core:
68 drop n xs = case xs of
73 _ -> drop (I# (n# -# 1#)) xs
75 Notice that the recursive call has an explicit constructor as argument.
76 Noticing this, we can make a specialised version of drop
78 RULE: drop (I# n#) xs ==> drop' n# xs
80 drop' n# xs = let n = I# n# in ...orig RHS...
82 Now the simplifier will apply the specialisation in the rhs of drop', giving
84 drop' n# xs = case xs of
88 _ -> drop (n# -# 1#) xs
92 We'd also like to catch cases where a parameter is carried along unchanged,
93 but evaluated each time round the loop:
95 f i n = if i>0 || i>n then i else f (i*2) n
97 Here f isn't strict in n, but we'd like to avoid evaluating it each iteration.
98 In Core, by the time we've w/wd (f is strict in i) we get
100 f i# n = case i# ># 0 of
102 True -> case n of n' { I# n# ->
105 True -> f (i# *# 2#) n'
107 At the call to f, we see that the argument, n is know to be (I# n#),
108 and n is evaluated elsewhere in the body of f, so we can play the same
114 We must be careful not to allocate the same constructor twice. Consider
115 f p = (...(case p of (a,b) -> e)...p...,
116 ...let t = (r,s) in ...t...(f t)...)
117 At the recursive call to f, we can see that t is a pair. But we do NOT want
118 to make a specialised copy:
119 f' a b = let p = (a,b) in (..., ...)
120 because now t is allocated by the caller, then r and s are passed to the
121 recursive call, which allocates the (r,s) pair again.
124 (a) the argument p is used in other than a case-scrutinsation way.
125 (b) the argument to the call is not a 'fresh' tuple; you have to
126 look into its unfolding to see that it's a tuple
128 Hence the "OR" part of Note [Good arguments] below.
130 ALTERNATIVE 2: pass both boxed and unboxed versions. This no longer saves
131 allocation, but does perhaps save evals. In the RULE we'd have
134 f (I# x#) = f' (I# x#) x#
136 If at the call site the (I# x) was an unfolding, then we'd have to
137 rely on CSE to eliminate the duplicate allocation.... This alternative
138 doesn't look attractive enough to pursue.
140 ALTERNATIVE 3: ignore the reboxing problem. The trouble is that
141 the conservative reboxing story prevents many useful functions from being
142 specialised. Example:
143 foo :: Maybe Int -> Int -> Int
145 foo x@(Just m) n = foo x (n-m)
146 Here the use of 'x' will clearly not require boxing in the specialised function.
148 The strictness analyser has the same problem, in fact. Example:
150 If we pass just 'a' and 'b' to the worker, it might need to rebox the
151 pair to create (a,b). A more sophisticated analysis might figure out
152 precisely the cases in which this could happen, but the strictness
153 analyser does no such analysis; it just passes 'a' and 'b', and hopes
156 So my current choice is to make SpecConstr similarly aggressive, and
157 ignore the bad potential of reboxing.
160 Note [Good arguments]
161 ~~~~~~~~~~~~~~~~~~~~~
164 * A self-recursive function. Ignore mutual recursion for now,
165 because it's less common, and the code is simpler for self-recursion.
169 a) At a recursive call, one or more parameters is an explicit
170 constructor application
172 That same parameter is scrutinised by a case somewhere in
173 the RHS of the function
177 b) At a recursive call, one or more parameters has an unfolding
178 that is an explicit constructor application
180 That same parameter is scrutinised by a case somewhere in
181 the RHS of the function
183 Those are the only uses of the parameter (see Note [Reboxing])
186 What to abstract over
187 ~~~~~~~~~~~~~~~~~~~~~
188 There's a bit of a complication with type arguments. If the call
191 f p = ...f ((:) [a] x xs)...
193 then our specialised function look like
195 f_spec x xs = let p = (:) [a] x xs in ....as before....
197 This only makes sense if either
198 a) the type variable 'a' is in scope at the top of f, or
199 b) the type variable 'a' is an argument to f (and hence fs)
201 Actually, (a) may hold for value arguments too, in which case
202 we may not want to pass them. Supose 'x' is in scope at f's
203 defn, but xs is not. Then we'd like
205 f_spec xs = let p = (:) [a] x xs in ....as before....
207 Similarly (b) may hold too. If x is already an argument at the
208 call, no need to pass it again.
210 Finally, if 'a' is not in scope at the call site, we could abstract
211 it as we do the term variables:
213 f_spec a x xs = let p = (:) [a] x xs in ...as before...
215 So the grand plan is:
217 * abstract the call site to a constructor-only pattern
218 e.g. C x (D (f p) (g q)) ==> C s1 (D s2 s3)
220 * Find the free variables of the abstracted pattern
222 * Pass these variables, less any that are in scope at
223 the fn defn. But see Note [Shadowing] below.
226 NOTICE that we only abstract over variables that are not in scope,
227 so we're in no danger of shadowing variables used in "higher up"
233 In this pass we gather up usage information that may mention variables
234 that are bound between the usage site and the definition site; or (more
235 seriously) may be bound to something different at the definition site.
238 f x = letrec g y v = let x = ...
241 Since 'x' is in scope at the call site, we may make a rewrite rule that
243 RULE forall a,b. g (a,b) x = ...
244 But this rule will never match, because it's really a different 'x' at
245 the call site -- and that difference will be manifest by the time the
246 simplifier gets to it. [A worry: the simplifier doesn't *guarantee*
247 no-shadowing, so perhaps it may not be distinct?]
249 Anyway, the rule isn't actually wrong, it's just not useful. One possibility
250 is to run deShadowBinds before running SpecConstr, but instead we run the
251 simplifier. That gives the simplest possible program for SpecConstr to
252 chew on; and it virtually guarantees no shadowing.
254 Note [Specialising for constant parameters]
255 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
256 This one is about specialising on a *constant* (but not necessarily
257 constructor) argument
259 foo :: Int -> (Int -> Int) -> Int
261 foo m f = foo (f m) (+1)
265 lvl_rmV :: GHC.Base.Int -> GHC.Base.Int
267 \ (ds_dlk :: GHC.Base.Int) ->
268 case ds_dlk of wild_alH { GHC.Base.I# x_alG ->
269 GHC.Base.I# (GHC.Prim.+# x_alG 1)
271 T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
274 \ (ww_sme :: GHC.Prim.Int#) (w_smg :: GHC.Base.Int -> GHC.Base.Int) ->
275 case ww_sme of ds_Xlw {
277 case w_smg (GHC.Base.I# ds_Xlw) of w1_Xmo { GHC.Base.I# ww1_Xmz ->
278 T.$wfoo ww1_Xmz lvl_rmV
283 The recursive call has lvl_rmV as its argument, so we could create a specialised copy
284 with that argument baked in; that is, not passed at all. Now it can perhaps be inlined.
286 When is this worth it? Call the constant 'lvl'
287 - If 'lvl' has an unfolding that is a constructor, see if the corresponding
288 parameter is scrutinised anywhere in the body.
290 - If 'lvl' has an unfolding that is a inlinable function, see if the corresponding
291 parameter is applied (...to enough arguments...?)
293 Also do this is if the function has RULES?
297 Note [Specialising for lambda parameters]
298 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
299 foo :: Int -> (Int -> Int) -> Int
301 foo m f = foo (f m) (\n -> n-m)
303 This is subtly different from the previous one in that we get an
304 explicit lambda as the argument:
306 T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
309 \ (ww_sm8 :: GHC.Prim.Int#) (w_sma :: GHC.Base.Int -> GHC.Base.Int) ->
310 case ww_sm8 of ds_Xlr {
312 case w_sma (GHC.Base.I# ds_Xlr) of w1_Xmf { GHC.Base.I# ww1_Xmq ->
315 (\ (n_ad3 :: GHC.Base.Int) ->
316 case n_ad3 of wild_alB { GHC.Base.I# x_alA ->
317 GHC.Base.I# (GHC.Prim.-# x_alA ds_Xlr)
323 I wonder if SpecConstr couldn't be extended to handle this? After all,
324 lambda is a sort of constructor for functions and perhaps it already
325 has most of the necessary machinery?
