3 1. Use a library type rather than an annotation for ForceSpecConstr
7 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
9 \section[SpecConstr]{Specialise over constructors}
12 -- The above warning supression flag is a temporary kludge.
13 -- While working on this module you are encouraged to remove it and fix
14 -- any warnings in the module. See
15 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
21 , SpecConstrAnnotation(..)
25 #include "HsVersions.h"
30 import CoreUnfold ( couldBeSmallEnoughToInline )
31 import CoreFVs ( exprsFreeVars )
33 import HscTypes ( ModGuts(..) )
34 import WwLib ( mkWorkerArgs )
38 import Type hiding( substTy )
40 import MkCore ( mkImpossibleExpr )
46 import DynFlags ( DynFlags(..) )
47 import StaticFlags ( opt_PprStyle_Debug )
48 import Maybes ( orElse, catMaybes, isJust, isNothing )
50 import DmdAnal ( both )
51 import Serialized ( deserializeWithData )
58 import Control.Monad ( zipWithM )
62 -- See Note [SpecConstrAnnotation]
64 type SpecConstrAnnotation = ()
66 import TyCon ( TyCon )
67 import GHC.Exts( SpecConstrAnnotation(..) )
71 -----------------------------------------------------
73 -----------------------------------------------------
78 drop n (x:xs) = drop (n-1) xs
80 After the first time round, we could pass n unboxed. This happens in
81 numerical code too. Here's what it looks like in Core:
83 drop n xs = case xs of
88 _ -> drop (I# (n# -# 1#)) xs
90 Notice that the recursive call has an explicit constructor as argument.
91 Noticing this, we can make a specialised version of drop
93 RULE: drop (I# n#) xs ==> drop' n# xs
95 drop' n# xs = let n = I# n# in ...orig RHS...
97 Now the simplifier will apply the specialisation in the rhs of drop', giving
99 drop' n# xs = case xs of
103 _ -> drop (n# -# 1#) xs
107 We'd also like to catch cases where a parameter is carried along unchanged,
108 but evaluated each time round the loop:
110 f i n = if i>0 || i>n then i else f (i*2) n
112 Here f isn't strict in n, but we'd like to avoid evaluating it each iteration.
113 In Core, by the time we've w/wd (f is strict in i) we get
115 f i# n = case i# ># 0 of
117 True -> case n of n' { I# n# ->
120 True -> f (i# *# 2#) n'
122 At the call to f, we see that the argument, n is know to be (I# n#),
123 and n is evaluated elsewhere in the body of f, so we can play the same
129 We must be careful not to allocate the same constructor twice. Consider
130 f p = (...(case p of (a,b) -> e)...p...,
131 ...let t = (r,s) in ...t...(f t)...)
132 At the recursive call to f, we can see that t is a pair. But we do NOT want
133 to make a specialised copy:
134 f' a b = let p = (a,b) in (..., ...)
135 because now t is allocated by the caller, then r and s are passed to the
136 recursive call, which allocates the (r,s) pair again.
139 (a) the argument p is used in other than a case-scrutinsation way.
140 (b) the argument to the call is not a 'fresh' tuple; you have to
141 look into its unfolding to see that it's a tuple
143 Hence the "OR" part of Note [Good arguments] below.
145 ALTERNATIVE 2: pass both boxed and unboxed versions. This no longer saves
146 allocation, but does perhaps save evals. In the RULE we'd have
149 f (I# x#) = f' (I# x#) x#
151 If at the call site the (I# x) was an unfolding, then we'd have to
152 rely on CSE to eliminate the duplicate allocation.... This alternative
153 doesn't look attractive enough to pursue.
155 ALTERNATIVE 3: ignore the reboxing problem. The trouble is that
156 the conservative reboxing story prevents many useful functions from being
157 specialised. Example:
158 foo :: Maybe Int -> Int -> Int
160 foo x@(Just m) n = foo x (n-m)
161 Here the use of 'x' will clearly not require boxing in the specialised function.
163 The strictness analyser has the same problem, in fact. Example:
165 If we pass just 'a' and 'b' to the worker, it might need to rebox the
166 pair to create (a,b). A more sophisticated analysis might figure out
167 precisely the cases in which this could happen, but the strictness
168 analyser does no such analysis; it just passes 'a' and 'b', and hopes
171 So my current choice is to make SpecConstr similarly aggressive, and
172 ignore the bad potential of reboxing.
175 Note [Good arguments]
176 ~~~~~~~~~~~~~~~~~~~~~
179 * A self-recursive function. Ignore mutual recursion for now,
180 because it's less common, and the code is simpler for self-recursion.
184 a) At a recursive call, one or more parameters is an explicit
185 constructor application
187 That same parameter is scrutinised by a case somewhere in
188 the RHS of the function
192 b) At a recursive call, one or more parameters has an unfolding
193 that is an explicit constructor application
195 That same parameter is scrutinised by a case somewhere in
196 the RHS of the function
198 Those are the only uses of the parameter (see Note [Reboxing])
201 What to abstract over
202 ~~~~~~~~~~~~~~~~~~~~~
203 There's a bit of a complication with type arguments. If the call
206 f p = ...f ((:) [a] x xs)...
208 then our specialised function look like
210 f_spec x xs = let p = (:) [a] x xs in ....as before....
212 This only makes sense if either
213 a) the type variable 'a' is in scope at the top of f, or
214 b) the type variable 'a' is an argument to f (and hence fs)
216 Actually, (a) may hold for value arguments too, in which case
217 we may not want to pass them. Supose 'x' is in scope at f's
218 defn, but xs is not. Then we'd like
220 f_spec xs = let p = (:) [a] x xs in ....as before....
222 Similarly (b) may hold too. If x is already an argument at the
223 call, no need to pass it again.
225 Finally, if 'a' is not in scope at the call site, we could abstract
226 it as we do the term variables:
228 f_spec a x xs = let p = (:) [a] x xs in ...as before...
230 So the grand plan is:
232 * abstract the call site to a constructor-only pattern
233 e.g. C x (D (f p) (g q)) ==> C s1 (D s2 s3)
235 * Find the free variables of the abstracted pattern
237 * Pass these variables, less any that are in scope at
238 the fn defn. But see Note [Shadowing] below.
241 NOTICE that we only abstract over variables that are not in scope,
242 so we're in no danger of shadowing variables used in "higher up"
248 In this pass we gather up usage information that may mention variables
249 that are bound between the usage site and the definition site; or (more
250 seriously) may be bound to something different at the definition site.
253 f x = letrec g y v = let x = ...
256 Since 'x' is in scope at the call site, we may make a rewrite rule that
258 RULE forall a,b. g (a,b) x = ...
259 But this rule will never match, because it's really a different 'x' at
260 the call site -- and that difference will be manifest by the time the
261 simplifier gets to it. [A worry: the simplifier doesn't *guarantee*
262 no-shadowing, so perhaps it may not be distinct?]
264 Anyway, the rule isn't actually wrong, it's just not useful. One possibility
265 is to run deShadowBinds before running SpecConstr, but instead we run the
266 simplifier. That gives the simplest possible program for SpecConstr to
267 chew on; and it virtually guarantees no shadowing.
269 Note [Specialising for constant parameters]
270 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
271 This one is about specialising on a *constant* (but not necessarily
272 constructor) argument
274 foo :: Int -> (Int -> Int) -> Int
276 foo m f = foo (f m) (+1)
280 lvl_rmV :: GHC.Base.Int -> GHC.Base.Int
282 \ (ds_dlk :: GHC.Base.Int) ->
283 case ds_dlk of wild_alH { GHC.Base.I# x_alG ->
284 GHC.Base.I# (GHC.Prim.+# x_alG 1)
286 T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
289 \ (ww_sme :: GHC.Prim.Int#) (w_smg :: GHC.Base.Int -> GHC.Base.Int) ->
290 case ww_sme of ds_Xlw {
292 case w_smg (GHC.Base.I# ds_Xlw) of w1_Xmo { GHC.Base.I# ww1_Xmz ->
293 T.$wfoo ww1_Xmz lvl_rmV
298 The recursive call has lvl_rmV as its argument, so we could create a specialised copy
299 with that argument baked in; that is, not passed at all. Now it can perhaps be inlined.
301 When is this worth it? Call the constant 'lvl'
302 - If 'lvl' has an unfolding that is a constructor, see if the corresponding
303 parameter is scrutinised anywhere in the body.
305 - If 'lvl' has an unfolding that is a inlinable function, see if the corresponding
306 parameter is applied (...to enough arguments...?)
