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 )
39 %************************************************************************
41 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
43 %************************************************************************
45 These notes describe how we implement specialisation to eliminate
48 The specialisation pass works on Core
49 syntax, complete with all the explicit dictionary application,
50 abstraction and construction as added by the type checker. The
51 existing type checker remains largely as it is.
53 One important thought: the {\em types} passed to an overloaded
54 function, and the {\em dictionaries} passed are mutually redundant.
55 If the same function is applied to the same type(s) then it is sure to
56 be applied to the same dictionary(s)---or rather to the same {\em
57 values}. (The arguments might look different but they will evaluate
60 Second important thought: we know that we can make progress by
61 treating dictionary arguments as static and worth specialising on. So
62 we can do without binding-time analysis, and instead specialise on
63 dictionary arguments and no others.
72 and suppose f is overloaded.
74 STEP 1: CALL-INSTANCE COLLECTION
76 We traverse <body>, accumulating all applications of f to types and
79 (Might there be partial applications, to just some of its types and
80 dictionaries? In principle yes, but in practice the type checker only
81 builds applications of f to all its types and dictionaries, so partial
82 applications could only arise as a result of transformation, and even
83 then I think it's unlikely. In any case, we simply don't accumulate such
84 partial applications.)
89 So now we have a collection of calls to f:
93 Notice that f may take several type arguments. To avoid ambiguity, we
94 say that f is called at type t1/t2 and t3/t4.
96 We take equivalence classes using equality of the *types* (ignoring
97 the dictionary args, which as mentioned previously are redundant).
99 STEP 3: SPECIALISATION
101 For each equivalence class, choose a representative (f t1 t2 d1 d2),
102 and create a local instance of f, defined thus:
104 f@t1/t2 = <f_rhs> t1 t2 d1 d2
106 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
107 of simplification will now result. However we don't actually *do* that
108 simplification. Rather, we leave it for the simplifier to do. If we
109 *did* do it, though, we'd get more call instances from the specialised
110 RHS. We can work out what they are by instantiating the call-instance
111 set from f's RHS with the types t1, t2.
113 Add this new id to f's IdInfo, to record that f has a specialised version.
115 Before doing any of this, check that f's IdInfo doesn't already
116 tell us about an existing instance of f at the required type/s.
117 (This might happen if specialisation was applied more than once, or
118 it might arise from user SPECIALIZE pragmas.)
122 Wait a minute! What if f is recursive? Then we can't just plug in
123 its right-hand side, can we?
125 But it's ok. The type checker *always* creates non-recursive definitions
126 for overloaded recursive functions. For example:
128 f x = f (x+x) -- Yes I know its silly
132 f a (d::Num a) = let p = +.sel a d
134 letrec fl (y::a) = fl (p y y)
138 We still have recusion for non-overloaded functions which we
139 speciailise, but the recursive call should get specialised to the
140 same recursive version.
146 All this is crystal clear when the function is applied to *constant
147 types*; that is, types which have no type variables inside. But what if
148 it is applied to non-constant types? Suppose we find a call of f at type
149 t1/t2. There are two possibilities:
151 (a) The free type variables of t1, t2 are in scope at the definition point
152 of f. In this case there's no problem, we proceed just as before. A common
153 example is as follows. Here's the Haskell:
158 After typechecking we have
160 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
161 in +.sel a d (f a d y) (f a d y)
163 Notice that the call to f is at type type "a"; a non-constant type.
164 Both calls to f are at the same type, so we can specialise to give:
166 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
167 in +.sel a d (f@a y) (f@a y)
170 (b) The other case is when the type variables in the instance types
171 are *not* in scope at the definition point of f. The example we are
172 working with above is a good case. There are two instances of (+.sel a d),
173 but "a" is not in scope at the definition of +.sel. Can we do anything?
174 Yes, we can "common them up", a sort of limited common sub-expression deal.
177 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
178 f@a (x::a) = +.sel@a x x
179 in +.sel@a (f@a y) (f@a y)
181 This can save work, and can't be spotted by the type checker, because
182 the two instances of +.sel weren't originally at the same type.
186 * There are quite a few variations here. For example, the defn of
187 +.sel could be floated ouside the \y, to attempt to gain laziness.
188 It certainly mustn't be floated outside the \d because the d has to
191 * We don't want to inline f_rhs in this case, because
192 that will duplicate code. Just commoning up the call is the point.
194 * Nothing gets added to +.sel's IdInfo.
196 * Don't bother unless the equivalence class has more than one item!
198 Not clear whether this is all worth it. It is of course OK to
199 simply discard call-instances when passing a big lambda.
201 Polymorphism 2 -- Overloading
203 Consider a function whose most general type is
205 f :: forall a b. Ord a => [a] -> b -> b
207 There is really no point in making a version of g at Int/Int and another
208 at Int/Bool, because it's only instancing the type variable "a" which
209 buys us any efficiency. Since g is completely polymorphic in b there
210 ain't much point in making separate versions of g for the different
213 That suggests that we should identify which of g's type variables
214 are constrained (like "a") and which are unconstrained (like "b").
215 Then when taking equivalence classes in STEP 2, we ignore the type args
216 corresponding to unconstrained type variable. In STEP 3 we make
217 polymorphic versions. Thus:
219 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
228 f a (d::Num a) = let g = ...
230 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
232 Here, g is only called at one type, but the dictionary isn't in scope at the
233 definition point for g. Usually the type checker would build a
234 definition for d1 which enclosed g, but the transformation system
235 might have moved d1's defn inward. Solution: float dictionary bindings
236 outwards along with call instances.
240 f x = let g p q = p==q
246 Before specialisation, leaving out type abstractions we have
248 f df x = let g :: Eq a => a -> a -> Bool
250 h :: Num a => a -> a -> (a, Bool)
251 h dh r s = let deq = eqFromNum dh
252 in (+ dh r s, g deq r s)
256 After specialising h we get a specialised version of h, like this:
258 h' r s = let deq = eqFromNum df
259 in (+ df r s, g deq r s)
261 But we can't naively make an instance for g from this, because deq is not in scope
262 at the defn of g. Instead, we have to float out the (new) defn of deq
263 to widen its scope. Notice that this floating can't be done in advance -- it only
264 shows up when specialisation is done.
266 User SPECIALIZE pragmas
267 ~~~~~~~~~~~~~~~~~~~~~~~
268 Specialisation pragmas can be digested by the type checker, and implemented
269 by adding extra definitions along with that of f, in the same way as before
271 f@t1/t2 = <f_rhs> t1 t2 d1 d2
273 Indeed the pragmas *have* to be dealt with by the type checker, because
274 only it knows how to build the dictionaries d1 and d2! For example
276 g :: Ord a => [a] -> [a]
277 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
279 Here, the specialised version of g is an application of g's rhs to the
280 Ord dictionary for (Tree Int), which only the type checker can conjure
281 up. There might not even *be* one, if (Tree Int) is not an instance of
282 Ord! (All the other specialision has suitable dictionaries to hand
285 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
286 it is buried in a complex (as-yet-un-desugared) binding group.