327 Furthermore, there's an immediate win, because you don't need to allocate the lamda
328 at the call site; and if perchance it's called in the recursive call, then you
329 may avoid allocating it altogether. Just like for constructors.
331 Looks cool, but probably rare...but it might be easy to implement.
334 Note [SpecConstr for casts]
335 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
338 data instance T Int = T Int
343 go (T n) = go (T (n-1))
345 The recursive call ends up looking like
346 go (T (I# ...) `cast` g)
347 So we want to spot the construtor application inside the cast.
348 That's why we have the Cast case in argToPat
350 Note [Local recursive groups]
351 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
352 For a *local* recursive group, we can see all the calls to the
353 function, so we seed the specialisation loop from the calls in the
354 body, not from the calls in the RHS. Consider:
356 bar m n = foo n (n,n) (n,n) (n,n) (n,n)
360 | n > 3000 = case p of { (p1,p2) -> foo (n-1) (p2,p1) q r s }
361 | n > 2000 = case q of { (q1,q2) -> foo (n-1) p (q2,q1) r s }
362 | n > 1000 = case r of { (r1,r2) -> foo (n-1) p q (r2,r1) s }
363 | otherwise = case s of { (s1,s2) -> foo (n-1) p q r (s2,s1) }
365 If we start with the RHSs of 'foo', we get lots and lots of specialisations,
366 most of which are not needed. But if we start with the (single) call
367 in the rhs of 'bar' we get exactly one fully-specialised copy, and all
368 the recursive calls go to this fully-specialised copy. Indeed, the original
369 function is later collected as dead code. This is very important in
370 specialising the loops arising from stream fusion, for example in NDP where
371 we were getting literally hundreds of (mostly unused) specialisations of
374 Note [Do not specialise diverging functions]
375 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
376 Specialising a function that just diverges is a waste of code.
377 Furthermore, it broke GHC (simpl014) thus:
379 f = \x. case x of (a,b) -> f x
380 If we specialise f we get
381 f = \x. case x of (a,b) -> fspec a b
382 But fspec doesn't have decent strictnes info. As it happened,
383 (f x) :: IO t, so the state hack applied and we eta expanded fspec,
384 and hence f. But now f's strictness is less than its arity, which
387 Note [Forcing specialisation]
388 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
389 With stream fusion and in other similar cases, we want to fully specialise
390 some (but not necessarily all!) loops regardless of their size and the
391 number of specialisations. We allow a library to specify this by annotating
392 a type with ForceSpecConstr and then adding a parameter of that type to the
393 loop. Here is a (simplified) example from the vector library:
395 data SPEC = SPEC | SPEC2
396 {-# ANN type SPEC ForceSpecConstr #-}
398 foldl :: (a -> b -> a) -> a -> Stream b -> a
400 foldl f z (Stream step s _) = foldl_loop SPEC z s
402 foldl_loop SPEC z s = case step s of
403 Yield x s' -> foldl_loop SPEC (f z x) s'
404 Skip -> foldl_loop SPEC z s'
407 SpecConstr will spot the SPEC parameter and always fully specialise
408 foldl_loop. Note that we can't just annotate foldl_loop since it isn't a
409 top-level function but even if we could, inlining etc. could easily drop the
410 annotation. We also have to prevent the SPEC argument from being removed by
411 w/w which is why SPEC is a sum type. This is all quite ugly; we ought to come
412 up with a better design.
414 ForceSpecConstr arguments are spotted in scExpr' and scTopBinds which then set
415 force_spec to True when calling specLoop. This flag makes specLoop and
416 specialise ignore specConstrCount and specConstrThreshold when deciding
417 whether to specialise a function.
419 -----------------------------------------------------
420 Stuff not yet handled
421 -----------------------------------------------------
423 Here are notes arising from Roman's work that I don't want to lose.
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 SpecConstr does no specialisation, because the second recursive call
435 looks like a boxed use of the argument. A pity.
437 $wfoo_sFw :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
439 \ (ww_sFo [Just L] :: GHC.Prim.Int#) (w_sFq [Just L] :: T.T GHC.Base.Int) ->
440 case ww_sFo of ds_Xw6 [Just L] {
442 case GHC.Prim.remInt# ds_Xw6 2 of wild1_aEF [Dead Just A] {
443 __DEFAULT -> $wfoo_sFw (GHC.Prim.-# ds_Xw6 1) w_sFq;
445 case w_sFq of wild_Xy [Just L] { T.T n_ad5 [Just U(L)] ->
446 case n_ad5 of wild1_aET [Just A] { GHC.Base.I# y_aES [Just L] ->
447 $wfoo_sFw (GHC.Prim.-# ds_Xw6 y_aES) wild_Xy
453 data a :*: b = !a :*: !b
456 foo :: (Int :*: T Int) -> Int
458 foo (x :*: t) | even x = case t of { T n -> foo ((x-n) :*: t) }
459 | otherwise = foo ((x-1) :*: t)
461 Very similar to the previous one, except that the parameters are now in
462 a strict tuple. Before SpecConstr, we have
464 $wfoo_sG3 :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
466 \ (ww_sFU [Just L] :: GHC.Prim.Int#) (ww_sFW [Just L] :: T.T
468 case ww_sFU of ds_Xws [Just L] {
470 case GHC.Prim.remInt# ds_Xws 2 of wild1_aEZ [Dead Just A] {
472 case ww_sFW of tpl_B2 [Just L] { T.T a_sFo [Just A] ->
473 $wfoo_sG3 (GHC.Prim.-# ds_Xws 1) tpl_B2 -- $wfoo1
476 case ww_sFW of wild_XB [Just A] { T.T n_ad7 [Just S(L)] ->
477 case n_ad7 of wild1_aFd [Just L] { GHC.Base.I# y_aFc [Just L] ->
478 $wfoo_sG3 (GHC.Prim.-# ds_Xws y_aFc) wild_XB -- $wfoo2
482 We get two specialisations:
483 "SC:$wfoo1" [0] __forall {a_sFB :: GHC.Base.Int sc_sGC :: GHC.Prim.Int#}
484 Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int a_sFB)
485 = Foo.$s$wfoo1 a_sFB sc_sGC ;
486 "SC:$wfoo2" [0] __forall {y_aFp :: GHC.Prim.Int# sc_sGC :: GHC.Prim.Int#}
487 Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int (GHC.Base.I# y_aFp))
488 = Foo.$s$wfoo y_aFp sc_sGC ;
490 But perhaps the first one isn't good. After all, we know that tpl_B2 is
491 a T (I# x) really, because T is strict and Int has one constructor. (We can't
492 unbox the strict fields, becuase T is polymorphic!)
494 %************************************************************************
496 \subsection{Annotations}
498 %************************************************************************
500 Annotating a type with NoSpecConstr will make SpecConstr not specialise
501 for arguments of that type.
