2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
4 \section[Specialise]{Stamping out overloading, and (optionally) polymorphism}
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
13 module Specialise ( specProgram ) where
15 #include "HsVersions.h"
17 import Id ( Id, idName, idType, mkUserLocal, idCoreRules,
18 idInlinePragma, setInlinePragma, setIdUnfolding,
19 isLocalId, idUnfolding )
20 import TcType ( Type, mkTyVarTy, tcSplitSigmaTy,
21 tyVarsOfTypes, tyVarsOfTheta, isClassPred,
22 tcCmpType, isUnLiftedType
24 import CoreSubst ( Subst, mkEmptySubst, extendTvSubstList, lookupIdSubst,
25 substBndr, substBndrs, substTy, substInScope,
26 cloneIdBndr, cloneIdBndrs, cloneRecIdBndrs,
29 import CoreUnfold ( mkUnfolding, mkInlineRule )
30 import SimplUtils ( interestingArg )
36 import CoreUtils ( exprIsTrivial, applyTypeToArgs, mkPiTypes )
37 import CoreFVs ( exprFreeVars, exprsFreeVars, idFreeVars )
38 import UniqSupply ( UniqSupply,
43 import MkId ( voidArgId, realWorldPrimId )
45 import Maybes ( catMaybes, isJust )
46 import BasicTypes ( Arity )
54 %************************************************************************
56 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
58 %************************************************************************
60 These notes describe how we implement specialisation to eliminate
63 The specialisation pass works on Core
64 syntax, complete with all the explicit dictionary application,
65 abstraction and construction as added by the type checker. The
66 existing type checker remains largely as it is.
68 One important thought: the {\em types} passed to an overloaded
69 function, and the {\em dictionaries} passed are mutually redundant.
70 If the same function is applied to the same type(s) then it is sure to
71 be applied to the same dictionary(s)---or rather to the same {\em
72 values}. (The arguments might look different but they will evaluate
75 Second important thought: we know that we can make progress by
76 treating dictionary arguments as static and worth specialising on. So
77 we can do without binding-time analysis, and instead specialise on
78 dictionary arguments and no others.
87 and suppose f is overloaded.
89 STEP 1: CALL-INSTANCE COLLECTION
91 We traverse <body>, accumulating all applications of f to types and
94 (Might there be partial applications, to just some of its types and
95 dictionaries? In principle yes, but in practice the type checker only
96 builds applications of f to all its types and dictionaries, so partial
97 applications could only arise as a result of transformation, and even
98 then I think it's unlikely. In any case, we simply don't accumulate such
99 partial applications.)
104 So now we have a collection of calls to f:
108 Notice that f may take several type arguments. To avoid ambiguity, we
109 say that f is called at type t1/t2 and t3/t4.
111 We take equivalence classes using equality of the *types* (ignoring
112 the dictionary args, which as mentioned previously are redundant).
114 STEP 3: SPECIALISATION
116 For each equivalence class, choose a representative (f t1 t2 d1 d2),
117 and create a local instance of f, defined thus:
119 f@t1/t2 = <f_rhs> t1 t2 d1 d2
121 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
122 of simplification will now result. However we don't actually *do* that
123 simplification. Rather, we leave it for the simplifier to do. If we
124 *did* do it, though, we'd get more call instances from the specialised
125 RHS. We can work out what they are by instantiating the call-instance
126 set from f's RHS with the types t1, t2.
128 Add this new id to f's IdInfo, to record that f has a specialised version.
130 Before doing any of this, check that f's IdInfo doesn't already
131 tell us about an existing instance of f at the required type/s.
132 (This might happen if specialisation was applied more than once, or
133 it might arise from user SPECIALIZE pragmas.)
137 Wait a minute! What if f is recursive? Then we can't just plug in
138 its right-hand side, can we?
140 But it's ok. The type checker *always* creates non-recursive definitions
141 for overloaded recursive functions. For example:
143 f x = f (x+x) -- Yes I know its silly
147 f a (d::Num a) = let p = +.sel a d
149 letrec fl (y::a) = fl (p y y)
153 We still have recusion for non-overloaded functions which we
154 speciailise, but the recursive call should get specialised to the
155 same recursive version.
161 All this is crystal clear when the function is applied to *constant
162 types*; that is, types which have no type variables inside. But what if
163 it is applied to non-constant types? Suppose we find a call of f at type
164 t1/t2. There are two possibilities:
166 (a) The free type variables of t1, t2 are in scope at the definition point
167 of f. In this case there's no problem, we proceed just as before. A common
168 example is as follows. Here's the Haskell:
173 After typechecking we have
175 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
176 in +.sel a d (f a d y) (f a d y)
178 Notice that the call to f is at type type "a"; a non-constant type.
179 Both calls to f are at the same type, so we can specialise to give:
181 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
182 in +.sel a d (f@a y) (f@a y)
185 (b) The other case is when the type variables in the instance types
186 are *not* in scope at the definition point of f. The example we are
187 working with above is a good case. There are two instances of (+.sel a d),
188 but "a" is not in scope at the definition of +.sel. Can we do anything?
189 Yes, we can "common them up", a sort of limited common sub-expression deal.
192 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
193 f@a (x::a) = +.sel@a x x
194 in +.sel@a (f@a y) (f@a y)
196 This can save work, and can't be spotted by the type checker, because
197 the two instances of +.sel weren't originally at the same type.
201 * There are quite a few variations here. For example, the defn of
202 +.sel could be floated ouside the \y, to attempt to gain laziness.
203 It certainly mustn't be floated outside the \d because the d has to
206 * We don't want to inline f_rhs in this case, because
207 that will duplicate code. Just commoning up the call is the point.
209 * Nothing gets added to +.sel's IdInfo.
211 * Don't bother unless the equivalence class has more than one item!
213 Not clear whether this is all worth it. It is of course OK to
214 simply discard call-instances when passing a big lambda.
216 Polymorphism 2 -- Overloading
218 Consider a function whose most general type is
220 f :: forall a b. Ord a => [a] -> b -> b
222 There is really no point in making a version of g at Int/Int and another
223 at Int/Bool, because it's only instancing the type variable "a" which
224 buys us any efficiency. Since g is completely polymorphic in b there
225 ain't much point in making separate versions of g for the different
228 That suggests that we should identify which of g's type variables
229 are constrained (like "a") and which are unconstrained (like "b").
230 Then when taking equivalence classes in STEP 2, we ignore the type args
231 corresponding to unconstrained type variable. In STEP 3 we make
232 polymorphic versions. Thus:
234 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
243 f a (d::Num a) = let g = ...