308 Also do this is if the function has RULES?
312 Note [Specialising for lambda parameters]
313 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
314 foo :: Int -> (Int -> Int) -> Int
316 foo m f = foo (f m) (\n -> n-m)
318 This is subtly different from the previous one in that we get an
319 explicit lambda as the argument:
321 T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
324 \ (ww_sm8 :: GHC.Prim.Int#) (w_sma :: GHC.Base.Int -> GHC.Base.Int) ->
325 case ww_sm8 of ds_Xlr {
327 case w_sma (GHC.Base.I# ds_Xlr) of w1_Xmf { GHC.Base.I# ww1_Xmq ->
330 (\ (n_ad3 :: GHC.Base.Int) ->
331 case n_ad3 of wild_alB { GHC.Base.I# x_alA ->
332 GHC.Base.I# (GHC.Prim.-# x_alA ds_Xlr)
338 I wonder if SpecConstr couldn't be extended to handle this? After all,
339 lambda is a sort of constructor for functions and perhaps it already
340 has most of the necessary machinery?
342 Furthermore, there's an immediate win, because you don't need to allocate the lamda
343 at the call site; and if perchance it's called in the recursive call, then you
344 may avoid allocating it altogether. Just like for constructors.
346 Looks cool, but probably rare...but it might be easy to implement.
349 Note [SpecConstr for casts]
350 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
353 data instance T Int = T Int
358 go (T n) = go (T (n-1))
360 The recursive call ends up looking like
361 go (T (I# ...) `cast` g)
362 So we want to spot the construtor application inside the cast.
363 That's why we have the Cast case in argToPat
365 Note [Local recursive groups]
366 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
367 For a *local* recursive group, we can see all the calls to the
368 function, so we seed the specialisation loop from the calls in the
369 body, not from the calls in the RHS. Consider:
371 bar m n = foo n (n,n) (n,n) (n,n) (n,n)
375 | n > 3000 = case p of { (p1,p2) -> foo (n-1) (p2,p1) q r s }
376 | n > 2000 = case q of { (q1,q2) -> foo (n-1) p (q2,q1) r s }
377 | n > 1000 = case r of { (r1,r2) -> foo (n-1) p q (r2,r1) s }
378 | otherwise = case s of { (s1,s2) -> foo (n-1) p q r (s2,s1) }
380 If we start with the RHSs of 'foo', we get lots and lots of specialisations,
381 most of which are not needed. But if we start with the (single) call
382 in the rhs of 'bar' we get exactly one fully-specialised copy, and all
383 the recursive calls go to this fully-specialised copy. Indeed, the original
384 function is later collected as dead code. This is very important in
385 specialising the loops arising from stream fusion, for example in NDP where
386 we were getting literally hundreds of (mostly unused) specialisations of
389 Note [Do not specialise diverging functions]
390 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
391 Specialising a function that just diverges is a waste of code.
392 Furthermore, it broke GHC (simpl014) thus:
394 f = \x. case x of (a,b) -> f x
395 If we specialise f we get
396 f = \x. case x of (a,b) -> fspec a b
397 But fspec doesn't have decent strictnes info. As it happened,
398 (f x) :: IO t, so the state hack applied and we eta expanded fspec,
399 and hence f. But now f's strictness is less than its arity, which
402 Note [SpecConstrAnnotation]
403 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
404 SpecConstrAnnotation is defined in GHC.Exts, and is only guaranteed to
405 be available in stage 2 (well, until the bootstrap compiler can be
406 guaranteed to have it)
408 So we define it to be () in stage1 (ie when GHCI is undefined), and
409 '#ifdef' out the code that uses it.
411 See also Note [Forcing specialisation]
413 Note [Forcing specialisation]
414 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
415 With stream fusion and in other similar cases, we want to fully specialise
416 some (but not necessarily all!) loops regardless of their size and the
417 number of specialisations. We allow a library to specify this by annotating
418 a type with ForceSpecConstr and then adding a parameter of that type to the
419 loop. Here is a (simplified) example from the vector library:
421 data SPEC = SPEC | SPEC2
422 {-# ANN type SPEC ForceSpecConstr #-}
424 foldl :: (a -> b -> a) -> a -> Stream b -> a
426 foldl f z (Stream step s _) = foldl_loop SPEC z s
428 foldl_loop !sPEC z s = case step s of
429 Yield x s' -> foldl_loop sPEC (f z x) s'
430 Skip -> foldl_loop sPEC z s'
433 SpecConstr will spot the SPEC parameter and always fully specialise
434 foldl_loop. Note that
436 * We have to prevent the SPEC argument from being removed by
437 w/w which is why (a) SPEC is a sum type, and (b) we have to seq on
440 * And lastly, the SPEC argument is ultimately eliminated by
441 SpecConstr itself so there is no runtime overhead.
443 This is all quite ugly; we ought to come up with a better design.
445 ForceSpecConstr arguments are spotted in scExpr' and scTopBinds which then set
446 sc_force to True when calling specLoop. This flag does three things:
447 * Ignore specConstrThreshold, to specialise functions of arbitrary size
449 * Ignore specConstrCount, to make arbitrary numbers of specialisations
451 * Specialise even for arguments that are not scrutinised in the loop
452 (see argToPat; Trac #4488)
454 This flag is inherited for nested non-recursive bindings (which are likely to
455 be join points and hence should be fully specialised) but reset for nested
458 What alternatives did I consider? Annotating the loop itself doesn't
459 work because (a) it is local and (b) it will be w/w'ed and I having
460 w/w propagating annotation somehow doesn't seem like a good idea. The
461 types of the loop arguments really seem to be the most persistent
464 Annotating the types that make up the loop state doesn't work,
465 either, because (a) it would prevent us from using types like Either
466 or tuples here, (b) we don't want to restrict the set of types that
467 can be used in Stream states and (c) some types are fixed by the user
468 (e.g., the accumulator here) but we still want to specialise as much
471 ForceSpecConstr is done by way of an annotation:
472 data SPEC = SPEC | SPEC2
473 {-# ANN type SPEC ForceSpecConstr #-}
474 But SPEC is the *only* type so annotated, so it'd be better to
475 use a particular library type.
477 Alternatives to ForceSpecConstr
478 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
479 Instead of giving the loop an extra argument of type SPEC, we
480 also considered *wrapping* arguments in SPEC, thus
481 data SPEC a = SPEC a | SPEC2
483 loop = \arg -> case arg of
485 case state of (x,y) -> ... loop (SPEC (x',y')) ...
487 The idea is that a SPEC argument says "specialise this argument
488 regardless of whether the function case-analyses it. But this
490 * SPEC must still be a sum type, else the strictness analyser
492 * But that means that 'loop' won't be strict in its real payload
493 This loss of strictness in turn screws up specialisation, because
494 we may end up with calls like
495 loop (SPEC (case z of (p,q) -> (q,p)))
496 Without the SPEC, if 'loop' was strict, the case would move out
497 and we'd see loop applied to a pair. But if 'loop' isn' strict
498 this doesn't look like a specialisable call.
502 The ignoreDataCon stuff allows you to say
503 {-# ANN type T NoSpecConstr #-}
504 to mean "don't specialise on arguments of this type. It was added
505 before we had ForceSpecConstr. Lacking ForceSpecConstr we specialised
506 regardless of size; and then we needed a way to turn that *off*. Now
507 that we have ForceSpecConstr, this NoSpecConstr is probably redundant.
508 (Used only for PArray.)
510 -----------------------------------------------------
511 Stuff not yet handled
512 -----------------------------------------------------
514 Here are notes arising from Roman's work that I don't want to lose.
520 foo :: Int -> T Int -> Int
522 foo x t | even x = case t of { T n -> foo (x-n) t }
523 | otherwise = foo (x-1) t
525 SpecConstr does no specialisation, because the second recursive call
526 looks like a boxed use of the argument. A pity.