504 data SpecConstrAnnotation = NoSpecConstr | ForceSpecConstr
505 deriving( Data, Typeable, Eq )
508 %************************************************************************
510 \subsection{Top level wrapper stuff}
512 %************************************************************************
515 specConstrProgram :: ModGuts -> CoreM ModGuts
516 specConstrProgram guts
518 dflags <- getDynFlags
519 us <- getUniqueSupplyM
520 annos <- getFirstAnnotations deserializeWithData guts
521 let binds' = fst $ initUs us (go (initScEnv dflags annos) (mg_binds guts))
522 return (guts { mg_binds = binds' })
525 go env (bind:binds) = do (env', bind') <- scTopBind env bind
526 binds' <- go env' binds
527 return (bind' : binds')
531 %************************************************************************
533 \subsection{Environment: goes downwards}
535 %************************************************************************
538 data ScEnv = SCE { sc_size :: Maybe Int, -- Size threshold
539 sc_count :: Maybe Int, -- Max # of specialisations for any one fn
540 -- See Note [Avoiding exponential blowup]
542 sc_subst :: Subst, -- Current substitution
543 -- Maps InIds to OutExprs
545 sc_how_bound :: HowBoundEnv,
546 -- Binds interesting non-top-level variables
547 -- Domain is OutVars (*after* applying the substitution)
550 -- Domain is OutIds (*after* applying the substitution)
551 -- Used even for top-level bindings (but not imported ones)
553 sc_annotations :: UniqFM SpecConstrAnnotation
556 ---------------------
557 -- As we go, we apply a substitution (sc_subst) to the current term
558 type InExpr = CoreExpr -- _Before_ applying the subst
561 type OutExpr = CoreExpr -- _After_ applying the subst
565 ---------------------
566 type HowBoundEnv = VarEnv HowBound -- Domain is OutVars
568 ---------------------
569 type ValueEnv = IdEnv Value -- Domain is OutIds
570 data Value = ConVal AltCon [CoreArg] -- _Saturated_ constructors
571 | LambdaVal -- Inlinable lambdas or PAPs
573 instance Outputable Value where
574 ppr (ConVal con args) = ppr con <+> interpp'SP args
575 ppr LambdaVal = ptext (sLit "<Lambda>")
577 ---------------------
578 initScEnv :: DynFlags -> UniqFM SpecConstrAnnotation -> ScEnv
579 initScEnv dflags anns
580 = SCE { sc_size = specConstrThreshold dflags,
581 sc_count = specConstrCount dflags,
582 sc_subst = emptySubst,
583 sc_how_bound = emptyVarEnv,
584 sc_vals = emptyVarEnv,
585 sc_annotations = anns }
587 data HowBound = RecFun -- These are the recursive functions for which
588 -- we seek interesting call patterns
590 | RecArg -- These are those functions' arguments, or their sub-components;
591 -- we gather occurrence information for these
593 instance Outputable HowBound where
594 ppr RecFun = text "RecFun"
595 ppr RecArg = text "RecArg"
597 lookupHowBound :: ScEnv -> Id -> Maybe HowBound
598 lookupHowBound env id = lookupVarEnv (sc_how_bound env) id
600 scSubstId :: ScEnv -> Id -> CoreExpr
601 scSubstId env v = lookupIdSubst (text "scSubstId") (sc_subst env) v
603 scSubstTy :: ScEnv -> Type -> Type
604 scSubstTy env ty = substTy (sc_subst env) ty
606 zapScSubst :: ScEnv -> ScEnv
607 zapScSubst env = env { sc_subst = zapSubstEnv (sc_subst env) }
609 extendScInScope :: ScEnv -> [Var] -> ScEnv
610 -- Bring the quantified variables into scope
611 extendScInScope env qvars = env { sc_subst = extendInScopeList (sc_subst env) qvars }
613 -- Extend the substitution
614 extendScSubst :: ScEnv -> Var -> OutExpr -> ScEnv
615 extendScSubst env var expr = env { sc_subst = extendSubst (sc_subst env) var expr }
617 extendScSubstList :: ScEnv -> [(Var,OutExpr)] -> ScEnv
618 extendScSubstList env prs = env { sc_subst = extendSubstList (sc_subst env) prs }
620 extendHowBound :: ScEnv -> [Var] -> HowBound -> ScEnv
621 extendHowBound env bndrs how_bound
622 = env { sc_how_bound = extendVarEnvList (sc_how_bound env)
623 [(bndr,how_bound) | bndr <- bndrs] }
625 extendBndrsWith :: HowBound -> ScEnv -> [Var] -> (ScEnv, [Var])
626 extendBndrsWith how_bound env bndrs
627 = (env { sc_subst = subst', sc_how_bound = hb_env' }, bndrs')
629 (subst', bndrs') = substBndrs (sc_subst env) bndrs
630 hb_env' = sc_how_bound env `extendVarEnvList`
631 [(bndr,how_bound) | bndr <- bndrs']
633 extendBndrWith :: HowBound -> ScEnv -> Var -> (ScEnv, Var)
634 extendBndrWith how_bound env bndr
635 = (env { sc_subst = subst', sc_how_bound = hb_env' }, bndr')
637 (subst', bndr') = substBndr (sc_subst env) bndr
638 hb_env' = extendVarEnv (sc_how_bound env) bndr' how_bound
640 extendRecBndrs :: ScEnv -> [Var] -> (ScEnv, [Var])
641 extendRecBndrs env bndrs = (env { sc_subst = subst' }, bndrs')
643 (subst', bndrs') = substRecBndrs (sc_subst env) bndrs
645 extendBndr :: ScEnv -> Var -> (ScEnv, Var)
646 extendBndr env bndr = (env { sc_subst = subst' }, bndr')
648 (subst', bndr') = substBndr (sc_subst env) bndr
650 extendValEnv :: ScEnv -> Id -> Maybe Value -> ScEnv
651 extendValEnv env _ Nothing = env
652 extendValEnv env id (Just cv) = env { sc_vals = extendVarEnv (sc_vals env) id cv }
654 extendCaseBndrs :: ScEnv -> Id -> AltCon -> [Var] -> (ScEnv, [Var])
658 -- we want to bind b, to (C x y)
659 -- NB1: Extends only the sc_vals part of the envt
660 -- NB2: Kill the dead-ness info on the pattern binders x,y, since
661 -- they are potentially made alive by the [b -> C x y] binding
662 extendCaseBndrs env case_bndr con alt_bndrs
663 | isDeadBinder case_bndr
666 = (env1, map zap alt_bndrs)
667 -- NB: We used to bind v too, if scrut = (Var v); but
668 -- the simplifer has already done this so it seems
669 -- redundant to do so here
671 -- Var v -> extendValEnv env1 v cval
674 zap v | isTyVar v = v -- See NB2 above
675 | otherwise = zapIdOccInfo v
676 env1 = extendValEnv env case_bndr cval
679 LitAlt {} -> Just (ConVal con [])
680 DataAlt {} -> Just (ConVal con vanilla_args)
682 vanilla_args = map Type (tyConAppArgs (idType case_bndr)) ++
683 varsToCoreExprs alt_bndrs
685 ignoreTyCon :: ScEnv -> TyCon -> Bool
686 ignoreTyCon env tycon
687 = lookupUFM (sc_annotations env) tycon == Just NoSpecConstr
689 ignoreType :: ScEnv -> Type -> Bool
691 = case splitTyConApp_maybe ty of
692 Just (tycon, _) -> ignoreTyCon env tycon
695 ignoreAltCon :: ScEnv -> AltCon -> Bool
696 ignoreAltCon env (DataAlt dc) = ignoreTyCon env (dataConTyCon dc)
697 ignoreAltCon env (LitAlt lit) = ignoreType env (literalType lit)
698 ignoreAltCon _ DEFAULT = True
700 forceSpecBndr :: ScEnv -> Var -> Bool
701 forceSpecBndr env var = forceSpecFunTy env . snd . splitForAllTys . varType $ var
703 forceSpecFunTy :: ScEnv -> Type -> Bool
704 forceSpecFunTy env = any (forceSpecArgTy env) . fst . splitFunTys
706 forceSpecArgTy :: ScEnv -> Type -> Bool
707 forceSpecArgTy env ty
708 | Just ty' <- coreView ty = forceSpecArgTy env ty'
710 forceSpecArgTy env ty
711 | Just (tycon, tys) <- splitTyConApp_maybe ty
713 = lookupUFM (sc_annotations env) tycon == Just ForceSpecConstr
714 || any (forceSpecArgTy env) tys
716 forceSpecArgTy _ _ = False
718 decreaseSpecCount :: ScEnv -> Int -> ScEnv
719 -- See Note [Avoiding exponential blowup]
720 decreaseSpecCount env n_specs
721 = env { sc_count = case sc_count env of
723 Just n -> Just (n `div` (n_specs + 1)) }
724 -- The "+1" takes account of the original function;
725 -- See Note [Avoiding exponential blowup]
728 Note [Avoiding exponential blowup]
729 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
730 The sc_count field of the ScEnv says how many times we are prepared to
731 duplicate a single function. But we must take care with recursive
732 specialiations. Consider
734 let $j1 = let $j2 = let $j3 = ...