245 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
247 Here, g is only called at one type, but the dictionary isn't in scope at the
248 definition point for g. Usually the type checker would build a
249 definition for d1 which enclosed g, but the transformation system
250 might have moved d1's defn inward. Solution: float dictionary bindings
251 outwards along with call instances.
255 f x = let g p q = p==q
261 Before specialisation, leaving out type abstractions we have
263 f df x = let g :: Eq a => a -> a -> Bool
265 h :: Num a => a -> a -> (a, Bool)
266 h dh r s = let deq = eqFromNum dh
267 in (+ dh r s, g deq r s)
271 After specialising h we get a specialised version of h, like this:
273 h' r s = let deq = eqFromNum df
274 in (+ df r s, g deq r s)
276 But we can't naively make an instance for g from this, because deq is not in scope
277 at the defn of g. Instead, we have to float out the (new) defn of deq
278 to widen its scope. Notice that this floating can't be done in advance -- it only
279 shows up when specialisation is done.
281 User SPECIALIZE pragmas
282 ~~~~~~~~~~~~~~~~~~~~~~~
283 Specialisation pragmas can be digested by the type checker, and implemented
284 by adding extra definitions along with that of f, in the same way as before
286 f@t1/t2 = <f_rhs> t1 t2 d1 d2
288 Indeed the pragmas *have* to be dealt with by the type checker, because
289 only it knows how to build the dictionaries d1 and d2! For example
291 g :: Ord a => [a] -> [a]
292 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
294 Here, the specialised version of g is an application of g's rhs to the
295 Ord dictionary for (Tree Int), which only the type checker can conjure
296 up. There might not even *be* one, if (Tree Int) is not an instance of
297 Ord! (All the other specialision has suitable dictionaries to hand
300 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
301 it is buried in a complex (as-yet-un-desugared) binding group.
304 f@t1/t2 = f* t1 t2 d1 d2
306 where f* is the Id f with an IdInfo which says "inline me regardless!".
307 Indeed all the specialisation could be done in this way.
308 That in turn means that the simplifier has to be prepared to inline absolutely
309 any in-scope let-bound thing.
312 Again, the pragma should permit polymorphism in unconstrained variables:
314 h :: Ord a => [a] -> b -> b
315 {-# SPECIALIZE h :: [Int] -> b -> b #-}
317 We *insist* that all overloaded type variables are specialised to ground types,
318 (and hence there can be no context inside a SPECIALIZE pragma).
319 We *permit* unconstrained type variables to be specialised to
321 - or left as a polymorphic type variable
322 but nothing in between. So
324 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
326 is *illegal*. (It can be handled, but it adds complication, and gains the
330 SPECIALISING INSTANCE DECLARATIONS
331 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
334 instance Foo a => Foo [a] where
336 {-# SPECIALIZE instance Foo [Int] #-}
338 The original instance decl creates a dictionary-function
341 dfun.Foo.List :: forall a. Foo a -> Foo [a]
343 The SPECIALIZE pragma just makes a specialised copy, just as for
344 ordinary function definitions:
346 dfun.Foo.List@Int :: Foo [Int]
347 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
349 The information about what instance of the dfun exist gets added to
350 the dfun's IdInfo in the same way as a user-defined function too.
353 Automatic instance decl specialisation?
354 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
355 Can instance decls be specialised automatically? It's tricky.
356 We could collect call-instance information for each dfun, but
357 then when we specialised their bodies we'd get new call-instances
358 for ordinary functions; and when we specialised their bodies, we might get
359 new call-instances of the dfuns, and so on. This all arises because of
360 the unrestricted mutual recursion between instance decls and value decls.
362 Still, there's no actual problem; it just means that we may not do all
363 the specialisation we could theoretically do.
365 Furthermore, instance decls are usually exported and used non-locally,
366 so we'll want to compile enough to get those specialisations done.
368 Lastly, there's no such thing as a local instance decl, so we can
369 survive solely by spitting out *usage* information, and then reading that
370 back in as a pragma when next compiling the file. So for now,
371 we only specialise instance decls in response to pragmas.
374 SPITTING OUT USAGE INFORMATION
375 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
377 To spit out usage information we need to traverse the code collecting
378 call-instance information for all imported (non-prelude?) functions
379 and data types. Then we equivalence-class it and spit it out.
381 This is done at the top-level when all the call instances which escape
382 must be for imported functions and data types.
384 *** Not currently done ***
387 Partial specialisation by pragmas
388 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
389 What about partial specialisation:
391 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
392 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
396 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
398 Seems quite reasonable. Similar things could be done with instance decls:
400 instance (Foo a, Foo b) => Foo (a,b) where
402 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
403 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
405 Ho hum. Things are complex enough without this. I pass.
408 Requirements for the simplifer
409 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
410 The simplifier has to be able to take advantage of the specialisation.
412 * When the simplifier finds an application of a polymorphic f, it looks in
413 f's IdInfo in case there is a suitable instance to call instead. This converts
415 f t1 t2 d1 d2 ===> f_t1_t2
417 Note that the dictionaries get eaten up too!
419 * Dictionary selection operations on constant dictionaries must be
422 +.sel Int d ===> +Int
424 The obvious way to do this is in the same way as other specialised
425 calls: +.sel has inside it some IdInfo which tells that if it's applied
426 to the type Int then it should eat a dictionary and transform to +Int.
428 In short, dictionary selectors need IdInfo inside them for constant
431 * Exactly the same applies if a superclass dictionary is being
434 Eq.sel Int d ===> dEqInt
436 * Something similar applies to dictionary construction too. Suppose
437 dfun.Eq.List is the function taking a dictionary for (Eq a) to
438 one for (Eq [a]). Then we want
440 dfun.Eq.List Int d ===> dEq.List_Int
442 Where does the Eq [Int] dictionary come from? It is built in
443 response to a SPECIALIZE pragma on the Eq [a] instance decl.
445 In short, dfun Ids need IdInfo with a specialisation for each
446 constant instance of their instance declaration.
448 All this uses a single mechanism: the SpecEnv inside an Id
451 What does the specialisation IdInfo look like?
452 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
454 The SpecEnv of an Id maps a list of types (the template) to an expression
458 For example, if f has this SpecInfo:
460 [Int, a] -> \d:Ord Int. f' a
462 it means that we can replace the call
464 f Int t ===> (\d. f' t)
466 This chucks one dictionary away and proceeds with the
467 specialised version of f, namely f'.