289 f@t1/t2 = f* t1 t2 d1 d2
291 where f* is the Id f with an IdInfo which says "inline me regardless!".
292 Indeed all the specialisation could be done in this way.
293 That in turn means that the simplifier has to be prepared to inline absolutely
294 any in-scope let-bound thing.
297 Again, the pragma should permit polymorphism in unconstrained variables:
299 h :: Ord a => [a] -> b -> b
300 {-# SPECIALIZE h :: [Int] -> b -> b #-}
302 We *insist* that all overloaded type variables are specialised to ground types,
303 (and hence there can be no context inside a SPECIALIZE pragma).
304 We *permit* unconstrained type variables to be specialised to
306 - or left as a polymorphic type variable
307 but nothing in between. So
309 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
311 is *illegal*. (It can be handled, but it adds complication, and gains the
315 SPECIALISING INSTANCE DECLARATIONS
316 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
319 instance Foo a => Foo [a] where
321 {-# SPECIALIZE instance Foo [Int] #-}
323 The original instance decl creates a dictionary-function
326 dfun.Foo.List :: forall a. Foo a -> Foo [a]
328 The SPECIALIZE pragma just makes a specialised copy, just as for
329 ordinary function definitions:
331 dfun.Foo.List@Int :: Foo [Int]
332 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
334 The information about what instance of the dfun exist gets added to
335 the dfun's IdInfo in the same way as a user-defined function too.
338 Automatic instance decl specialisation?
339 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
340 Can instance decls be specialised automatically? It's tricky.
341 We could collect call-instance information for each dfun, but
342 then when we specialised their bodies we'd get new call-instances
343 for ordinary functions; and when we specialised their bodies, we might get
344 new call-instances of the dfuns, and so on. This all arises because of
345 the unrestricted mutual recursion between instance decls and value decls.
347 Still, there's no actual problem; it just means that we may not do all
348 the specialisation we could theoretically do.
350 Furthermore, instance decls are usually exported and used non-locally,
351 so we'll want to compile enough to get those specialisations done.
353 Lastly, there's no such thing as a local instance decl, so we can
354 survive solely by spitting out *usage* information, and then reading that
355 back in as a pragma when next compiling the file. So for now,
356 we only specialise instance decls in response to pragmas.
359 SPITTING OUT USAGE INFORMATION
360 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
362 To spit out usage information we need to traverse the code collecting
363 call-instance information for all imported (non-prelude?) functions
364 and data types. Then we equivalence-class it and spit it out.
366 This is done at the top-level when all the call instances which escape
367 must be for imported functions and data types.
369 *** Not currently done ***
372 Partial specialisation by pragmas
373 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
374 What about partial specialisation:
376 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
377 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
381 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
383 Seems quite reasonable. Similar things could be done with instance decls:
385 instance (Foo a, Foo b) => Foo (a,b) where
387 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
388 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
390 Ho hum. Things are complex enough without this. I pass.
393 Requirements for the simplifer
394 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
395 The simplifier has to be able to take advantage of the specialisation.
397 * When the simplifier finds an application of a polymorphic f, it looks in
398 f's IdInfo in case there is a suitable instance to call instead. This converts
400 f t1 t2 d1 d2 ===> f_t1_t2
402 Note that the dictionaries get eaten up too!
404 * Dictionary selection operations on constant dictionaries must be
407 +.sel Int d ===> +Int
409 The obvious way to do this is in the same way as other specialised
410 calls: +.sel has inside it some IdInfo which tells that if it's applied
411 to the type Int then it should eat a dictionary and transform to +Int.
413 In short, dictionary selectors need IdInfo inside them for constant
416 * Exactly the same applies if a superclass dictionary is being
419 Eq.sel Int d ===> dEqInt
421 * Something similar applies to dictionary construction too. Suppose
422 dfun.Eq.List is the function taking a dictionary for (Eq a) to
423 one for (Eq [a]). Then we want
425 dfun.Eq.List Int d ===> dEq.List_Int
427 Where does the Eq [Int] dictionary come from? It is built in
428 response to a SPECIALIZE pragma on the Eq [a] instance decl.
430 In short, dfun Ids need IdInfo with a specialisation for each
431 constant instance of their instance declaration.
433 All this uses a single mechanism: the SpecEnv inside an Id
436 What does the specialisation IdInfo look like?
437 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
439 The SpecEnv of an Id maps a list of types (the template) to an expression
443 For example, if f has this SpecInfo:
445 [Int, a] -> \d:Ord Int. f' a
447 it means that we can replace the call
449 f Int t ===> (\d. f' t)
451 This chucks one dictionary away and proceeds with the
452 specialised version of f, namely f'.
455 What can't be done this way?
456 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
457 There is no way, post-typechecker, to get a dictionary for (say)
458 Eq a from a dictionary for Eq [a]. So if we find
462 we can't transform to
467 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
469 Of course, we currently have no way to automatically derive
470 eqList, nor to connect it to the Eq [a] instance decl, but you
471 can imagine that it might somehow be possible. Taking advantage
472 of this is permanently ruled out.
474 Still, this is no great hardship, because we intend to eliminate
475 overloading altogether anyway!
477 A note about non-tyvar dictionaries
478 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
479 Some Ids have types like
481 forall a,b,c. Eq a -> Ord [a] -> tau
483 This seems curious at first, because we usually only have dictionary
484 args whose types are of the form (C a) where a is a type variable.
485 But this doesn't hold for the functions arising from instance decls,
486 which sometimes get arguements with types of form (C (T a)) for some
489 Should we specialise wrt this compound-type dictionary? We used to say
491 "This is a heuristic judgement, as indeed is the fact that we
492 specialise wrt only dictionaries. We choose *not* to specialise
493 wrt compound dictionaries because at the moment the only place
494 they show up is in instance decls, where they are simply plugged
495 into a returned dictionary. So nothing is gained by specialising
498 But it is simpler and more uniform to specialise wrt these dicts too;
499 and in future GHC is likely to support full fledged type signatures
501 f :: Eq [(a,b)] => ...
504 %************************************************************************
506 \subsubsection{The new specialiser}
508 %************************************************************************
510 Our basic game plan is this. For let(rec) bound function
511 f :: (C a, D c) => (a,b,c,d) -> Bool
513 * Find any specialised calls of f, (f ts ds), where
514 ts are the type arguments t1 .. t4, and
515 ds are the dictionary arguments d1 .. d2.
517 * Add a new definition for f1 (say):
519 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
521 Note that we abstract over the unconstrained type arguments.
525 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
527 to the specialisations of f. This will be used by the
528 simplifier to replace calls
529 (f t1 t2 t3 t4) da db
531 (\d1 d1 -> f1 t2 t4) da db
533 All the stuff about how many dictionaries to discard, and what types
534 to apply the specialised function to, are handled by the fact that the
535 SpecEnv contains a template for the result of the specialisation.
537 We don't build *partial* specialisations for f. For example:
539 f :: Eq a => a -> a -> Bool
540 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
542 Here, little is gained by making a specialised copy of f.
543 There's a distinct danger that the specialised version would
544 first build a dictionary for (Eq b, Eq c), and then select the (==)
545 method from it! Even if it didn't, not a great deal is saved.
