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"
20 import CoreUnfold ( mkUnfolding, mkInlineRule )
25 import CoreUtils ( exprIsTrivial, applyTypeToArgs, mkPiTypes )
26 import CoreFVs ( exprFreeVars, exprsFreeVars, idFreeVars )
27 import UniqSupply ( UniqSupply, UniqSM, initUs_, MonadUnique(..) )
29 import MkId ( voidArgId, realWorldPrimId )
31 import Maybes ( catMaybes, isJust )
32 import BasicTypes ( Arity )
40 %************************************************************************
42 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
44 %************************************************************************
46 These notes describe how we implement specialisation to eliminate
49 The specialisation pass works on Core
50 syntax, complete with all the explicit dictionary application,
51 abstraction and construction as added by the type checker. The
52 existing type checker remains largely as it is.
54 One important thought: the {\em types} passed to an overloaded
55 function, and the {\em dictionaries} passed are mutually redundant.
56 If the same function is applied to the same type(s) then it is sure to
57 be applied to the same dictionary(s)---or rather to the same {\em
58 values}. (The arguments might look different but they will evaluate
61 Second important thought: we know that we can make progress by
62 treating dictionary arguments as static and worth specialising on. So
63 we can do without binding-time analysis, and instead specialise on
64 dictionary arguments and no others.
73 and suppose f is overloaded.
75 STEP 1: CALL-INSTANCE COLLECTION
77 We traverse <body>, accumulating all applications of f to types and
80 (Might there be partial applications, to just some of its types and
81 dictionaries? In principle yes, but in practice the type checker only
82 builds applications of f to all its types and dictionaries, so partial
83 applications could only arise as a result of transformation, and even
84 then I think it's unlikely. In any case, we simply don't accumulate such
85 partial applications.)
90 So now we have a collection of calls to f:
94 Notice that f may take several type arguments. To avoid ambiguity, we
95 say that f is called at type t1/t2 and t3/t4.
97 We take equivalence classes using equality of the *types* (ignoring
98 the dictionary args, which as mentioned previously are redundant).
100 STEP 3: SPECIALISATION
102 For each equivalence class, choose a representative (f t1 t2 d1 d2),
103 and create a local instance of f, defined thus:
105 f@t1/t2 = <f_rhs> t1 t2 d1 d2
107 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
108 of simplification will now result. However we don't actually *do* that
109 simplification. Rather, we leave it for the simplifier to do. If we
110 *did* do it, though, we'd get more call instances from the specialised
111 RHS. We can work out what they are by instantiating the call-instance
112 set from f's RHS with the types t1, t2.
114 Add this new id to f's IdInfo, to record that f has a specialised version.
116 Before doing any of this, check that f's IdInfo doesn't already
117 tell us about an existing instance of f at the required type/s.
118 (This might happen if specialisation was applied more than once, or
119 it might arise from user SPECIALIZE pragmas.)
123 Wait a minute! What if f is recursive? Then we can't just plug in
124 its right-hand side, can we?
126 But it's ok. The type checker *always* creates non-recursive definitions
127 for overloaded recursive functions. For example:
129 f x = f (x+x) -- Yes I know its silly
133 f a (d::Num a) = let p = +.sel a d
135 letrec fl (y::a) = fl (p y y)
139 We still have recusion for non-overloaded functions which we
140 speciailise, but the recursive call should get specialised to the
141 same recursive version.
147 All this is crystal clear when the function is applied to *constant
148 types*; that is, types which have no type variables inside. But what if
149 it is applied to non-constant types? Suppose we find a call of f at type
150 t1/t2. There are two possibilities:
152 (a) The free type variables of t1, t2 are in scope at the definition point
153 of f. In this case there's no problem, we proceed just as before. A common
154 example is as follows. Here's the Haskell:
159 After typechecking we have
161 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
162 in +.sel a d (f a d y) (f a d y)
164 Notice that the call to f is at type type "a"; a non-constant type.
165 Both calls to f are at the same type, so we can specialise to give:
167 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
168 in +.sel a d (f@a y) (f@a y)
171 (b) The other case is when the type variables in the instance types
172 are *not* in scope at the definition point of f. The example we are
173 working with above is a good case. There are two instances of (+.sel a d),
174 but "a" is not in scope at the definition of +.sel. Can we do anything?
175 Yes, we can "common them up", a sort of limited common sub-expression deal.
178 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
179 f@a (x::a) = +.sel@a x x
180 in +.sel@a (f@a y) (f@a y)
182 This can save work, and can't be spotted by the type checker, because
183 the two instances of +.sel weren't originally at the same type.
187 * There are quite a few variations here. For example, the defn of
188 +.sel could be floated ouside the \y, to attempt to gain laziness.
189 It certainly mustn't be floated outside the \d because the d has to
192 * We don't want to inline f_rhs in this case, because
193 that will duplicate code. Just commoning up the call is the point.
195 * Nothing gets added to +.sel's IdInfo.
197 * Don't bother unless the equivalence class has more than one item!
199 Not clear whether this is all worth it. It is of course OK to
200 simply discard call-instances when passing a big lambda.
202 Polymorphism 2 -- Overloading
204 Consider a function whose most general type is
206 f :: forall a b. Ord a => [a] -> b -> b
208 There is really no point in making a version of g at Int/Int and another
209 at Int/Bool, because it's only instancing the type variable "a" which
210 buys us any efficiency. Since g is completely polymorphic in b there
211 ain't much point in making separate versions of g for the different
214 That suggests that we should identify which of g's type variables
215 are constrained (like "a") and which are unconstrained (like "b").
216 Then when taking equivalence classes in STEP 2, we ignore the type args
217 corresponding to unconstrained type variable. In STEP 3 we make
218 polymorphic versions. Thus:
220 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
229 f a (d::Num a) = let g = ...
231 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
233 Here, g is only called at one type, but the dictionary isn't in scope at the
234 definition point for g. Usually the type checker would build a
235 definition for d1 which enclosed g, but the transformation system
236 might have moved d1's defn inward. Solution: float dictionary bindings
237 outwards along with call instances.
241 f x = let g p q = p==q
247 Before specialisation, leaving out type abstractions we have
249 f df x = let g :: Eq a => a -> a -> Bool
251 h :: Num a => a -> a -> (a, Bool)
252 h dh r s = let deq = eqFromNum dh
253 in (+ dh r s, g deq r s)
257 After specialising h we get a specialised version of h, like this:
259 h' r s = let deq = eqFromNum df
260 in (+ df r s, g deq r s)
262 But we can't naively make an instance for g from this, because deq is not in scope
263 at the defn of g. Instead, we have to float out the (new) defn of deq
264 to widen its scope. Notice that this floating can't be done in advance -- it only
265 shows up when specialisation is done.