470 What can't be done this way?
471 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
472 There is no way, post-typechecker, to get a dictionary for (say)
473 Eq a from a dictionary for Eq [a]. So if we find
477 we can't transform to
482 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
484 Of course, we currently have no way to automatically derive
485 eqList, nor to connect it to the Eq [a] instance decl, but you
486 can imagine that it might somehow be possible. Taking advantage
487 of this is permanently ruled out.
489 Still, this is no great hardship, because we intend to eliminate
490 overloading altogether anyway!
492 A note about non-tyvar dictionaries
493 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
494 Some Ids have types like
496 forall a,b,c. Eq a -> Ord [a] -> tau
498 This seems curious at first, because we usually only have dictionary
499 args whose types are of the form (C a) where a is a type variable.
500 But this doesn't hold for the functions arising from instance decls,
501 which sometimes get arguements with types of form (C (T a)) for some
504 Should we specialise wrt this compound-type dictionary? We used to say
506 "This is a heuristic judgement, as indeed is the fact that we
507 specialise wrt only dictionaries. We choose *not* to specialise
508 wrt compound dictionaries because at the moment the only place
509 they show up is in instance decls, where they are simply plugged
510 into a returned dictionary. So nothing is gained by specialising
513 But it is simpler and more uniform to specialise wrt these dicts too;
514 and in future GHC is likely to support full fledged type signatures
516 f :: Eq [(a,b)] => ...
519 %************************************************************************
521 \subsubsection{The new specialiser}
523 %************************************************************************
525 Our basic game plan is this. For let(rec) bound function
526 f :: (C a, D c) => (a,b,c,d) -> Bool
528 * Find any specialised calls of f, (f ts ds), where
529 ts are the type arguments t1 .. t4, and
530 ds are the dictionary arguments d1 .. d2.
532 * Add a new definition for f1 (say):
534 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
536 Note that we abstract over the unconstrained type arguments.
540 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
542 to the specialisations of f. This will be used by the
543 simplifier to replace calls
544 (f t1 t2 t3 t4) da db
546 (\d1 d1 -> f1 t2 t4) da db
548 All the stuff about how many dictionaries to discard, and what types
549 to apply the specialised function to, are handled by the fact that the
550 SpecEnv contains a template for the result of the specialisation.
552 We don't build *partial* specialisations for f. For example:
554 f :: Eq a => a -> a -> Bool
555 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
557 Here, little is gained by making a specialised copy of f.
558 There's a distinct danger that the specialised version would
559 first build a dictionary for (Eq b, Eq c), and then select the (==)
560 method from it! Even if it didn't, not a great deal is saved.
562 We do, however, generate polymorphic, but not overloaded, specialisations:
564 f :: Eq a => [a] -> b -> b -> b
565 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
567 Hence, the invariant is this:
569 *** no specialised version is overloaded ***
572 %************************************************************************
574 \subsubsection{The exported function}
576 %************************************************************************
579 specProgram :: UniqSupply -> [CoreBind] -> [CoreBind]
580 specProgram us binds = initSM us (do (binds', uds') <- go binds
581 return (dumpAllDictBinds uds' binds'))
583 -- We need to start with a Subst that knows all the things
584 -- that are in scope, so that the substitution engine doesn't
585 -- accidentally re-use a unique that's already in use
586 -- Easiest thing is to do it all at once, as if all the top-level
587 -- decls were mutually recursive
588 top_subst = mkEmptySubst (mkInScopeSet (mkVarSet (bindersOfBinds binds)))
590 go [] = return ([], emptyUDs)
591 go (bind:binds) = do (binds', uds) <- go binds
592 (bind', uds') <- specBind top_subst bind uds
593 return (bind' ++ binds', uds')
596 %************************************************************************
598 \subsubsection{@specExpr@: the main function}
600 %************************************************************************
603 specVar :: Subst -> Id -> CoreExpr
604 specVar subst v = lookupIdSubst subst v
606 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
607 -- We carry a substitution down:
608 -- a) we must clone any binding that might float outwards,
609 -- to avoid name clashes
610 -- b) we carry a type substitution to use when analysing
611 -- the RHS of specialised bindings (no type-let!)
613 ---------------- First the easy cases --------------------
614 specExpr subst (Type ty) = return (Type (substTy subst ty), emptyUDs)
615 specExpr subst (Var v) = return (specVar subst v, emptyUDs)
616 specExpr _ (Lit lit) = return (Lit lit, emptyUDs)
617 specExpr subst (Cast e co) = do
618 (e', uds) <- specExpr subst e
619 return ((Cast e' (substTy subst co)), uds)
620 specExpr subst (Note note body) = do
621 (body', uds) <- specExpr subst body
622 return (Note (specNote subst note) body', uds)
625 ---------------- Applications might generate a call instance --------------------
626 specExpr subst expr@(App {})
629 go (App fun arg) args = do (arg', uds_arg) <- specExpr subst arg
630 (fun', uds_app) <- go fun (arg':args)
631 return (App fun' arg', uds_arg `plusUDs` uds_app)
633 go (Var f) args = case specVar subst f of
634 Var f' -> return (Var f', mkCallUDs f' args)
635 e' -> return (e', emptyUDs) -- I don't expect this!