742 If we specialise $j1 then in each specialisation (as well as the original)
743 we can specialise $j2, and similarly $j3. Even if we make just *one*
744 specialisation of each, becuase we also have the original we'll get 2^n
745 copies of $j3, which is not good.
747 So when recursively specialising we divide the sc_count by the number of
748 copies we are making at this level, including the original.
751 %************************************************************************
753 \subsection{Usage information: flows upwards}
755 %************************************************************************
760 scu_calls :: CallEnv, -- Calls
761 -- The functions are a subset of the
762 -- RecFuns in the ScEnv
764 scu_occs :: !(IdEnv ArgOcc) -- Information on argument occurrences
765 } -- The domain is OutIds
767 type CallEnv = IdEnv [Call]
768 type Call = (ValueEnv, [CoreArg])
769 -- The arguments of the call, together with the
770 -- env giving the constructor bindings at the call site
773 nullUsage = SCU { scu_calls = emptyVarEnv, scu_occs = emptyVarEnv }
775 combineCalls :: CallEnv -> CallEnv -> CallEnv
776 combineCalls = plusVarEnv_C (++)
778 combineUsage :: ScUsage -> ScUsage -> ScUsage
779 combineUsage u1 u2 = SCU { scu_calls = combineCalls (scu_calls u1) (scu_calls u2),
780 scu_occs = plusVarEnv_C combineOcc (scu_occs u1) (scu_occs u2) }
782 combineUsages :: [ScUsage] -> ScUsage
783 combineUsages [] = nullUsage
784 combineUsages us = foldr1 combineUsage us
786 lookupOcc :: ScUsage -> OutVar -> (ScUsage, ArgOcc)
787 lookupOcc (SCU { scu_calls = sc_calls, scu_occs = sc_occs }) bndr
788 = (SCU {scu_calls = sc_calls, scu_occs = delVarEnv sc_occs bndr},
789 lookupVarEnv sc_occs bndr `orElse` NoOcc)
791 lookupOccs :: ScUsage -> [OutVar] -> (ScUsage, [ArgOcc])
792 lookupOccs (SCU { scu_calls = sc_calls, scu_occs = sc_occs }) bndrs
793 = (SCU {scu_calls = sc_calls, scu_occs = delVarEnvList sc_occs bndrs},
794 [lookupVarEnv sc_occs b `orElse` NoOcc | b <- bndrs])
796 data ArgOcc = NoOcc -- Doesn't occur at all; or a type argument
797 | UnkOcc -- Used in some unknown way
799 | ScrutOcc (UniqFM [ArgOcc]) -- See Note [ScrutOcc]
801 | BothOcc -- Definitely taken apart, *and* perhaps used in some other way
805 An occurrence of ScrutOcc indicates that the thing, or a `cast` version of the thing,
806 is *only* taken apart or applied.
808 Functions, literal: ScrutOcc emptyUFM
809 Data constructors: ScrutOcc subs,
811 where (subs :: UniqFM [ArgOcc]) gives usage of the *pattern-bound* components,
812 The domain of the UniqFM is the Unique of the data constructor
814 The [ArgOcc] is the occurrences of the *pattern-bound* components
815 of the data structure. E.g.
816 data T a = forall b. MkT a b (b->a)
817 A pattern binds b, x::a, y::b, z::b->a, but not 'a'!
821 instance Outputable ArgOcc where
822 ppr (ScrutOcc xs) = ptext (sLit "scrut-occ") <> ppr xs
823 ppr UnkOcc = ptext (sLit "unk-occ")
824 ppr BothOcc = ptext (sLit "both-occ")
825 ppr NoOcc = ptext (sLit "no-occ")
827 -- Experimentally, this vesion of combineOcc makes ScrutOcc "win", so
828 -- that if the thing is scrutinised anywhere then we get to see that
829 -- in the overall result, even if it's also used in a boxed way
830 -- This might be too agressive; see Note [Reboxing] Alternative 3
831 combineOcc :: ArgOcc -> ArgOcc -> ArgOcc
832 combineOcc NoOcc occ = occ
833 combineOcc occ NoOcc = occ
834 combineOcc (ScrutOcc xs) (ScrutOcc ys) = ScrutOcc (plusUFM_C combineOccs xs ys)
835 combineOcc _occ (ScrutOcc ys) = ScrutOcc ys
836 combineOcc (ScrutOcc xs) _occ = ScrutOcc xs
837 combineOcc UnkOcc UnkOcc = UnkOcc
838 combineOcc _ _ = BothOcc
840 combineOccs :: [ArgOcc] -> [ArgOcc] -> [ArgOcc]
841 combineOccs xs ys = zipWithEqual "combineOccs" combineOcc xs ys
843 setScrutOcc :: ScEnv -> ScUsage -> OutExpr -> ArgOcc -> ScUsage
844 -- _Overwrite_ the occurrence info for the scrutinee, if the scrutinee
845 -- is a variable, and an interesting variable
846 setScrutOcc env usg (Cast e _) occ = setScrutOcc env usg e occ
847 setScrutOcc env usg (Note _ e) occ = setScrutOcc env usg e occ
848 setScrutOcc env usg (Var v) occ
849 | Just RecArg <- lookupHowBound env v = usg { scu_occs = extendVarEnv (scu_occs usg) v occ }
851 setScrutOcc _env usg _other _occ -- Catch-all
854 conArgOccs :: ArgOcc -> AltCon -> [ArgOcc]
855 -- Find usage of components of data con; returns [UnkOcc...] if unknown
856 -- See Note [ScrutOcc] for the extra UnkOccs in the vanilla datacon case
858 conArgOccs (ScrutOcc fm) (DataAlt dc)
859 | Just pat_arg_occs <- lookupUFM fm dc
860 = [UnkOcc | _ <- dataConUnivTyVars dc] ++ pat_arg_occs
862 conArgOccs _other _con = repeat UnkOcc
865 %************************************************************************
867 \subsection{The main recursive function}
869 %************************************************************************
871 The main recursive function gathers up usage information, and
872 creates specialised versions of functions.