267 User SPECIALIZE pragmas
268 ~~~~~~~~~~~~~~~~~~~~~~~
269 Specialisation pragmas can be digested by the type checker, and implemented
270 by adding extra definitions along with that of f, in the same way as before
272 f@t1/t2 = <f_rhs> t1 t2 d1 d2
274 Indeed the pragmas *have* to be dealt with by the type checker, because
275 only it knows how to build the dictionaries d1 and d2! For example
277 g :: Ord a => [a] -> [a]
278 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
280 Here, the specialised version of g is an application of g's rhs to the
281 Ord dictionary for (Tree Int), which only the type checker can conjure
282 up. There might not even *be* one, if (Tree Int) is not an instance of
283 Ord! (All the other specialision has suitable dictionaries to hand
286 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
287 it is buried in a complex (as-yet-un-desugared) binding group.
290 f@t1/t2 = f* t1 t2 d1 d2
292 where f* is the Id f with an IdInfo which says "inline me regardless!".
293 Indeed all the specialisation could be done in this way.
294 That in turn means that the simplifier has to be prepared to inline absolutely
295 any in-scope let-bound thing.
298 Again, the pragma should permit polymorphism in unconstrained variables:
300 h :: Ord a => [a] -> b -> b
301 {-# SPECIALIZE h :: [Int] -> b -> b #-}
303 We *insist* that all overloaded type variables are specialised to ground types,
304 (and hence there can be no context inside a SPECIALIZE pragma).
305 We *permit* unconstrained type variables to be specialised to
307 - or left as a polymorphic type variable
308 but nothing in between. So
310 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
312 is *illegal*. (It can be handled, but it adds complication, and gains the
316 SPECIALISING INSTANCE DECLARATIONS
317 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
320 instance Foo a => Foo [a] where
322 {-# SPECIALIZE instance Foo [Int] #-}
324 The original instance decl creates a dictionary-function
327 dfun.Foo.List :: forall a. Foo a -> Foo [a]
329 The SPECIALIZE pragma just makes a specialised copy, just as for
330 ordinary function definitions:
332 dfun.Foo.List@Int :: Foo [Int]
333 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
335 The information about what instance of the dfun exist gets added to
336 the dfun's IdInfo in the same way as a user-defined function too.
339 Automatic instance decl specialisation?
340 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
341 Can instance decls be specialised automatically? It's tricky.
342 We could collect call-instance information for each dfun, but
343 then when we specialised their bodies we'd get new call-instances
344 for ordinary functions; and when we specialised their bodies, we might get
345 new call-instances of the dfuns, and so on. This all arises because of
346 the unrestricted mutual recursion between instance decls and value decls.
348 Still, there's no actual problem; it just means that we may not do all
349 the specialisation we could theoretically do.
351 Furthermore, instance decls are usually exported and used non-locally,
352 so we'll want to compile enough to get those specialisations done.
354 Lastly, there's no such thing as a local instance decl, so we can
355 survive solely by spitting out *usage* information, and then reading that
356 back in as a pragma when next compiling the file. So for now,
357 we only specialise instance decls in response to pragmas.
360 SPITTING OUT USAGE INFORMATION
361 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
363 To spit out usage information we need to traverse the code collecting
364 call-instance information for all imported (non-prelude?) functions
365 and data types. Then we equivalence-class it and spit it out.
367 This is done at the top-level when all the call instances which escape
368 must be for imported functions and data types.
370 *** Not currently done ***
373 Partial specialisation by pragmas
374 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
375 What about partial specialisation:
377 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
378 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
382 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
384 Seems quite reasonable. Similar things could be done with instance decls:
386 instance (Foo a, Foo b) => Foo (a,b) where
388 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
389 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
391 Ho hum. Things are complex enough without this. I pass.
394 Requirements for the simplifer
395 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
396 The simplifier has to be able to take advantage of the specialisation.
398 * When the simplifier finds an application of a polymorphic f, it looks in
399 f's IdInfo in case there is a suitable instance to call instead. This converts
401 f t1 t2 d1 d2 ===> f_t1_t2
403 Note that the dictionaries get eaten up too!
405 * Dictionary selection operations on constant dictionaries must be
408 +.sel Int d ===> +Int
410 The obvious way to do this is in the same way as other specialised
411 calls: +.sel has inside it some IdInfo which tells that if it's applied
412 to the type Int then it should eat a dictionary and transform to +Int.
414 In short, dictionary selectors need IdInfo inside them for constant
417 * Exactly the same applies if a superclass dictionary is being
420 Eq.sel Int d ===> dEqInt
422 * Something similar applies to dictionary construction too. Suppose
423 dfun.Eq.List is the function taking a dictionary for (Eq a) to
424 one for (Eq [a]). Then we want
426 dfun.Eq.List Int d ===> dEq.List_Int
428 Where does the Eq [Int] dictionary come from? It is built in
429 response to a SPECIALIZE pragma on the Eq [a] instance decl.
431 In short, dfun Ids need IdInfo with a specialisation for each
432 constant instance of their instance declaration.
434 All this uses a single mechanism: the SpecEnv inside an Id
437 What does the specialisation IdInfo look like?
438 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
440 The SpecEnv of an Id maps a list of types (the template) to an expression
444 For example, if f has this SpecInfo:
446 [Int, a] -> \d:Ord Int. f' a
448 it means that we can replace the call
450 f Int t ===> (\d. f' t)
452 This chucks one dictionary away and proceeds with the
453 specialised version of f, namely f'.
456 What can't be done this way?
457 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
458 There is no way, post-typechecker, to get a dictionary for (say)
459 Eq a from a dictionary for Eq [a]. So if we find
463 we can't transform to
468 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
470 Of course, we currently have no way to automatically derive
471 eqList, nor to connect it to the Eq [a] instance decl, but you
472 can imagine that it might somehow be possible. Taking advantage
473 of this is permanently ruled out.
475 Still, this is no great hardship, because we intend to eliminate
476 overloading altogether anyway!
478 A note about non-tyvar dictionaries
479 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
480 Some Ids have types like
482 forall a,b,c. Eq a -> Ord [a] -> tau
484 This seems curious at first, because we usually only have dictionary
485 args whose types are of the form (C a) where a is a type variable.
486 But this doesn't hold for the functions arising from instance decls,
487 which sometimes get arguements with types of form (C (T a)) for some
490 Should we specialise wrt this compound-type dictionary? We used to say
492 "This is a heuristic judgement, as indeed is the fact that we
493 specialise wrt only dictionaries. We choose *not* to specialise
494 wrt compound dictionaries because at the moment the only place
495 they show up is in instance decls, where they are simply plugged
496 into a returned dictionary. So nothing is gained by specialising
499 But it is simpler and more uniform to specialise wrt these dicts too;
500 and in future GHC is likely to support full fledged type signatures
502 f :: Eq [(a,b)] => ...
505 %************************************************************************
507 \subsubsection{The new specialiser}
509 %************************************************************************
511 Our basic game plan is this. For let(rec) bound function
512 f :: (C a, D c) => (a,b,c,d) -> Bool
514 * Find any specialised calls of f, (f ts ds), where
515 ts are the type arguments t1 .. t4, and
516 ds are the dictionary arguments d1 .. d2.
