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
17 #include "HsVersions.h"
22 import CoreUnfold ( couldBeSmallEnoughToInline )
23 import CoreFVs ( exprsFreeVars )
24 import WwLib ( mkWorkerArgs )
25 import DataCon ( dataConRepArity, dataConUnivTyVars )
28 import Type hiding( substTy )
34 import OccName ( mkSpecOcc )
35 import DynFlags ( DynFlags(..) )
36 import StaticFlags ( opt_PprStyle_Debug )
37 import StaticFlags ( opt_SpecInlineJoinPoints )
38 import BasicTypes ( Activation(..) )
39 import Maybes ( orElse, catMaybes, isJust, isNothing )
41 import List ( nubBy, partition )
47 import Control.Monad ( zipWithM )
50 -----------------------------------------------------
52 -----------------------------------------------------
57 drop n (x:xs) = drop (n-1) xs
59 After the first time round, we could pass n unboxed. This happens in
60 numerical code too. Here's what it looks like in Core:
62 drop n xs = case xs of
67 _ -> drop (I# (n# -# 1#)) xs
69 Notice that the recursive call has an explicit constructor as argument.
70 Noticing this, we can make a specialised version of drop
72 RULE: drop (I# n#) xs ==> drop' n# xs
74 drop' n# xs = let n = I# n# in ...orig RHS...
76 Now the simplifier will apply the specialisation in the rhs of drop', giving
78 drop' n# xs = case xs of
82 _ -> drop (n# -# 1#) xs
86 We'd also like to catch cases where a parameter is carried along unchanged,
87 but evaluated each time round the loop:
89 f i n = if i>0 || i>n then i else f (i*2) n
91 Here f isn't strict in n, but we'd like to avoid evaluating it each iteration.
92 In Core, by the time we've w/wd (f is strict in i) we get
94 f i# n = case i# ># 0 of
96 True -> case n of n' { I# n# ->
99 True -> f (i# *# 2#) n'
101 At the call to f, we see that the argument, n is know to be (I# n#),
102 and n is evaluated elsewhere in the body of f, so we can play the same
108 We must be careful not to allocate the same constructor twice. Consider
109 f p = (...(case p of (a,b) -> e)...p...,
110 ...let t = (r,s) in ...t...(f t)...)
111 At the recursive call to f, we can see that t is a pair. But we do NOT want
112 to make a specialised copy:
113 f' a b = let p = (a,b) in (..., ...)
114 because now t is allocated by the caller, then r and s are passed to the
115 recursive call, which allocates the (r,s) pair again.
118 (a) the argument p is used in other than a case-scrutinsation way.
119 (b) the argument to the call is not a 'fresh' tuple; you have to
120 look into its unfolding to see that it's a tuple
122 Hence the "OR" part of Note [Good arguments] below.
124 ALTERNATIVE 2: pass both boxed and unboxed versions. This no longer saves
125 allocation, but does perhaps save evals. In the RULE we'd have
128 f (I# x#) = f' (I# x#) x#
130 If at the call site the (I# x) was an unfolding, then we'd have to
131 rely on CSE to eliminate the duplicate allocation.... This alternative
132 doesn't look attractive enough to pursue.
134 ALTERNATIVE 3: ignore the reboxing problem. The trouble is that
135 the conservative reboxing story prevents many useful functions from being
136 specialised. Example:
137 foo :: Maybe Int -> Int -> Int
139 foo x@(Just m) n = foo x (n-m)
140 Here the use of 'x' will clearly not require boxing in the specialised function.
142 The strictness analyser has the same problem, in fact. Example:
144 If we pass just 'a' and 'b' to the worker, it might need to rebox the
145 pair to create (a,b). A more sophisticated analysis might figure out
146 precisely the cases in which this could happen, but the strictness
147 analyser does no such analysis; it just passes 'a' and 'b', and hopes
150 So my current choice is to make SpecConstr similarly aggressive, and
151 ignore the bad potential of reboxing.
154 Note [Good arguments]
155 ~~~~~~~~~~~~~~~~~~~~~
158 * A self-recursive function. Ignore mutual recursion for now,
159 because it's less common, and the code is simpler for self-recursion.
163 a) At a recursive call, one or more parameters is an explicit
164 constructor application
166 That same parameter is scrutinised by a case somewhere in
167 the RHS of the function
171 b) At a recursive call, one or more parameters has an unfolding
172 that is an explicit constructor application
174 That same parameter is scrutinised by a case somewhere in
175 the RHS of the function
177 Those are the only uses of the parameter (see Note [Reboxing])
180 What to abstract over
181 ~~~~~~~~~~~~~~~~~~~~~
182 There's a bit of a complication with type arguments. If the call
185 f p = ...f ((:) [a] x xs)...
187 then our specialised function look like
189 f_spec x xs = let p = (:) [a] x xs in ....as before....
191 This only makes sense if either
192 a) the type variable 'a' is in scope at the top of f, or
193 b) the type variable 'a' is an argument to f (and hence fs)
195 Actually, (a) may hold for value arguments too, in which case
196 we may not want to pass them. Supose 'x' is in scope at f's
197 defn, but xs is not. Then we'd like
199 f_spec xs = let p = (:) [a] x xs in ....as before....
201 Similarly (b) may hold too. If x is already an argument at the
202 call, no need to pass it again.
204 Finally, if 'a' is not in scope at the call site, we could abstract
205 it as we do the term variables:
207 f_spec a x xs = let p = (:) [a] x xs in ...as before...
209 So the grand plan is:
211 * abstract the call site to a constructor-only pattern
212 e.g. C x (D (f p) (g q)) ==> C s1 (D s2 s3)
214 * Find the free variables of the abstracted pattern
216 * Pass these variables, less any that are in scope at
217 the fn defn. But see Note [Shadowing] below.
220 NOTICE that we only abstract over variables that are not in scope,
221 so we're in no danger of shadowing variables used in "higher up"
227 In this pass we gather up usage information that may mention variables
228 that are bound between the usage site and the definition site; or (more
229 seriously) may be bound to something different at the definition site.
232 f x = letrec g y v = let x = ...
235 Since 'x' is in scope at the call site, we may make a rewrite rule that
237 RULE forall a,b. g (a,b) x = ...
238 But this rule will never match, because it's really a different 'x' at
239 the call site -- and that difference will be manifest by the time the
240 simplifier gets to it. [A worry: the simplifier doesn't *guarantee*
241 no-shadowing, so perhaps it may not be distinct?]
243 Anyway, the rule isn't actually wrong, it's just not useful. One possibility
244 is to run deShadowBinds before running SpecConstr, but instead we run the
245 simplifier. That gives the simplest possible program for SpecConstr to
246 chew on; and it virtually guarantees no shadowing.