528 $wfoo_sFw :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
530 \ (ww_sFo [Just L] :: GHC.Prim.Int#) (w_sFq [Just L] :: T.T GHC.Base.Int) ->
531 case ww_sFo of ds_Xw6 [Just L] {
533 case GHC.Prim.remInt# ds_Xw6 2 of wild1_aEF [Dead Just A] {
534 __DEFAULT -> $wfoo_sFw (GHC.Prim.-# ds_Xw6 1) w_sFq;
536 case w_sFq of wild_Xy [Just L] { T.T n_ad5 [Just U(L)] ->
537 case n_ad5 of wild1_aET [Just A] { GHC.Base.I# y_aES [Just L] ->
538 $wfoo_sFw (GHC.Prim.-# ds_Xw6 y_aES) wild_Xy
544 data a :*: b = !a :*: !b
547 foo :: (Int :*: T Int) -> Int
549 foo (x :*: t) | even x = case t of { T n -> foo ((x-n) :*: t) }
550 | otherwise = foo ((x-1) :*: t)
552 Very similar to the previous one, except that the parameters are now in
553 a strict tuple. Before SpecConstr, we have
555 $wfoo_sG3 :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
557 \ (ww_sFU [Just L] :: GHC.Prim.Int#) (ww_sFW [Just L] :: T.T
559 case ww_sFU of ds_Xws [Just L] {
561 case GHC.Prim.remInt# ds_Xws 2 of wild1_aEZ [Dead Just A] {
563 case ww_sFW of tpl_B2 [Just L] { T.T a_sFo [Just A] ->
564 $wfoo_sG3 (GHC.Prim.-# ds_Xws 1) tpl_B2 -- $wfoo1
567 case ww_sFW of wild_XB [Just A] { T.T n_ad7 [Just S(L)] ->
568 case n_ad7 of wild1_aFd [Just L] { GHC.Base.I# y_aFc [Just L] ->
569 $wfoo_sG3 (GHC.Prim.-# ds_Xws y_aFc) wild_XB -- $wfoo2
573 We get two specialisations:
574 "SC:$wfoo1" [0] __forall {a_sFB :: GHC.Base.Int sc_sGC :: GHC.Prim.Int#}
575 Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int a_sFB)
576 = Foo.$s$wfoo1 a_sFB sc_sGC ;
577 "SC:$wfoo2" [0] __forall {y_aFp :: GHC.Prim.Int# sc_sGC :: GHC.Prim.Int#}
578 Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int (GHC.Base.I# y_aFp))
579 = Foo.$s$wfoo y_aFp sc_sGC ;
581 But perhaps the first one isn't good. After all, we know that tpl_B2 is
582 a T (I# x) really, because T is strict and Int has one constructor. (We can't
583 unbox the strict fields, becuase T is polymorphic!)
585 %************************************************************************
587 \subsection{Top level wrapper stuff}
589 %************************************************************************
592 specConstrProgram :: ModGuts -> CoreM ModGuts
593 specConstrProgram guts
595 dflags <- getDynFlags
596 us <- getUniqueSupplyM
597 annos <- getFirstAnnotations deserializeWithData guts
598 let binds' = fst $ initUs us (go (initScEnv dflags annos) (mg_binds guts))
599 return (guts { mg_binds = binds' })
602 go env (bind:binds) = do (env', bind') <- scTopBind env bind
603 binds' <- go env' binds
604 return (bind' : binds')
608 %************************************************************************
610 \subsection{Environment: goes downwards}
612 %************************************************************************
615 data ScEnv = SCE { sc_size :: Maybe Int, -- Size threshold
616 sc_count :: Maybe Int, -- Max # of specialisations for any one fn
617 -- See Note [Avoiding exponential blowup]
618 sc_force :: Bool, -- Force specialisation?
619 -- See Note [Forcing specialisation]
621 sc_subst :: Subst, -- Current substitution
622 -- Maps InIds to OutExprs
624 sc_how_bound :: HowBoundEnv,
625 -- Binds interesting non-top-level variables
626 -- Domain is OutVars (*after* applying the substitution)
629 -- Domain is OutIds (*after* applying the substitution)
630 -- Used even for top-level bindings (but not imported ones)
632 sc_annotations :: UniqFM SpecConstrAnnotation
635 ---------------------
636 -- As we go, we apply a substitution (sc_subst) to the current term
637 type InExpr = CoreExpr -- _Before_ applying the subst
640 type OutExpr = CoreExpr -- _After_ applying the subst
644 ---------------------
645 type HowBoundEnv = VarEnv HowBound -- Domain is OutVars
647 ---------------------
648 type ValueEnv = IdEnv Value -- Domain is OutIds
649 data Value = ConVal AltCon [CoreArg] -- _Saturated_ constructors
650 -- The AltCon is never DEFAULT
651 | LambdaVal -- Inlinable lambdas or PAPs
653 instance Outputable Value where
654 ppr (ConVal con args) = ppr con <+> interpp'SP args
655 ppr LambdaVal = ptext (sLit "<Lambda>")
657 ---------------------
658 initScEnv :: DynFlags -> UniqFM SpecConstrAnnotation -> ScEnv
659 initScEnv dflags anns
660 = SCE { sc_size = specConstrThreshold dflags,
661 sc_count = specConstrCount dflags,
663 sc_subst = emptySubst,
664 sc_how_bound = emptyVarEnv,
665 sc_vals = emptyVarEnv,
666 sc_annotations = anns }
668 data HowBound = RecFun -- These are the recursive functions for which
669 -- we seek interesting call patterns
671 | RecArg -- These are those functions' arguments, or their sub-components;
672 -- we gather occurrence information for these
674 instance Outputable HowBound where
675 ppr RecFun = text "RecFun"
676 ppr RecArg = text "RecArg"
678 scForce :: ScEnv -> Bool -> ScEnv
679 scForce env b = env { sc_force = b }
681 lookupHowBound :: ScEnv -> Id -> Maybe HowBound
682 lookupHowBound env id = lookupVarEnv (sc_how_bound env) id
684 scSubstId :: ScEnv -> Id -> CoreExpr
685 scSubstId env v = lookupIdSubst (text "scSubstId") (sc_subst env) v
687 scSubstTy :: ScEnv -> Type -> Type
688 scSubstTy env ty = substTy (sc_subst env) ty
690 zapScSubst :: ScEnv -> ScEnv
691 zapScSubst env = env { sc_subst = zapSubstEnv (sc_subst env) }
693 extendScInScope :: ScEnv -> [Var] -> ScEnv
694 -- Bring the quantified variables into scope
695 extendScInScope env qvars = env { sc_subst = extendInScopeList (sc_subst env) qvars }
697 -- Extend the substitution
698 extendScSubst :: ScEnv -> Var -> OutExpr -> ScEnv
699 extendScSubst env var expr = env { sc_subst = extendSubst (sc_subst env) var expr }
701 extendScSubstList :: ScEnv -> [(Var,OutExpr)] -> ScEnv
702 extendScSubstList env prs = env { sc_subst = extendSubstList (sc_subst env) prs }
704 extendHowBound :: ScEnv -> [Var] -> HowBound -> ScEnv
705 extendHowBound env bndrs how_bound
706 = env { sc_how_bound = extendVarEnvList (sc_how_bound env)
707 [(bndr,how_bound) | bndr <- bndrs] }
709 extendBndrsWith :: HowBound -> ScEnv -> [Var] -> (ScEnv, [Var])
710 extendBndrsWith how_bound env bndrs
711 = (env { sc_subst = subst', sc_how_bound = hb_env' }, bndrs')
713 (subst', bndrs') = substBndrs (sc_subst env) bndrs
714 hb_env' = sc_how_bound env `extendVarEnvList`
715 [(bndr,how_bound) | bndr <- bndrs']
717 extendBndrWith :: HowBound -> ScEnv -> Var -> (ScEnv, Var)
718 extendBndrWith how_bound env bndr
719 = (env { sc_subst = subst', sc_how_bound = hb_env' }, bndr')
721 (subst', bndr') = substBndr (sc_subst env) bndr
722 hb_env' = extendVarEnv (sc_how_bound env) bndr' how_bound
724 extendRecBndrs :: ScEnv -> [Var] -> (ScEnv, [Var])
725 extendRecBndrs env bndrs = (env { sc_subst = subst' }, bndrs')
727 (subst', bndrs') = substRecBndrs (sc_subst env) bndrs
729 extendBndr :: ScEnv -> Var -> (ScEnv, Var)
730 extendBndr env bndr = (env { sc_subst = subst' }, bndr')
732 (subst', bndr') = substBndr (sc_subst env) bndr
734 extendValEnv :: ScEnv -> Id -> Maybe Value -> ScEnv
735 extendValEnv env _ Nothing = env
736 extendValEnv env id (Just cv) = env { sc_vals = extendVarEnv (sc_vals env) id cv }
738 extendCaseBndrs :: ScEnv -> OutExpr -> OutId -> AltCon -> [Var] -> (ScEnv, [Var])
742 -- we want to bind b, to (C x y)
743 -- NB1: Extends only the sc_vals part of the envt
744 -- NB2: Kill the dead-ness info on the pattern binders x,y, since
745 -- they are potentially made alive by the [b -> C x y] binding