518 * Add a new definition for f1 (say):
520 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
522 Note that we abstract over the unconstrained type arguments.
526 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
528 to the specialisations of f. This will be used by the
529 simplifier to replace calls
530 (f t1 t2 t3 t4) da db
532 (\d1 d1 -> f1 t2 t4) da db
534 All the stuff about how many dictionaries to discard, and what types
535 to apply the specialised function to, are handled by the fact that the
536 SpecEnv contains a template for the result of the specialisation.
538 We don't build *partial* specialisations for f. For example:
540 f :: Eq a => a -> a -> Bool
541 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
543 Here, little is gained by making a specialised copy of f.
544 There's a distinct danger that the specialised version would
545 first build a dictionary for (Eq b, Eq c), and then select the (==)
546 method from it! Even if it didn't, not a great deal is saved.
548 We do, however, generate polymorphic, but not overloaded, specialisations:
550 f :: Eq a => [a] -> b -> b -> b
551 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
553 Hence, the invariant is this:
555 *** no specialised version is overloaded ***
558 %************************************************************************
560 \subsubsection{The exported function}
562 %************************************************************************
565 specProgram :: UniqSupply -> [CoreBind] -> [CoreBind]
566 specProgram us binds = initSM us $
567 do { (binds', uds') <- go binds
568 ; return (wrapDictBinds (ud_binds uds') binds') }
570 -- We need to start with a Subst that knows all the things
571 -- that are in scope, so that the substitution engine doesn't
572 -- accidentally re-use a unique that's already in use
573 -- Easiest thing is to do it all at once, as if all the top-level
574 -- decls were mutually recursive
575 top_subst = mkEmptySubst (mkInScopeSet (mkVarSet (bindersOfBinds binds)))
577 go [] = return ([], emptyUDs)
578 go (bind:binds) = do (binds', uds) <- go binds
579 (bind', uds') <- specBind top_subst bind uds
580 return (bind' ++ binds', uds')
583 %************************************************************************
585 \subsubsection{@specExpr@: the main function}
587 %************************************************************************
590 specVar :: Subst -> Id -> CoreExpr
591 specVar subst v = lookupIdSubst subst v
593 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
594 -- We carry a substitution down:
595 -- a) we must clone any binding that might float outwards,
596 -- to avoid name clashes
597 -- b) we carry a type substitution to use when analysing
598 -- the RHS of specialised bindings (no type-let!)
600 ---------------- First the easy cases --------------------
601 specExpr subst (Type ty) = return (Type (CoreSubst.substTy subst ty), emptyUDs)
602 specExpr subst (Var v) = return (specVar subst v, emptyUDs)
603 specExpr _ (Lit lit) = return (Lit lit, emptyUDs)
604 specExpr subst (Cast e co) = do
605 (e', uds) <- specExpr subst e
606 return ((Cast e' (CoreSubst.substTy subst co)), uds)
607 specExpr subst (Note note body) = do
608 (body', uds) <- specExpr subst body
609 return (Note (specNote subst note) body', uds)
612 ---------------- Applications might generate a call instance --------------------
613 specExpr subst expr@(App {})
616 go (App fun arg) args = do (arg', uds_arg) <- specExpr subst arg
617 (fun', uds_app) <- go fun (arg':args)
618 return (App fun' arg', uds_arg `plusUDs` uds_app)
620 go (Var f) args = case specVar subst f of
621 Var f' -> return (Var f', mkCallUDs f' args)
622 e' -> return (e', emptyUDs) -- I don't expect this!
623 go other _ = specExpr subst other
625 ---------------- Lambda/case require dumping of usage details --------------------
626 specExpr subst e@(Lam _ _) = do
627 (body', uds) <- specExpr subst' body
628 let (free_uds, dumped_dbs) = dumpUDs bndrs' uds
629 return (mkLams bndrs' (wrapDictBindsE dumped_dbs body'), free_uds)
631 (bndrs, body) = collectBinders e
632 (subst', bndrs') = substBndrs subst bndrs
633 -- More efficient to collect a group of binders together all at once
634 -- and we don't want to split a lambda group with dumped bindings
636 specExpr subst (Case scrut case_bndr ty alts) = do
637 (scrut', uds_scrut) <- specExpr subst scrut
638 (alts', uds_alts) <- mapAndCombineSM spec_alt alts
639 return (Case scrut' case_bndr' (CoreSubst.substTy subst ty) alts',
640 uds_scrut `plusUDs` uds_alts)
642 (subst_alt, case_bndr') = substBndr subst case_bndr
643 -- No need to clone case binder; it can't float like a let(rec)
645 spec_alt (con, args, rhs) = do
646 (rhs', uds) <- specExpr subst_rhs rhs
647 let (free_uds, dumped_dbs) = dumpUDs args' uds
648 return ((con, args', wrapDictBindsE dumped_dbs rhs'), free_uds)
650 (subst_rhs, args') = substBndrs subst_alt args
652 ---------------- Finally, let is the interesting case --------------------
653 specExpr subst (Let bind body) = do
655 (rhs_subst, body_subst, bind') <- cloneBindSM subst bind
657 -- Deal with the body
658 (body', body_uds) <- specExpr body_subst body
660 -- Deal with the bindings
661 (binds', uds) <- specBind rhs_subst bind' body_uds
664 return (foldr Let body' binds', uds)
666 -- Must apply the type substitution to coerceions
667 specNote :: Subst -> Note -> Note
668 specNote _ note = note
671 %************************************************************************
673 \subsubsection{Dealing with a binding}
675 %************************************************************************
678 specBind :: Subst -- Use this for RHSs
680 -> UsageDetails -- Info on how the scope of the binding
681 -> SpecM ([CoreBind], -- New bindings
682 UsageDetails) -- And info to pass upstream
684 -- Returned UsageDetails:
685 -- No calls for binders of this bind
686 specBind rhs_subst (NonRec fn rhs) body_uds
687 = do { (rhs', rhs_uds) <- specExpr rhs_subst rhs
688 ; (fn', spec_defns, body_uds1) <- specDefn rhs_subst body_uds fn rhs
690 ; let pairs = spec_defns ++ [(fn', rhs')]
691 -- fn' mentions the spec_defns in its rules,
692 -- so put the latter first
694 combined_uds = body_uds1 `plusUDs` rhs_uds
695 -- This way round a call in rhs_uds of a function f
696 -- at type T will override a call of f at T in body_uds1; and
697 -- that is good because it'll tend to keep "earlier" calls
698 -- See Note [Specialisation of dictionary functions]
700 (free_uds, dump_dbs, float_all) = dumpBindUDs [fn] combined_uds
701 -- See Note [From non-recursive to recursive]
703 final_binds | isEmptyBag dump_dbs = [NonRec b r | (b,r) <- pairs]
704 | otherwise = [Rec (flattenDictBinds dump_dbs pairs)]
707 -- Rather than discard the calls mentioning the bound variables
708 -- we float this binding along with the others
709 return ([], free_uds `snocDictBinds` final_binds)
711 -- No call in final_uds mentions bound variables,
712 -- so we can just leave the binding here
713 return (final_binds, free_uds) }
716 specBind rhs_subst (Rec pairs) body_uds
717 -- Note [Specialising a recursive group]
718 = do { let (bndrs,rhss) = unzip pairs
719 ; (rhss', rhs_uds) <- mapAndCombineSM (specExpr rhs_subst) rhss
720 ; let scope_uds = body_uds `plusUDs` rhs_uds
721 -- Includes binds and calls arising from rhss
723 ; (bndrs1, spec_defns1, uds1) <- specDefns rhs_subst scope_uds pairs
725 ; (bndrs3, spec_defns3, uds3)
726 <- if null spec_defns1 -- Common case: no specialisation
727 then return (bndrs1, [], uds1)
728 else do { -- Specialisation occurred; do it again
729 (bndrs2, spec_defns2, uds2)
730 <- specDefns rhs_subst uds1 (bndrs1 `zip` rhss)
731 ; return (bndrs2, spec_defns2 ++ spec_defns1, uds2) }
733 ; let (final_uds, dumped_dbs, float_all) = dumpBindUDs bndrs uds3
734 bind = Rec (flattenDictBinds dumped_dbs $
735 spec_defns3 ++ zip bndrs3 rhss')
738 return ([], final_uds `snocDictBind` bind)
740 return ([bind], final_uds) }
743 ---------------------------
745 -> UsageDetails -- Info on how it is used in its scope
746 -> [(Id,CoreExpr)] -- The things being bound and their un-processed RHS
747 -> SpecM ([Id], -- Original Ids with RULES added
748 [(Id,CoreExpr)], -- Extra, specialised bindings
749 UsageDetails) -- Stuff to fling upwards from the specialised versions
751 -- Specialise a list of bindings (the contents of a Rec), but flowing usages
752 -- upwards binding by binding. Example: { f = ...g ...; g = ...f .... }
753 -- Then if the input CallDetails has a specialised call for 'g', whose specialisation
754 -- in turn generates a specialised call for 'f', we catch that in this one sweep.