636 go other _ = specExpr subst other
638 ---------------- Lambda/case require dumping of usage details --------------------
639 specExpr subst e@(Lam _ _) = do
640 (body', uds) <- specExpr subst' body
641 let (filtered_uds, body'') = dumpUDs bndrs' uds body'
642 return (mkLams bndrs' body'', filtered_uds)
644 (bndrs, body) = collectBinders e
645 (subst', bndrs') = substBndrs subst bndrs
646 -- More efficient to collect a group of binders together all at once
647 -- and we don't want to split a lambda group with dumped bindings
649 specExpr subst (Case scrut case_bndr ty alts) = do
650 (scrut', uds_scrut) <- specExpr subst scrut
651 (alts', uds_alts) <- mapAndCombineSM spec_alt alts
652 return (Case scrut' case_bndr' (substTy subst ty) alts', uds_scrut `plusUDs` uds_alts)
654 (subst_alt, case_bndr') = substBndr subst case_bndr
655 -- No need to clone case binder; it can't float like a let(rec)
657 spec_alt (con, args, rhs) = do
658 (rhs', uds) <- specExpr subst_rhs rhs
659 let (uds', rhs'') = dumpUDs args uds rhs'
660 return ((con, args', rhs''), uds')
662 (subst_rhs, args') = substBndrs subst_alt args
664 ---------------- Finally, let is the interesting case --------------------
665 specExpr subst (Let bind body) = do
667 (rhs_subst, body_subst, bind') <- cloneBindSM subst bind
669 -- Deal with the body
670 (body', body_uds) <- specExpr body_subst body
672 -- Deal with the bindings
673 (binds', uds) <- specBind rhs_subst bind' body_uds
676 return (foldr Let body' binds', uds)
678 -- Must apply the type substitution to coerceions
679 specNote :: Subst -> Note -> Note
680 specNote _ note = note
683 %************************************************************************
685 \subsubsection{Dealing with a binding}
687 %************************************************************************
690 specBind :: Subst -- Use this for RHSs
692 -> UsageDetails -- Info on how the scope of the binding
693 -> SpecM ([CoreBind], -- New bindings
694 UsageDetails) -- And info to pass upstream
696 specBind rhs_subst bind body_uds
697 = do { (bind', bind_uds) <- specBindItself rhs_subst bind (calls body_uds)
698 ; return (finishSpecBind bind' bind_uds body_uds) }
700 finishSpecBind :: CoreBind -> UsageDetails -> UsageDetails -> ([CoreBind], UsageDetails)
702 (MkUD { dict_binds = rhs_dbs, calls = rhs_calls, ud_fvs = rhs_fvs })
703 (MkUD { dict_binds = body_dbs, calls = body_calls, ud_fvs = body_fvs })
704 | not (mkVarSet bndrs `intersectsVarSet` all_fvs)
705 -- Common case 1: the bound variables are not
706 -- mentioned in the dictionary bindings
707 = ([bind], MkUD { dict_binds = body_dbs `unionBags` rhs_dbs
708 -- It's important that the `unionBags` is this way round,
709 -- because body_uds may bind dictionaries that are
710 -- used in the calls passed to specDefn. So the
711 -- dictionary bindings in rhs_uds may mention
712 -- dictionaries bound in body_uds.
714 , ud_fvs = all_fvs })
716 | case bind of { NonRec {} -> True; Rec {} -> False }
717 -- Common case 2: no specialisation happened, and binding
718 -- is non-recursive. But the binding may be
719 -- mentioned in body_dbs, so we should put it first
720 = ([], MkUD { dict_binds = rhs_dbs `unionBags` ((bind, b_fvs) `consBag` body_dbs)
722 , ud_fvs = all_fvs `unionVarSet` b_fvs })
724 | otherwise -- General case: make a huge Rec (sigh)
725 = ([], MkUD { dict_binds = unitBag (Rec all_db_prs, all_db_fvs)
727 , ud_fvs = all_fvs `unionVarSet` b_fvs })
729 all_fvs = rhs_fvs `unionVarSet` body_fvs
730 all_calls = zapCalls bndrs (rhs_calls `unionCalls` body_calls)
732 bndrs = bindersOf bind
733 b_fvs = bind_fvs bind
735 (all_db_prs, all_db_fvs) = add (bind, b_fvs) $
736 foldrBag add ([], emptyVarSet) $
737 rhs_dbs `unionBags` body_dbs
738 add (NonRec b r, b_fvs) (prs, fvs) = ((b,r) : prs, b_fvs `unionVarSet` fvs)
739 add (Rec b_prs, b_fvs) (prs, fvs) = (b_prs ++ prs, b_fvs `unionVarSet` fvs)
741 ---------------------------
742 specBindItself :: Subst -> CoreBind -> CallDetails -> SpecM (CoreBind, UsageDetails)
744 -- specBindItself deals with the RHS, specialising it according
745 -- to the calls found in the body (if any)
746 specBindItself rhs_subst (NonRec fn rhs) call_info
747 = do { (rhs', rhs_uds) <- specExpr rhs_subst rhs -- Do RHS of original fn
748 ; (fn', spec_defns, spec_uds) <- specDefn rhs_subst call_info fn rhs
749 ; if null spec_defns then
750 return (NonRec fn rhs', rhs_uds)
752 return (Rec ((fn',rhs') : spec_defns), rhs_uds `plusUDs` spec_uds) }
753 -- bndr' mentions the spec_defns in its SpecEnv
754 -- Not sure why we couln't just put the spec_defns first
756 specBindItself rhs_subst (Rec pairs) call_info
757 -- Note [Specialising a recursive group]
758 = do { let (bndrs,rhss) = unzip pairs
759 ; (rhss', rhs_uds) <- mapAndCombineSM (specExpr rhs_subst) rhss
760 ; let all_calls = call_info `unionCalls` calls rhs_uds
761 ; (bndrs1, spec_defns1, spec_uds1) <- specDefns rhs_subst all_calls pairs
763 ; if null spec_defns1 then -- Common case: no specialisation
764 return (Rec (bndrs `zip` rhss'), rhs_uds)
765 else do -- Specialisation occurred; do it again
766 { (bndrs2, spec_defns2, spec_uds2) <-
767 -- pprTrace "specB" (ppr bndrs $$ ppr rhs_uds) $
768 specDefns rhs_subst (calls spec_uds1) (bndrs1 `zip` rhss)
770 ; let all_defns = spec_defns1 ++ spec_defns2 ++ zip bndrs2 rhss'
772 ; return (Rec all_defns, rhs_uds `plusUDs` spec_uds1 `plusUDs` spec_uds2) } }
775 ---------------------------
777 -> CallDetails -- Info on how it is used in its scope
778 -> [(Id,CoreExpr)] -- The things being bound and their un-processed RHS
779 -> SpecM ([Id], -- Original Ids with RULES added
780 [(Id,CoreExpr)], -- Extra, specialised bindings
781 UsageDetails) -- Stuff to fling upwards from the specialised versions
783 -- Specialise a list of bindings (the contents of a Rec), but flowing usages
784 -- upwards binding by binding. Example: { f = ...g ...; g = ...f .... }
785 -- Then if the input CallDetails has a specialised call for 'g', whose specialisation
786 -- in turn generates a specialised call for 'f', we catch that in this one sweep.
787 -- But not vice versa (it's a fixpoint problem).