746 extendCaseBndrs env scrut case_bndr con alt_bndrs
749 live_case_bndr = not (isDeadBinder case_bndr)
750 env1 | Var v <- scrut = extendValEnv env v cval
751 | otherwise = env -- See Note [Add scrutinee to ValueEnv too]
752 env2 | live_case_bndr = extendValEnv env1 case_bndr cval
755 alt_bndrs' | case scrut of { Var {} -> True; _ -> live_case_bndr }
762 LitAlt {} -> Just (ConVal con [])
763 DataAlt {} -> Just (ConVal con vanilla_args)
765 vanilla_args = map Type (tyConAppArgs (idType case_bndr)) ++
766 varsToCoreExprs alt_bndrs
768 zap v | isTyCoVar v = v -- See NB2 above
769 | otherwise = zapIdOccInfo v
772 decreaseSpecCount :: ScEnv -> Int -> ScEnv
773 -- See Note [Avoiding exponential blowup]
774 decreaseSpecCount env n_specs
775 = env { sc_count = case sc_count env of
777 Just n -> Just (n `div` (n_specs + 1)) }
778 -- The "+1" takes account of the original function;
779 -- See Note [Avoiding exponential blowup]
781 ---------------------------------------------------
782 -- See Note [SpecConstrAnnotation]
783 ignoreType :: ScEnv -> Type -> Bool
784 ignoreDataCon :: ScEnv -> DataCon -> Bool
785 forceSpecBndr :: ScEnv -> Var -> Bool
787 ignoreType _ _ = False
788 ignoreDataCon _ _ = False
789 forceSpecBndr _ _ = False
793 ignoreDataCon env dc = ignoreTyCon env (dataConTyCon dc)
796 = case splitTyConApp_maybe ty of
797 Just (tycon, _) -> ignoreTyCon env tycon
800 ignoreTyCon :: ScEnv -> TyCon -> Bool
801 ignoreTyCon env tycon
802 = lookupUFM (sc_annotations env) tycon == Just NoSpecConstr
804 forceSpecBndr env var = forceSpecFunTy env . snd . splitForAllTys . varType $ var
806 forceSpecFunTy :: ScEnv -> Type -> Bool
807 forceSpecFunTy env = any (forceSpecArgTy env) . fst . splitFunTys
809 forceSpecArgTy :: ScEnv -> Type -> Bool
810 forceSpecArgTy env ty
811 | Just ty' <- coreView ty = forceSpecArgTy env ty'
813 forceSpecArgTy env ty
814 | Just (tycon, tys) <- splitTyConApp_maybe ty
816 = lookupUFM (sc_annotations env) tycon == Just ForceSpecConstr
817 || any (forceSpecArgTy env) tys
819 forceSpecArgTy _ _ = False
823 Note [Add scrutinee to ValueEnv too]
824 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
829 By the time we get to the call (f y), the ValueEnv
830 will have a binding for y, and for c
833 BUT that's not enough! Looking at the call (f y) we
834 see that y is pair (a,b), but we also need to know what 'b' is.
835 So in extendCaseBndrs we must *also* add the binding
837 else we lose a useful specialisation for f. This is necessary even
838 though the simplifier has systematically replaced uses of 'x' with 'y'
839 and 'b' with 'c' in the code. The use of 'b' in the ValueEnv came
840 from outside the case. See Trac #4908 for the live example.
842 Note [Avoiding exponential blowup]
843 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
844 The sc_count field of the ScEnv says how many times we are prepared to
845 duplicate a single function. But we must take care with recursive
846 specialiations. Consider
848 let $j1 = let $j2 = let $j3 = ...
856 If we specialise $j1 then in each specialisation (as well as the original)
857 we can specialise $j2, and similarly $j3. Even if we make just *one*
858 specialisation of each, becuase we also have the original we'll get 2^n
859 copies of $j3, which is not good.
861 So when recursively specialising we divide the sc_count by the number of
862 copies we are making at this level, including the original.
865 %************************************************************************
867 \subsection{Usage information: flows upwards}
869 %************************************************************************
874 scu_calls :: CallEnv, -- Calls
875 -- The functions are a subset of the
876 -- RecFuns in the ScEnv
878 scu_occs :: !(IdEnv ArgOcc) -- Information on argument occurrences
879 } -- The domain is OutIds
881 type CallEnv = IdEnv [Call]
882 type Call = (ValueEnv, [CoreArg])
883 -- The arguments of the call, together with the
884 -- env giving the constructor bindings at the call site
887 nullUsage = SCU { scu_calls = emptyVarEnv, scu_occs = emptyVarEnv }
889 combineCalls :: CallEnv -> CallEnv -> CallEnv
890 combineCalls = plusVarEnv_C (++)
892 combineUsage :: ScUsage -> ScUsage -> ScUsage
893 combineUsage u1 u2 = SCU { scu_calls = combineCalls (scu_calls u1) (scu_calls u2),
894 scu_occs = plusVarEnv_C combineOcc (scu_occs u1) (scu_occs u2) }
896 combineUsages :: [ScUsage] -> ScUsage
897 combineUsages [] = nullUsage
898 combineUsages us = foldr1 combineUsage us
900 lookupOccs :: ScUsage -> [OutVar] -> (ScUsage, [ArgOcc])
901 lookupOccs (SCU { scu_calls = sc_calls, scu_occs = sc_occs }) bndrs
902 = (SCU {scu_calls = sc_calls, scu_occs = delVarEnvList sc_occs bndrs},
903 [lookupVarEnv sc_occs b `orElse` NoOcc | b <- bndrs])
905 data ArgOcc = NoOcc -- Doesn't occur at all; or a type argument
906 | UnkOcc -- Used in some unknown way
908 | ScrutOcc -- See Note [ScrutOcc]
909 (DataConEnv [ArgOcc]) -- How the sub-components are used
911 type DataConEnv a = UniqFM a -- Keyed by DataCon
915 An occurrence of ScrutOcc indicates that the thing, or a `cast` version of the thing,
916 is *only* taken apart or applied.
918 Functions, literal: ScrutOcc emptyUFM
919 Data constructors: ScrutOcc subs,
921 where (subs :: UniqFM [ArgOcc]) gives usage of the *pattern-bound* components,
922 The domain of the UniqFM is the Unique of the data constructor
924 The [ArgOcc] is the occurrences of the *pattern-bound* components
925 of the data structure. E.g.
926 data T a = forall b. MkT a b (b->a)
927 A pattern binds b, x::a, y::b, z::b->a, but not 'a'!
931 instance Outputable ArgOcc where
932 ppr (ScrutOcc xs) = ptext (sLit "scrut-occ") <> ppr xs
933 ppr UnkOcc = ptext (sLit "unk-occ")
934 ppr NoOcc = ptext (sLit "no-occ")
936 evalScrutOcc :: ArgOcc
937 evalScrutOcc = ScrutOcc emptyUFM
939 -- Experimentally, this vesion of combineOcc makes ScrutOcc "win", so
940 -- that if the thing is scrutinised anywhere then we get to see that
941 -- in the overall result, even if it's also used in a boxed way
942 -- This might be too agressive; see Note [Reboxing] Alternative 3
943 combineOcc :: ArgOcc -> ArgOcc -> ArgOcc
944 combineOcc NoOcc occ = occ
945 combineOcc occ NoOcc = occ
946 combineOcc (ScrutOcc xs) (ScrutOcc ys) = ScrutOcc (plusUFM_C combineOccs xs ys)
947 combineOcc UnkOcc (ScrutOcc ys) = ScrutOcc ys
948 combineOcc (ScrutOcc xs) UnkOcc = ScrutOcc xs
949 combineOcc UnkOcc UnkOcc = UnkOcc
951 combineOccs :: [ArgOcc] -> [ArgOcc] -> [ArgOcc]
952 combineOccs xs ys = zipWithEqual "combineOccs" combineOcc xs ys
954 setScrutOcc :: ScEnv -> ScUsage -> OutExpr -> ArgOcc -> ScUsage
955 -- _Overwrite_ the occurrence info for the scrutinee, if the scrutinee
956 -- is a variable, and an interesting variable
957 setScrutOcc env usg (Cast e _) occ = setScrutOcc env usg e occ
958 setScrutOcc env usg (Note _ e) occ = setScrutOcc env usg e occ
959 setScrutOcc env usg (Var v) occ
960 | Just RecArg <- lookupHowBound env v = usg { scu_occs = extendVarEnv (scu_occs usg) v occ }
962 setScrutOcc _env usg _other _occ -- Catch-all
966 %************************************************************************
968 \subsection{The main recursive function}
970 %************************************************************************
972 The main recursive function gathers up usage information, and
973 creates specialised versions of functions.