547 We do, however, generate polymorphic, but not overloaded, specialisations:
549 f :: Eq a => [a] -> b -> b -> b
550 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
552 Hence, the invariant is this:
554 *** no specialised version is overloaded ***
557 %************************************************************************
559 \subsubsection{The exported function}
561 %************************************************************************
564 specProgram :: UniqSupply -> [CoreBind] -> [CoreBind]
565 specProgram us binds = initSM us $
566 do { (binds', uds') <- go binds
567 ; return (wrapDictBinds (ud_binds uds') binds') }
569 -- We need to start with a Subst that knows all the things
570 -- that are in scope, so that the substitution engine doesn't
571 -- accidentally re-use a unique that's already in use
572 -- Easiest thing is to do it all at once, as if all the top-level
573 -- decls were mutually recursive
574 top_subst = mkEmptySubst (mkInScopeSet (mkVarSet (bindersOfBinds binds)))
576 go [] = return ([], emptyUDs)
577 go (bind:binds) = do (binds', uds) <- go binds
578 (bind', uds') <- specBind top_subst bind uds
579 return (bind' ++ binds', uds')
582 %************************************************************************
584 \subsubsection{@specExpr@: the main function}
586 %************************************************************************
589 specVar :: Subst -> Id -> CoreExpr
590 specVar subst v = lookupIdSubst subst v
592 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
593 -- We carry a substitution down:
594 -- a) we must clone any binding that might float outwards,
595 -- to avoid name clashes
596 -- b) we carry a type substitution to use when analysing
597 -- the RHS of specialised bindings (no type-let!)
599 ---------------- First the easy cases --------------------
600 specExpr subst (Type ty) = return (Type (CoreSubst.substTy subst ty), emptyUDs)
601 specExpr subst (Var v) = return (specVar subst v, emptyUDs)
602 specExpr _ (Lit lit) = return (Lit lit, emptyUDs)
603 specExpr subst (Cast e co) = do
604 (e', uds) <- specExpr subst e
605 return ((Cast e' (CoreSubst.substTy subst co)), uds)
606 specExpr subst (Note note body) = do
607 (body', uds) <- specExpr subst body
608 return (Note (specNote subst note) body', uds)
611 ---------------- Applications might generate a call instance --------------------
612 specExpr subst expr@(App {})
615 go (App fun arg) args = do (arg', uds_arg) <- specExpr subst arg
616 (fun', uds_app) <- go fun (arg':args)
617 return (App fun' arg', uds_arg `plusUDs` uds_app)
619 go (Var f) args = case specVar subst f of
620 Var f' -> return (Var f', mkCallUDs f' args)
621 e' -> return (e', emptyUDs) -- I don't expect this!
622 go other _ = specExpr subst other
624 ---------------- Lambda/case require dumping of usage details --------------------
625 specExpr subst e@(Lam _ _) = do
626 (body', uds) <- specExpr subst' body
627 let (free_uds, dumped_dbs) = dumpUDs bndrs' uds
628 return (mkLams bndrs' (wrapDictBindsE dumped_dbs body'), free_uds)
630 (bndrs, body) = collectBinders e
631 (subst', bndrs') = substBndrs subst bndrs
632 -- More efficient to collect a group of binders together all at once
633 -- and we don't want to split a lambda group with dumped bindings
635 specExpr subst (Case scrut case_bndr ty alts) = do
636 (scrut', uds_scrut) <- specExpr subst scrut
637 (alts', uds_alts) <- mapAndCombineSM spec_alt alts
638 return (Case scrut' case_bndr' (CoreSubst.substTy subst ty) alts',
639 uds_scrut `plusUDs` uds_alts)
641 (subst_alt, case_bndr') = substBndr subst case_bndr
642 -- No need to clone case binder; it can't float like a let(rec)
644 spec_alt (con, args, rhs) = do
645 (rhs', uds) <- specExpr subst_rhs rhs
646 let (free_uds, dumped_dbs) = dumpUDs args' uds
647 return ((con, args', wrapDictBindsE dumped_dbs rhs'), free_uds)
649 (subst_rhs, args') = substBndrs subst_alt args
651 ---------------- Finally, let is the interesting case --------------------
652 specExpr subst (Let bind body) = do
654 (rhs_subst, body_subst, bind') <- cloneBindSM subst bind
656 -- Deal with the body
657 (body', body_uds) <- specExpr body_subst body
659 -- Deal with the bindings
660 (binds', uds) <- specBind rhs_subst bind' body_uds
663 return (foldr Let body' binds', uds)
665 -- Must apply the type substitution to coerceions
666 specNote :: Subst -> Note -> Note
667 specNote _ note = note
670 %************************************************************************
672 \subsubsection{Dealing with a binding}
674 %************************************************************************
677 specBind :: Subst -- Use this for RHSs
679 -> UsageDetails -- Info on how the scope of the binding
680 -> SpecM ([CoreBind], -- New bindings
681 UsageDetails) -- And info to pass upstream
683 -- Returned UsageDetails:
684 -- No calls for binders of this bind
685 specBind rhs_subst (NonRec fn rhs) body_uds
686 = do { (rhs', rhs_uds) <- specExpr rhs_subst rhs
687 ; (fn', spec_defns, body_uds1) <- specDefn rhs_subst body_uds fn rhs
689 ; let pairs = spec_defns ++ [(fn', rhs')]
690 -- fn' mentions the spec_defns in its rules,
691 -- so put the latter first
693 combined_uds = body_uds1 `plusUDs` rhs_uds
694 -- This way round a call in rhs_uds of a function f
695 -- at type T will override a call of f at T in body_uds1; and
696 -- that is good because it'll tend to keep "earlier" calls
697 -- See Note [Specialisation of dictionary functions]
699 (free_uds, dump_dbs, float_all) = dumpBindUDs [fn] combined_uds
700 -- See Note [From non-recursive to recursive]
702 final_binds | isEmptyBag dump_dbs = [NonRec b r | (b,r) <- pairs]
703 | otherwise = [Rec (flattenDictBinds dump_dbs pairs)]
706 -- Rather than discard the calls mentioning the bound variables
707 -- we float this binding along with the others
708 return ([], free_uds `snocDictBinds` final_binds)
710 -- No call in final_uds mentions bound variables,
711 -- so we can just leave the binding here
712 return (final_binds, free_uds) }
715 specBind rhs_subst (Rec pairs) body_uds
716 -- Note [Specialising a recursive group]
717 = do { let (bndrs,rhss) = unzip pairs
718 ; (rhss', rhs_uds) <- mapAndCombineSM (specExpr rhs_subst) rhss
719 ; let scope_uds = body_uds `plusUDs` rhs_uds
720 -- Includes binds and calls arising from rhss
722 ; (bndrs1, spec_defns1, uds1) <- specDefns rhs_subst scope_uds pairs
724 ; (bndrs3, spec_defns3, uds3)
725 <- if null spec_defns1 -- Common case: no specialisation
726 then return (bndrs1, [], uds1)
727 else do { -- Specialisation occurred; do it again
728 (bndrs2, spec_defns2, uds2)
729 <- specDefns rhs_subst uds1 (bndrs1 `zip` rhss)
730 ; return (bndrs2, spec_defns2 ++ spec_defns1, uds2) }
732 ; let (final_uds, dumped_dbs, float_all) = dumpBindUDs bndrs uds3
733 bind = Rec (flattenDictBinds dumped_dbs $
734 spec_defns3 ++ zip bndrs3 rhss')
737 return ([], final_uds `snocDictBind` bind)
739 return ([bind], final_uds) }
742 ---------------------------
744 -> UsageDetails -- Info on how it is used in its scope
745 -> [(Id,CoreExpr)] -- The things being bound and their un-processed RHS
746 -> SpecM ([Id], -- Original Ids with RULES added
747 [(Id,CoreExpr)], -- Extra, specialised bindings
748 UsageDetails) -- Stuff to fling upwards from the specialised versions
750 -- Specialise a list of bindings (the contents of a Rec), but flowing usages
751 -- upwards binding by binding. Example: { f = ...g ...; g = ...f .... }
752 -- Then if the input CallDetails has a specialised call for 'g', whose specialisation
753 -- in turn generates a specialised call for 'f', we catch that in this one sweep.