755 -- But not vice versa (it's a fixpoint problem).
757 specDefns _subst uds []
758 = return ([], [], uds)
759 specDefns subst uds ((bndr,rhs):pairs)
760 = do { (bndrs1, spec_defns1, uds1) <- specDefns subst uds pairs
761 ; (bndr1, spec_defns2, uds2) <- specDefn subst uds1 bndr rhs
762 ; return (bndr1 : bndrs1, spec_defns1 ++ spec_defns2, uds2) }
764 ---------------------------
766 -> UsageDetails -- Info on how it is used in its scope
767 -> Id -> CoreExpr -- The thing being bound and its un-processed RHS
768 -> SpecM (Id, -- Original Id with added RULES
769 [(Id,CoreExpr)], -- Extra, specialised bindings
770 UsageDetails) -- Stuff to fling upwards from the specialised versions
772 specDefn subst body_uds fn rhs
773 -- The first case is the interesting one
774 | rhs_tyvars `lengthIs` n_tyvars -- Rhs of fn's defn has right number of big lambdas
775 && rhs_ids `lengthAtLeast` n_dicts -- and enough dict args
776 && notNull calls_for_me -- And there are some calls to specialise
778 -- && not (certainlyWillInline (idUnfolding fn)) -- And it's not small
779 -- See Note [Inline specialisation] for why we do not
780 -- switch off specialisation for inline functions
782 = do { -- Make a specialised version for each call in calls_for_me
783 stuff <- mapM spec_call calls_for_me
784 ; let (spec_defns, spec_uds, spec_rules) = unzip3 (catMaybes stuff)
785 fn' = addIdSpecialisations fn spec_rules
786 final_uds = body_uds_without_me `plusUDs` plusUDList spec_uds
787 -- It's important that the `plusUDs` is this way
788 -- round, because body_uds_without_me may bind
789 -- dictionaries that are used in calls_for_me passed
790 -- to specDefn. So the dictionary bindings in
791 -- spec_uds may mention dictionaries bound in
792 -- body_uds_without_me
794 ; return (fn', spec_defns, final_uds) }
796 | otherwise -- No calls or RHS doesn't fit our preconceptions
797 = WARN( notNull calls_for_me, ptext (sLit "Missed specialisation opportunity for") <+> ppr fn )
798 -- Note [Specialisation shape]
799 return (fn, [], body_uds_without_me)
803 fn_arity = idArity fn
804 fn_unf = idUnfolding fn
805 (tyvars, theta, _) = tcSplitSigmaTy fn_type
806 n_tyvars = length tyvars
807 n_dicts = length theta
808 inline_act = idInlineActivation fn
810 -- Figure out whether the function has an INLINE pragma
811 -- See Note [Inline specialisations]
812 fn_has_inline_rule :: Maybe (InlineRuleInfo, Arity) -- Gives arity of the *specialised* inline rule
814 | Just inl <- isInlineRule_maybe fn_unf
816 InlWrapper _ -> Just (InlUnSat, spec_arity)
817 _ -> Just (inl, spec_arity)
818 | otherwise = Nothing
820 spec_arity = unfoldingArity fn_unf - n_dicts
822 (rhs_tyvars, rhs_ids, rhs_body) = collectTyAndValBinders rhs
824 (body_uds_without_me, calls_for_me) = callsForMe fn body_uds
826 rhs_dict_ids = take n_dicts rhs_ids
827 body = mkLams (drop n_dicts rhs_ids) rhs_body
828 -- Glue back on the non-dict lambdas
830 already_covered :: [CoreExpr] -> Bool
831 already_covered args -- Note [Specialisations already covered]
832 = isJust (lookupRule (const True) (substInScope subst)
833 fn args (idCoreRules fn))
835 mk_ty_args :: [Maybe Type] -> [CoreExpr]
836 mk_ty_args call_ts = zipWithEqual "spec_call" mk_ty_arg rhs_tyvars call_ts
838 mk_ty_arg rhs_tyvar Nothing = Type (mkTyVarTy rhs_tyvar)
839 mk_ty_arg _ (Just ty) = Type ty
841 ----------------------------------------------------------
842 -- Specialise to one particular call pattern
843 spec_call :: CallInfo -- Call instance
844 -> SpecM (Maybe ((Id,CoreExpr), -- Specialised definition
845 UsageDetails, -- Usage details from specialised body
846 CoreRule)) -- Info for the Id's SpecEnv
847 spec_call (CallKey call_ts, (call_ds, _))
848 = ASSERT( call_ts `lengthIs` n_tyvars && call_ds `lengthIs` n_dicts )
850 -- Suppose f's defn is f = /\ a b c -> \ d1 d2 -> rhs
851 -- Supppose the call is for f [Just t1, Nothing, Just t3] [dx1, dx2]
853 -- Construct the new binding
854 -- f1 = SUBST[a->t1,c->t3, d1->d1', d2->d2'] (/\ b -> rhs)
855 -- PLUS the usage-details
856 -- { d1' = dx1; d2' = dx2 }
857 -- where d1', d2' are cloned versions of d1,d2, with the type substitution
858 -- applied. These auxiliary bindings just avoid duplication of dx1, dx2
860 -- Note that the substitution is applied to the whole thing.