976 scExpr, scExpr' :: ScEnv -> CoreExpr -> UniqSM (ScUsage, CoreExpr)
977 -- The unique supply is needed when we invent
978 -- a new name for the specialised function and its args
980 scExpr env e = scExpr' env e
983 scExpr' env (Var v) = case scSubstId env v of
984 Var v' -> return (varUsage env v' UnkOcc, Var v')
985 e' -> scExpr (zapScSubst env) e'
987 scExpr' env (Type t) = return (nullUsage, Type (scSubstTy env t))
988 scExpr' _ e@(Lit {}) = return (nullUsage, e)
989 scExpr' env (Note n e) = do (usg,e') <- scExpr env e
990 return (usg, Note n e')
991 scExpr' env (Cast e co) = do (usg, e') <- scExpr env e
992 return (usg, Cast e' (scSubstTy env co))
993 scExpr' env e@(App _ _) = scApp env (collectArgs e)
994 scExpr' env (Lam b e) = do let (env', b') = extendBndr env b
995 (usg, e') <- scExpr env' e
996 return (usg, Lam b' e')
998 scExpr' env (Case scrut b ty alts)
999 = do { (scrut_usg, scrut') <- scExpr env scrut
1000 ; case isValue (sc_vals env) scrut' of
1001 Just (ConVal con args) -> sc_con_app con args scrut'
1002 _other -> sc_vanilla scrut_usg scrut'
1005 sc_con_app con args scrut' -- Known constructor; simplify
1006 = do { let (_, bs, rhs) = findAlt con alts
1007 `orElse` (DEFAULT, [], mkImpossibleExpr (coreAltsType alts))
1008 alt_env' = extendScSubstList env ((b,scrut') : bs `zip` trimConArgs con args)
1009 ; scExpr alt_env' rhs }
1011 sc_vanilla scrut_usg scrut' -- Normal case
1012 = do { let (alt_env,b') = extendBndrWith RecArg env b
1013 -- Record RecArg for the components
1015 ; (alt_usgs, alt_occs, alts')
1016 <- mapAndUnzip3M (sc_alt alt_env scrut' b') alts
1018 ; let scrut_occ = foldr1 combineOcc alt_occs -- Never empty
1019 scrut_usg' = setScrutOcc env scrut_usg scrut' scrut_occ
1020 -- The combined usage of the scrutinee is given
1021 -- by scrut_occ, which is passed to scScrut, which
1022 -- in turn treats a bare-variable scrutinee specially
1024 ; return (foldr combineUsage scrut_usg' alt_usgs,
1025 Case scrut' b' (scSubstTy env ty) alts') }
1027 sc_alt env scrut' b' (con,bs,rhs)
1028 = do { let (env1, bs1) = extendBndrsWith RecArg env bs
1029 (env2, bs2) = extendCaseBndrs env1 scrut' b' con bs1
1030 ; (usg, rhs') <- scExpr env2 rhs
1031 ; let (usg', b_occ:arg_occs) = lookupOccs usg (b':bs2)
1032 scrut_occ = case con of
1033 DataAlt dc -> ScrutOcc (unitUFM dc arg_occs)
1034 _ -> ScrutOcc emptyUFM
1035 ; return (usg', b_occ `combineOcc` scrut_occ, (con, bs2, rhs')) }
1037 scExpr' env (Let (NonRec bndr rhs) body)
1038 | isTyCoVar bndr -- Type-lets may be created by doBeta
1039 = scExpr' (extendScSubst env bndr rhs) body
1042 = do { let (body_env, bndr') = extendBndr env bndr
1043 ; (rhs_usg, rhs_info) <- scRecRhs env (bndr',rhs)
1045 ; let body_env2 = extendHowBound body_env [bndr'] RecFun
1046 -- Note [Local let bindings]
1047 RI _ rhs' _ _ _ = rhs_info
1048 body_env3 = extendValEnv body_env2 bndr' (isValue (sc_vals env) rhs')
1050 ; (body_usg, body') <- scExpr body_env3 body
1052 -- NB: For non-recursive bindings we inherit sc_force flag from
1053 -- the parent function (see Note [Forcing specialisation])
1054 ; (spec_usg, specs) <- specialise env
1055 (scu_calls body_usg)
1057 (SI [] 0 (Just rhs_usg))
1059 ; return (body_usg { scu_calls = scu_calls body_usg `delVarEnv` bndr' }
1060 `combineUsage` rhs_usg `combineUsage` spec_usg,
1061 mkLets [NonRec b r | (b,r) <- specInfoBinds rhs_info specs] body')
1065 -- A *local* recursive group: see Note [Local recursive groups]
1066 scExpr' env (Let (Rec prs) body)
1067 = do { let (bndrs,rhss) = unzip prs
1068 (rhs_env1,bndrs') = extendRecBndrs env bndrs
1069 rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun
1070 force_spec = any (forceSpecBndr env) bndrs'
1071 -- Note [Forcing specialisation]
1073 ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss)
1074 ; (body_usg, body') <- scExpr rhs_env2 body
1076 -- NB: start specLoop from body_usg
1077 ; (spec_usg, specs) <- specLoop (scForce rhs_env2 force_spec)
1078 (scu_calls body_usg) rhs_infos nullUsage
1079 [SI [] 0 (Just usg) | usg <- rhs_usgs]
1080 -- Do not unconditionally generate specialisations from rhs_usgs
1081 -- Instead use them only if we find an unspecialised call
1082 -- See Note [Local recursive groups]
1084 ; let rhs_usg = combineUsages rhs_usgs
1085 all_usg = spec_usg `combineUsage` rhs_usg `combineUsage` body_usg
1086 bind' = Rec (concat (zipWith specInfoBinds rhs_infos specs))
1088 ; return (all_usg { scu_calls = scu_calls all_usg `delVarEnvList` bndrs' },
1092 Note [Local let bindings]
1093 ~~~~~~~~~~~~~~~~~~~~~~~~~
1094 It is not uncommon to find this
1096 let $j = \x. <blah> in ...$j True...$j True...
1098 Here $j is an arbitrary let-bound function, but it often comes up for
1099 join points. We might like to specialise $j for its call patterns.
1100 Notice the difference from a letrec, where we look for call patterns
1101 in the *RHS* of the function. Here we look for call patterns in the
1104 At one point I predicated this on the RHS mentioning the outer
1105 recursive function, but that's not essential and might even be
1106 harmful. I'm not sure.
1110 scApp :: ScEnv -> (InExpr, [InExpr]) -> UniqSM (ScUsage, CoreExpr)
1112 scApp env (Var fn, args) -- Function is a variable
1113 = ASSERT( not (null args) )
1114 do { args_w_usgs <- mapM (scExpr env) args
1115 ; let (arg_usgs, args') = unzip args_w_usgs
1116 arg_usg = combineUsages arg_usgs
1117 ; case scSubstId env fn of
1118 fn'@(Lam {}) -> scExpr (zapScSubst env) (doBeta fn' args')
1119 -- Do beta-reduction and try again
1121 Var fn' -> return (arg_usg `combineUsage` mk_fn_usg fn' args',
1122 mkApps (Var fn') args')
1124 other_fn' -> return (arg_usg, mkApps other_fn' args') }
1125 -- NB: doing this ignores any usage info from the substituted
1126 -- function, but I don't think that matters. If it does
1129 doBeta :: OutExpr -> [OutExpr] -> OutExpr
1130 -- ToDo: adjust for System IF
1131 doBeta (Lam bndr body) (arg : args) = Let (NonRec bndr arg) (doBeta body args)
1132 doBeta fn args = mkApps fn args
1135 = case lookupHowBound env fn' of
1136 Just RecFun -> SCU { scu_calls = unitVarEnv fn' [(sc_vals env, args')]
1137 , scu_occs = emptyVarEnv }
1138 Just RecArg -> SCU { scu_calls = emptyVarEnv
1139 , scu_occs = unitVarEnv fn' evalScrutOcc }
1140 Nothing -> nullUsage
1142 -- The function is almost always a variable, but not always.