754 -- But not vice versa (it's a fixpoint problem).
756 specDefns _subst uds []
757 = return ([], [], uds)
758 specDefns subst uds ((bndr,rhs):pairs)
759 = do { (bndrs1, spec_defns1, uds1) <- specDefns subst uds pairs
760 ; (bndr1, spec_defns2, uds2) <- specDefn subst uds1 bndr rhs
761 ; return (bndr1 : bndrs1, spec_defns1 ++ spec_defns2, uds2) }
763 ---------------------------
765 -> UsageDetails -- Info on how it is used in its scope
766 -> Id -> CoreExpr -- The thing being bound and its un-processed RHS
767 -> SpecM (Id, -- Original Id with added RULES
768 [(Id,CoreExpr)], -- Extra, specialised bindings
769 UsageDetails) -- Stuff to fling upwards from the specialised versions
771 specDefn subst body_uds fn rhs
772 -- The first case is the interesting one
773 | rhs_tyvars `lengthIs` n_tyvars -- Rhs of fn's defn has right number of big lambdas
774 && rhs_ids `lengthAtLeast` n_dicts -- and enough dict args
775 && notNull calls_for_me -- And there are some calls to specialise
777 -- && not (certainlyWillInline (idUnfolding fn)) -- And it's not small
778 -- See Note [Inline specialisation] for why we do not
779 -- switch off specialisation for inline functions
781 = do { -- Make a specialised version for each call in calls_for_me
782 stuff <- mapM spec_call calls_for_me
783 ; let (spec_defns, spec_uds, spec_rules) = unzip3 (catMaybes stuff)
784 fn' = addIdSpecialisations fn spec_rules
785 final_uds = body_uds_without_me `plusUDs` plusUDList spec_uds
786 -- It's important that the `plusUDs` is this way
787 -- round, because body_uds_without_me may bind
788 -- dictionaries that are used in calls_for_me passed
789 -- to specDefn. So the dictionary bindings in
790 -- spec_uds may mention dictionaries bound in
791 -- body_uds_without_me
793 ; return (fn', spec_defns, final_uds) }
795 | otherwise -- No calls or RHS doesn't fit our preconceptions
796 = WARN( notNull calls_for_me, ptext (sLit "Missed specialisation opportunity for") <+> ppr fn )
797 -- Note [Specialisation shape]
798 return (fn, [], body_uds_without_me)
802 fn_arity = idArity fn
803 fn_unf = realIdUnfolding fn -- Ignore loop-breaker-ness here
804 (tyvars, theta, _) = tcSplitSigmaTy fn_type
805 n_tyvars = length tyvars
806 n_dicts = length theta
807 inline_act = idInlineActivation fn
809 -- Figure out whether the function has an INLINE pragma
810 -- See Note [Inline specialisations]
811 fn_has_inline_rule :: Maybe Bool -- Derive sat-flag from existing thing
812 fn_has_inline_rule = case isInlineRule_maybe fn_unf of
813 Just (_,sat) -> Just sat
816 spec_arity = unfoldingArity fn_unf - n_dicts -- Arity of the *specialised* inline rule
818 (rhs_tyvars, rhs_ids, rhs_body) = collectTyAndValBinders rhs
820 (body_uds_without_me, calls_for_me) = callsForMe fn body_uds
822 rhs_dict_ids = take n_dicts rhs_ids
823 body = mkLams (drop n_dicts rhs_ids) rhs_body
824 -- Glue back on the non-dict lambdas
826 already_covered :: [CoreExpr] -> Bool
827 already_covered args -- Note [Specialisations already covered]
828 = isJust (lookupRule (const True) realIdUnfolding
830 fn args (idCoreRules fn))
832 mk_ty_args :: [Maybe Type] -> [CoreExpr]
833 mk_ty_args call_ts = zipWithEqual "spec_call" mk_ty_arg rhs_tyvars call_ts
835 mk_ty_arg rhs_tyvar Nothing = Type (mkTyVarTy rhs_tyvar)
836 mk_ty_arg _ (Just ty) = Type ty
838 ----------------------------------------------------------
839 -- Specialise to one particular call pattern
840 spec_call :: CallInfo -- Call instance
841 -> SpecM (Maybe ((Id,CoreExpr), -- Specialised definition
842 UsageDetails, -- Usage details from specialised body
843 CoreRule)) -- Info for the Id's SpecEnv
844 spec_call (CallKey call_ts, (call_ds, _))
845 = ASSERT( call_ts `lengthIs` n_tyvars && call_ds `lengthIs` n_dicts )
847 -- Suppose f's defn is f = /\ a b c -> \ d1 d2 -> rhs
848 -- Supppose the call is for f [Just t1, Nothing, Just t3] [dx1, dx2]
850 -- Construct the new binding
851 -- f1 = SUBST[a->t1,c->t3, d1->d1', d2->d2'] (/\ b -> rhs)
852 -- PLUS the usage-details
853 -- { d1' = dx1; d2' = dx2 }
854 -- where d1', d2' are cloned versions of d1,d2, with the type substitution
855 -- applied. These auxiliary bindings just avoid duplication of dx1, dx2
857 -- Note that the substitution is applied to the whole thing.