248 Note [Specialising for constant parameters]
249 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
250 This one is about specialising on a *constant* (but not necessarily
251 constructor) argument
253 foo :: Int -> (Int -> Int) -> Int
255 foo m f = foo (f m) (+1)
259 lvl_rmV :: GHC.Base.Int -> GHC.Base.Int
261 \ (ds_dlk :: GHC.Base.Int) ->
262 case ds_dlk of wild_alH { GHC.Base.I# x_alG ->
263 GHC.Base.I# (GHC.Prim.+# x_alG 1)
265 T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
268 \ (ww_sme :: GHC.Prim.Int#) (w_smg :: GHC.Base.Int -> GHC.Base.Int) ->
269 case ww_sme of ds_Xlw {
271 case w_smg (GHC.Base.I# ds_Xlw) of w1_Xmo { GHC.Base.I# ww1_Xmz ->
272 T.$wfoo ww1_Xmz lvl_rmV
277 The recursive call has lvl_rmV as its argument, so we could create a specialised copy
278 with that argument baked in; that is, not passed at all. Now it can perhaps be inlined.
280 When is this worth it? Call the constant 'lvl'
281 - If 'lvl' has an unfolding that is a constructor, see if the corresponding
282 parameter is scrutinised anywhere in the body.
284 - If 'lvl' has an unfolding that is a inlinable function, see if the corresponding
285 parameter is applied (...to enough arguments...?)
287 Also do this is if the function has RULES?
291 Note [Specialising for lambda parameters]
292 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
293 foo :: Int -> (Int -> Int) -> Int
295 foo m f = foo (f m) (\n -> n-m)
297 This is subtly different from the previous one in that we get an
298 explicit lambda as the argument:
300 T.$wfoo :: GHC.Prim.Int# -> (GHC.Base.Int -> GHC.Base.Int) ->
303 \ (ww_sm8 :: GHC.Prim.Int#) (w_sma :: GHC.Base.Int -> GHC.Base.Int) ->
304 case ww_sm8 of ds_Xlr {
306 case w_sma (GHC.Base.I# ds_Xlr) of w1_Xmf { GHC.Base.I# ww1_Xmq ->
309 (\ (n_ad3 :: GHC.Base.Int) ->
310 case n_ad3 of wild_alB { GHC.Base.I# x_alA ->
311 GHC.Base.I# (GHC.Prim.-# x_alA ds_Xlr)
317 I wonder if SpecConstr couldn't be extended to handle this? After all,
318 lambda is a sort of constructor for functions and perhaps it already
319 has most of the necessary machinery?
321 Furthermore, there's an immediate win, because you don't need to allocate the lamda
322 at the call site; and if perchance it's called in the recursive call, then you
323 may avoid allocating it altogether. Just like for constructors.
325 Looks cool, but probably rare...but it might be easy to implement.
328 Note [SpecConstr for casts]
329 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
332 data instance T Int = T Int
337 go (T n) = go (T (n-1))
339 The recursive call ends up looking like
340 go (T (I# ...) `cast` g)
341 So we want to spot the construtor application inside the cast.
342 That's why we have the Cast case in argToPat
344 Note [Local recursive groups]
345 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
346 For a *local* recursive group, we can see all the calls to the
347 function, so we seed the specialisation loop from the calls in the
348 body, not from the calls in the RHS. Consider:
350 bar m n = foo n (n,n) (n,n) (n,n) (n,n)
354 | n > 3000 = case p of { (p1,p2) -> foo (n-1) (p2,p1) q r s }
355 | n > 2000 = case q of { (q1,q2) -> foo (n-1) p (q2,q1) r s }
356 | n > 1000 = case r of { (r1,r2) -> foo (n-1) p q (r2,r1) s }
357 | otherwise = case s of { (s1,s2) -> foo (n-1) p q r (s2,s1) }
359 If we start with the RHSs of 'foo', we get lots and lots of specialisations,
360 most of which are not needed. But if we start with the (single) call
361 in the rhs of 'bar' we get exactly one fully-specialised copy, and all
362 the recursive calls go to this fully-specialised copy. Indeed, the original
363 function is later collected as dead code. This is very important in
364 specialising the loops arising from stream fusion, for example in NDP where
365 we were getting literally hundreds of (mostly unused) specialisations of
368 Note [Do not specialise diverging functions]
369 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
370 Specialising a function that just diverges is a waste of code.
371 Furthermore, it broke GHC (simpl014) thus:
373 f = \x. case x of (a,b) -> f x
374 If we specialise f we get
375 f = \x. case x of (a,b) -> fspec a b
376 But fspec doesn't have decent strictnes info. As it happened,
377 (f x) :: IO t, so the state hack applied and we eta expanded fspec,
378 and hence f. But now f's strictness is less than its arity, which
381 -----------------------------------------------------
382 Stuff not yet handled
383 -----------------------------------------------------
385 Here are notes arising from Roman's work that I don't want to lose.
391 foo :: Int -> T Int -> Int
393 foo x t | even x = case t of { T n -> foo (x-n) t }
394 | otherwise = foo (x-1) t
396 SpecConstr does no specialisation, because the second recursive call
397 looks like a boxed use of the argument. A pity.
399 $wfoo_sFw :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
401 \ (ww_sFo [Just L] :: GHC.Prim.Int#) (w_sFq [Just L] :: T.T GHC.Base.Int) ->
402 case ww_sFo of ds_Xw6 [Just L] {
404 case GHC.Prim.remInt# ds_Xw6 2 of wild1_aEF [Dead Just A] {
405 __DEFAULT -> $wfoo_sFw (GHC.Prim.-# ds_Xw6 1) w_sFq;
407 case w_sFq of wild_Xy [Just L] { T.T n_ad5 [Just U(L)] ->
408 case n_ad5 of wild1_aET [Just A] { GHC.Base.I# y_aES [Just L] ->
409 $wfoo_sFw (GHC.Prim.-# ds_Xw6 y_aES) wild_Xy
415 data a :*: b = !a :*: !b
418 foo :: (Int :*: T Int) -> Int
420 foo (x :*: t) | even x = case t of { T n -> foo ((x-n) :*: t) }
421 | otherwise = foo ((x-1) :*: t)
423 Very similar to the previous one, except that the parameters are now in
424 a strict tuple. Before SpecConstr, we have
426 $wfoo_sG3 :: GHC.Prim.Int# -> T.T GHC.Base.Int -> GHC.Prim.Int#
428 \ (ww_sFU [Just L] :: GHC.Prim.Int#) (ww_sFW [Just L] :: T.T
430 case ww_sFU of ds_Xws [Just L] {
432 case GHC.Prim.remInt# ds_Xws 2 of wild1_aEZ [Dead Just A] {
434 case ww_sFW of tpl_B2 [Just L] { T.T a_sFo [Just A] ->
435 $wfoo_sG3 (GHC.Prim.-# ds_Xws 1) tpl_B2 -- $wfoo1
438 case ww_sFW of wild_XB [Just A] { T.T n_ad7 [Just S(L)] ->
439 case n_ad7 of wild1_aFd [Just L] { GHC.Base.I# y_aFc [Just L] ->
440 $wfoo_sG3 (GHC.Prim.-# ds_Xws y_aFc) wild_XB -- $wfoo2
444 We get two specialisations:
445 "SC:$wfoo1" [0] __forall {a_sFB :: GHC.Base.Int sc_sGC :: GHC.Prim.Int#}
446 Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int a_sFB)
447 = Foo.$s$wfoo1 a_sFB sc_sGC ;
448 "SC:$wfoo2" [0] __forall {y_aFp :: GHC.Prim.Int# sc_sGC :: GHC.Prim.Int#}
449 Foo.$wfoo sc_sGC (Foo.T @ GHC.Base.Int (GHC.Base.I# y_aFp))
450 = Foo.$s$wfoo y_aFp sc_sGC ;
452 But perhaps the first one isn't good. After all, we know that tpl_B2 is
453 a T (I# x) really, because T is strict and Int has one constructor. (We can't
454 unbox the strict fields, becuase T is polymorphic!)