1143 -- In particular, if this pass follows float-in,
1144 -- which it may, we can get
1145 -- (let f = ...f... in f) arg1 arg2
1146 scApp env (other_fn, args)
1147 = do { (fn_usg, fn') <- scExpr env other_fn
1148 ; (arg_usgs, args') <- mapAndUnzipM (scExpr env) args
1149 ; return (combineUsages arg_usgs `combineUsage` fn_usg, mkApps fn' args') }
1151 ----------------------
1152 scTopBind :: ScEnv -> CoreBind -> UniqSM (ScEnv, CoreBind)
1153 scTopBind env (Rec prs)
1154 | Just threshold <- sc_size env
1156 , not (all (couldBeSmallEnoughToInline threshold) rhss)
1157 -- No specialisation
1158 = do { let (rhs_env,bndrs') = extendRecBndrs env bndrs
1159 ; (_, rhss') <- mapAndUnzipM (scExpr rhs_env) rhss
1160 ; return (rhs_env, Rec (bndrs' `zip` rhss')) }
1161 | otherwise -- Do specialisation
1162 = do { let (rhs_env1,bndrs') = extendRecBndrs env bndrs
1163 rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun
1165 ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss)
1166 ; let rhs_usg = combineUsages rhs_usgs
1168 ; (_, specs) <- specLoop (scForce rhs_env2 force_spec)
1169 (scu_calls rhs_usg) rhs_infos nullUsage
1170 [SI [] 0 Nothing | _ <- bndrs]
1172 ; return (rhs_env1, -- For the body of the letrec, delete the RecFun business
1173 Rec (concat (zipWith specInfoBinds rhs_infos specs))) }
1175 (bndrs,rhss) = unzip prs
1176 force_spec = any (forceSpecBndr env) bndrs
1177 -- Note [Forcing specialisation]
1179 scTopBind env (NonRec bndr rhs)
1180 = do { (_, rhs') <- scExpr env rhs
1181 ; let (env1, bndr') = extendBndr env bndr
1182 env2 = extendValEnv env1 bndr' (isValue (sc_vals env) rhs')
1183 ; return (env2, NonRec bndr' rhs') }
1185 ----------------------
1186 scRecRhs :: ScEnv -> (OutId, InExpr) -> UniqSM (ScUsage, RhsInfo)
1187 scRecRhs env (bndr,rhs)
1188 = do { let (arg_bndrs,body) = collectBinders rhs
1189 (body_env, arg_bndrs') = extendBndrsWith RecArg env arg_bndrs
1190 ; (body_usg, body') <- scExpr body_env body
1191 ; let (rhs_usg, arg_occs) = lookupOccs body_usg arg_bndrs'
1192 ; return (rhs_usg, RI bndr (mkLams arg_bndrs' body')
1193 arg_bndrs body arg_occs) }
1194 -- The arg_occs says how the visible,
1195 -- lambda-bound binders of the RHS are used
1196 -- (including the TyVar binders)
1197 -- Two pats are the same if they match both ways
1199 ----------------------
1200 specInfoBinds :: RhsInfo -> SpecInfo -> [(Id,CoreExpr)]
1201 specInfoBinds (RI fn new_rhs _ _ _) (SI specs _ _)
1202 = [(id,rhs) | OS _ _ id rhs <- specs] ++
1203 -- First the specialised bindings
1205 [(fn `addIdSpecialisations` rules, new_rhs)]
1206 -- And now the original binding
1208 rules = [r | OS _ r _ _ <- specs]
1210 ----------------------
1211 varUsage :: ScEnv -> OutVar -> ArgOcc -> ScUsage
1213 | Just RecArg <- lookupHowBound env v = SCU { scu_calls = emptyVarEnv
1214 , scu_occs = unitVarEnv v use }
1215 | otherwise = nullUsage
1219 %************************************************************************
1221 The specialiser itself
1223 %************************************************************************
1226 data RhsInfo = RI OutId -- The binder
1227 OutExpr -- The new RHS
1228 [InVar] InExpr -- The *original* RHS (\xs.body)
1229 -- Note [Specialise original body]
1230 [ArgOcc] -- Info on how the xs occur in body
1232 data SpecInfo = SI [OneSpec] -- The specialisations we have generated
1234 Int -- Length of specs; used for numbering them
1236 (Maybe ScUsage) -- Nothing => we have generated specialisations
1237 -- from calls in the *original* RHS
1238 -- Just cs => we haven't, and this is the usage
1239 -- of the original RHS
1240 -- See Note [Local recursive groups]
1242 -- One specialisation: Rule plus definition
1243 data OneSpec = OS CallPat -- Call pattern that generated this specialisation
1244 CoreRule -- Rule connecting original id with the specialisation
1245 OutId OutExpr -- Spec id + its rhs
1251 -> ScUsage -> [SpecInfo] -- One per binder; acccumulating parameter
1252 -> UniqSM (ScUsage, [SpecInfo]) -- ...ditto...
1254 specLoop env all_calls rhs_infos usg_so_far specs_so_far
1255 = do { specs_w_usg <- zipWithM (specialise env all_calls) rhs_infos specs_so_far
1256 ; let (new_usg_s, all_specs) = unzip specs_w_usg
1257 new_usg = combineUsages new_usg_s
1258 new_calls = scu_calls new_usg
1259 all_usg = usg_so_far `combineUsage` new_usg
1260 ; if isEmptyVarEnv new_calls then
1261 return (all_usg, all_specs)
1263 specLoop env new_calls rhs_infos all_usg all_specs }
1267 -> CallEnv -- Info on calls
1269 -> SpecInfo -- Original RHS plus patterns dealt with
1270 -> UniqSM (ScUsage, SpecInfo) -- New specialised versions and their usage
1272 -- Note: this only generates *specialised* bindings
1273 -- The original binding is added by specInfoBinds
1275 -- Note: the rhs here is the optimised version of the original rhs
1276 -- So when we make a specialised copy of the RHS, we're starting
1277 -- from an RHS whose nested functions have been optimised already.
1279 specialise env bind_calls (RI fn _ arg_bndrs body arg_occs)
1280 spec_info@(SI specs spec_count mb_unspec)
1281 | not (isBottomingId fn) -- Note [Do not specialise diverging functions]
1282 , not (isNeverActive (idInlineActivation fn)) -- See Note [Transfer activation]
1283 , notNull arg_bndrs -- Only specialise functions
1284 , Just all_calls <- lookupVarEnv bind_calls fn
1285 = do { (boring_call, pats) <- callsToPats env specs arg_occs all_calls
1286 -- ; pprTrace "specialise" (vcat [ ppr fn <+> text "with" <+> int (length pats) <+> text "good patterns"
1287 -- , text "arg_occs" <+> ppr arg_occs
1288 -- , text "calls" <+> ppr all_calls
1289 -- , text "good pats" <+> ppr pats]) $
1292 -- Bale out if too many specialisations
1293 ; let n_pats = length pats
1294 spec_count' = n_pats + spec_count
1295 ; case sc_count env of
1296 Just max | not (sc_force env) && spec_count' > max
1297 -> pprTrace "SpecConstr" msg $
1298 return (nullUsage, spec_info)
1300 msg = vcat [ sep [ ptext (sLit "Function") <+> quotes (ppr fn)
1301 , nest 2 (ptext (sLit "has") <+>
1302 speakNOf spec_count' (ptext (sLit "call pattern")) <> comma <+>
1303 ptext (sLit "but the limit is") <+> int max) ]
1304 , ptext (sLit "Use -fspec-constr-count=n to set the bound")
1306 extra | not opt_PprStyle_Debug = ptext (sLit "Use -dppr-debug to see specialisations")
1307 | otherwise = ptext (sLit "Specialisations:") <+> ppr (pats ++ [p | OS p _ _ _ <- specs])
1309 _normal_case -> do {
1311 let spec_env = decreaseSpecCount env n_pats
1312 ; (spec_usgs, new_specs) <- mapAndUnzipM (spec_one spec_env fn arg_bndrs body)
1313 (pats `zip` [spec_count..])
1314 -- See Note [Specialise original body]
1316 ; let spec_usg = combineUsages spec_usgs
1317 (new_usg, mb_unspec')
1319 Just rhs_usg | boring_call -> (spec_usg `combineUsage` rhs_usg, Nothing)
1320 _ -> (spec_usg, mb_unspec)
1322 ; return (new_usg, SI (new_specs ++ specs) spec_count' mb_unspec') } }
1324 = return (nullUsage, spec_info) -- The boring case
1327 ---------------------
1329 -> OutId -- Function
1330 -> [InVar] -- Lambda-binders of RHS; should match patterns
1331 -> InExpr -- Body of the original function
1333 -> UniqSM (ScUsage, OneSpec) -- Rule and binding
1335 -- spec_one creates a specialised copy of the function, together
1336 -- with a rule for using it. I'm very proud of how short this
1337 -- function is, considering what it does :-).