875 scExpr, scExpr' :: ScEnv -> CoreExpr -> UniqSM (ScUsage, CoreExpr)
876 -- The unique supply is needed when we invent
877 -- a new name for the specialised function and its args
879 scExpr env e = scExpr' env e
882 scExpr' env (Var v) = case scSubstId env v of
883 Var v' -> return (varUsage env v' UnkOcc, Var v')
884 e' -> scExpr (zapScSubst env) e'
886 scExpr' env (Type t) = return (nullUsage, Type (scSubstTy env t))
887 scExpr' _ e@(Lit {}) = return (nullUsage, e)
888 scExpr' env (Note n e) = do (usg,e') <- scExpr env e
889 return (usg, Note n e')
890 scExpr' env (Cast e co) = do (usg, e') <- scExpr env e
891 return (usg, Cast e' (scSubstTy env co))
892 scExpr' env e@(App _ _) = scApp env (collectArgs e)
893 scExpr' env (Lam b e) = do let (env', b') = extendBndr env b
894 (usg, e') <- scExpr env' e
895 return (usg, Lam b' e')
897 scExpr' env (Case scrut b ty alts)
898 = do { (scrut_usg, scrut') <- scExpr env scrut
899 ; case isValue (sc_vals env) scrut' of
900 Just (ConVal con args) -> sc_con_app con args scrut'
901 _other -> sc_vanilla scrut_usg scrut'
904 sc_con_app con args scrut' -- Known constructor; simplify
905 = do { let (_, bs, rhs) = findAlt con alts
906 `orElse` (DEFAULT, [], mkImpossibleExpr (coreAltsType alts))
907 alt_env' = extendScSubstList env ((b,scrut') : bs `zip` trimConArgs con args)
908 ; scExpr alt_env' rhs }
910 sc_vanilla scrut_usg scrut' -- Normal case
911 = do { let (alt_env,b') = extendBndrWith RecArg env b
912 -- Record RecArg for the components
914 ; (alt_usgs, alt_occs, alts')
915 <- mapAndUnzip3M (sc_alt alt_env scrut' b') alts
917 ; let (alt_usg, b_occ) = lookupOcc (combineUsages alt_usgs) b'
918 scrut_occ = foldr combineOcc b_occ alt_occs
919 scrut_usg' = setScrutOcc env scrut_usg scrut' scrut_occ
920 -- The combined usage of the scrutinee is given
921 -- by scrut_occ, which is passed to scScrut, which
922 -- in turn treats a bare-variable scrutinee specially
924 ; return (alt_usg `combineUsage` scrut_usg',
925 Case scrut' b' (scSubstTy env ty) alts') }
927 sc_alt env _scrut' b' (con,bs,rhs)
928 = do { let (env1, bs1) = extendBndrsWith RecArg env bs
929 (env2, bs2) = extendCaseBndrs env1 b' con bs1
930 ; (usg,rhs') <- scExpr env2 rhs
931 ; let (usg', arg_occs) = lookupOccs usg bs2
932 scrut_occ = case con of
933 DataAlt dc -> ScrutOcc (unitUFM dc arg_occs)
934 _ -> ScrutOcc emptyUFM
935 ; return (usg', scrut_occ, (con, bs2, rhs')) }
937 scExpr' env (Let (NonRec bndr rhs) body)
938 | isTyVar bndr -- Type-lets may be created by doBeta
939 = scExpr' (extendScSubst env bndr rhs) body
941 | otherwise -- Note [Local let bindings]
942 = do { let (body_env, bndr') = extendBndr env bndr
943 body_env2 = extendHowBound body_env [bndr'] RecFun
944 ; (body_usg, body') <- scExpr body_env2 body
946 ; (rhs_usg, rhs_info) <- scRecRhs env (bndr',rhs)
948 -- NB: We don't use the ForceSpecConstr mechanism (see
949 -- Note [Forcing specialisation]) for non-recursive bindings
950 -- at the moment. I'm not sure if this is the right thing to do.
951 ; let force_spec = False
952 ; (spec_usg, specs) <- specialise env force_spec
955 (SI [] 0 (Just rhs_usg))
957 ; return (body_usg { scu_calls = scu_calls body_usg `delVarEnv` bndr' }
958 `combineUsage` spec_usg,
959 mkLets [NonRec b r | (b,r) <- specInfoBinds rhs_info specs] body')
963 -- A *local* recursive group: see Note [Local recursive groups]
964 scExpr' env (Let (Rec prs) body)
965 = do { let (bndrs,rhss) = unzip prs
966 (rhs_env1,bndrs') = extendRecBndrs env bndrs
967 rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun
968 force_spec = any (forceSpecBndr env) bndrs'
969 -- Note [Forcing specialisation]
971 ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss)
972 ; (body_usg, body') <- scExpr rhs_env2 body
974 -- NB: start specLoop from body_usg
975 ; (spec_usg, specs) <- specLoop rhs_env2 force_spec
976 (scu_calls body_usg) rhs_infos nullUsage
977 [SI [] 0 (Just usg) | usg <- rhs_usgs]
978 -- Do not unconditionally use rhs_usgs.
979 -- Instead use them only if we find an unspecialised call
980 -- See Note [Local recursive groups]
982 ; let all_usg = spec_usg `combineUsage` body_usg
983 bind' = Rec (concat (zipWith specInfoBinds rhs_infos specs))
985 ; return (all_usg { scu_calls = scu_calls all_usg `delVarEnvList` bndrs' },
989 Note [Local let bindings]
990 ~~~~~~~~~~~~~~~~~~~~~~~~~
991 It is not uncommon to find this
993 let $j = \x. <blah> in ...$j True...$j True...
995 Here $j is an arbitrary let-bound function, but it often comes up for
996 join points. We might like to specialise $j for its call patterns.
997 Notice the difference from a letrec, where we look for call patterns
998 in the *RHS* of the function. Here we look for call patterns in the
1001 At one point I predicated this on the RHS mentioning the outer
1002 recursive function, but that's not essential and might even be
1003 harmful. I'm not sure.
1007 scApp :: ScEnv -> (InExpr, [InExpr]) -> UniqSM (ScUsage, CoreExpr)
1009 scApp env (Var fn, args) -- Function is a variable
1010 = ASSERT( not (null args) )
1011 do { args_w_usgs <- mapM (scExpr env) args
1012 ; let (arg_usgs, args') = unzip args_w_usgs
1013 arg_usg = combineUsages arg_usgs
1014 ; case scSubstId env fn of
1015 fn'@(Lam {}) -> scExpr (zapScSubst env) (doBeta fn' args')
1016 -- Do beta-reduction and try again
1018 Var fn' -> return (arg_usg `combineUsage` fn_usg, mkApps (Var fn') args')
1020 fn_usg = case lookupHowBound env fn' of
1021 Just RecFun -> SCU { scu_calls = unitVarEnv fn' [(sc_vals env, args')],
1022 scu_occs = emptyVarEnv }
1023 Just RecArg -> SCU { scu_calls = emptyVarEnv,
1024 scu_occs = unitVarEnv fn' (ScrutOcc emptyUFM) }
1025 Nothing -> nullUsage
1028 other_fn' -> return (arg_usg, mkApps other_fn' args') }
1029 -- NB: doing this ignores any usage info from the substituted
1030 -- function, but I don't think that matters. If it does
1033 doBeta :: OutExpr -> [OutExpr] -> OutExpr
1034 -- ToDo: adjust for System IF
1035 doBeta (Lam bndr body) (arg : args) = Let (NonRec bndr arg) (doBeta body args)
1036 doBeta fn args = mkApps fn args
1038 -- The function is almost always a variable, but not always.