458 %************************************************************************
460 \subsection{Top level wrapper stuff}
462 %************************************************************************
465 specConstrProgram :: DynFlags -> UniqSupply -> [CoreBind] -> [CoreBind]
466 specConstrProgram dflags us binds = fst $ initUs us (go (initScEnv dflags) binds)
469 go env (bind:binds) = do (env', bind') <- scTopBind env bind
470 binds' <- go env' binds
471 return (bind' : binds')
475 %************************************************************************
477 \subsection{Environment: goes downwards}
479 %************************************************************************
482 data ScEnv = SCE { sc_size :: Maybe Int, -- Size threshold
483 sc_count :: Maybe Int, -- Max # of specialisations for any one fn
485 sc_subst :: Subst, -- Current substitution
486 -- Maps InIds to OutExprs
488 sc_how_bound :: HowBoundEnv,
489 -- Binds interesting non-top-level variables
490 -- Domain is OutVars (*after* applying the substitution)
493 -- Domain is OutIds (*after* applying the substitution)
494 -- Used even for top-level bindings (but not imported ones)
497 ---------------------
498 -- As we go, we apply a substitution (sc_subst) to the current term
499 type InExpr = CoreExpr -- _Before_ applying the subst
501 type OutExpr = CoreExpr -- _After_ applying the subst
505 ---------------------
506 type HowBoundEnv = VarEnv HowBound -- Domain is OutVars
508 ---------------------
509 type ValueEnv = IdEnv Value -- Domain is OutIds
510 data Value = ConVal AltCon [CoreArg] -- _Saturated_ constructors
511 | LambdaVal -- Inlinable lambdas or PAPs
513 instance Outputable Value where
514 ppr (ConVal con args) = ppr con <+> interpp'SP args
515 ppr LambdaVal = ptext (sLit "<Lambda>")
517 ---------------------
518 initScEnv :: DynFlags -> ScEnv
520 = SCE { sc_size = specConstrThreshold dflags,
521 sc_count = specConstrCount dflags,
522 sc_subst = emptySubst,
523 sc_how_bound = emptyVarEnv,
524 sc_vals = emptyVarEnv }
526 data HowBound = RecFun -- These are the recursive functions for which
527 -- we seek interesting call patterns
529 | RecArg -- These are those functions' arguments, or their sub-components;
530 -- we gather occurrence information for these
532 instance Outputable HowBound where
533 ppr RecFun = text "RecFun"
534 ppr RecArg = text "RecArg"
536 lookupHowBound :: ScEnv -> Id -> Maybe HowBound
537 lookupHowBound env id = lookupVarEnv (sc_how_bound env) id
539 scSubstId :: ScEnv -> Id -> CoreExpr
540 scSubstId env v = lookupIdSubst (sc_subst env) v
542 scSubstTy :: ScEnv -> Type -> Type
543 scSubstTy env ty = substTy (sc_subst env) ty
545 zapScSubst :: ScEnv -> ScEnv
546 zapScSubst env = env { sc_subst = zapSubstEnv (sc_subst env) }
548 extendScInScope :: ScEnv -> [Var] -> ScEnv
549 -- Bring the quantified variables into scope
550 extendScInScope env qvars = env { sc_subst = extendInScopeList (sc_subst env) qvars }
552 -- Extend the substitution
553 extendScSubst :: ScEnv -> Var -> OutExpr -> ScEnv
554 extendScSubst env var expr = env { sc_subst = extendSubst (sc_subst env) var expr }
556 extendScSubstList :: ScEnv -> [(Var,OutExpr)] -> ScEnv
557 extendScSubstList env prs = env { sc_subst = extendSubstList (sc_subst env) prs }
559 extendHowBound :: ScEnv -> [Var] -> HowBound -> ScEnv
560 extendHowBound env bndrs how_bound
561 = env { sc_how_bound = extendVarEnvList (sc_how_bound env)
562 [(bndr,how_bound) | bndr <- bndrs] }
564 extendBndrsWith :: HowBound -> ScEnv -> [Var] -> (ScEnv, [Var])
565 extendBndrsWith how_bound env bndrs
566 = (env { sc_subst = subst', sc_how_bound = hb_env' }, bndrs')
568 (subst', bndrs') = substBndrs (sc_subst env) bndrs
569 hb_env' = sc_how_bound env `extendVarEnvList`
570 [(bndr,how_bound) | bndr <- bndrs']
572 extendBndrWith :: HowBound -> ScEnv -> Var -> (ScEnv, Var)
573 extendBndrWith how_bound env bndr
574 = (env { sc_subst = subst', sc_how_bound = hb_env' }, bndr')
576 (subst', bndr') = substBndr (sc_subst env) bndr
577 hb_env' = extendVarEnv (sc_how_bound env) bndr' how_bound
579 extendRecBndrs :: ScEnv -> [Var] -> (ScEnv, [Var])
580 extendRecBndrs env bndrs = (env { sc_subst = subst' }, bndrs')
582 (subst', bndrs') = substRecBndrs (sc_subst env) bndrs
584 extendBndr :: ScEnv -> Var -> (ScEnv, Var)
585 extendBndr env bndr = (env { sc_subst = subst' }, bndr')
587 (subst', bndr') = substBndr (sc_subst env) bndr
589 extendValEnv :: ScEnv -> Id -> Maybe Value -> ScEnv
590 extendValEnv env _ Nothing = env
591 extendValEnv env id (Just cv) = env { sc_vals = extendVarEnv (sc_vals env) id cv }
593 extendCaseBndrs :: ScEnv -> Id -> AltCon -> [Var] -> (ScEnv, [Var])
597 -- we want to bind b, to (C x y)
598 -- NB1: Extends only the sc_vals part of the envt
599 -- NB2: Kill the dead-ness info on the pattern binders x,y, since
600 -- they are potentially made alive by the [b -> C x y] binding
601 extendCaseBndrs env case_bndr con alt_bndrs
602 | isDeadBinder case_bndr
605 = (env1, map zap alt_bndrs)
606 -- NB: We used to bind v too, if scrut = (Var v); but