1343 f = /\b \y::[(a,b)] -> ....f (b,c) ((:) (a,(b,c)) (x,v) (h w))...
1344 [c::*, v::(b,c) are presumably bound by the (...) part]
1346 f_spec = /\ b c \ v::(b,c) hw::[(a,(b,c))] ->
1347 (...entire body of f...) [b -> (b,c),
1348 y -> ((:) (a,(b,c)) (x,v) hw)]
1350 RULE: forall b::* c::*, -- Note, *not* forall a, x
1354 f (b,c) ((:) (a,(b,c)) (x,v) hw) = f_spec b c v hw
1357 spec_one env fn arg_bndrs body (call_pat@(qvars, pats), rule_number)
1358 = do { spec_uniq <- getUniqueUs
1359 ; let spec_env = extendScSubstList (extendScInScope env qvars)
1360 (arg_bndrs `zip` pats)
1362 fn_loc = nameSrcSpan fn_name
1363 spec_occ = mkSpecOcc (nameOccName fn_name)
1364 rule_name = mkFastString ("SC:" ++ showSDoc (ppr fn <> int rule_number))
1365 spec_name = mkInternalName spec_uniq spec_occ fn_loc
1366 -- ; pprTrace "{spec_one" (ppr (sc_count env) <+> ppr fn <+> ppr pats <+> text "-->" <+> ppr spec_name) $
1369 -- Specialise the body
1370 ; (spec_usg, spec_body) <- scExpr spec_env body
1372 -- ; pprTrace "done spec_one}" (ppr fn) $
1375 -- And build the results
1376 ; let spec_id = mkLocalId spec_name (mkPiTypes spec_lam_args body_ty)
1377 `setIdStrictness` spec_str -- See Note [Transfer strictness]
1378 `setIdArity` count isId spec_lam_args
1379 spec_str = calcSpecStrictness fn spec_lam_args pats
1380 (spec_lam_args, spec_call_args) = mkWorkerArgs qvars body_ty
1381 -- Usual w/w hack to avoid generating
1382 -- a spec_rhs of unlifted type and no args
1384 spec_rhs = mkLams spec_lam_args spec_body
1385 body_ty = exprType spec_body
1386 rule_rhs = mkVarApps (Var spec_id) spec_call_args
1387 inline_act = idInlineActivation fn
1388 rule = mkRule True {- Auto -} True {- Local -}
1389 rule_name inline_act fn_name qvars pats rule_rhs
1390 -- See Note [Transfer activation]
1391 ; return (spec_usg, OS call_pat rule spec_id spec_rhs) }
1393 calcSpecStrictness :: Id -- The original function
1394 -> [Var] -> [CoreExpr] -- Call pattern
1395 -> StrictSig -- Strictness of specialised thing
1396 -- See Note [Transfer strictness]
1397 calcSpecStrictness fn qvars pats
1398 = StrictSig (mkTopDmdType spec_dmds TopRes)
1400 spec_dmds = [ lookupVarEnv dmd_env qv `orElse` lazyDmd | qv <- qvars, isId qv ]
1401 StrictSig (DmdType _ dmds _) = idStrictness fn
1403 dmd_env = go emptyVarEnv dmds pats
1405 go env ds (Type {} : pats) = go env ds pats
1406 go env (d:ds) (pat : pats) = go (go_one env d pat) ds pats
1409 go_one env d (Var v) = extendVarEnv_C both env v d
1410 go_one env (Box d) e = go_one env d e
1411 go_one env (Eval (Prod ds)) e
1412 | (Var _, args) <- collectArgs e = go env ds args
1413 go_one env _ _ = env
1417 Note [Specialise original body]
1418 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1419 The RhsInfo for a binding keeps the *original* body of the binding. We
1420 must specialise that, *not* the result of applying specExpr to the RHS
1421 (which is also kept in RhsInfo). Otherwise we end up specialising a
1422 specialised RHS, and that can lead directly to exponential behaviour.
1424 Note [Transfer activation]
1425 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1426 This note is for SpecConstr, but exactly the same thing
1427 happens in the overloading specialiser; see
1428 Note [Auto-specialisation and RULES] in Specialise.
1430 In which phase should the specialise-constructor rules be active?
1431 Originally I made them always-active, but Manuel found that this
1432 defeated some clever user-written rules. Then I made them active only
1433 in Phase 0; after all, currently, the specConstr transformation is
1434 only run after the simplifier has reached Phase 0, but that meant
1435 that specialisations didn't fire inside wrappers; see test
1436 simplCore/should_compile/spec-inline.
1438 So now I just use the inline-activation of the parent Id, as the
1439 activation for the specialiation RULE, just like the main specialiser;
1441 This in turn means there is no point in specialising NOINLINE things,
1442 so we test for that.
1444 Note [Transfer strictness]
1445 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1446 We must transfer strictness information from the original function to
1447 the specialised one. Suppose, for example
1450 and a RULE f (a:as) b = f_spec a as b
1452 Now we want f_spec to have strictess LLS, otherwise we'll use call-by-need
1453 when calling f_spec instead of call-by-value. And that can result in
1454 unbounded worsening in space (cf the classic foldl vs foldl')
1456 See Trac #3437 for a good example.
1458 The function calcSpecStrictness performs the calculation.
1461 %************************************************************************
1463 \subsection{Argument analysis}
1465 %************************************************************************
1467 This code deals with analysing call-site arguments to see whether
1468 they are constructor applications.
1472 type CallPat = ([Var], [CoreExpr]) -- Quantified variables and arguments
1474 callsToPats :: ScEnv -> [OneSpec] -> [ArgOcc] -> [Call] -> UniqSM (Bool, [CallPat])
1475 -- Result has no duplicate patterns,
1476 -- nor ones mentioned in done_pats
1477 -- Bool indicates that there was at least one boring pattern
1478 callsToPats env done_specs bndr_occs calls
1479 = do { mb_pats <- mapM (callToPats env bndr_occs) calls
1481 ; let good_pats :: [CallPat]
1482 good_pats = catMaybes mb_pats
1483 done_pats = [p | OS p _ _ _ <- done_specs]
1484 is_done p = any (samePat p) done_pats
1486 ; return (any isNothing mb_pats,
1487 filterOut is_done (nubBy samePat good_pats)) }
1489 callToPats :: ScEnv -> [ArgOcc] -> Call -> UniqSM (Maybe CallPat)
1490 -- The [Var] is the variables to quantify over in the rule
1491 -- Type variables come first, since they may scope
1492 -- over the following term variables
1493 -- The [CoreExpr] are the argument patterns for the rule
1494 callToPats env bndr_occs (con_env, args)
1495 | length args < length bndr_occs -- Check saturated
1498 = do { let in_scope = substInScope (sc_subst env)
1499 ; (interesting, pats) <- argsToPats env in_scope con_env args bndr_occs
1500 ; let pat_fvs = varSetElems (exprsFreeVars pats)
1501 qvars = filterOut (`elemInScopeSet` in_scope) pat_fvs
1502 -- Quantify over variables that are not in sccpe
1504 -- See Note [Shadowing] at the top
1506 (tvs, ids) = partition isTyCoVar qvars
1508 -- Put the type variables first; the type of a term
1509 -- variable may mention a type variable
1511 ; -- pprTrace "callToPats" (ppr args $$ ppr prs $$ ppr bndr_occs) $
1513 then return (Just (qvars', pats))
1514 else return Nothing }
1516 -- argToPat takes an actual argument, and returns an abstracted
1517 -- version, consisting of just the "constructor skeleton" of the
1518 -- argument, with non-constructor sub-expression replaced by new
1519 -- placeholder variables. For example:
1520 -- C a (D (f x) (g y)) ==> C p1 (D p2 p3)
1523 -> InScopeSet -- What's in scope at the fn defn site
1524 -> ValueEnv -- ValueEnv at the call site
1525 -> CoreArg -- A call arg (or component thereof)
1527 -> UniqSM (Bool, CoreArg)
1529 -- Returns (interesting, pat),
1530 -- where pat is the pattern derived from the argument
1531 -- interesting=True if the pattern is non-trivial (not a variable or type)
1532 -- E.g. x:xs --> (True, x:xs)
1533 -- f xs --> (False, w) where w is a fresh wildcard
1534 -- (f xs, 'c') --> (True, (w, 'c')) where w is a fresh wildcard
1535 -- \x. x+y --> (True, \x. x+y)
1536 -- lvl7 --> (True, lvl7) if lvl7 is bound
1537 -- somewhere further out
1539 argToPat _env _in_scope _val_env arg@(Type {}) _arg_occ
1540 = return (False, arg)
1542 argToPat env in_scope val_env (Note _ arg) arg_occ
1543 = argToPat env in_scope val_env arg arg_occ
1544 -- Note [Notes in call patterns]
1545 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1546 -- Ignore Notes. In particular, we want to ignore any InlineMe notes
1547 -- Perhaps we should not ignore profiling notes, but I'm going to
1548 -- ride roughshod over them all for now.