1039 -- In particular, if this pass follows float-in,
1040 -- which it may, we can get
1041 -- (let f = ...f... in f) arg1 arg2
1042 scApp env (other_fn, args)
1043 = do { (fn_usg, fn') <- scExpr env other_fn
1044 ; (arg_usgs, args') <- mapAndUnzipM (scExpr env) args
1045 ; return (combineUsages arg_usgs `combineUsage` fn_usg, mkApps fn' args') }
1047 ----------------------
1048 scTopBind :: ScEnv -> CoreBind -> UniqSM (ScEnv, CoreBind)
1049 scTopBind env (Rec prs)
1050 | Just threshold <- sc_size env
1052 , not (all (couldBeSmallEnoughToInline threshold) rhss)
1053 -- No specialisation
1054 = do { let (rhs_env,bndrs') = extendRecBndrs env bndrs
1055 ; (_, rhss') <- mapAndUnzipM (scExpr rhs_env) rhss
1056 ; return (rhs_env, Rec (bndrs' `zip` rhss')) }
1057 | otherwise -- Do specialisation
1058 = do { let (rhs_env1,bndrs') = extendRecBndrs env bndrs
1059 rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun
1061 ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss)
1062 ; let rhs_usg = combineUsages rhs_usgs
1064 ; (_, specs) <- specLoop rhs_env2 force_spec
1065 (scu_calls rhs_usg) rhs_infos nullUsage
1066 [SI [] 0 Nothing | _ <- bndrs]
1068 ; return (rhs_env1, -- For the body of the letrec, delete the RecFun business
1069 Rec (concat (zipWith specInfoBinds rhs_infos specs))) }
1071 (bndrs,rhss) = unzip prs
1072 force_spec = any (forceSpecBndr env) bndrs
1073 -- Note [Forcing specialisation]
1075 scTopBind env (NonRec bndr rhs)
1076 = do { (_, rhs') <- scExpr env rhs
1077 ; let (env1, bndr') = extendBndr env bndr
1078 env2 = extendValEnv env1 bndr' (isValue (sc_vals env) rhs')
1079 ; return (env2, NonRec bndr' rhs') }
1081 ----------------------
1082 scRecRhs :: ScEnv -> (OutId, InExpr) -> UniqSM (ScUsage, RhsInfo)
1083 scRecRhs env (bndr,rhs)
1084 = do { let (arg_bndrs,body) = collectBinders rhs
1085 (body_env, arg_bndrs') = extendBndrsWith RecArg env arg_bndrs
1086 ; (body_usg, body') <- scExpr body_env body
1087 ; let (rhs_usg, arg_occs) = lookupOccs body_usg arg_bndrs'
1088 ; return (rhs_usg, RI bndr (mkLams arg_bndrs' body')
1089 arg_bndrs body arg_occs) }
1090 -- The arg_occs says how the visible,
1091 -- lambda-bound binders of the RHS are used
1092 -- (including the TyVar binders)
1093 -- Two pats are the same if they match both ways
1095 ----------------------
1096 specInfoBinds :: RhsInfo -> SpecInfo -> [(Id,CoreExpr)]
1097 specInfoBinds (RI fn new_rhs _ _ _) (SI specs _ _)
1098 = [(id,rhs) | OS _ _ id rhs <- specs] ++
1099 [(fn `addIdSpecialisations` rules, new_rhs)]
1101 rules = [r | OS _ r _ _ <- specs]
1103 ----------------------
1104 varUsage :: ScEnv -> OutVar -> ArgOcc -> ScUsage
1106 | Just RecArg <- lookupHowBound env v = SCU { scu_calls = emptyVarEnv
1107 , scu_occs = unitVarEnv v use }
1108 | otherwise = nullUsage
1112 %************************************************************************
1114 The specialiser itself
1116 %************************************************************************
1119 data RhsInfo = RI OutId -- The binder
1120 OutExpr -- The new RHS
1121 [InVar] InExpr -- The *original* RHS (\xs.body)
1122 -- Note [Specialise original body]
1123 [ArgOcc] -- Info on how the xs occur in body
1125 data SpecInfo = SI [OneSpec] -- The specialisations we have generated
1127 Int -- Length of specs; used for numbering them
1129 (Maybe ScUsage) -- Nothing => we have generated specialisations
1130 -- from calls in the *original* RHS
1131 -- Just cs => we haven't, and this is the usage
1132 -- of the original RHS
1133 -- See Note [Local recursive groups]
1135 -- One specialisation: Rule plus definition
1136 data OneSpec = OS CallPat -- Call pattern that generated this specialisation
1137 CoreRule -- Rule connecting original id with the specialisation
1138 OutId OutExpr -- Spec id + its rhs
1142 -> Bool -- force specialisation?
1143 -- Note [Forcing specialisation]
1146 -> ScUsage -> [SpecInfo] -- One per binder; acccumulating parameter
1147 -> UniqSM (ScUsage, [SpecInfo]) -- ...ditto...
1148 specLoop env force_spec all_calls rhs_infos usg_so_far specs_so_far
1149 = do { specs_w_usg <- zipWithM (specialise env force_spec all_calls) rhs_infos specs_so_far
1150 ; let (new_usg_s, all_specs) = unzip specs_w_usg
1151 new_usg = combineUsages new_usg_s
1152 new_calls = scu_calls new_usg
1153 all_usg = usg_so_far `combineUsage` new_usg
1154 ; if isEmptyVarEnv new_calls then
1155 return (all_usg, all_specs)
1157 specLoop env force_spec new_calls rhs_infos all_usg all_specs }
1161 -> Bool -- force specialisation?
1162 -- Note [Forcing specialisation]
1163 -> CallEnv -- Info on calls
1165 -> SpecInfo -- Original RHS plus patterns dealt with
1166 -> UniqSM (ScUsage, SpecInfo) -- New specialised versions and their usage
1168 -- Note: the rhs here is the optimised version of the original rhs
1169 -- So when we make a specialised copy of the RHS, we're starting
1170 -- from an RHS whose nested functions have been optimised already.
1172 specialise env force_spec bind_calls (RI fn _ arg_bndrs body arg_occs)
1173 spec_info@(SI specs spec_count mb_unspec)
1174 | not (isBottomingId fn) -- Note [Do not specialise diverging functions]
1175 , notNull arg_bndrs -- Only specialise functions
1176 , Just all_calls <- lookupVarEnv bind_calls fn
1177 = do { (boring_call, pats) <- callsToPats env specs arg_occs all_calls
1178 -- ; pprTrace "specialise" (vcat [ ppr fn <+> text "with" <+> int (length pats) <+> text "good patterns"
1179 -- , text "arg_occs" <+> ppr arg_occs
1180 -- , text "calls" <+> ppr all_calls
1181 -- , text "good pats" <+> ppr pats]) $
1184 -- Bale out if too many specialisations
1185 ; let n_pats = length pats
1186 spec_count' = n_pats + spec_count
1187 ; case sc_count env of
1188 Just max | not force_spec && spec_count' > max
1189 -> pprTrace "SpecConstr" msg $
1190 return (nullUsage, spec_info)
1192 msg = vcat [ sep [ ptext (sLit "Function") <+> quotes (ppr fn)
1193 , nest 2 (ptext (sLit "has") <+>
1194 speakNOf spec_count' (ptext (sLit "call pattern")) <> comma <+>
1195 ptext (sLit "but the limit is") <+> int max) ]
1196 , ptext (sLit "Use -fspec-constr-count=n to set the bound")
1198 extra | not opt_PprStyle_Debug = ptext (sLit "Use -dppr-debug to see specialisations")
1199 | otherwise = ptext (sLit "Specialisations:") <+> ppr (pats ++ [p | OS p _ _ _ <- specs])
1201 _normal_case -> do {
1203 let spec_env = decreaseSpecCount env n_pats
1204 ; (spec_usgs, new_specs) <- mapAndUnzipM (spec_one spec_env fn arg_bndrs body)
1205 (pats `zip` [spec_count..])
1206 -- See Note [Specialise original body]
1208 ; let spec_usg = combineUsages spec_usgs
1209 (new_usg, mb_unspec')
1211 Just rhs_usg | boring_call -> (spec_usg `combineUsage` rhs_usg, Nothing)
1212 _ -> (spec_usg, mb_unspec)
1214 ; return (new_usg, SI (new_specs ++ specs) spec_count' mb_unspec') } }
1216 = return (nullUsage, spec_info) -- The boring case
1219 ---------------------
1221 -> OutId -- Function
1222 -> [InVar] -- Lambda-binders of RHS; should match patterns
1223 -> InExpr -- Body of the original function
1225 -> UniqSM (ScUsage, OneSpec) -- Rule and binding
1227 -- spec_one creates a specialised copy of the function, together
1228 -- with a rule for using it. I'm very proud of how short this
1229 -- function is, considering what it does :-).
1235 f = /\b \y::[(a,b)] -> ....f (b,c) ((:) (a,(b,c)) (x,v) (h w))...