607 -- the simplifer has already done this so it seems
608 -- redundant to do so here
610 -- Var v -> extendValEnv env1 v cval
613 zap v | isTyVar v = v -- See NB2 above
614 | otherwise = zapIdOccInfo v
615 env1 = extendValEnv env case_bndr cval
618 LitAlt {} -> Just (ConVal con [])
619 DataAlt {} -> Just (ConVal con vanilla_args)
621 vanilla_args = map Type (tyConAppArgs (idType case_bndr)) ++
622 varsToCoreExprs alt_bndrs
626 %************************************************************************
628 \subsection{Usage information: flows upwards}
630 %************************************************************************
635 scu_calls :: CallEnv, -- Calls
636 -- The functions are a subset of the
637 -- RecFuns in the ScEnv
639 scu_occs :: !(IdEnv ArgOcc) -- Information on argument occurrences
640 } -- The domain is OutIds
642 type CallEnv = IdEnv [Call]
643 type Call = (ValueEnv, [CoreArg])
644 -- The arguments of the call, together with the
645 -- env giving the constructor bindings at the call site
648 nullUsage = SCU { scu_calls = emptyVarEnv, scu_occs = emptyVarEnv }
650 combineCalls :: CallEnv -> CallEnv -> CallEnv
651 combineCalls = plusVarEnv_C (++)
653 combineUsage :: ScUsage -> ScUsage -> ScUsage
654 combineUsage u1 u2 = SCU { scu_calls = combineCalls (scu_calls u1) (scu_calls u2),
655 scu_occs = plusVarEnv_C combineOcc (scu_occs u1) (scu_occs u2) }
657 combineUsages :: [ScUsage] -> ScUsage
658 combineUsages [] = nullUsage
659 combineUsages us = foldr1 combineUsage us
661 lookupOcc :: ScUsage -> OutVar -> (ScUsage, ArgOcc)
662 lookupOcc (SCU { scu_calls = sc_calls, scu_occs = sc_occs }) bndr
663 = (SCU {scu_calls = sc_calls, scu_occs = delVarEnv sc_occs bndr},
664 lookupVarEnv sc_occs bndr `orElse` NoOcc)
666 lookupOccs :: ScUsage -> [OutVar] -> (ScUsage, [ArgOcc])
667 lookupOccs (SCU { scu_calls = sc_calls, scu_occs = sc_occs }) bndrs
668 = (SCU {scu_calls = sc_calls, scu_occs = delVarEnvList sc_occs bndrs},
669 [lookupVarEnv sc_occs b `orElse` NoOcc | b <- bndrs])
671 data ArgOcc = NoOcc -- Doesn't occur at all; or a type argument
672 | UnkOcc -- Used in some unknown way
674 | ScrutOcc (UniqFM [ArgOcc]) -- See Note [ScrutOcc]
676 | BothOcc -- Definitely taken apart, *and* perhaps used in some other way
680 An occurrence of ScrutOcc indicates that the thing, or a `cast` version of the thing,
681 is *only* taken apart or applied.
683 Functions, literal: ScrutOcc emptyUFM
684 Data constructors: ScrutOcc subs,
686 where (subs :: UniqFM [ArgOcc]) gives usage of the *pattern-bound* components,
687 The domain of the UniqFM is the Unique of the data constructor
689 The [ArgOcc] is the occurrences of the *pattern-bound* components
690 of the data structure. E.g.
691 data T a = forall b. MkT a b (b->a)
692 A pattern binds b, x::a, y::b, z::b->a, but not 'a'!
696 instance Outputable ArgOcc where
697 ppr (ScrutOcc xs) = ptext (sLit "scrut-occ") <> ppr xs
698 ppr UnkOcc = ptext (sLit "unk-occ")
699 ppr BothOcc = ptext (sLit "both-occ")
700 ppr NoOcc = ptext (sLit "no-occ")
702 -- Experimentally, this vesion of combineOcc makes ScrutOcc "win", so
703 -- that if the thing is scrutinised anywhere then we get to see that
704 -- in the overall result, even if it's also used in a boxed way
705 -- This might be too agressive; see Note [Reboxing] Alternative 3
706 combineOcc :: ArgOcc -> ArgOcc -> ArgOcc
707 combineOcc NoOcc occ = occ
708 combineOcc occ NoOcc = occ
709 combineOcc (ScrutOcc xs) (ScrutOcc ys) = ScrutOcc (plusUFM_C combineOccs xs ys)
710 combineOcc _occ (ScrutOcc ys) = ScrutOcc ys
711 combineOcc (ScrutOcc xs) _occ = ScrutOcc xs
712 combineOcc UnkOcc UnkOcc = UnkOcc
713 combineOcc _ _ = BothOcc
715 combineOccs :: [ArgOcc] -> [ArgOcc] -> [ArgOcc]
716 combineOccs xs ys = zipWithEqual "combineOccs" combineOcc xs ys
718 setScrutOcc :: ScEnv -> ScUsage -> OutExpr -> ArgOcc -> ScUsage
719 -- _Overwrite_ the occurrence info for the scrutinee, if the scrutinee
720 -- is a variable, and an interesting variable
721 setScrutOcc env usg (Cast e _) occ = setScrutOcc env usg e occ
722 setScrutOcc env usg (Note _ e) occ = setScrutOcc env usg e occ
723 setScrutOcc env usg (Var v) occ
724 | Just RecArg <- lookupHowBound env v = usg { scu_occs = extendVarEnv (scu_occs usg) v occ }
726 setScrutOcc _env usg _other _occ -- Catch-all
729 conArgOccs :: ArgOcc -> AltCon -> [ArgOcc]
730 -- Find usage of components of data con; returns [UnkOcc...] if unknown
731 -- See Note [ScrutOcc] for the extra UnkOccs in the vanilla datacon case
733 conArgOccs (ScrutOcc fm) (DataAlt dc)
734 | Just pat_arg_occs <- lookupUFM fm dc
735 = [UnkOcc | _ <- dataConUnivTyVars dc] ++ pat_arg_occs
737 conArgOccs _other _con = repeat UnkOcc
740 %************************************************************************
742 \subsection{The main recursive function}
744 %************************************************************************
746 The main recursive function gathers up usage information, and
747 creates specialised versions of functions.