861 -- This is convenient, but just slightly fragile. Notably:
862 -- * There had better be no name clashes in a/b/c
864 -- poly_tyvars = [b] in the example above
865 -- spec_tyvars = [a,c]
866 -- ty_args = [t1,b,t3]
867 poly_tyvars = [tv | (tv, Nothing) <- rhs_tyvars `zip` call_ts]
868 spec_tv_binds = [(tv,ty) | (tv, Just ty) <- rhs_tyvars `zip` call_ts]
869 spec_ty_args = map snd spec_tv_binds
870 ty_args = mk_ty_args call_ts
871 rhs_subst = CoreSubst.extendTvSubstList subst spec_tv_binds
873 ; (rhs_subst1, inst_dict_ids) <- cloneDictBndrs rhs_subst rhs_dict_ids
874 -- Clone rhs_dicts, including instantiating their types
876 ; let (rhs_subst2, dx_binds) = bindAuxiliaryDicts rhs_subst1 $
877 (my_zipEqual rhs_dict_ids inst_dict_ids call_ds)
878 inst_args = ty_args ++ map Var inst_dict_ids
880 ; if already_covered inst_args then
883 { -- Figure out the type of the specialised function
884 let body_ty = applyTypeToArgs rhs fn_type inst_args
885 (lam_args, app_args) -- Add a dummy argument if body_ty is unlifted
886 | isUnLiftedType body_ty -- C.f. WwLib.mkWorkerArgs
887 = (poly_tyvars ++ [voidArgId], poly_tyvars ++ [realWorldPrimId])
888 | otherwise = (poly_tyvars, poly_tyvars)
889 spec_id_ty = mkPiTypes lam_args body_ty
891 ; spec_f <- newSpecIdSM fn spec_id_ty
892 ; let spec_f_w_arity = setIdArity spec_f (max 0 (fn_arity - n_dicts))
893 -- Adding arity information just propagates it a bit faster
894 -- See Note [Arity decrease] in Simplify
896 ; (spec_rhs, rhs_uds) <- specExpr rhs_subst2 (mkLams lam_args body)
898 -- The rule to put in the function's specialisation is:
899 -- forall b, d1',d2'. f t1 b t3 d1' d2' = f1 b
900 rule_name = mkFastString ("SPEC " ++ showSDoc (ppr fn <+> ppr spec_ty_args))
901 spec_env_rule = mkLocalRule
903 inline_act -- Note [Auto-specialisation and RULES]
905 (poly_tyvars ++ inst_dict_ids)
907 (mkVarApps (Var spec_f_w_arity) app_args)
909 -- Add the { d1' = dx1; d2' = dx2 } usage stuff
910 final_uds = foldr consDictBind rhs_uds dx_binds
912 -- See Note [Inline specialisations]
913 final_spec_f | Just (inl, spec_arity) <- fn_has_inline_rule
914 = spec_f_w_arity `setInlineActivation` inline_act
915 `setIdUnfolding` mkInlineRule inl spec_rhs spec_arity
916 -- I'm not sure this should be unconditionally InlSat
919 ; return (Just ((final_spec_f, spec_rhs), final_uds, spec_env_rule)) } }
922 | debugIsOn && not (equalLength xs ys && equalLength ys zs)
923 = pprPanic "my_zipEqual" (vcat [ ppr xs, ppr ys
924 , ppr fn <+> ppr call_ts
925 , ppr (idType fn), ppr theta
926 , ppr n_dicts, ppr rhs_dict_ids
928 | otherwise = zip3 xs ys zs
932 -> [(DictId,DictId,CoreExpr)] -- (orig_dict, inst_dict, dx)
933 -> (Subst, -- Substitute for all orig_dicts
934 [CoreBind]) -- Auxiliary bindings
935 -- Bind any dictionary arguments to fresh names, to preserve sharing
936 -- Substitution already substitutes orig_dict -> inst_dict
937 bindAuxiliaryDicts subst triples = go subst [] triples
939 go subst binds [] = (subst, binds)
940 go subst binds ((d, dx_id, dx) : pairs)
941 | exprIsTrivial dx = go (extendIdSubst subst d dx) binds pairs
942 -- No auxiliary binding necessary
943 | otherwise = go subst_w_unf (NonRec dx_id dx : binds) pairs
945 dx_id1 = dx_id `setIdUnfolding` mkUnfolding False dx
946 subst_w_unf = extendIdSubst subst d (Var dx_id1)
947 -- Important! We're going to substitute dx_id1 for d
948 -- and we want it to look "interesting", else we won't gather *any*
949 -- consequential calls. E.g.
951 -- If we specialise f for a call (f (dfun dNumInt)), we'll get
952 -- a consequent call (g d') with an auxiliary definition
954 -- We want that consequent call to look interesting
957 Note [From non-recursive to recursive]
958 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
959 Even in the non-recursive case, if any dict-binds depend on 'fn' we might
960 have built a recursive knot
963 MkUD { ud_binds = d7 = MkD ..f..
964 , ud_calls = ...(f T d7)... }
968 Rec { fs x = <blah>[T/a, d7/d]
973 Here the recursion is only through the RULE.
976 Note [Specialisation of dictionary functions]
977 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
978 Here is a nasty example that bit us badly: see Trac #3591
980 dfun a d = MkD a d (meth d)
986 None of these definitions is recursive. What happened was that we
987 generated a specialisation:
989 RULE forall d. dfun T d = dT
990 dT = (MkD a d (meth d)) [T/a, d1/d]
993 But now we use the RULE on the RHS of d2, to get
995 d2 = dT = MkD d1 (meth d1)
998 and now d1 is bottom! The problem is that when specialising 'dfun' we
999 should first dump "below" the binding all floated dictionary bindings
1000 that mention 'dfun' itself. So d2 and d3 (and hence d1) must be
1001 placed below 'dfun', and thus unavailable to it when specialising
1002 'dfun'. That in turn means that the call (dfun T d1) must be
1003 discarded. On the other hand, the call (dfun T d4) is fine, assuming
1004 d4 doesn't mention dfun.
1008 class C a where { foo,bar :: [a] -> [a] }
1010 instance C Int where
1014 r_bar :: C a => [a] -> [a]
1015 r_bar xs = bar (xs ++ xs)
1019 r_bar a (c::C a) (xs::[a]) = bar a d (xs ++ xs)
1021 Rec { $fCInt :: C Int = MkC foo_help reverse
1022 foo_help (xs::[Int]) = r_bar Int $fCInt xs }
1024 The call (r_bar $fCInt) mentions $fCInt,
1025 which mentions foo_help,
1026 which mentions r_bar
1027 But we DO want to specialise r_bar at Int:
1029 Rec { $fCInt :: C Int = MkC foo_help reverse
1030 foo_help (xs::[Int]) = r_bar Int $fCInt xs
1032 r_bar a (c::C a) (xs::[a]) = bar a d (xs ++ xs)
1033 RULE r_bar Int _ = r_bar_Int
1035 r_bar_Int xs = bar Int $fCInt (xs ++ xs)
1038 Note that, because of its RULE, r_bar joins the recursive
1039 group. (In this case it'll unravel a short moment later.)
1042 Conclusion: we catch the nasty case using filter_dfuns in
1043 callsForMe To be honest I'm not 100% certain that this is 100%
1044 right, but it works. Sigh.
1047 Note [Specialising a recursive group]
1048 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1050 let rec { f x = ...g x'...