1236 [c::*, v::(b,c) are presumably bound by the (...) part]
1238 f_spec = /\ b c \ v::(b,c) hw::[(a,(b,c))] ->
1239 (...entire body of f...) [b -> (b,c),
1240 y -> ((:) (a,(b,c)) (x,v) hw)]
1242 RULE: forall b::* c::*, -- Note, *not* forall a, x
1246 f (b,c) ((:) (a,(b,c)) (x,v) hw) = f_spec b c v hw
1249 spec_one env fn arg_bndrs body (call_pat@(qvars, pats), rule_number)
1250 = do { spec_uniq <- getUniqueUs
1251 ; let spec_env = extendScSubstList (extendScInScope env qvars)
1252 (arg_bndrs `zip` pats)
1254 fn_loc = nameSrcSpan fn_name
1255 spec_occ = mkSpecOcc (nameOccName fn_name)
1256 rule_name = mkFastString ("SC:" ++ showSDoc (ppr fn <> int rule_number))
1257 spec_name = mkInternalName spec_uniq spec_occ fn_loc
1258 -- ; pprTrace "{spec_one" (ppr (sc_count env) <+> ppr fn <+> ppr pats <+> text "-->" <+> ppr spec_name) $
1261 -- Specialise the body
1262 ; (spec_usg, spec_body) <- scExpr spec_env body
1264 -- ; pprTrace "done spec_one}" (ppr fn) $
1267 -- And build the results
1268 ; let spec_id = mkLocalId spec_name (mkPiTypes spec_lam_args body_ty)
1269 `setIdStrictness` spec_str -- See Note [Transfer strictness]
1270 `setIdArity` count isId spec_lam_args
1271 spec_str = calcSpecStrictness fn spec_lam_args pats
1272 (spec_lam_args, spec_call_args) = mkWorkerArgs qvars body_ty
1273 -- Usual w/w hack to avoid generating
1274 -- a spec_rhs of unlifted type and no args
1276 spec_rhs = mkLams spec_lam_args spec_body
1277 body_ty = exprType spec_body
1278 rule_rhs = mkVarApps (Var spec_id) spec_call_args
1279 inline_act = idInlineActivation fn
1280 rule = mkLocalRule rule_name inline_act fn_name qvars pats rule_rhs
1281 ; return (spec_usg, OS call_pat rule spec_id spec_rhs) }
1283 calcSpecStrictness :: Id -- The original function
1284 -> [Var] -> [CoreExpr] -- Call pattern
1285 -> StrictSig -- Strictness of specialised thing
1286 -- See Note [Transfer strictness]
1287 calcSpecStrictness fn qvars pats
1288 = StrictSig (mkTopDmdType spec_dmds TopRes)
1290 spec_dmds = [ lookupVarEnv dmd_env qv `orElse` lazyDmd | qv <- qvars, isId qv ]
1291 StrictSig (DmdType _ dmds _) = idStrictness fn
1293 dmd_env = go emptyVarEnv dmds pats
1295 go env ds (Type {} : pats) = go env ds pats
1296 go env (d:ds) (pat : pats) = go (go_one env d pat) ds pats
1299 go_one env d (Var v) = extendVarEnv_C both env v d
1300 go_one env (Box d) e = go_one env d e
1301 go_one env (Eval (Prod ds)) e
1302 | (Var _, args) <- collectArgs e = go env ds args
1303 go_one env _ _ = env
1307 Note [Specialise original body]
1308 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1309 The RhsInfo for a binding keeps the *original* body of the binding. We
1310 must specialise that, *not* the result of applying specExpr to the RHS
1311 (which is also kept in RhsInfo). Otherwise we end up specialising a
1312 specialised RHS, and that can lead directly to exponential behaviour.
1314 Note [Transfer activation]
1315 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1316 In which phase should the specialise-constructor rules be active?
1317 Originally I made them always-active, but Manuel found that this
1318 defeated some clever user-written rules. Then I made them active only
1319 in Phase 0; after all, currently, the specConstr transformation is
1320 only run after the simplifier has reached Phase 0, but that meant
1321 that specialisations didn't fire inside wrappers; see test
1322 simplCore/should_compile/spec-inline.
1324 So now I just use the inline-activation of the parent Id, as the
1325 activation for the specialiation RULE, just like the main specialiser;
1326 see Note [Auto-specialisation and RULES] in Specialise.
1329 Note [Transfer strictness]
1330 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1331 We must transfer strictness information from the original function to
1332 the specialised one. Suppose, for example
1335 and a RULE f (a:as) b = f_spec a as b
1337 Now we want f_spec to have strictess LLS, otherwise we'll use call-by-need
1338 when calling f_spec instead of call-by-value. And that can result in
1339 unbounded worsening in space (cf the classic foldl vs foldl')
1341 See Trac #3437 for a good example.
1343 The function calcSpecStrictness performs the calculation.
1346 %************************************************************************
1348 \subsection{Argument analysis}
1350 %************************************************************************
1352 This code deals with analysing call-site arguments to see whether
1353 they are constructor applications.
1357 type CallPat = ([Var], [CoreExpr]) -- Quantified variables and arguments
1360 callsToPats :: ScEnv -> [OneSpec] -> [ArgOcc] -> [Call] -> UniqSM (Bool, [CallPat])
1361 -- Result has no duplicate patterns,
1362 -- nor ones mentioned in done_pats
1363 -- Bool indicates that there was at least one boring pattern
1364 callsToPats env done_specs bndr_occs calls
1365 = do { mb_pats <- mapM (callToPats env bndr_occs) calls
1367 ; let good_pats :: [([Var], [CoreArg])]
1368 good_pats = catMaybes mb_pats
1369 done_pats = [p | OS p _ _ _ <- done_specs]
1370 is_done p = any (samePat p) done_pats
1372 ; return (any isNothing mb_pats,
1373 filterOut is_done (nubBy samePat good_pats)) }
1375 callToPats :: ScEnv -> [ArgOcc] -> Call -> UniqSM (Maybe CallPat)
1376 -- The [Var] is the variables to quantify over in the rule
1377 -- Type variables come first, since they may scope
1378 -- over the following term variables
1379 -- The [CoreExpr] are the argument patterns for the rule
1380 callToPats env bndr_occs (con_env, args)
1381 | length args < length bndr_occs -- Check saturated
1384 = do { let in_scope = substInScope (sc_subst env)
1385 ; prs <- argsToPats env in_scope con_env (args `zip` bndr_occs)
1386 ; let (interesting_s, pats) = unzip prs
1387 pat_fvs = varSetElems (exprsFreeVars pats)
1388 qvars = filterOut (`elemInScopeSet` in_scope) pat_fvs
1389 -- Quantify over variables that are not in sccpe
1391 -- See Note [Shadowing] at the top
1393 (tvs, ids) = partition isTyVar qvars
1395 -- Put the type variables first; the type of a term
1396 -- variable may mention a type variable
1398 ; -- pprTrace "callToPats" (ppr args $$ ppr prs $$ ppr bndr_occs) $
1400 then return (Just (qvars', pats))
1401 else return Nothing }
1403 -- argToPat takes an actual argument, and returns an abstracted
1404 -- version, consisting of just the "constructor skeleton" of the
1405 -- argument, with non-constructor sub-expression replaced by new
1406 -- placeholder variables. For example:
1407 -- C a (D (f x) (g y)) ==> C p1 (D p2 p3)
1410 -> InScopeSet -- What's in scope at the fn defn site
1411 -> ValueEnv -- ValueEnv at the call site
1412 -> CoreArg -- A call arg (or component thereof)
1414 -> UniqSM (Bool, CoreArg)
1415 -- Returns (interesting, pat),
1416 -- where pat is the pattern derived from the argument
1417 -- intersting=True if the pattern is non-trivial (not a variable or type)
1418 -- E.g. x:xs --> (True, x:xs)
1419 -- f xs --> (False, w) where w is a fresh wildcard
1420 -- (f xs, 'c') --> (True, (w, 'c')) where w is a fresh wildcard
1421 -- \x. x+y --> (True, \x. x+y)
1422 -- lvl7 --> (True, lvl7) if lvl7 is bound
1423 -- somewhere further out
1425 argToPat _env _in_scope _val_env arg@(Type {}) _arg_occ
1426 = return (False, arg)
1428 argToPat env in_scope val_env (Note _ arg) arg_occ
1429 = argToPat env in_scope val_env arg arg_occ
1430 -- Note [Notes in call patterns]
1431 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1432 -- Ignore Notes. In particular, we want to ignore any InlineMe notes
1433 -- Perhaps we should not ignore profiling notes, but I'm going to
1434 -- ride roughshod over them all for now.