750 scExpr, scExpr' :: ScEnv -> CoreExpr -> UniqSM (ScUsage, CoreExpr)
751 -- The unique supply is needed when we invent
752 -- a new name for the specialised function and its args
754 scExpr env e = scExpr' env e
757 scExpr' env (Var v) = case scSubstId env v of
758 Var v' -> return (varUsage env v' UnkOcc, Var v')
759 e' -> scExpr (zapScSubst env) e'
761 scExpr' env (Type t) = return (nullUsage, Type (scSubstTy env t))
762 scExpr' _ e@(Lit {}) = return (nullUsage, e)
763 scExpr' env (Note n e) = do (usg,e') <- scExpr env e
764 return (usg, Note n e')
765 scExpr' env (Cast e co) = do (usg, e') <- scExpr env e
766 return (usg, Cast e' (scSubstTy env co))
767 scExpr' env e@(App _ _) = scApp env (collectArgs e)
768 scExpr' env (Lam b e) = do let (env', b') = extendBndr env b
769 (usg, e') <- scExpr env' e
770 return (usg, Lam b' e')
772 scExpr' env (Case scrut b ty alts)
773 = do { (scrut_usg, scrut') <- scExpr env scrut
774 ; case isValue (sc_vals env) scrut' of
775 Just (ConVal con args) -> sc_con_app con args scrut'
776 _other -> sc_vanilla scrut_usg scrut'
779 sc_con_app con args scrut' -- Known constructor; simplify
780 = do { let (_, bs, rhs) = findAlt con alts
781 alt_env' = extendScSubstList env ((b,scrut') : bs `zip` trimConArgs con args)
782 ; scExpr alt_env' rhs }
784 sc_vanilla scrut_usg scrut' -- Normal case
785 = do { let (alt_env,b') = extendBndrWith RecArg env b
786 -- Record RecArg for the components
788 ; (alt_usgs, alt_occs, alts')
789 <- mapAndUnzip3M (sc_alt alt_env scrut' b') alts
791 ; let (alt_usg, b_occ) = lookupOcc (combineUsages alt_usgs) b'
792 scrut_occ = foldr combineOcc b_occ alt_occs
793 scrut_usg' = setScrutOcc env scrut_usg scrut' scrut_occ
794 -- The combined usage of the scrutinee is given
795 -- by scrut_occ, which is passed to scScrut, which
796 -- in turn treats a bare-variable scrutinee specially
798 ; return (alt_usg `combineUsage` scrut_usg',
799 Case scrut' b' (scSubstTy env ty) alts') }
801 sc_alt env _scrut' b' (con,bs,rhs)
802 = do { let (env1, bs1) = extendBndrsWith RecArg env bs
803 (env2, bs2) = extendCaseBndrs env1 b' con bs1
804 ; (usg,rhs') <- scExpr env2 rhs
805 ; let (usg', arg_occs) = lookupOccs usg bs2
806 scrut_occ = case con of
807 DataAlt dc -> ScrutOcc (unitUFM dc arg_occs)
808 _ -> ScrutOcc emptyUFM
809 ; return (usg', scrut_occ, (con, bs2, rhs')) }
811 scExpr' env (Let (NonRec bndr rhs) body)
812 | isTyVar bndr -- Type-lets may be created by doBeta
813 = scExpr' (extendScSubst env bndr rhs) body
815 = do { let (body_env, bndr') = extendBndr env bndr
816 ; (rhs_usg, (_, args', rhs_body', _)) <- scRecRhs env (bndr',rhs)
817 ; let rhs' = mkLams args' rhs_body'
819 ; if not opt_SpecInlineJoinPoints || null args' || isEmptyVarEnv (scu_calls rhs_usg) then do
821 let body_env2 = extendValEnv body_env bndr' (isValue (sc_vals env) rhs')
822 -- Record if the RHS is a value
823 ; (body_usg, body') <- scExpr body_env2 body
824 ; return (body_usg `combineUsage` rhs_usg, Let (NonRec bndr' rhs') body') }
825 else -- For now, just brutally inline the join point
826 do { let body_env2 = extendScSubst env bndr rhs'
827 ; scExpr body_env2 body } }
831 do { -- Join-point case
832 let body_env2 = extendHowBound body_env [bndr'] RecFun
833 -- If the RHS of this 'let' contains calls
834 -- to recursive functions that we're trying
835 -- to specialise, then treat this let too
836 -- as one to specialise
837 ; (body_usg, body') <- scExpr body_env2 body
839 ; (spec_usg, _, specs) <- specialise env (scu_calls body_usg) ([], rhs_info)
841 ; return (body_usg { scu_calls = scu_calls body_usg `delVarEnv` bndr' }
842 `combineUsage` rhs_usg `combineUsage` spec_usg,
843 mkLets [NonRec b r | (b,r) <- specInfoBinds rhs_info specs] body')
847 -- A *local* recursive group: see Note [Local recursive groups]
848 scExpr' env (Let (Rec prs) body)
849 = do { let (bndrs,rhss) = unzip prs
850 (rhs_env1,bndrs') = extendRecBndrs env bndrs
851 rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun
853 ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss)
854 ; (body_usg, body') <- scExpr rhs_env2 body
856 -- NB: start specLoop from body_usg
857 ; (spec_usg, specs) <- specLoop rhs_env2 (scu_calls body_usg) rhs_infos nullUsage
858 [SI [] 0 (Just usg) | usg <- rhs_usgs]
860 ; let all_usg = spec_usg `combineUsage` body_usg
861 bind' = Rec (concat (zipWith specInfoBinds rhs_infos specs))
863 ; return (all_usg { scu_calls = scu_calls all_usg `delVarEnvList` bndrs' },
866 -----------------------------------
867 scApp :: ScEnv -> (InExpr, [InExpr]) -> UniqSM (ScUsage, CoreExpr)
869 scApp env (Var fn, args) -- Function is a variable
870 = ASSERT( not (null args) )
871 do { args_w_usgs <- mapM (scExpr env) args
872 ; let (arg_usgs, args') = unzip args_w_usgs
873 arg_usg = combineUsages arg_usgs
874 ; case scSubstId env fn of
875 fn'@(Lam {}) -> scExpr (zapScSubst env) (doBeta fn' args')
876 -- Do beta-reduction and try again
878 Var fn' -> return (arg_usg `combineUsage` fn_usg, mkApps (Var fn') args')
880 fn_usg = case lookupHowBound env fn' of
881 Just RecFun -> SCU { scu_calls = unitVarEnv fn' [(sc_vals env, args')],
882 scu_occs = emptyVarEnv }
883 Just RecArg -> SCU { scu_calls = emptyVarEnv,
884 scu_occs = unitVarEnv fn' (ScrutOcc emptyUFM) }
888 other_fn' -> return (arg_usg, mkApps other_fn' args') }
889 -- NB: doing this ignores any usage info from the substituted
890 -- function, but I don't think that matters. If it does
893 doBeta :: OutExpr -> [OutExpr] -> OutExpr
894 -- ToDo: adjust for System IF
895 doBeta (Lam bndr body) (arg : args) = Let (NonRec bndr arg) (doBeta body args)
896 doBeta fn args = mkApps fn args
898 -- The function is almost always a variable, but not always.