1051 ; g y = ...f y'.... }
1053 Here we specialise 'f' at Char; but that is very likely to lead to
1054 a specialisation of 'g' at Char. We must do the latter, else the
1055 whole point of specialisation is lost.
1057 But we do not want to keep iterating to a fixpoint, because in the
1058 presence of polymorphic recursion we might generate an infinite number
1061 So we use the following heuristic:
1062 * Arrange the rec block in dependency order, so far as possible
1063 (the occurrence analyser already does this)
1065 * Specialise it much like a sequence of lets
1067 * Then go through the block a second time, feeding call-info from
1068 the RHSs back in the bottom, as it were
1070 In effect, the ordering maxmimises the effectiveness of each sweep,
1071 and we do just two sweeps. This should catch almost every case of
1072 monomorphic recursion -- the exception could be a very knotted-up
1073 recursion with multiple cycles tied up together.
1075 This plan is implemented in the Rec case of specBindItself.
1077 Note [Specialisations already covered]
1078 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1079 We obviously don't want to generate two specialisations for the same
1080 argument pattern. There are two wrinkles
1082 1. We do the already-covered test in specDefn, not when we generate
1083 the CallInfo in mkCallUDs. We used to test in the latter place, but
1084 we now iterate the specialiser somewhat, and the Id at the call site
1085 might therefore not have all the RULES that we can see in specDefn
1087 2. What about two specialisations where the second is an *instance*
1088 of the first? If the more specific one shows up first, we'll generate
1089 specialisations for both. If the *less* specific one shows up first,
1090 we *don't* currently generate a specialisation for the more specific
1091 one. (See the call to lookupRule in already_covered.) Reasons:
1092 (a) lookupRule doesn't say which matches are exact (bad reason)
1093 (b) if the earlier specialisation is user-provided, it's
1094 far from clear that we should auto-specialise further
1096 Note [Auto-specialisation and RULES]
1097 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1099 g :: Num a => a -> a
1102 f :: (Int -> Int) -> Int
1104 {-# RULE f g = 0 #-}
1106 Suppose that auto-specialisation makes a specialised version of
1107 g::Int->Int That version won't appear in the LHS of the RULE for f.
1108 So if the specialisation rule fires too early, the rule for f may
1111 It might be possible to add new rules, to "complete" the rewrite system.
1113 RULE forall d. g Int d = g_spec
1117 But that's a bit complicated. For now we ask the programmer's help,
1118 by *copying the INLINE activation pragma* to the auto-specialised rule.
1119 So if g says {-# NOINLINE[2] g #-}, then the auto-spec rule will also
1120 not be active until phase 2.
1123 Note [Specialisation shape]
1124 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1125 We only specialise a function if it has visible top-level lambdas
1126 corresponding to its overloading. E.g. if
1127 f :: forall a. Eq a => ....
1128 then its body must look like
1131 Reason: when specialising the body for a call (f ty dexp), we want to
1132 substitute dexp for d, and pick up specialised calls in the body of f.
1134 This doesn't always work. One example I came across was this:
1135 newtype Gen a = MkGen{ unGen :: Int -> a }
1137 choose :: Eq a => a -> Gen a
1138 choose n = MkGen (\r -> n)
1140 oneof = choose (1::Int)
1142 It's a silly exapmle, but we get
1143 choose = /\a. g `cast` co
1144 where choose doesn't have any dict arguments. Thus far I have not
1145 tried to fix this (wait till there's a real example).
1148 Note [Inline specialisations]
1149 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1150 We transfer to the specialised function any INLINE stuff from the
1151 original. This means (a) the Activation in the IdInfo, and (b) any
1152 InlineMe on the RHS. We do not, however, transfer the RuleMatchInfo
1153 since we do not expect the specialisation to occur in rewrite rules.
1155 This is a change (Jun06). Previously the idea is that the point of
1156 inlining was precisely to specialise the function at its call site,
1157 and that's not so important for the specialised copies. But
1158 *pragma-directed* specialisation now takes place in the
1159 typechecker/desugarer, with manually specified INLINEs. The
1160 specialiation here is automatic. It'd be very odd if a function
1161 marked INLINE was specialised (because of some local use), and then
1162 forever after (including importing modules) the specialised version
1163 wasn't INLINEd. After all, the programmer said INLINE!
1165 You might wonder why we don't just not specialise INLINE functions.
1166 It's because even INLINE functions are sometimes not inlined, when
1167 they aren't applied to interesting arguments. But perhaps the type
1168 arguments alone are enough to specialise (even though the args are too
1169 boring to trigger inlining), and it's certainly better to call the
1170 specialised version.
1172 A case in point is dictionary functions, which are current marked
1173 INLINE, but which are worth specialising.
1176 %************************************************************************
1178 \subsubsection{UsageDetails and suchlike}
1180 %************************************************************************
1185 ud_binds :: !(Bag DictBind),
1186 -- Floated dictionary bindings
1187 -- The order is important;
1188 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
1189 -- (Remember, Bags preserve order in GHC.)
1191 ud_calls :: !CallDetails
1193 -- INVARIANT: suppose bs = bindersOf ud_binds
1194 -- Then 'calls' may *mention* 'bs',
1195 -- but there should be no calls *for* bs
1198 instance Outputable UsageDetails where
1199 ppr (MkUD { ud_binds = dbs, ud_calls = calls })
1200 = ptext (sLit "MkUD") <+> braces (sep (punctuate comma
1201 [ptext (sLit "binds") <+> equals <+> ppr dbs,
1202 ptext (sLit "calls") <+> equals <+> ppr calls]))
1204 type DictBind = (CoreBind, VarSet)
1205 -- The set is the free vars of the binding
1206 -- both tyvars and dicts
1208 type DictExpr = CoreExpr
1210 emptyUDs :: UsageDetails
1211 emptyUDs = MkUD { ud_binds = emptyBag, ud_calls = emptyVarEnv }
1213 ------------------------------------------------------------
1214 type CallDetails = IdEnv CallInfoSet
1215 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
1217 -- CallInfo uses a FiniteMap, thereby ensuring that
1218 -- we record only one call instance for any key
1220 -- The list of types and dictionaries is guaranteed to
1221 -- match the type of f
1222 type CallInfoSet = FiniteMap CallKey ([DictExpr], VarSet)
1223 -- Range is dict args and the vars of the whole
1224 -- call (including tyvars)
1225 -- [*not* include the main id itself, of course]
1227 type CallInfo = (CallKey, ([DictExpr], VarSet))
1229 instance Outputable CallKey where
1230 ppr (CallKey ts) = ppr ts
1232 -- Type isn't an instance of Ord, so that we can control which
1233 -- instance we use. That's tiresome here. Oh well
1234 instance Eq CallKey where
1235 k1 == k2 = case k1 `compare` k2 of { EQ -> True; _ -> False }
1237 instance Ord CallKey where
1238 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
1240 cmp Nothing Nothing = EQ
1241 cmp Nothing (Just _) = LT
1242 cmp (Just _) Nothing = GT
1243 cmp (Just t1) (Just t2) = tcCmpType t1 t2
1245 unionCalls :: CallDetails -> CallDetails -> CallDetails
1246 unionCalls c1 c2 = plusVarEnv_C plusFM c1 c2
1248 -- plusCalls :: UsageDetails -> CallDetails -> UsageDetails
1249 -- plusCalls uds call_ds = uds { ud_calls = ud_calls uds `unionCalls` call_ds }
1251 callDetailsFVs :: CallDetails -> VarSet
1252 callDetailsFVs calls = foldVarEnv (unionVarSet . callInfoFVs) emptyVarSet calls
1254 callInfoFVs :: CallInfoSet -> VarSet
1255 callInfoFVs call_info = foldFM (\_ (_,fv) vs -> unionVarSet fv vs) emptyVarSet call_info
1257 ------------------------------------------------------------
1258 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> UsageDetails
1259 singleCall id tys dicts
1260 = MkUD {ud_binds = emptyBag,
1261 ud_calls = unitVarEnv id (unitFM (CallKey tys) (dicts, call_fvs)) }
1263 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
1264 tys_fvs = tyVarsOfTypes (catMaybes tys)
1265 -- The type args (tys) are guaranteed to be part of the dictionary
1266 -- types, because they are just the constrained types,
1267 -- and the dictionary is therefore sure to be bound
1268 -- inside the binding for any type variables free in the type;
1269 -- hence it's safe to neglect tyvars free in tys when making
1270 -- the free-var set for this call
1271 -- BUT I don't trust this reasoning; play safe and include tys_fvs
1273 -- We don't include the 'id' itself.