858 -- This is convenient, but just slightly fragile. Notably:
859 -- * There had better be no name clashes in a/b/c
861 -- poly_tyvars = [b] in the example above
862 -- spec_tyvars = [a,c]
863 -- ty_args = [t1,b,t3]
864 poly_tyvars = [tv | (tv, Nothing) <- rhs_tyvars `zip` call_ts]
865 spec_tv_binds = [(tv,ty) | (tv, Just ty) <- rhs_tyvars `zip` call_ts]
866 spec_ty_args = map snd spec_tv_binds
867 ty_args = mk_ty_args call_ts
868 rhs_subst = CoreSubst.extendTvSubstList subst spec_tv_binds
870 ; (rhs_subst1, inst_dict_ids) <- cloneDictBndrs rhs_subst rhs_dict_ids
871 -- Clone rhs_dicts, including instantiating their types
873 ; let (rhs_subst2, dx_binds) = bindAuxiliaryDicts rhs_subst1 $
874 (my_zipEqual rhs_dict_ids inst_dict_ids call_ds)
875 inst_args = ty_args ++ map Var inst_dict_ids
877 ; if already_covered inst_args then
880 { -- Figure out the type of the specialised function
881 let body_ty = applyTypeToArgs rhs fn_type inst_args
882 (lam_args, app_args) -- Add a dummy argument if body_ty is unlifted
883 | isUnLiftedType body_ty -- C.f. WwLib.mkWorkerArgs
884 = (poly_tyvars ++ [voidArgId], poly_tyvars ++ [realWorldPrimId])
885 | otherwise = (poly_tyvars, poly_tyvars)
886 spec_id_ty = mkPiTypes lam_args body_ty
888 ; spec_f <- newSpecIdSM fn spec_id_ty
889 ; let spec_f_w_arity = setIdArity spec_f (max 0 (fn_arity - n_dicts))
890 -- Adding arity information just propagates it a bit faster
891 -- See Note [Arity decrease] in Simplify
893 ; (spec_rhs, rhs_uds) <- specExpr rhs_subst2 (mkLams lam_args body)
895 -- The rule to put in the function's specialisation is:
896 -- forall b, d1',d2'. f t1 b t3 d1' d2' = f1 b
897 rule_name = mkFastString ("SPEC " ++ showSDoc (ppr fn <+> ppr spec_ty_args))
898 spec_env_rule = mkLocalRule
900 inline_act -- Note [Auto-specialisation and RULES]
902 (poly_tyvars ++ inst_dict_ids)
904 (mkVarApps (Var spec_f_w_arity) app_args)
906 -- Add the { d1' = dx1; d2' = dx2 } usage stuff
907 final_uds = foldr consDictBind rhs_uds dx_binds
909 -- See Note [Inline specialisations]
910 final_spec_f | Just sat <- fn_has_inline_rule
911 = spec_f_w_arity `setInlineActivation` inline_act
912 `setIdUnfolding` mkInlineRule sat spec_rhs spec_arity
913 -- I'm not sure this should be unconditionally InlSat
916 ; return (Just ((final_spec_f, spec_rhs), final_uds, spec_env_rule)) } }
919 | debugIsOn && not (equalLength xs ys && equalLength ys zs)
920 = pprPanic "my_zipEqual" (vcat [ ppr xs, ppr ys
921 , ppr fn <+> ppr call_ts
922 , ppr (idType fn), ppr theta
923 , ppr n_dicts, ppr rhs_dict_ids
925 | otherwise = zip3 xs ys zs
929 -> [(DictId,DictId,CoreExpr)] -- (orig_dict, inst_dict, dx)
930 -> (Subst, -- Substitute for all orig_dicts
931 [CoreBind]) -- Auxiliary bindings
932 -- Bind any dictionary arguments to fresh names, to preserve sharing
933 -- Substitution already substitutes orig_dict -> inst_dict
934 bindAuxiliaryDicts subst triples = go subst [] triples
936 go subst binds [] = (subst, binds)
937 go subst binds ((d, dx_id, dx) : pairs)
938 | exprIsTrivial dx = go (extendIdSubst subst d dx) binds pairs
939 -- No auxiliary binding necessary
940 | otherwise = go subst_w_unf (NonRec dx_id dx : binds) pairs
942 dx_id1 = dx_id `setIdUnfolding` mkUnfolding False dx
943 subst_w_unf = extendIdSubst subst d (Var dx_id1)
944 -- Important! We're going to substitute dx_id1 for d
945 -- and we want it to look "interesting", else we won't gather *any*
946 -- consequential calls. E.g.
948 -- If we specialise f for a call (f (dfun dNumInt)), we'll get
949 -- a consequent call (g d') with an auxiliary definition
951 -- We want that consequent call to look interesting
954 Note [From non-recursive to recursive]
955 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
956 Even in the non-recursive case, if any dict-binds depend on 'fn' we might
957 have built a recursive knot
960 MkUD { ud_binds = d7 = MkD ..f..
961 , ud_calls = ...(f T d7)... }
965 Rec { fs x = <blah>[T/a, d7/d]
970 Here the recursion is only through the RULE.
973 Note [Specialisation of dictionary functions]
974 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
975 Here is a nasty example that bit us badly: see Trac #3591
977 dfun a d = MkD a d (meth d)
983 None of these definitions is recursive. What happened was that we
984 generated a specialisation:
986 RULE forall d. dfun T d = dT
987 dT = (MkD a d (meth d)) [T/a, d1/d]
990 But now we use the RULE on the RHS of d2, to get
992 d2 = dT = MkD d1 (meth d1)
995 and now d1 is bottom! The problem is that when specialising 'dfun' we
996 should first dump "below" the binding all floated dictionary bindings
997 that mention 'dfun' itself. So d2 and d3 (and hence d1) must be
998 placed below 'dfun', and thus unavailable to it when specialising
999 'dfun'. That in turn means that the call (dfun T d1) must be
1000 discarded. On the other hand, the call (dfun T d4) is fine, assuming
1001 d4 doesn't mention dfun.
1005 class C a where { foo,bar :: [a] -> [a] }
1007 instance C Int where
1011 r_bar :: C a => [a] -> [a]
1012 r_bar xs = bar (xs ++ xs)
1016 r_bar a (c::C a) (xs::[a]) = bar a d (xs ++ xs)
1018 Rec { $fCInt :: C Int = MkC foo_help reverse
1019 foo_help (xs::[Int]) = r_bar Int $fCInt xs }
1021 The call (r_bar $fCInt) mentions $fCInt,
1022 which mentions foo_help,
1023 which mentions r_bar
1024 But we DO want to specialise r_bar at Int:
1026 Rec { $fCInt :: C Int = MkC foo_help reverse
1027 foo_help (xs::[Int]) = r_bar Int $fCInt xs
1029 r_bar a (c::C a) (xs::[a]) = bar a d (xs ++ xs)
1030 RULE r_bar Int _ = r_bar_Int
1032 r_bar_Int xs = bar Int $fCInt (xs ++ xs)
1035 Note that, because of its RULE, r_bar joins the recursive
1036 group. (In this case it'll unravel a short moment later.)
1039 Conclusion: we catch the nasty case using filter_dfuns in
1040 callsForMe To be honest I'm not 100% certain that this is 100%
1041 right, but it works. Sigh.
1044 Note [Specialising a recursive group]
1045 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1047 let rec { f x = ...g x'...
1048 ; g y = ...f y'.... }
1050 Here we specialise 'f' at Char; but that is very likely to lead to
1051 a specialisation of 'g' at Char. We must do the latter, else the
1052 whole point of specialisation is lost.