899 -- In particular, if this pass follows float-in,
900 -- which it may, we can get
901 -- (let f = ...f... in f) arg1 arg2
902 scApp env (other_fn, args)
903 = do { (fn_usg, fn') <- scExpr env other_fn
904 ; (arg_usgs, args') <- mapAndUnzipM (scExpr env) args
905 ; return (combineUsages arg_usgs `combineUsage` fn_usg, mkApps fn' args') }
907 ----------------------
908 scTopBind :: ScEnv -> CoreBind -> UniqSM (ScEnv, CoreBind)
909 scTopBind env (Rec prs)
910 | Just threshold <- sc_size env
911 , not (all (couldBeSmallEnoughToInline threshold) rhss)
913 = do { let (rhs_env,bndrs') = extendRecBndrs env bndrs
914 ; (_, rhss') <- mapAndUnzipM (scExpr rhs_env) rhss
915 ; return (rhs_env, Rec (bndrs' `zip` rhss')) }
916 | otherwise -- Do specialisation
917 = do { let (rhs_env1,bndrs') = extendRecBndrs env bndrs
918 rhs_env2 = extendHowBound rhs_env1 bndrs' RecFun
920 ; (rhs_usgs, rhs_infos) <- mapAndUnzipM (scRecRhs rhs_env2) (bndrs' `zip` rhss)
921 ; let rhs_usg = combineUsages rhs_usgs
923 ; (_, specs) <- specLoop rhs_env2 (scu_calls rhs_usg) rhs_infos nullUsage
924 [SI [] 0 Nothing | _ <- bndrs]
926 ; return (rhs_env1, -- For the body of the letrec, delete the RecFun business
927 Rec (concat (zipWith specInfoBinds rhs_infos specs))) }
929 (bndrs,rhss) = unzip prs
931 scTopBind env (NonRec bndr rhs)
932 = do { (_, rhs') <- scExpr env rhs
933 ; let (env1, bndr') = extendBndr env bndr
934 env2 = extendValEnv env1 bndr' (isValue (sc_vals env) rhs')
935 ; return (env2, NonRec bndr' rhs') }
937 ----------------------
938 scRecRhs :: ScEnv -> (OutId, InExpr) -> UniqSM (ScUsage, RhsInfo)
939 scRecRhs env (bndr,rhs)
940 = do { let (arg_bndrs,body) = collectBinders rhs
941 (body_env, arg_bndrs') = extendBndrsWith RecArg env arg_bndrs
942 ; (body_usg, body') <- scExpr body_env body
943 ; let (rhs_usg, arg_occs) = lookupOccs body_usg arg_bndrs'
944 ; return (rhs_usg, (bndr, arg_bndrs', body', arg_occs)) }
946 -- The arg_occs says how the visible,
947 -- lambda-bound binders of the RHS are used
948 -- (including the TyVar binders)
949 -- Two pats are the same if they match both ways
951 ----------------------
952 specInfoBinds :: RhsInfo -> SpecInfo -> [(Id,CoreExpr)]
953 specInfoBinds (fn, args, body, _) (SI specs _ _)
954 = [(id,rhs) | OS _ _ id rhs <- specs] ++
955 [(fn `addIdSpecialisations` rules, mkLams args body)]
957 rules = [r | OS _ r _ _ <- specs]
959 ----------------------
960 varUsage :: ScEnv -> OutVar -> ArgOcc -> ScUsage
962 | Just RecArg <- lookupHowBound env v = SCU { scu_calls = emptyVarEnv
963 , scu_occs = unitVarEnv v use }
964 | otherwise = nullUsage
968 %************************************************************************
970 The specialiser itself
972 %************************************************************************
975 type RhsInfo = (OutId, [OutVar], OutExpr, [ArgOcc])
976 -- Info about the *original* RHS of a binding we are specialising
977 -- Original binding f = \xs.body
978 -- Plus info about usage of arguments
980 data SpecInfo = SI [OneSpec] -- The specialisations we have generated
981 Int -- Length of specs; used for numbering them
982 (Maybe ScUsage) -- Nothing => we have generated specialisations
983 -- from calls in the *original* RHS
984 -- Just cs => we haven't, and this is the usage
985 -- of the original RHS
987 -- One specialisation: Rule plus definition
988 data OneSpec = OS CallPat -- Call pattern that generated this specialisation
989 CoreRule -- Rule connecting original id with the specialisation
990 OutId OutExpr -- Spec id + its rhs
996 -> ScUsage -> [SpecInfo] -- One per binder; acccumulating parameter
997 -> UniqSM (ScUsage, [SpecInfo]) -- ...ditto...
998 specLoop env all_calls rhs_infos usg_so_far specs_so_far
999 = do { specs_w_usg <- zipWithM (specialise env all_calls) rhs_infos specs_so_far
1000 ; let (new_usg_s, all_specs) = unzip specs_w_usg
1001 new_usg = combineUsages new_usg_s
1002 new_calls = scu_calls new_usg
1003 all_usg = usg_so_far `combineUsage` new_usg
1004 ; if isEmptyVarEnv new_calls then
1005 return (all_usg, all_specs)
1007 specLoop env new_calls rhs_infos all_usg all_specs }
1011 -> CallEnv -- Info on calls
1013 -> SpecInfo -- Original RHS plus patterns dealt with
1014 -> UniqSM (ScUsage, SpecInfo) -- New specialised versions and their usage
1016 -- Note: the rhs here is the optimised version of the original rhs
1017 -- So when we make a specialised copy of the RHS, we're starting
1018 -- from an RHS whose nested functions have been optimised already.
1020 specialise env bind_calls (fn, arg_bndrs, body, arg_occs)
1021 spec_info@(SI specs spec_count mb_unspec)
1022 | not (isBottomingId fn) -- Note [Do not specialise diverging functions]
1023 , notNull arg_bndrs -- Only specialise functions
1024 , Just all_calls <- lookupVarEnv bind_calls fn
1025 = do { (boring_call, pats) <- callsToPats env specs arg_occs all_calls
1026 -- ; pprTrace "specialise" (vcat [ppr fn <+> ppr arg_occs,
1027 -- text "calls" <+> ppr all_calls,
1028 -- text "good pats" <+> ppr pats]) $
1031 -- Bale out if too many specialisations
1032 -- Rather a hacky way to do so, but it'll do for now
1033 ; let spec_count' = length pats + spec_count
1034 ; case sc_count env of
1035 Just max | spec_count' > max
1036 -> WARN( True, msg ) return (nullUsage, spec_info)
1038 msg = vcat [ sep [ ptext (sLit "SpecConstr: specialisation of") <+> quotes (ppr fn)
1039 , nest 2 (ptext (sLit "limited by bound of")) <+> int max ]
1040 , ptext (sLit "Use -fspec-constr-count=n to set the bound")
1042 extra | not opt_PprStyle_Debug = ptext (sLit "Use -dppr-debug to see specialisations")
1043 | otherwise = ptext (sLit "Specialisations:") <+> ppr (pats ++ [p | OS p _ _ _ <- specs])
1045 _normal_case -> do {
1047 (spec_usgs, new_specs) <- mapAndUnzipM (spec_one env fn arg_bndrs body)
1048 (pats `zip` [spec_count..])
1050 ; let spec_usg = combineUsages spec_usgs
1051 (new_usg, mb_unspec')
1053 Just rhs_usg | boring_call -> (spec_usg `combineUsage` rhs_usg, Nothing)
1054 _ -> (spec_usg, mb_unspec)
1056 ; return (new_usg, SI (new_specs ++ specs) spec_count' mb_unspec') } }
1058 = return (nullUsage, spec_info) -- The boring case
1061 ---------------------
1063 -> OutId -- Function
1064 -> [Var] -- Lambda-binders of RHS; should match patterns
1065 -> CoreExpr -- Body of the original function
1067 -> UniqSM (ScUsage, OneSpec) -- Rule and binding
1069 -- spec_one creates a specialised copy of the function, together
1070 -- with a rule for using it. I'm very proud of how short this
1071 -- function is, considering what it does :-).