789 specDefns _subst _call_info []
790 = return ([], [], emptyUDs)
791 specDefns subst call_info ((bndr,rhs):pairs)
792 = do { (bndrs', spec_defns, spec_uds) <- specDefns subst call_info pairs
793 ; let all_calls = call_info `unionCalls` calls spec_uds
794 ; (bndr', spec_defns1, spec_uds1) <- specDefn subst all_calls bndr rhs
795 ; return (bndr' : bndrs',
796 spec_defns1 ++ spec_defns,
797 spec_uds1 `plusUDs` spec_uds) }
799 ---------------------------
801 -> CallDetails -- Info on how it is used in its scope
802 -> Id -> CoreExpr -- The thing being bound and its un-processed RHS
803 -> SpecM (Id, -- Original Id with added RULES
804 [(Id,CoreExpr)], -- Extra, specialised bindings
805 UsageDetails) -- Stuff to fling upwards from the specialised versions
807 specDefn subst calls fn rhs
808 -- The first case is the interesting one
809 | rhs_tyvars `lengthIs` n_tyvars -- Rhs of fn's defn has right number of big lambdas
810 && rhs_ids `lengthAtLeast` n_dicts -- and enough dict args
811 && notNull calls_for_me -- And there are some calls to specialise
813 -- && not (certainlyWillInline (idUnfolding fn)) -- And it's not small
814 -- See Note [Inline specialisation] for why we do not
815 -- switch off specialisation for inline functions
817 = do { -- Make a specialised version for each call in calls_for_me
818 stuff <- mapM spec_call calls_for_me
819 ; let (spec_defns, spec_uds, spec_rules) = unzip3 (catMaybes stuff)
820 fn' = addIdSpecialisations fn spec_rules
821 ; return (fn', spec_defns, plusUDList spec_uds) }
823 | otherwise -- No calls or RHS doesn't fit our preconceptions
824 = WARN( notNull calls_for_me, ptext (sLit "Missed specialisation opportunity for") <+> ppr fn )
825 -- Note [Specialisation shape]
826 return (fn, [], emptyUDs)
830 (tyvars, theta, _) = tcSplitSigmaTy fn_type
831 n_tyvars = length tyvars
832 n_dicts = length theta
833 inline_prag = idInlinePragma fn
835 -- Figure out whether the function has an INLINE pragma
836 -- See Note [Inline specialisations]
837 fn_has_inline_rule :: Maybe Arity -- Gives arity of the *specialised* inline rule
838 fn_has_inline_rule = case idUnfolding fn of
839 InlineRule { uf_arity = arity } -> Just (arity - n_dicts)
842 (rhs_tyvars, rhs_ids, rhs_body) = collectTyAndValBinders rhs
844 rhs_dict_ids = take n_dicts rhs_ids
845 body = mkLams (drop n_dicts rhs_ids) rhs_body
846 -- Glue back on the non-dict lambdas
848 calls_for_me = case lookupFM calls fn of
850 Just cs -> fmToList cs
852 already_covered :: [CoreExpr] -> Bool
853 already_covered args -- Note [Specialisations already covered]
854 = isJust (lookupRule (const True) (substInScope subst)
855 fn args (idCoreRules fn))
857 mk_ty_args :: [Maybe Type] -> [CoreExpr]
858 mk_ty_args call_ts = zipWithEqual "spec_call" mk_ty_arg rhs_tyvars call_ts
860 mk_ty_arg rhs_tyvar Nothing = Type (mkTyVarTy rhs_tyvar)
861 mk_ty_arg _ (Just ty) = Type ty
863 ----------------------------------------------------------
864 -- Specialise to one particular call pattern
865 spec_call :: (CallKey, ([DictExpr], VarSet)) -- Call instance
866 -> SpecM (Maybe ((Id,CoreExpr), -- Specialised definition
867 UsageDetails, -- Usage details from specialised body
868 CoreRule)) -- Info for the Id's SpecEnv
869 spec_call (CallKey call_ts, (call_ds, _))
870 = ASSERT( call_ts `lengthIs` n_tyvars && call_ds `lengthIs` n_dicts )
872 -- Suppose f's defn is f = /\ a b c -> \ d1 d2 -> rhs
873 -- Supppose the call is for f [Just t1, Nothing, Just t3] [dx1, dx2]
875 -- Construct the new binding
876 -- f1 = SUBST[a->t1,c->t3, d1->d1', d2->d2'] (/\ b -> rhs)
877 -- PLUS the usage-details
878 -- { d1' = dx1; d2' = dx2 }
879 -- where d1', d2' are cloned versions of d1,d2, with the type substitution
880 -- applied. These auxiliary bindings just avoid duplication of dx1, dx2
882 -- Note that the substitution is applied to the whole thing.
883 -- This is convenient, but just slightly fragile. Notably:
884 -- * There had better be no name clashes in a/b/c
886 -- poly_tyvars = [b] in the example above
887 -- spec_tyvars = [a,c]
888 -- ty_args = [t1,b,t3]
889 poly_tyvars = [tv | (tv, Nothing) <- rhs_tyvars `zip` call_ts]
890 spec_tv_binds = [(tv,ty) | (tv, Just ty) <- rhs_tyvars `zip` call_ts]
891 spec_ty_args = map snd spec_tv_binds
892 ty_args = mk_ty_args call_ts
893 rhs_subst = extendTvSubstList subst spec_tv_binds
895 ; (rhs_subst1, inst_dict_ids) <- cloneDictBndrs rhs_subst rhs_dict_ids
896 -- Clone rhs_dicts, including instantiating their types
898 ; let (rhs_subst2, dx_binds) = bindAuxiliaryDicts rhs_subst1 $
899 (my_zipEqual rhs_dict_ids inst_dict_ids call_ds)
900 inst_args = ty_args ++ map Var inst_dict_ids
902 ; if already_covered inst_args then
905 { -- Figure out the type of the specialised function
906 let body_ty = applyTypeToArgs rhs fn_type inst_args
907 (lam_args, app_args) -- Add a dummy argument if body_ty is unlifted
908 | isUnLiftedType body_ty -- C.f. WwLib.mkWorkerArgs
909 = (poly_tyvars ++ [voidArgId], poly_tyvars ++ [realWorldPrimId])
910 | otherwise = (poly_tyvars, poly_tyvars)
911 spec_id_ty = mkPiTypes lam_args body_ty
913 ; spec_f <- newSpecIdSM fn spec_id_ty
914 ; (spec_rhs, rhs_uds) <- specExpr rhs_subst2 (mkLams lam_args body)
916 -- The rule to put in the function's specialisation is:
917 -- forall b, d1',d2'. f t1 b t3 d1' d2' = f1 b
918 rule_name = mkFastString ("SPEC " ++ showSDoc (ppr fn <+> ppr spec_ty_args))
919 spec_env_rule = mkLocalRule
921 inline_prag -- Note [Auto-specialisation and RULES]
923 (poly_tyvars ++ inst_dict_ids)
925 (mkVarApps (Var spec_f) app_args)
927 -- Add the { d1' = dx1; d2' = dx2 } usage stuff
928 final_uds = foldr addDictBind rhs_uds dx_binds
930 -- See Note [Inline specialisations]
931 final_spec_f | Just spec_arity <- fn_has_inline_rule
932 = spec_f `setInlinePragma` inline_prag
933 `setIdUnfolding` mkInlineRule spec_rhs spec_arity
936 ; return (Just ((final_spec_f, spec_rhs), final_uds, spec_env_rule)) } }