1054 But we do not want to keep iterating to a fixpoint, because in the
1055 presence of polymorphic recursion we might generate an infinite number
1058 So we use the following heuristic:
1059 * Arrange the rec block in dependency order, so far as possible
1060 (the occurrence analyser already does this)
1062 * Specialise it much like a sequence of lets
1064 * Then go through the block a second time, feeding call-info from
1065 the RHSs back in the bottom, as it were
1067 In effect, the ordering maxmimises the effectiveness of each sweep,
1068 and we do just two sweeps. This should catch almost every case of
1069 monomorphic recursion -- the exception could be a very knotted-up
1070 recursion with multiple cycles tied up together.
1072 This plan is implemented in the Rec case of specBindItself.
1074 Note [Specialisations already covered]
1075 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1076 We obviously don't want to generate two specialisations for the same
1077 argument pattern. There are two wrinkles
1079 1. We do the already-covered test in specDefn, not when we generate
1080 the CallInfo in mkCallUDs. We used to test in the latter place, but
1081 we now iterate the specialiser somewhat, and the Id at the call site
1082 might therefore not have all the RULES that we can see in specDefn
1084 2. What about two specialisations where the second is an *instance*
1085 of the first? If the more specific one shows up first, we'll generate
1086 specialisations for both. If the *less* specific one shows up first,
1087 we *don't* currently generate a specialisation for the more specific
1088 one. (See the call to lookupRule in already_covered.) Reasons:
1089 (a) lookupRule doesn't say which matches are exact (bad reason)
1090 (b) if the earlier specialisation is user-provided, it's
1091 far from clear that we should auto-specialise further
1093 Note [Auto-specialisation and RULES]
1094 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1096 g :: Num a => a -> a
1099 f :: (Int -> Int) -> Int
1101 {-# RULE f g = 0 #-}
1103 Suppose that auto-specialisation makes a specialised version of
1104 g::Int->Int That version won't appear in the LHS of the RULE for f.
1105 So if the specialisation rule fires too early, the rule for f may
1108 It might be possible to add new rules, to "complete" the rewrite system.
1110 RULE forall d. g Int d = g_spec
1114 But that's a bit complicated. For now we ask the programmer's help,
1115 by *copying the INLINE activation pragma* to the auto-specialised rule.
1116 So if g says {-# NOINLINE[2] g #-}, then the auto-spec rule will also
1117 not be active until phase 2.
1120 Note [Specialisation shape]
1121 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1122 We only specialise a function if it has visible top-level lambdas
1123 corresponding to its overloading. E.g. if
1124 f :: forall a. Eq a => ....
1125 then its body must look like
1128 Reason: when specialising the body for a call (f ty dexp), we want to
1129 substitute dexp for d, and pick up specialised calls in the body of f.
1131 This doesn't always work. One example I came across was this:
1132 newtype Gen a = MkGen{ unGen :: Int -> a }
1134 choose :: Eq a => a -> Gen a
1135 choose n = MkGen (\r -> n)
1137 oneof = choose (1::Int)
1139 It's a silly exapmle, but we get
1140 choose = /\a. g `cast` co
1141 where choose doesn't have any dict arguments. Thus far I have not
1142 tried to fix this (wait till there's a real example).
1145 Note [Inline specialisations]
1146 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1147 We transfer to the specialised function any INLINE stuff from the
1148 original. This means (a) the Activation in the IdInfo, and (b) any
1149 InlineMe on the RHS. We do not, however, transfer the RuleMatchInfo
1150 since we do not expect the specialisation to occur in rewrite rules.
1152 This is a change (Jun06). Previously the idea is that the point of
1153 inlining was precisely to specialise the function at its call site,
1154 and that's not so important for the specialised copies. But
1155 *pragma-directed* specialisation now takes place in the
1156 typechecker/desugarer, with manually specified INLINEs. The
1157 specialiation here is automatic. It'd be very odd if a function
1158 marked INLINE was specialised (because of some local use), and then
1159 forever after (including importing modules) the specialised version
1160 wasn't INLINEd. After all, the programmer said INLINE!
1162 You might wonder why we don't just not specialise INLINE functions.
1163 It's because even INLINE functions are sometimes not inlined, when
1164 they aren't applied to interesting arguments. But perhaps the type
1165 arguments alone are enough to specialise (even though the args are too
1166 boring to trigger inlining), and it's certainly better to call the
1167 specialised version.
1169 A case in point is dictionary functions, which are current marked
1170 INLINE, but which are worth specialising.
1173 %************************************************************************
1175 \subsubsection{UsageDetails and suchlike}
1177 %************************************************************************
1182 ud_binds :: !(Bag DictBind),
1183 -- Floated dictionary bindings
1184 -- The order is important;
1185 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
1186 -- (Remember, Bags preserve order in GHC.)
1188 ud_calls :: !CallDetails
1190 -- INVARIANT: suppose bs = bindersOf ud_binds
1191 -- Then 'calls' may *mention* 'bs',
1192 -- but there should be no calls *for* bs
1195 instance Outputable UsageDetails where
1196 ppr (MkUD { ud_binds = dbs, ud_calls = calls })
1197 = ptext (sLit "MkUD") <+> braces (sep (punctuate comma
1198 [ptext (sLit "binds") <+> equals <+> ppr dbs,
1199 ptext (sLit "calls") <+> equals <+> ppr calls]))
1201 type DictBind = (CoreBind, VarSet)
1202 -- The set is the free vars of the binding
1203 -- both tyvars and dicts
1205 type DictExpr = CoreExpr
1207 emptyUDs :: UsageDetails
1208 emptyUDs = MkUD { ud_binds = emptyBag, ud_calls = emptyVarEnv }
1210 ------------------------------------------------------------
1211 type CallDetails = IdEnv CallInfoSet
1212 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
1214 -- CallInfo uses a FiniteMap, thereby ensuring that
1215 -- we record only one call instance for any key
1217 -- The list of types and dictionaries is guaranteed to
1218 -- match the type of f
1219 type CallInfoSet = FiniteMap CallKey ([DictExpr], VarSet)
1220 -- Range is dict args and the vars of the whole
1221 -- call (including tyvars)
1222 -- [*not* include the main id itself, of course]
1224 type CallInfo = (CallKey, ([DictExpr], VarSet))
1226 instance Outputable CallKey where
1227 ppr (CallKey ts) = ppr ts
1229 -- Type isn't an instance of Ord, so that we can control which
1230 -- instance we use. That's tiresome here. Oh well
1231 instance Eq CallKey where
1232 k1 == k2 = case k1 `compare` k2 of { EQ -> True; _ -> False }
1234 instance Ord CallKey where
1235 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
1237 cmp Nothing Nothing = EQ
1238 cmp Nothing (Just _) = LT
1239 cmp (Just _) Nothing = GT
1240 cmp (Just t1) (Just t2) = tcCmpType t1 t2
1242 unionCalls :: CallDetails -> CallDetails -> CallDetails
1243 unionCalls c1 c2 = plusVarEnv_C plusFM c1 c2
1245 -- plusCalls :: UsageDetails -> CallDetails -> UsageDetails
1246 -- plusCalls uds call_ds = uds { ud_calls = ud_calls uds `unionCalls` call_ds }
1248 callDetailsFVs :: CallDetails -> VarSet
1249 callDetailsFVs calls = foldVarEnv (unionVarSet . callInfoFVs) emptyVarSet calls
1251 callInfoFVs :: CallInfoSet -> VarSet
1252 callInfoFVs call_info = foldFM (\_ (_,fv) vs -> unionVarSet fv vs) emptyVarSet call_info
1254 ------------------------------------------------------------
1255 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> UsageDetails
1256 singleCall id tys dicts
1257 = MkUD {ud_binds = emptyBag,
1258 ud_calls = unitVarEnv id (unitFM (CallKey tys) (dicts, call_fvs)) }
1260 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
1261 tys_fvs = tyVarsOfTypes (catMaybes tys)
1262 -- The type args (tys) are guaranteed to be part of the dictionary
1263 -- types, because they are just the constrained types,
1264 -- and the dictionary is therefore sure to be bound
1265 -- inside the binding for any type variables free in the type;
1266 -- hence it's safe to neglect tyvars free in tys when making
1267 -- the free-var set for this call
1268 -- BUT I don't trust this reasoning; play safe and include tys_fvs
1270 -- We don't include the 'id' itself.