1077 f = /\b \y::[(a,b)] -> ....f (b,c) ((:) (a,(b,c)) (x,v) (h w))...
1078 [c::*, v::(b,c) are presumably bound by the (...) part]
1080 f_spec = /\ b c \ v::(b,c) hw::[(a,(b,c))] ->
1081 (...entire body of f...) [b -> (b,c),
1082 y -> ((:) (a,(b,c)) (x,v) hw)]
1084 RULE: forall b::* c::*, -- Note, *not* forall a, x
1088 f (b,c) ((:) (a,(b,c)) (x,v) hw) = f_spec b c v hw
1091 spec_one env fn arg_bndrs body (call_pat@(qvars, pats), rule_number)
1092 = do { -- Specialise the body
1093 let spec_env = extendScSubstList (extendScInScope env qvars)
1094 (arg_bndrs `zip` pats)
1095 ; (spec_usg, spec_body) <- scExpr spec_env body
1097 -- ; pprTrace "spec_one" (ppr fn <+> vcat [text "pats" <+> ppr pats,
1098 -- text "calls" <+> (ppr (scu_calls spec_usg))])
1101 -- And build the results
1102 ; spec_uniq <- getUniqueUs
1103 ; let (spec_lam_args, spec_call_args) = mkWorkerArgs qvars body_ty
1104 -- Usual w/w hack to avoid generating
1105 -- a spec_rhs of unlifted type and no args
1108 fn_loc = nameSrcSpan fn_name
1109 spec_occ = mkSpecOcc (nameOccName fn_name)
1110 rule_name = mkFastString ("SC:" ++ showSDoc (ppr fn <> int rule_number))
1111 spec_rhs = mkLams spec_lam_args spec_body
1112 spec_id = mkUserLocal spec_occ spec_uniq (mkPiTypes spec_lam_args body_ty) fn_loc
1113 body_ty = exprType spec_body
1114 rule_rhs = mkVarApps (Var spec_id) spec_call_args
1115 rule = mkLocalRule rule_name specConstrActivation fn_name qvars pats rule_rhs
1116 ; return (spec_usg, OS call_pat rule spec_id spec_rhs) }
1118 -- In which phase should the specialise-constructor rules be active?
1119 -- Originally I made them always-active, but Manuel found that
1120 -- this defeated some clever user-written rules. So Plan B
1121 -- is to make them active only in Phase 0; after all, currently,
1122 -- the specConstr transformation is only run after the simplifier
1123 -- has reached Phase 0. In general one would want it to be
1124 -- flag-controllable, but for now I'm leaving it baked in
1126 specConstrActivation :: Activation
1127 specConstrActivation = ActiveAfter 0 -- Baked in; see comments above
1130 %************************************************************************
1132 \subsection{Argument analysis}
1134 %************************************************************************
1136 This code deals with analysing call-site arguments to see whether
1137 they are constructor applications.
1141 type CallPat = ([Var], [CoreExpr]) -- Quantified variables and arguments
1144 callsToPats :: ScEnv -> [OneSpec] -> [ArgOcc] -> [Call] -> UniqSM (Bool, [CallPat])
1145 -- Result has no duplicate patterns,
1146 -- nor ones mentioned in done_pats
1147 -- Bool indicates that there was at least one boring pattern
1148 callsToPats env done_specs bndr_occs calls
1149 = do { mb_pats <- mapM (callToPats env bndr_occs) calls
1151 ; let good_pats :: [([Var], [CoreArg])]
1152 good_pats = catMaybes mb_pats
1153 done_pats = [p | OS p _ _ _ <- done_specs]
1154 is_done p = any (samePat p) done_pats
1156 ; return (any isNothing mb_pats,
1157 filterOut is_done (nubBy samePat good_pats)) }
1159 callToPats :: ScEnv -> [ArgOcc] -> Call -> UniqSM (Maybe CallPat)
1160 -- The [Var] is the variables to quantify over in the rule
1161 -- Type variables come first, since they may scope
1162 -- over the following term variables
1163 -- The [CoreExpr] are the argument patterns for the rule
1164 callToPats env bndr_occs (con_env, args)
1165 | length args < length bndr_occs -- Check saturated
1168 = do { let in_scope = substInScope (sc_subst env)
1169 ; prs <- argsToPats in_scope con_env (args `zip` bndr_occs)
1170 ; let (interesting_s, pats) = unzip prs
1171 pat_fvs = varSetElems (exprsFreeVars pats)
1172 qvars = filterOut (`elemInScopeSet` in_scope) pat_fvs
1173 -- Quantify over variables that are not in sccpe
1175 -- See Note [Shadowing] at the top
1177 (tvs, ids) = partition isTyVar qvars
1179 -- Put the type variables first; the type of a term
1180 -- variable may mention a type variable
1182 ; -- pprTrace "callToPats" (ppr args $$ ppr prs $$ ppr bndr_occs) $
1184 then return (Just (qvars', pats))
1185 else return Nothing }
1187 -- argToPat takes an actual argument, and returns an abstracted
1188 -- version, consisting of just the "constructor skeleton" of the
1189 -- argument, with non-constructor sub-expression replaced by new
1190 -- placeholder variables. For example:
1191 -- C a (D (f x) (g y)) ==> C p1 (D p2 p3)
1193 argToPat :: InScopeSet -- What's in scope at the fn defn site
1194 -> ValueEnv -- ValueEnv at the call site
1195 -> CoreArg -- A call arg (or component thereof)
1197 -> UniqSM (Bool, CoreArg)
1198 -- Returns (interesting, pat),
1199 -- where pat is the pattern derived from the argument
1200 -- intersting=True if the pattern is non-trivial (not a variable or type)
1201 -- E.g. x:xs --> (True, x:xs)
1202 -- f xs --> (False, w) where w is a fresh wildcard
1203 -- (f xs, 'c') --> (True, (w, 'c')) where w is a fresh wildcard
1204 -- \x. x+y --> (True, \x. x+y)
1205 -- lvl7 --> (True, lvl7) if lvl7 is bound
1206 -- somewhere further out
1208 argToPat _in_scope _val_env arg@(Type {}) _arg_occ
1209 = return (False, arg)
1211 argToPat in_scope val_env (Note _ arg) arg_occ
1212 = argToPat in_scope val_env arg arg_occ
1213 -- Note [Notes in call patterns]
1214 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1215 -- Ignore Notes. In particular, we want to ignore any InlineMe notes
1216 -- Perhaps we should not ignore profiling notes, but I'm going to
1217 -- ride roughshod over them all for now.