1435 --- See Note [Notes in RULE matching] in Rules
1437 argToPat env in_scope val_env (Let _ arg) arg_occ
1438 = argToPat env in_scope val_env arg arg_occ
1439 -- Look through let expressions
1440 -- e.g. f (let v = rhs in \y -> ...v...)
1441 -- Here we can specialise for f (\y -> ...)
1442 -- because the rule-matcher will look through the let.
1444 argToPat env in_scope val_env (Cast arg co) arg_occ
1445 | not (ignoreType env ty2)
1446 = do { (interesting, arg') <- argToPat env in_scope val_env arg arg_occ
1447 ; if not interesting then
1450 { -- Make a wild-card pattern for the coercion
1452 ; let co_name = mkSysTvName uniq (fsLit "sg")
1453 co_var = mkCoVar co_name (mkCoKind ty1 ty2)
1454 ; return (interesting, Cast arg' (mkTyVarTy co_var)) } }
1456 (ty1, ty2) = coercionKind co
1460 {- Disabling lambda specialisation for now
1461 It's fragile, and the spec_loop can be infinite
1462 argToPat in_scope val_env arg arg_occ
1464 = return (True, arg)
1466 is_value_lam (Lam v e) -- Spot a value lambda, even if
1467 | isId v = True -- it is inside a type lambda
1468 | otherwise = is_value_lam e
1469 is_value_lam other = False
1472 -- Check for a constructor application
1473 -- NB: this *precedes* the Var case, so that we catch nullary constrs
1474 argToPat env in_scope val_env arg arg_occ
1475 | Just (ConVal dc args) <- isValue val_env arg
1476 , not (ignoreAltCon env dc)
1478 ScrutOcc _ -> True -- Used only by case scrutinee
1479 BothOcc -> case arg of -- Used elsewhere
1480 App {} -> True -- see Note [Reboxing]
1482 _other -> False -- No point; the arg is not decomposed
1483 = do { args' <- argsToPats env in_scope val_env (args `zip` conArgOccs arg_occ dc)
1484 ; return (True, mk_con_app dc (map snd args')) }
1486 -- Check if the argument is a variable that
1487 -- is in scope at the function definition site
1488 -- It's worth specialising on this if
1489 -- (a) it's used in an interesting way in the body
1490 -- (b) we know what its value is
1491 argToPat env in_scope val_env (Var v) arg_occ
1492 | case arg_occ of { UnkOcc -> False; _other -> True }, -- (a)
1494 not (ignoreType env (varType v))
1495 = return (True, Var v)
1498 | isLocalId v = v `elemInScopeSet` in_scope
1499 && isJust (lookupVarEnv val_env v)
1500 -- Local variables have values in val_env
1501 | otherwise = isValueUnfolding (idUnfolding v)
1502 -- Imports have unfoldings
1504 -- I'm really not sure what this comment means
1505 -- And by not wild-carding we tend to get forall'd
1506 -- variables that are in soope, which in turn can
1507 -- expose the weakness in let-matching
1508 -- See Note [Matching lets] in Rules
1510 -- Check for a variable bound inside the function.
1511 -- Don't make a wild-card, because we may usefully share
1512 -- e.g. f a = let x = ... in f (x,x)
1513 -- NB: this case follows the lambda and con-app cases!!
1514 -- argToPat _in_scope _val_env (Var v) _arg_occ
1515 -- = return (False, Var v)
1516 -- SLPJ : disabling this to avoid proliferation of versions
1517 -- also works badly when thinking about seeding the loop
1518 -- from the body of the let
1519 -- f x y = letrec g z = ... in g (x,y)
1520 -- We don't want to specialise for that *particular* x,y
1522 -- The default case: make a wild-card
1523 argToPat _env _in_scope _val_env arg _arg_occ
1524 = wildCardPat (exprType arg)
1526 wildCardPat :: Type -> UniqSM (Bool, CoreArg)
1527 wildCardPat ty = do { uniq <- getUniqueUs
1528 ; let id = mkSysLocal (fsLit "sc") uniq ty
1529 ; return (False, Var id) }
1531 argsToPats :: ScEnv -> InScopeSet -> ValueEnv
1532 -> [(CoreArg, ArgOcc)]
1533 -> UniqSM [(Bool, CoreArg)]
1534 argsToPats env in_scope val_env args
1537 do_one (arg,occ) = argToPat env in_scope val_env arg occ
1542 isValue :: ValueEnv -> CoreExpr -> Maybe Value
1543 isValue _env (Lit lit)
1544 = Just (ConVal (LitAlt lit) [])
1547 | Just stuff <- lookupVarEnv env v
1548 = Just stuff -- You might think we could look in the idUnfolding here
1549 -- but that doesn't take account of which branch of a
1550 -- case we are in, which is the whole point
1552 | not (isLocalId v) && isCheapUnfolding unf
1553 = isValue env (unfoldingTemplate unf)
1556 -- However we do want to consult the unfolding
1557 -- as well, for let-bound constructors!
1559 isValue env (Lam b e)
1560 | isTyVar b = case isValue env e of
1561 Just _ -> Just LambdaVal
1563 | otherwise = Just LambdaVal
1565 isValue _env expr -- Maybe it's a constructor application
1566 | (Var fun, args) <- collectArgs expr
1567 = case isDataConWorkId_maybe fun of
1569 Just con | args `lengthAtLeast` dataConRepArity con
1570 -- Check saturated; might be > because the
1571 -- arity excludes type args
1572 -> Just (ConVal (DataAlt con) args)
1574 _other | valArgCount args < idArity fun
1575 -- Under-applied function
1576 -> Just LambdaVal -- Partial application
1580 isValue _env _expr = Nothing
1582 mk_con_app :: AltCon -> [CoreArg] -> CoreExpr
1583 mk_con_app (LitAlt lit) [] = Lit lit
1584 mk_con_app (DataAlt con) args = mkConApp con args
1585 mk_con_app _other _args = panic "SpecConstr.mk_con_app"
1587 samePat :: CallPat -> CallPat -> Bool
1588 samePat (vs1, as1) (vs2, as2)
1591 same (Var v1) (Var v2)
1592 | v1 `elem` vs1 = v2 `elem` vs2
1593 | v2 `elem` vs2 = False
1594 | otherwise = v1 == v2
1596 same (Lit l1) (Lit l2) = l1==l2
1597 same (App f1 a1) (App f2 a2) = same f1 f2 && same a1 a2
1599 same (Type {}) (Type {}) = True -- Note [Ignore type differences]
1600 same (Note _ e1) e2 = same e1 e2 -- Ignore casts and notes
1601 same (Cast e1 _) e2 = same e1 e2
1602 same e1 (Note _ e2) = same e1 e2
1603 same e1 (Cast e2 _) = same e1 e2
1605 same e1 e2 = WARN( bad e1 || bad e2, ppr e1 $$ ppr e2)
1606 False -- Let, lambda, case should not occur
1607 bad (Case {}) = True
1613 Note [Ignore type differences]
1614 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1615 We do not want to generate specialisations where the call patterns
1616 differ only in their type arguments! Not only is it utterly useless,
1617 but it also means that (with polymorphic recursion) we can generate
1618 an infinite number of specialisations. Example is Data.Sequence.adjustTree,