1272 mkCallUDs :: Id -> [CoreExpr] -> UsageDetails
1274 | not (isLocalId f) -- Imported from elsewhere
1275 || null theta -- Not overloaded
1276 || not (all isClassPred theta)
1277 -- Only specialise if all overloading is on class params.
1278 -- In ptic, with implicit params, the type args
1279 -- *don't* say what the value of the implicit param is!
1280 || not (spec_tys `lengthIs` n_tyvars)
1281 || not ( dicts `lengthIs` n_dicts)
1282 || not (any interestingDict dicts) -- Note [Interesting dictionary arguments]
1283 -- See also Note [Specialisations already covered]
1284 = -- pprTrace "mkCallUDs: discarding" (vcat [ppr f, ppr args, ppr n_tyvars, ppr n_dicts, ppr (map interestingDict dicts)])
1285 emptyUDs -- Not overloaded, or no specialisation wanted
1288 = -- pprTrace "mkCallUDs: keeping" (vcat [ppr f, ppr args, ppr n_tyvars, ppr n_dicts, ppr (map interestingDict dicts)])
1289 singleCall f spec_tys dicts
1291 (tyvars, theta, _) = tcSplitSigmaTy (idType f)
1292 constrained_tyvars = tyVarsOfTheta theta
1293 n_tyvars = length tyvars
1294 n_dicts = length theta
1296 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
1297 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
1300 | tyvar `elemVarSet` constrained_tyvars = Just ty
1301 | otherwise = Nothing
1304 Note [Interesting dictionary arguments]
1305 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1307 \a.\d:Eq a. let f = ... in ...(f d)...
1308 There really is not much point in specialising f wrt the dictionary d,
1309 because the code for the specialised f is not improved at all, because
1310 d is lambda-bound. We simply get junk specialisations.
1312 What is "interesting"? Just that it has *some* structure.
1315 interestingDict :: CoreExpr -> Bool
1316 -- A dictionary argument is interesting if it has *some* structure
1317 interestingDict (Var v) = hasSomeUnfolding (idUnfolding v)
1318 || isDataConWorkId v
1319 interestingDict (Type _) = False
1320 interestingDict (App fn (Type _)) = interestingDict fn
1321 interestingDict (Note _ a) = interestingDict a
1322 interestingDict (Cast e _) = interestingDict e
1323 interestingDict _ = True
1327 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
1328 plusUDs (MkUD {ud_binds = db1, ud_calls = calls1})
1329 (MkUD {ud_binds = db2, ud_calls = calls2})
1330 = MkUD { ud_binds = db1 `unionBags` db2
1331 , ud_calls = calls1 `unionCalls` calls2 }
1333 plusUDList :: [UsageDetails] -> UsageDetails
1334 plusUDList = foldr plusUDs emptyUDs
1336 -----------------------------
1337 _dictBindBndrs :: Bag DictBind -> [Id]
1338 _dictBindBndrs dbs = foldrBag ((++) . bindersOf . fst) [] dbs
1340 mkDB :: CoreBind -> DictBind
1341 mkDB bind = (bind, bind_fvs bind)
1343 bind_fvs :: CoreBind -> VarSet
1344 bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
1345 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1348 rhs_fvs = unionVarSets (map pair_fvs prs)
1350 pair_fvs :: (Id, CoreExpr) -> VarSet
1351 pair_fvs (bndr, rhs) = exprFreeVars rhs `unionVarSet` idFreeVars bndr
1352 -- Don't forget variables mentioned in the
1353 -- rules of the bndr. C.f. OccAnal.addRuleUsage
1354 -- Also tyvars mentioned in its type; they may not appear in the RHS
1358 flattenDictBinds :: Bag DictBind -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
1359 flattenDictBinds dbs pairs
1360 = foldrBag add pairs dbs
1362 add (NonRec b r,_) pairs = (b,r) : pairs
1363 add (Rec prs1, _) pairs = prs1 ++ pairs
1365 snocDictBinds :: UsageDetails -> [CoreBind] -> UsageDetails
1366 -- Add ud_binds to the tail end of the bindings in uds
1367 snocDictBinds uds dbs
1368 = uds { ud_binds = ud_binds uds `unionBags`
1369 foldr (consBag . mkDB) emptyBag dbs }
1371 consDictBind :: CoreBind -> UsageDetails -> UsageDetails
1372 consDictBind bind uds = uds { ud_binds = mkDB bind `consBag` ud_binds uds }
1374 snocDictBind :: UsageDetails -> CoreBind -> UsageDetails
1375 snocDictBind uds bind = uds { ud_binds = ud_binds uds `snocBag` mkDB bind }
1377 wrapDictBinds :: Bag DictBind -> [CoreBind] -> [CoreBind]
1378 wrapDictBinds dbs binds
1379 = foldrBag add binds dbs
1381 add (bind,_) binds = bind : binds
1383 wrapDictBindsE :: Bag DictBind -> CoreExpr -> CoreExpr
1384 wrapDictBindsE dbs expr
1385 = foldrBag add expr dbs
1387 add (bind,_) expr = Let bind expr
1389 ----------------------
1390 dumpUDs :: [CoreBndr] -> UsageDetails -> (UsageDetails, Bag DictBind)
1391 -- Used at a lambda or case binder; just dump anything mentioning the binder
1392 dumpUDs bndrs uds@(MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
1393 | null bndrs = (uds, emptyBag) -- Common in case alternatives
1394 | otherwise = (free_uds, dump_dbs)
1396 free_uds = MkUD { ud_binds = free_dbs, ud_calls = free_calls }
1397 bndr_set = mkVarSet bndrs
1398 (free_dbs, dump_dbs, dump_set) = splitDictBinds orig_dbs bndr_set
1399 free_calls = deleteCallsMentioning dump_set $ -- Drop calls mentioning bndr_set on the floor
1400 deleteCallsFor bndrs orig_calls -- Discard calls for bndr_set; there should be
1401 -- no calls for any of the dicts in dump_dbs
1403 dumpBindUDs :: [CoreBndr] -> UsageDetails -> (UsageDetails, Bag DictBind, Bool)
1404 -- Used at a lambda or case binder; just dump anything mentioning the binder
1405 dumpBindUDs bndrs (MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
1406 = (free_uds, dump_dbs, float_all)
1408 free_uds = MkUD { ud_binds = free_dbs, ud_calls = free_calls }
1409 bndr_set = mkVarSet bndrs
1410 (free_dbs, dump_dbs, dump_set) = splitDictBinds orig_dbs bndr_set
1411 free_calls = deleteCallsFor bndrs orig_calls
1412 float_all = dump_set `intersectsVarSet` callDetailsFVs free_calls