1218 --- See Note [Notes in RULE matching] in Rules
1220 argToPat in_scope val_env (Let _ arg) arg_occ
1221 = argToPat in_scope val_env arg arg_occ
1222 -- Look through let expressions
1223 -- e.g. f (let v = rhs in \y -> ...v...)
1224 -- Here we can specialise for f (\y -> ...)
1225 -- because the rule-matcher will look through the let.
1227 argToPat in_scope val_env (Cast arg co) arg_occ
1228 = do { (interesting, arg') <- argToPat in_scope val_env arg arg_occ
1229 ; let (ty1,ty2) = coercionKind co
1230 ; if not interesting then
1233 { -- Make a wild-card pattern for the coercion
1235 ; let co_name = mkSysTvName uniq (fsLit "sg")
1236 co_var = mkCoVar co_name (mkCoKind ty1 ty2)
1237 ; return (interesting, Cast arg' (mkTyVarTy co_var)) } }
1239 {- Disabling lambda specialisation for now
1240 It's fragile, and the spec_loop can be infinite
1241 argToPat in_scope val_env arg arg_occ
1243 = return (True, arg)
1245 is_value_lam (Lam v e) -- Spot a value lambda, even if
1246 | isId v = True -- it is inside a type lambda
1247 | otherwise = is_value_lam e
1248 is_value_lam other = False
1251 -- Check for a constructor application
1252 -- NB: this *precedes* the Var case, so that we catch nullary constrs
1253 argToPat in_scope val_env arg arg_occ
1254 | Just (ConVal dc args) <- isValue val_env arg
1256 ScrutOcc _ -> True -- Used only by case scrutinee
1257 BothOcc -> case arg of -- Used elsewhere
1258 App {} -> True -- see Note [Reboxing]
1260 _other -> False -- No point; the arg is not decomposed
1261 = do { args' <- argsToPats in_scope val_env (args `zip` conArgOccs arg_occ dc)
1262 ; return (True, mk_con_app dc (map snd args')) }
1264 -- Check if the argument is a variable that
1265 -- is in scope at the function definition site
1266 -- It's worth specialising on this if
1267 -- (a) it's used in an interesting way in the body
1268 -- (b) we know what its value is
1269 argToPat in_scope val_env (Var v) arg_occ
1270 | case arg_occ of { UnkOcc -> False; _other -> True }, -- (a)
1272 = return (True, Var v)
1275 | isLocalId v = v `elemInScopeSet` in_scope
1276 && isJust (lookupVarEnv val_env v)
1277 -- Local variables have values in val_env
1278 | otherwise = isValueUnfolding (idUnfolding v)
1279 -- Imports have unfoldings
1281 -- I'm really not sure what this comment means
1282 -- And by not wild-carding we tend to get forall'd
1283 -- variables that are in soope, which in turn can
1284 -- expose the weakness in let-matching
1285 -- See Note [Matching lets] in Rules
1287 -- Check for a variable bound inside the function.
1288 -- Don't make a wild-card, because we may usefully share
1289 -- e.g. f a = let x = ... in f (x,x)
1290 -- NB: this case follows the lambda and con-app cases!!
1291 -- argToPat _in_scope _val_env (Var v) _arg_occ
1292 -- = return (False, Var v)
1293 -- SLPJ : disabling this to avoid proliferation of versions
1294 -- also works badly when thinking about seeding the loop
1295 -- from the body of the let
1296 -- f x y = letrec g z = ... in g (x,y)
1297 -- We don't want to specialise for that *particular* x,y
1299 -- The default case: make a wild-card
1300 argToPat _in_scope _val_env arg _arg_occ
1301 = wildCardPat (exprType arg)
1303 wildCardPat :: Type -> UniqSM (Bool, CoreArg)
1304 wildCardPat ty = do { uniq <- getUniqueUs
1305 ; let id = mkSysLocal (fsLit "sc") uniq ty
1306 ; return (False, Var id) }
1308 argsToPats :: InScopeSet -> ValueEnv
1309 -> [(CoreArg, ArgOcc)]
1310 -> UniqSM [(Bool, CoreArg)]
1311 argsToPats in_scope val_env args
1314 do_one (arg,occ) = argToPat in_scope val_env arg occ
1319 isValue :: ValueEnv -> CoreExpr -> Maybe Value
1320 isValue _env (Lit lit)
1321 = Just (ConVal (LitAlt lit) [])
1324 | Just stuff <- lookupVarEnv env v
1325 = Just stuff -- You might think we could look in the idUnfolding here
1326 -- but that doesn't take account of which branch of a
1327 -- case we are in, which is the whole point
1329 | not (isLocalId v) && isCheapUnfolding unf
1330 = isValue env (unfoldingTemplate unf)
1333 -- However we do want to consult the unfolding
1334 -- as well, for let-bound constructors!
1336 isValue env (Lam b e)
1337 | isTyVar b = case isValue env e of
1338 Just _ -> Just LambdaVal
1340 | otherwise = Just LambdaVal
1342 isValue _env expr -- Maybe it's a constructor application
1343 | (Var fun, args) <- collectArgs expr
1344 = case isDataConWorkId_maybe fun of
1346 Just con | args `lengthAtLeast` dataConRepArity con
1347 -- Check saturated; might be > because the
1348 -- arity excludes type args
1349 -> Just (ConVal (DataAlt con) args)
1351 _other | valArgCount args < idArity fun
1352 -- Under-applied function
1353 -> Just LambdaVal -- Partial application
1357 isValue _env _expr = Nothing
1359 mk_con_app :: AltCon -> [CoreArg] -> CoreExpr
1360 mk_con_app (LitAlt lit) [] = Lit lit
1361 mk_con_app (DataAlt con) args = mkConApp con args
1362 mk_con_app _other _args = panic "SpecConstr.mk_con_app"
1364 samePat :: CallPat -> CallPat -> Bool
1365 samePat (vs1, as1) (vs2, as2)
1368 same (Var v1) (Var v2)
1369 | v1 `elem` vs1 = v2 `elem` vs2
1370 | v2 `elem` vs2 = False
1371 | otherwise = v1 == v2
1373 same (Lit l1) (Lit l2) = l1==l2
1374 same (App f1 a1) (App f2 a2) = same f1 f2 && same a1 a2
1376 same (Type {}) (Type {}) = True -- Note [Ignore type differences]
1377 same (Note _ e1) e2 = same e1 e2 -- Ignore casts and notes
1378 same (Cast e1 _) e2 = same e1 e2
1379 same e1 (Note _ e2) = same e1 e2
1380 same e1 (Cast e2 _) = same e1 e2
1382 same e1 e2 = WARN( bad e1 || bad e2, ppr e1 $$ ppr e2)
1383 False -- Let, lambda, case should not occur
1384 bad (Case {}) = True
1390 Note [Ignore type differences]
1391 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1392 We do not want to generate specialisations where the call patterns
1393 differ only in their type arguments! Not only is it utterly useless,
1394 but it also means that (with polymorphic recursion) we can generate
1395 an infinite number of specialisations. Example is Data.Sequence.adjustTree,