939 | debugIsOn && not (equalLength xs ys && equalLength ys zs)
940 = pprPanic "my_zipEqual" (vcat [ ppr xs, ppr ys
941 , ppr fn <+> ppr call_ts
942 , ppr (idType fn), ppr theta
943 , ppr n_dicts, ppr rhs_dict_ids
945 | otherwise = zip3 xs ys zs
949 -> [(DictId,DictId,CoreExpr)] -- (orig_dict, inst_dict, dx)
950 -> (Subst, -- Substitute for all orig_dicts
951 [(DictId, CoreExpr)]) -- Auxiliary bindings
952 -- Bind any dictionary arguments to fresh names, to preserve sharing
953 -- Substitution already substitutes orig_dict -> inst_dict
954 bindAuxiliaryDicts subst triples = go subst [] triples
956 go subst binds [] = (subst, binds)
957 go subst binds ((d, dx_id, dx) : pairs)
958 | exprIsTrivial dx = go (extendIdSubst subst d dx) binds pairs
959 -- No auxiliary binding necessary
960 | otherwise = go subst_w_unf ((dx_id,dx) : binds) pairs
962 dx_id1 = dx_id `setIdUnfolding` mkUnfolding False dx
963 subst_w_unf = extendIdSubst subst d (Var dx_id1)
964 -- Important! We're going to substitute dx_id1 for d
965 -- and we want it to look "interesting", else we won't gather *any*
966 -- consequential calls. E.g.
968 -- If we specialise f for a call (f (dfun dNumInt)), we'll get
969 -- a consequent call (g d') with an auxiliary definition
971 -- We want that consequent call to look interesting
974 Note [Specialising a recursive group]
975 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
977 let rec { f x = ...g x'...
978 ; g y = ...f y'.... }
980 Here we specialise 'f' at Char; but that is very likely to lead to
981 a specialisation of 'g' at Char. We must do the latter, else the
982 whole point of specialisation is lost.
984 But we do not want to keep iterating to a fixpoint, because in the
985 presence of polymorphic recursion we might generate an infinite number
988 So we use the following heuristic:
989 * Arrange the rec block in dependency order, so far as possible
990 (the occurrence analyser already does this)
992 * Specialise it much like a sequence of lets
994 * Then go through the block a second time, feeding call-info from
995 the RHSs back in the bottom, as it were
997 In effect, the ordering maxmimises the effectiveness of each sweep,
998 and we do just two sweeps. This should catch almost every case of
999 monomorphic recursion -- the exception could be a very knotted-up
1000 recursion with multiple cycles tied up together.
1002 This plan is implemented in the Rec case of specBindItself.
1004 Note [Specialisations already covered]
1005 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1006 We obviously don't want to generate two specialisations for the same
1007 argument pattern. There are two wrinkles
1009 1. We do the already-covered test in specDefn, not when we generate
1010 the CallInfo in mkCallUDs. We used to test in the latter place, but
1011 we now iterate the specialiser somewhat, and the Id at the call site
1012 might therefore not have all the RULES that we can see in specDefn
1014 2. What about two specialisations where the second is an *instance*
1015 of the first? If the more specific one shows up first, we'll generate
1016 specialisations for both. If the *less* specific one shows up first,
1017 we *don't* currently generate a specialisation for the more specific
1018 one. (See the call to lookupRule in already_covered.) Reasons:
1019 (a) lookupRule doesn't say which matches are exact (bad reason)
1020 (b) if the earlier specialisation is user-provided, it's
1021 far from clear that we should auto-specialise further
1023 Note [Auto-specialisation and RULES]
1024 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1026 g :: Num a => a -> a
1029 f :: (Int -> Int) -> Int
1031 {-# RULE f g = 0 #-}
1033 Suppose that auto-specialisation makes a specialised version of
1034 g::Int->Int That version won't appear in the LHS of the RULE for f.
1035 So if the specialisation rule fires too early, the rule for f may
1038 It might be possible to add new rules, to "complete" the rewrite system.
1040 RULE forall d. g Int d = g_spec
1044 But that's a bit complicated. For now we ask the programmer's help,
1045 by *copying the INLINE activation pragma* to the auto-specialised rule.
1046 So if g says {-# NOINLINE[2] g #-}, then the auto-spec rule will also
1047 not be active until phase 2.
1050 Note [Specialisation shape]
1051 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1052 We only specialise a function if it has visible top-level lambdas
1053 corresponding to its overloading. E.g. if
1054 f :: forall a. Eq a => ....
1055 then its body must look like
1058 Reason: when specialising the body for a call (f ty dexp), we want to
1059 substitute dexp for d, and pick up specialised calls in the body of f.
1061 This doesn't always work. One example I came across was this:
1062 newtype Gen a = MkGen{ unGen :: Int -> a }
1064 choose :: Eq a => a -> Gen a
1065 choose n = MkGen (\r -> n)
1067 oneof = choose (1::Int)
1069 It's a silly exapmle, but we get
1070 choose = /\a. g `cast` co
1071 where choose doesn't have any dict arguments. Thus far I have not
1072 tried to fix this (wait till there's a real example).
1075 Note [Inline specialisations]
1076 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1077 We transfer to the specialised function any INLINE stuff from the
1078 original. This means (a) the Activation in the IdInfo, and (b) any
1079 InlineMe on the RHS.
1081 This is a change (Jun06). Previously the idea is that the point of
1082 inlining was precisely to specialise the function at its call site,
1083 and that's not so important for the specialised copies. But
1084 *pragma-directed* specialisation now takes place in the
1085 typechecker/desugarer, with manually specified INLINEs. The
1086 specialiation here is automatic. It'd be very odd if a function
1087 marked INLINE was specialised (because of some local use), and then
1088 forever after (including importing modules) the specialised version
1089 wasn't INLINEd. After all, the programmer said INLINE!