1414 callsForMe :: Id -> UsageDetails -> (UsageDetails, [CallInfo])
1415 callsForMe fn (MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
1416 = -- pprTrace ("callsForMe")
1418 -- text "Orig dbs =" <+> ppr (_dictBindBndrs orig_dbs),
1419 -- text "Orig calls =" <+> ppr orig_calls,
1420 -- text "Dep set =" <+> ppr dep_set,
1421 -- text "Calls for me =" <+> ppr calls_for_me]) $
1422 (uds_without_me, calls_for_me)
1424 uds_without_me = MkUD { ud_binds = orig_dbs, ud_calls = delVarEnv orig_calls fn }
1425 calls_for_me = case lookupVarEnv orig_calls fn of
1427 Just cs -> filter_dfuns (fmToList cs)
1429 dep_set = foldlBag go (unitVarSet fn) orig_dbs
1430 go dep_set (db,fvs) | fvs `intersectsVarSet` dep_set
1431 = extendVarSetList dep_set (bindersOf db)
1434 -- Note [Specialisation of dictionary functions]
1435 filter_dfuns | isDFunId fn = filter ok_call
1436 | otherwise = \cs -> cs
1438 ok_call (_, (_,fvs)) = not (fvs `intersectsVarSet` dep_set)
1440 ----------------------
1441 splitDictBinds :: Bag DictBind -> IdSet -> (Bag DictBind, Bag DictBind, IdSet)
1442 -- Returns (free_dbs, dump_dbs, dump_set)
1443 splitDictBinds dbs bndr_set
1444 = foldlBag split_db (emptyBag, emptyBag, bndr_set) dbs
1445 -- Important that it's foldl not foldr;
1446 -- we're accumulating the set of dumped ids in dump_set
1448 split_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1449 | dump_idset `intersectsVarSet` fvs -- Dump it
1450 = (free_dbs, dump_dbs `snocBag` db,
1451 extendVarSetList dump_idset (bindersOf bind))
1453 | otherwise -- Don't dump it
1454 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1457 ----------------------
1458 deleteCallsMentioning :: VarSet -> CallDetails -> CallDetails
1459 -- Remove calls *mentioning* bs
1460 deleteCallsMentioning bs calls
1461 = mapVarEnv filter_calls calls
1463 filter_calls :: CallInfoSet -> CallInfoSet
1464 filter_calls = filterFM (\_ (_, fvs) -> not (fvs `intersectsVarSet` bs))
1466 deleteCallsFor :: [Id] -> CallDetails -> CallDetails
1467 -- Remove calls *for* bs
1468 deleteCallsFor bs calls = delVarEnvList calls bs
1472 %************************************************************************
1474 \subsubsection{Boring helper functions}
1476 %************************************************************************
1479 type SpecM a = UniqSM a
1481 initSM :: UniqSupply -> SpecM a -> a
1484 mapAndCombineSM :: (a -> SpecM (b, UsageDetails)) -> [a] -> SpecM ([b], UsageDetails)
1485 mapAndCombineSM _ [] = return ([], emptyUDs)
1486 mapAndCombineSM f (x:xs) = do (y, uds1) <- f x
1487 (ys, uds2) <- mapAndCombineSM f xs
1488 return (y:ys, uds1 `plusUDs` uds2)
1490 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1491 -- Clone the binders of the bind; return new bind with the cloned binders
1492 -- Return the substitution to use for RHSs, and the one to use for the body
1493 cloneBindSM subst (NonRec bndr rhs) = do
1494 us <- getUniqueSupplyM
1495 let (subst', bndr') = cloneIdBndr subst us bndr
1496 return (subst, subst', NonRec bndr' rhs)
1498 cloneBindSM subst (Rec pairs) = do
1499 us <- getUniqueSupplyM
1500 let (subst', bndrs') = cloneRecIdBndrs subst us (map fst pairs)
1501 return (subst', subst', Rec (bndrs' `zip` map snd pairs))
1503 cloneDictBndrs :: Subst -> [CoreBndr] -> SpecM (Subst, [CoreBndr])
1504 cloneDictBndrs subst bndrs
1505 = do { us <- getUniqueSupplyM
1506 ; return (cloneIdBndrs subst us bndrs) }
1508 newSpecIdSM :: Id -> Type -> SpecM Id
1509 -- Give the new Id a similar occurrence name to the old one
1510 newSpecIdSM old_id new_ty
1511 = do { uniq <- getUniqueM
1513 name = idName old_id
1514 new_occ = mkSpecOcc (nameOccName name)
1515 new_id = mkUserLocal new_occ uniq new_ty (getSrcSpan name)
1520 Old (but interesting) stuff about unboxed bindings
1521 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1523 What should we do when a value is specialised to a *strict* unboxed value?
1525 map_*_* f (x:xs) = let h = f x
1529 Could convert let to case:
1531 map_*_Int# f (x:xs) = case f x of h# ->
1535 This may be undesirable since it forces evaluation here, but the value
1536 may not be used in all branches of the body. In the general case this
1537 transformation is impossible since the mutual recursion in a letrec
1538 cannot be expressed as a case.
1540 There is also a problem with top-level unboxed values, since our
1541 implementation cannot handle unboxed values at the top level.
1543 Solution: Lift the binding of the unboxed value and extract it when it
1546 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1551 Now give it to the simplifier and the _Lifting will be optimised away.
1553 The benfit is that we have given the specialised "unboxed" values a
1554 very simplep lifted semantics and then leave it up to the simplifier to
1555 optimise it --- knowing that the overheads will be removed in nearly
1558 In particular, the value will only be evaluted in the branches of the
1559 program which use it, rather than being forced at the point where the
1560 value is bound. For example:
1562 filtermap_*_* p f (x:xs)
1569 filtermap_*_Int# p f (x:xs)
1570 = let h = case (f x) of h# -> _Lift h#
1573 True -> case h of _Lift h#
1577 The binding for h can still be inlined in the one branch and the
1578 _Lifting eliminated.
1581 Question: When won't the _Lifting be eliminated?
1583 Answer: When they at the top-level (where it is necessary) or when
1584 inlining would duplicate work (or possibly code depending on
1585 options). However, the _Lifting will still be eliminated if the
1586 strictness analyser deems the lifted binding strict.