1275 mkCallUDs :: Id -> [CoreExpr] -> UsageDetails
1277 | not (isLocalId f) -- Imported from elsewhere
1278 || null theta -- Not overloaded
1279 || not (all isClassPred theta)
1280 -- Only specialise if all overloading is on class params.
1281 -- In ptic, with implicit params, the type args
1282 -- *don't* say what the value of the implicit param is!
1283 || not (spec_tys `lengthIs` n_tyvars)
1284 || not ( dicts `lengthIs` n_dicts)
1285 || not (any interestingDict dicts) -- Note [Interesting dictionary arguments]
1286 -- See also Note [Specialisations already covered]
1287 = -- pprTrace "mkCallUDs: discarding" (vcat [ppr f, ppr args, ppr n_tyvars, ppr n_dicts, ppr (map interestingDict dicts)])
1288 emptyUDs -- Not overloaded, or no specialisation wanted
1291 = -- pprTrace "mkCallUDs: keeping" (vcat [ppr f, ppr args, ppr n_tyvars, ppr n_dicts, ppr (map interestingDict dicts)])
1292 singleCall f spec_tys dicts
1294 (tyvars, theta, _) = tcSplitSigmaTy (idType f)
1295 constrained_tyvars = tyVarsOfTheta theta
1296 n_tyvars = length tyvars
1297 n_dicts = length theta
1299 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
1300 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
1303 | tyvar `elemVarSet` constrained_tyvars = Just ty
1304 | otherwise = Nothing
1307 Note [Interesting dictionary arguments]
1308 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1310 \a.\d:Eq a. let f = ... in ...(f d)...
1311 There really is not much point in specialising f wrt the dictionary d,
1312 because the code for the specialised f is not improved at all, because
1313 d is lambda-bound. We simply get junk specialisations.
1315 What is "interesting"? Just that it has *some* structure.
1318 interestingDict :: CoreExpr -> Bool
1319 -- A dictionary argument is interesting if it has *some* structure
1320 interestingDict (Var v) = hasSomeUnfolding (idUnfolding v)
1321 || isDataConWorkId v
1322 interestingDict (Type _) = False
1323 interestingDict (App fn (Type _)) = interestingDict fn
1324 interestingDict (Note _ a) = interestingDict a
1325 interestingDict (Cast e _) = interestingDict e
1326 interestingDict _ = True
1330 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
1331 plusUDs (MkUD {ud_binds = db1, ud_calls = calls1})
1332 (MkUD {ud_binds = db2, ud_calls = calls2})
1333 = MkUD { ud_binds = db1 `unionBags` db2
1334 , ud_calls = calls1 `unionCalls` calls2 }
1336 plusUDList :: [UsageDetails] -> UsageDetails
1337 plusUDList = foldr plusUDs emptyUDs
1339 -----------------------------
1340 _dictBindBndrs :: Bag DictBind -> [Id]
1341 _dictBindBndrs dbs = foldrBag ((++) . bindersOf . fst) [] dbs
1343 mkDB :: CoreBind -> DictBind
1344 mkDB bind = (bind, bind_fvs bind)
1346 bind_fvs :: CoreBind -> VarSet
1347 bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
1348 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1351 rhs_fvs = unionVarSets (map pair_fvs prs)
1353 pair_fvs :: (Id, CoreExpr) -> VarSet
1354 pair_fvs (bndr, rhs) = exprFreeVars rhs `unionVarSet` idFreeVars bndr
1355 -- Don't forget variables mentioned in the
1356 -- rules of the bndr. C.f. OccAnal.addRuleUsage
1357 -- Also tyvars mentioned in its type; they may not appear in the RHS
1361 flattenDictBinds :: Bag DictBind -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
1362 flattenDictBinds dbs pairs
1363 = foldrBag add pairs dbs
1365 add (NonRec b r,_) pairs = (b,r) : pairs
1366 add (Rec prs1, _) pairs = prs1 ++ pairs
1368 snocDictBinds :: UsageDetails -> [CoreBind] -> UsageDetails
1369 -- Add ud_binds to the tail end of the bindings in uds
1370 snocDictBinds uds dbs
1371 = uds { ud_binds = ud_binds uds `unionBags`
1372 foldr (consBag . mkDB) emptyBag dbs }
1374 consDictBind :: CoreBind -> UsageDetails -> UsageDetails
1375 consDictBind bind uds = uds { ud_binds = mkDB bind `consBag` ud_binds uds }
1377 snocDictBind :: UsageDetails -> CoreBind -> UsageDetails
1378 snocDictBind uds bind = uds { ud_binds = ud_binds uds `snocBag` mkDB bind }
1380 wrapDictBinds :: Bag DictBind -> [CoreBind] -> [CoreBind]
1381 wrapDictBinds dbs binds
1382 = foldrBag add binds dbs
1384 add (bind,_) binds = bind : binds
1386 wrapDictBindsE :: Bag DictBind -> CoreExpr -> CoreExpr
1387 wrapDictBindsE dbs expr
1388 = foldrBag add expr dbs
1390 add (bind,_) expr = Let bind expr
1392 ----------------------
1393 dumpUDs :: [CoreBndr] -> UsageDetails -> (UsageDetails, Bag DictBind)
1394 -- Used at a lambda or case binder; just dump anything mentioning the binder
1395 dumpUDs bndrs uds@(MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
1396 | null bndrs = (uds, emptyBag) -- Common in case alternatives
1397 | otherwise = (free_uds, dump_dbs)
1399 free_uds = MkUD { ud_binds = free_dbs, ud_calls = free_calls }
1400 bndr_set = mkVarSet bndrs
1401 (free_dbs, dump_dbs, dump_set) = splitDictBinds orig_dbs bndr_set
1402 free_calls = deleteCallsMentioning dump_set $ -- Drop calls mentioning bndr_set on the floor
1403 deleteCallsFor bndrs orig_calls -- Discard calls for bndr_set; there should be
1404 -- no calls for any of the dicts in dump_dbs
1406 dumpBindUDs :: [CoreBndr] -> UsageDetails -> (UsageDetails, Bag DictBind, Bool)
1407 -- Used at a lambda or case binder; just dump anything mentioning the binder
1408 dumpBindUDs bndrs (MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
1409 = (free_uds, dump_dbs, float_all)
1411 free_uds = MkUD { ud_binds = free_dbs, ud_calls = free_calls }
1412 bndr_set = mkVarSet bndrs
1413 (free_dbs, dump_dbs, dump_set) = splitDictBinds orig_dbs bndr_set
1414 free_calls = deleteCallsFor bndrs orig_calls
1415 float_all = dump_set `intersectsVarSet` callDetailsFVs free_calls