1091 You might wonder why we don't just not specialise INLINE functions.
1092 It's because even INLINE functions are sometimes not inlined, when
1093 they aren't applied to interesting arguments. But perhaps the type
1094 arguments alone are enough to specialise (even though the args are too
1095 boring to trigger inlining), and it's certainly better to call the
1096 specialised version.
1098 A case in point is dictionary functions, which are current marked
1099 INLINE, but which are worth specialising.
1102 %************************************************************************
1104 \subsubsection{UsageDetails and suchlike}
1106 %************************************************************************
1111 dict_binds :: !(Bag DictBind),
1112 -- Floated dictionary bindings
1113 -- The order is important;
1114 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
1115 -- (Remember, Bags preserve order in GHC.)
1117 calls :: !CallDetails,
1119 ud_fvs :: !VarSet -- A superset of the variables mentioned in
1120 -- either dict_binds or calls
1123 instance Outputable UsageDetails where
1124 ppr (MkUD { dict_binds = dbs, calls = calls, ud_fvs = fvs })
1125 = ptext (sLit "MkUD") <+> braces (sep (punctuate comma
1126 [ptext (sLit "binds") <+> equals <+> ppr dbs,
1127 ptext (sLit "calls") <+> equals <+> ppr calls,
1128 ptext (sLit "fvs") <+> equals <+> ppr fvs]))
1130 type DictBind = (CoreBind, VarSet)
1131 -- The set is the free vars of the binding
1132 -- both tyvars and dicts
1134 type DictExpr = CoreExpr
1136 emptyUDs :: UsageDetails
1137 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM, ud_fvs = emptyVarSet }
1139 ------------------------------------------------------------
1140 type CallDetails = FiniteMap Id CallInfo
1141 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
1143 -- CallInfo uses a FiniteMap, thereby ensuring that
1144 -- we record only one call instance for any key
1146 -- The list of types and dictionaries is guaranteed to
1147 -- match the type of f
1148 type CallInfo = FiniteMap CallKey ([DictExpr], VarSet)
1149 -- Range is dict args and the vars of the whole
1150 -- call (including tyvars)
1151 -- [*not* include the main id itself, of course]
1153 instance Outputable CallKey where
1154 ppr (CallKey ts) = ppr ts
1156 -- Type isn't an instance of Ord, so that we can control which
1157 -- instance we use. That's tiresome here. Oh well
1158 instance Eq CallKey where
1159 k1 == k2 = case k1 `compare` k2 of { EQ -> True; _ -> False }
1161 instance Ord CallKey where
1162 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
1164 cmp Nothing Nothing = EQ
1165 cmp Nothing (Just _) = LT
1166 cmp (Just _) Nothing = GT
1167 cmp (Just t1) (Just t2) = tcCmpType t1 t2
1169 unionCalls :: CallDetails -> CallDetails -> CallDetails
1170 unionCalls c1 c2 = plusFM_C plusFM c1 c2
1172 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> UsageDetails
1173 singleCall id tys dicts
1174 = MkUD {dict_binds = emptyBag,
1175 calls = unitFM id (unitFM (CallKey tys) (dicts, call_fvs)),
1178 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
1179 tys_fvs = tyVarsOfTypes (catMaybes tys)
1180 -- The type args (tys) are guaranteed to be part of the dictionary
1181 -- types, because they are just the constrained types,
1182 -- and the dictionary is therefore sure to be bound
1183 -- inside the binding for any type variables free in the type;
1184 -- hence it's safe to neglect tyvars free in tys when making
1185 -- the free-var set for this call
1186 -- BUT I don't trust this reasoning; play safe and include tys_fvs
1188 -- We don't include the 'id' itself.
1190 mkCallUDs :: Id -> [CoreExpr] -> UsageDetails
1192 | not (isLocalId f) -- Imported from elsewhere
1193 || null theta -- Not overloaded
1194 || not (all isClassPred theta)
1195 -- Only specialise if all overloading is on class params.
1196 -- In ptic, with implicit params, the type args
1197 -- *don't* say what the value of the implicit param is!
1198 || not (spec_tys `lengthIs` n_tyvars)
1199 || not ( dicts `lengthIs` n_dicts)
1200 || not (any interestingArg dicts) -- Note [Interesting dictionary arguments]
1201 -- See also Note [Specialisations already covered]
1202 = -- pprTrace "mkCallUDs: discarding" (vcat [ppr f, ppr args, ppr n_tyvars, ppr n_dicts, ppr (map interestingArg dicts)])
1203 emptyUDs -- Not overloaded, or no specialisation wanted
1206 = -- pprTrace "mkCallUDs: keeping" (vcat [ppr f, ppr args, ppr n_tyvars, ppr n_dicts, ppr (map interestingArg dicts)])
1207 singleCall f spec_tys dicts
1209 (tyvars, theta, _) = tcSplitSigmaTy (idType f)
1210 constrained_tyvars = tyVarsOfTheta theta
1211 n_tyvars = length tyvars
1212 n_dicts = length theta
1214 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
1215 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
1218 | tyvar `elemVarSet` constrained_tyvars = Just ty
1219 | otherwise = Nothing
1222 Note [Interesting dictionary arguments]
1223 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1225 \a.\d:Eq a. let f = ... in ...(f d)...
1226 There really is not much point in specialising f wrt the dictionary d,
1227 because the code for the specialised f is not improved at all, because
1228 d is lambda-bound. We simply get junk specialisations.
1230 We re-use the function SimplUtils.interestingArg function to determine
1231 what sort of dictionary arguments have *some* information in them.