1549 --- See Note [Notes in RULE matching] in Rules
1551 argToPat env in_scope val_env (Let _ arg) arg_occ
1552 = argToPat env in_scope val_env arg arg_occ
1553 -- See Note [Matching lets] in Rule.lhs
1554 -- Look through let expressions
1555 -- e.g. f (let v = rhs in (v,w))
1556 -- Here we can specialise for f (v,w)
1557 -- because the rule-matcher will look through the let.
1559 {- Disabled; see Note [Matching cases] in Rule.lhs
1560 argToPat env in_scope val_env (Case scrut _ _ [(_, _, rhs)]) arg_occ
1561 | exprOkForSpeculation scrut -- See Note [Matching cases] in Rule.hhs
1562 = argToPat env in_scope val_env rhs arg_occ
1565 argToPat env in_scope val_env (Cast arg co) arg_occ
1566 | isIdentityCoercion co -- Substitution in the SpecConstr itself
1567 -- can lead to identity coercions
1568 = argToPat env in_scope val_env arg arg_occ
1569 | not (ignoreType env ty2)
1570 = do { (interesting, arg') <- argToPat env in_scope val_env arg arg_occ
1571 ; if not interesting then
1574 { -- Make a wild-card pattern for the coercion
1576 ; let co_name = mkSysTvName uniq (fsLit "sg")
1577 co_var = mkCoVar co_name (mkCoKind ty1 ty2)
1578 ; return (interesting, Cast arg' (mkTyVarTy co_var)) } }
1580 (ty1, ty2) = coercionKind co
1584 {- Disabling lambda specialisation for now
1585 It's fragile, and the spec_loop can be infinite
1586 argToPat in_scope val_env arg arg_occ
1588 = return (True, arg)
1590 is_value_lam (Lam v e) -- Spot a value lambda, even if
1591 | isId v = True -- it is inside a type lambda
1592 | otherwise = is_value_lam e
1593 is_value_lam other = False
1596 -- Check for a constructor application
1597 -- NB: this *precedes* the Var case, so that we catch nullary constrs
1598 argToPat env in_scope val_env arg arg_occ
1599 | Just (ConVal (DataAlt dc) args) <- isValue val_env arg
1600 , not (ignoreDataCon env dc) -- See Note [NoSpecConstr]
1601 , Just arg_occs <- mb_scrut dc
1602 = do { let (ty_args, rest_args) = splitAtList (dataConUnivTyVars dc) args
1603 ; (_, args') <- argsToPats env in_scope val_env rest_args arg_occs
1605 mkConApp dc (ty_args ++ args')) }
1607 mb_scrut dc = case arg_occ of
1609 | Just occs <- lookupUFM bs dc
1610 -> Just (occs) -- See Note [Reboxing]
1611 _other | sc_force env -> Just (repeat UnkOcc)
1612 | otherwise -> Nothing
1614 -- Check if the argument is a variable that
1615 -- (a) is used in an interesting way in the body
1616 -- (b) we know what its value is
1617 -- In that case it counts as "interesting"
1618 argToPat env in_scope val_env (Var v) arg_occ
1619 | sc_force env || case arg_occ of { UnkOcc -> False; _other -> True }, -- (a)
1621 not (ignoreType env (varType v))
1622 = return (True, Var v)
1625 | isLocalId v = v `elemInScopeSet` in_scope
1626 && isJust (lookupVarEnv val_env v)
1627 -- Local variables have values in val_env
1628 | otherwise = isValueUnfolding (idUnfolding v)
1629 -- Imports have unfoldings
1631 -- I'm really not sure what this comment means
1632 -- And by not wild-carding we tend to get forall'd
1633 -- variables that are in soope, which in turn can
1634 -- expose the weakness in let-matching
1635 -- See Note [Matching lets] in Rules
1637 -- Check for a variable bound inside the function.
1638 -- Don't make a wild-card, because we may usefully share
1639 -- e.g. f a = let x = ... in f (x,x)
1640 -- NB: this case follows the lambda and con-app cases!!
1641 -- argToPat _in_scope _val_env (Var v) _arg_occ
1642 -- = return (False, Var v)
1643 -- SLPJ : disabling this to avoid proliferation of versions
1644 -- also works badly when thinking about seeding the loop
1645 -- from the body of the let
1646 -- f x y = letrec g z = ... in g (x,y)
1647 -- We don't want to specialise for that *particular* x,y
1649 -- The default case: make a wild-card
1650 argToPat _env _in_scope _val_env arg _arg_occ
1651 = wildCardPat (exprType arg)
1653 wildCardPat :: Type -> UniqSM (Bool, CoreArg)
1655 = do { uniq <- getUniqueUs
1656 ; let id = mkSysLocal (fsLit "sc") uniq ty
1657 ; return (False, Var id) }
1659 argsToPats :: ScEnv -> InScopeSet -> ValueEnv
1660 -> [CoreArg] -> [ArgOcc] -- Should be same length
1661 -> UniqSM (Bool, [CoreArg])
1662 argsToPats env in_scope val_env args occs
1663 = do { stuff <- zipWithM (argToPat env in_scope val_env) args occs
1664 ; let (interesting_s, args') = unzip stuff
1665 ; return (or interesting_s, args') }
1670 isValue :: ValueEnv -> CoreExpr -> Maybe Value
1671 isValue _env (Lit lit)
1672 = Just (ConVal (LitAlt lit) [])
1675 | Just stuff <- lookupVarEnv env v
1676 = Just stuff -- You might think we could look in the idUnfolding here
1677 -- but that doesn't take account of which branch of a
1678 -- case we are in, which is the whole point
1680 | not (isLocalId v) && isCheapUnfolding unf
1681 = isValue env (unfoldingTemplate unf)
1684 -- However we do want to consult the unfolding
1685 -- as well, for let-bound constructors!
1687 isValue env (Lam b e)
1688 | isTyCoVar b = case isValue env e of
1689 Just _ -> Just LambdaVal
1691 | otherwise = Just LambdaVal
1693 isValue _env expr -- Maybe it's a constructor application
1694 | (Var fun, args) <- collectArgs expr
1695 = case isDataConWorkId_maybe fun of
1697 Just con | args `lengthAtLeast` dataConRepArity con
1698 -- Check saturated; might be > because the
1699 -- arity excludes type args
1700 -> Just (ConVal (DataAlt con) args)
1702 _other | valArgCount args < idArity fun
1703 -- Under-applied function
1704 -> Just LambdaVal -- Partial application
1708 isValue _env _expr = Nothing
1710 samePat :: CallPat -> CallPat -> Bool
1711 samePat (vs1, as1) (vs2, as2)
1714 same (Var v1) (Var v2)
1715 | v1 `elem` vs1 = v2 `elem` vs2
1716 | v2 `elem` vs2 = False
1717 | otherwise = v1 == v2
1719 same (Lit l1) (Lit l2) = l1==l2
1720 same (App f1 a1) (App f2 a2) = same f1 f2 && same a1 a2
1722 same (Type {}) (Type {}) = True -- Note [Ignore type differences]
1723 same (Note _ e1) e2 = same e1 e2 -- Ignore casts and notes
1724 same (Cast e1 _) e2 = same e1 e2
1725 same e1 (Note _ e2) = same e1 e2
1726 same e1 (Cast e2 _) = same e1 e2
1728 same e1 e2 = WARN( bad e1 || bad e2, ppr e1 $$ ppr e2)
1729 False -- Let, lambda, case should not occur
1730 bad (Case {}) = True
1736 Note [Ignore type differences]
1737 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1738 We do not want to generate specialisations where the call patterns
1739 differ only in their type arguments! Not only is it utterly useless,
1740 but it also means that (with polymorphic recursion) we can generate
1741 an infinite number of specialisations. Example is Data.Sequence.adjustTree,