1417 callsForMe :: Id -> UsageDetails -> (UsageDetails, [CallInfo])
1418 callsForMe fn (MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
1419 = -- pprTrace ("callsForMe")
1421 -- text "Orig dbs =" <+> ppr (_dictBindBndrs orig_dbs),
1422 -- text "Orig calls =" <+> ppr orig_calls,
1423 -- text "Dep set =" <+> ppr dep_set,
1424 -- text "Calls for me =" <+> ppr calls_for_me]) $
1425 (uds_without_me, calls_for_me)
1427 uds_without_me = MkUD { ud_binds = orig_dbs, ud_calls = delVarEnv orig_calls fn }
1428 calls_for_me = case lookupVarEnv orig_calls fn of
1430 Just cs -> filter_dfuns (fmToList cs)
1432 dep_set = foldlBag go (unitVarSet fn) orig_dbs
1433 go dep_set (db,fvs) | fvs `intersectsVarSet` dep_set
1434 = extendVarSetList dep_set (bindersOf db)
1437 -- Note [Specialisation of dictionary functions]
1438 filter_dfuns | isDFunId fn = filter ok_call
1439 | otherwise = \cs -> cs
1441 ok_call (_, (_,fvs)) = not (fvs `intersectsVarSet` dep_set)
1443 ----------------------
1444 splitDictBinds :: Bag DictBind -> IdSet -> (Bag DictBind, Bag DictBind, IdSet)
1445 -- Returns (free_dbs, dump_dbs, dump_set)
1446 splitDictBinds dbs bndr_set
1447 = foldlBag split_db (emptyBag, emptyBag, bndr_set) dbs
1448 -- Important that it's foldl not foldr;
1449 -- we're accumulating the set of dumped ids in dump_set
1451 split_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1452 | dump_idset `intersectsVarSet` fvs -- Dump it
1453 = (free_dbs, dump_dbs `snocBag` db,
1454 extendVarSetList dump_idset (bindersOf bind))
1456 | otherwise -- Don't dump it
1457 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1460 ----------------------
1461 deleteCallsMentioning :: VarSet -> CallDetails -> CallDetails
1462 -- Remove calls *mentioning* bs
1463 deleteCallsMentioning bs calls
1464 = mapVarEnv filter_calls calls
1466 filter_calls :: CallInfoSet -> CallInfoSet
1467 filter_calls = filterFM (\_ (_, fvs) -> not (fvs `intersectsVarSet` bs))
1469 deleteCallsFor :: [Id] -> CallDetails -> CallDetails
1470 -- Remove calls *for* bs
1471 deleteCallsFor bs calls = delVarEnvList calls bs
1475 %************************************************************************
1477 \subsubsection{Boring helper functions}
1479 %************************************************************************
1482 type SpecM a = UniqSM a
1484 initSM :: UniqSupply -> SpecM a -> a
1487 mapAndCombineSM :: (a -> SpecM (b, UsageDetails)) -> [a] -> SpecM ([b], UsageDetails)
1488 mapAndCombineSM _ [] = return ([], emptyUDs)
1489 mapAndCombineSM f (x:xs) = do (y, uds1) <- f x
1490 (ys, uds2) <- mapAndCombineSM f xs
1491 return (y:ys, uds1 `plusUDs` uds2)
1493 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1494 -- Clone the binders of the bind; return new bind with the cloned binders
1495 -- Return the substitution to use for RHSs, and the one to use for the body
1496 cloneBindSM subst (NonRec bndr rhs) = do
1497 us <- getUniqueSupplyM
1498 let (subst', bndr') = cloneIdBndr subst us bndr
1499 return (subst, subst', NonRec bndr' rhs)
1501 cloneBindSM subst (Rec pairs) = do
1502 us <- getUniqueSupplyM
1503 let (subst', bndrs') = cloneRecIdBndrs subst us (map fst pairs)
1504 return (subst', subst', Rec (bndrs' `zip` map snd pairs))
1506 cloneDictBndrs :: Subst -> [CoreBndr] -> SpecM (Subst, [CoreBndr])
1507 cloneDictBndrs subst bndrs
1508 = do { us <- getUniqueSupplyM
1509 ; return (cloneIdBndrs subst us bndrs) }
1511 newSpecIdSM :: Id -> Type -> SpecM Id
1512 -- Give the new Id a similar occurrence name to the old one
1513 newSpecIdSM old_id new_ty
1514 = do { uniq <- getUniqueM
1516 name = idName old_id
1517 new_occ = mkSpecOcc (nameOccName name)
1518 new_id = mkUserLocal new_occ uniq new_ty (getSrcSpan name)
1523 Old (but interesting) stuff about unboxed bindings
1524 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1526 What should we do when a value is specialised to a *strict* unboxed value?
1528 map_*_* f (x:xs) = let h = f x
1532 Could convert let to case:
1534 map_*_Int# f (x:xs) = case f x of h# ->
1538 This may be undesirable since it forces evaluation here, but the value
1539 may not be used in all branches of the body. In the general case this
1540 transformation is impossible since the mutual recursion in a letrec
1541 cannot be expressed as a case.
1543 There is also a problem with top-level unboxed values, since our
1544 implementation cannot handle unboxed values at the top level.
1546 Solution: Lift the binding of the unboxed value and extract it when it
1549 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1554 Now give it to the simplifier and the _Lifting will be optimised away.
1556 The benfit is that we have given the specialised "unboxed" values a
1557 very simplep lifted semantics and then leave it up to the simplifier to
1558 optimise it --- knowing that the overheads will be removed in nearly
1561 In particular, the value will only be evaluted in the branches of the
1562 program which use it, rather than being forced at the point where the
1563 value is bound. For example:
1565 filtermap_*_* p f (x:xs)
1572 filtermap_*_Int# p f (x:xs)
1573 = let h = case (f x) of h# -> _Lift h#
1576 True -> case h of _Lift h#
1580 The binding for h can still be inlined in the one branch and the
1581 _Lifting eliminated.
1584 Question: When won't the _Lifting be eliminated?
1586 Answer: When they at the top-level (where it is necessary) or when
1587 inlining would duplicate work (or possibly code depending on
1588 options). However, the _Lifting will still be eliminated if the
1589 strictness analyser deems the lifted binding strict.