1235 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
1236 plusUDs (MkUD {dict_binds = db1, calls = calls1, ud_fvs = fvs1})
1237 (MkUD {dict_binds = db2, calls = calls2, ud_fvs = fvs2})
1238 = MkUD {dict_binds = d, calls = c, ud_fvs = fvs1 `unionVarSet` fvs2}
1240 d = db1 `unionBags` db2
1241 c = calls1 `unionCalls` calls2
1243 plusUDList :: [UsageDetails] -> UsageDetails
1244 plusUDList = foldr plusUDs emptyUDs
1246 -- zapCalls deletes calls to ids from uds
1247 zapCalls :: [Id] -> CallDetails -> CallDetails
1248 zapCalls ids calls = delListFromFM calls ids
1250 mkDB :: CoreBind -> DictBind
1251 mkDB bind = (bind, bind_fvs bind)
1253 bind_fvs :: CoreBind -> VarSet
1254 bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
1255 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1258 rhs_fvs = unionVarSets (map pair_fvs prs)
1260 pair_fvs :: (Id, CoreExpr) -> VarSet
1261 pair_fvs (bndr, rhs) = exprFreeVars rhs `unionVarSet` idFreeVars bndr
1262 -- Don't forget variables mentioned in the
1263 -- rules of the bndr. C.f. OccAnal.addRuleUsage
1264 -- Also tyvars mentioned in its type; they may not appear in the RHS
1268 addDictBind :: (Id,CoreExpr) -> UsageDetails -> UsageDetails
1269 addDictBind (dict,rhs) uds
1270 = uds { dict_binds = db `consBag` dict_binds uds
1271 , ud_fvs = ud_fvs uds `unionVarSet` fvs }
1273 db@(_, fvs) = mkDB (NonRec dict rhs)
1275 dumpAllDictBinds :: UsageDetails -> [CoreBind] -> [CoreBind]
1276 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
1277 = foldrBag add binds dbs
1279 add (bind,_) binds = bind : binds
1281 dumpUDs :: [CoreBndr]
1282 -> UsageDetails -> CoreExpr
1283 -> (UsageDetails, CoreExpr)
1284 dumpUDs bndrs (MkUD { dict_binds = orig_dbs
1285 , calls = orig_calls
1286 , ud_fvs = fvs}) body
1287 = (new_uds, foldrBag add_let body dump_dbs)
1288 -- This may delete fewer variables
1289 -- than in priciple possible
1292 MkUD { dict_binds = free_dbs
1293 , calls = free_calls
1294 , ud_fvs = fvs `minusVarSet` bndr_set}
1296 bndr_set = mkVarSet bndrs
1297 add_let (bind,_) body = Let bind body
1299 (free_dbs, dump_dbs, dump_set)
1300 = foldlBag dump_db (emptyBag, emptyBag, bndr_set) orig_dbs
1301 -- Important that it's foldl not foldr;
1302 -- we're accumulating the set of dumped ids in dump_set
1304 free_calls = filterCalls dump_set orig_calls
1306 dump_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1307 | dump_idset `intersectsVarSet` fvs -- Dump it
1308 = (free_dbs, dump_dbs `snocBag` db,
1309 extendVarSetList dump_idset (bindersOf bind))
1311 | otherwise -- Don't dump it
1312 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1314 filterCalls :: VarSet -> CallDetails -> CallDetails
1315 -- Remove any calls that mention the variables
1316 filterCalls bs calls
1317 = mapFM (\_ cs -> filter_calls cs) $
1318 filterFM (\k _ -> not (k `elemVarSet` bs)) calls
1320 filter_calls :: CallInfo -> CallInfo
1321 filter_calls = filterFM (\_ (_, fvs) -> not (fvs `intersectsVarSet` bs))
1325 %************************************************************************
1327 \subsubsection{Boring helper functions}
1329 %************************************************************************
1332 type SpecM a = UniqSM a
1334 initSM :: UniqSupply -> SpecM a -> a
1337 mapAndCombineSM :: (a -> SpecM (b, UsageDetails)) -> [a] -> SpecM ([b], UsageDetails)
1338 mapAndCombineSM _ [] = return ([], emptyUDs)
1339 mapAndCombineSM f (x:xs) = do (y, uds1) <- f x
1340 (ys, uds2) <- mapAndCombineSM f xs
1341 return (y:ys, uds1 `plusUDs` uds2)
1343 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1344 -- Clone the binders of the bind; return new bind with the cloned binders
1345 -- Return the substitution to use for RHSs, and the one to use for the body
1346 cloneBindSM subst (NonRec bndr rhs) = do
1347 us <- getUniqueSupplyM
1348 let (subst', bndr') = cloneIdBndr subst us bndr
1349 return (subst, subst', NonRec bndr' rhs)
1351 cloneBindSM subst (Rec pairs) = do
1352 us <- getUniqueSupplyM
1353 let (subst', bndrs') = cloneRecIdBndrs subst us (map fst pairs)
1354 return (subst', subst', Rec (bndrs' `zip` map snd pairs))
1356 cloneDictBndrs :: Subst -> [CoreBndr] -> SpecM (Subst, [CoreBndr])
1357 cloneDictBndrs subst bndrs
1358 = do { us <- getUniqueSupplyM
1359 ; return (cloneIdBndrs subst us bndrs) }
1361 newSpecIdSM :: Id -> Type -> SpecM Id
1362 -- Give the new Id a similar occurrence name to the old one
1363 newSpecIdSM old_id new_ty
1364 = do { uniq <- getUniqueM
1366 name = idName old_id
1367 new_occ = mkSpecOcc (nameOccName name)
1368 new_id = mkUserLocal new_occ uniq new_ty (getSrcSpan name)
1373 Old (but interesting) stuff about unboxed bindings
1374 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1376 What should we do when a value is specialised to a *strict* unboxed value?
1378 map_*_* f (x:xs) = let h = f x
1382 Could convert let to case:
1384 map_*_Int# f (x:xs) = case f x of h# ->
1388 This may be undesirable since it forces evaluation here, but the value
1389 may not be used in all branches of the body. In the general case this
1390 transformation is impossible since the mutual recursion in a letrec
1391 cannot be expressed as a case.
1393 There is also a problem with top-level unboxed values, since our
1394 implementation cannot handle unboxed values at the top level.
1396 Solution: Lift the binding of the unboxed value and extract it when it
1399 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1404 Now give it to the simplifier and the _Lifting will be optimised away.
1406 The benfit is that we have given the specialised "unboxed" values a
1407 very simplep lifted semantics and then leave it up to the simplifier to
1408 optimise it --- knowing that the overheads will be removed in nearly
1411 In particular, the value will only be evaluted in the branches of the
1412 program which use it, rather than being forced at the point where the
1413 value is bound. For example:
1415 filtermap_*_* p f (x:xs)
1422 filtermap_*_Int# p f (x:xs)
1423 = let h = case (f x) of h# -> _Lift h#
1426 True -> case h of _Lift h#
1430 The binding for h can still be inlined in the one branch and the
1431 _Lifting eliminated.
1434 Question: When won't the _Lifting be eliminated?
1436 Answer: When they at the top-level (where it is necessary) or when
1437 inlining would duplicate work (or possibly code depending on
1438 options). However, the _Lifting will still be eliminated if the
1439 strictness analyser deems the lifted binding strict.