2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
4 \section[Specialise]{Stamping out overloading, and (optionally) polymorphism}
7 -- The above warning supression flag is a temporary kludge.
8 -- While working on this module you are encouraged to remove it and fix
9 -- any warnings in the module. See
10 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
13 module Specialise ( specProgram ) where
15 #include "HsVersions.h"
17 import DynFlags ( DynFlags, DynFlag(..) )
18 import Id ( Id, idName, idType, mkUserLocal,
19 idInlinePragma, setInlinePragma )
20 import TcType ( Type, mkTyVarTy, tcSplitSigmaTy,
21 tyVarsOfTypes, tyVarsOfTheta, isClassPred,
22 tcCmpType, isUnLiftedType
24 import CoreSubst ( Subst, mkEmptySubst, extendTvSubstList, lookupIdSubst,
25 substBndr, substBndrs, substTy, substInScope,
26 cloneIdBndr, cloneIdBndrs, cloneRecIdBndrs
31 import CoreUtils ( applyTypeToArgs, mkPiTypes )
32 import CoreFVs ( exprFreeVars, exprsFreeVars, idFreeVars )
33 import CoreTidy ( tidyRules )
34 import CoreLint ( showPass, endPass )
35 import Rules ( addIdSpecialisations, mkLocalRule, lookupRule, emptyRuleBase, rulesOfBinds )
36 import PprCore ( pprRules )
37 import UniqSupply ( UniqSupply,
42 import MkId ( voidArgId, realWorldPrimId )
44 import Maybes ( catMaybes, maybeToBool )
45 import ErrUtils ( dumpIfSet_dyn )
53 %************************************************************************
55 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
57 %************************************************************************
59 These notes describe how we implement specialisation to eliminate
62 The specialisation pass works on Core
63 syntax, complete with all the explicit dictionary application,
64 abstraction and construction as added by the type checker. The
65 existing type checker remains largely as it is.
67 One important thought: the {\em types} passed to an overloaded
68 function, and the {\em dictionaries} passed are mutually redundant.
69 If the same function is applied to the same type(s) then it is sure to
70 be applied to the same dictionary(s)---or rather to the same {\em
71 values}. (The arguments might look different but they will evaluate
74 Second important thought: we know that we can make progress by
75 treating dictionary arguments as static and worth specialising on. So
76 we can do without binding-time analysis, and instead specialise on
77 dictionary arguments and no others.
86 and suppose f is overloaded.
88 STEP 1: CALL-INSTANCE COLLECTION
90 We traverse <body>, accumulating all applications of f to types and
93 (Might there be partial applications, to just some of its types and
94 dictionaries? In principle yes, but in practice the type checker only
95 builds applications of f to all its types and dictionaries, so partial
96 applications could only arise as a result of transformation, and even
97 then I think it's unlikely. In any case, we simply don't accumulate such
98 partial applications.)
103 So now we have a collection of calls to f:
107 Notice that f may take several type arguments. To avoid ambiguity, we
108 say that f is called at type t1/t2 and t3/t4.
110 We take equivalence classes using equality of the *types* (ignoring
111 the dictionary args, which as mentioned previously are redundant).
113 STEP 3: SPECIALISATION
115 For each equivalence class, choose a representative (f t1 t2 d1 d2),
116 and create a local instance of f, defined thus:
118 f@t1/t2 = <f_rhs> t1 t2 d1 d2
120 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
121 of simplification will now result. However we don't actually *do* that
122 simplification. Rather, we leave it for the simplifier to do. If we
123 *did* do it, though, we'd get more call instances from the specialised
124 RHS. We can work out what they are by instantiating the call-instance
125 set from f's RHS with the types t1, t2.
127 Add this new id to f's IdInfo, to record that f has a specialised version.
129 Before doing any of this, check that f's IdInfo doesn't already
130 tell us about an existing instance of f at the required type/s.
131 (This might happen if specialisation was applied more than once, or
132 it might arise from user SPECIALIZE pragmas.)
136 Wait a minute! What if f is recursive? Then we can't just plug in
137 its right-hand side, can we?
139 But it's ok. The type checker *always* creates non-recursive definitions
140 for overloaded recursive functions. For example:
142 f x = f (x+x) -- Yes I know its silly
146 f a (d::Num a) = let p = +.sel a d
148 letrec fl (y::a) = fl (p y y)
152 We still have recusion for non-overloaded functions which we
153 speciailise, but the recursive call should get specialised to the
154 same recursive version.
160 All this is crystal clear when the function is applied to *constant
161 types*; that is, types which have no type variables inside. But what if
162 it is applied to non-constant types? Suppose we find a call of f at type
163 t1/t2. There are two possibilities:
165 (a) The free type variables of t1, t2 are in scope at the definition point
166 of f. In this case there's no problem, we proceed just as before. A common
167 example is as follows. Here's the Haskell:
172 After typechecking we have
174 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
175 in +.sel a d (f a d y) (f a d y)
177 Notice that the call to f is at type type "a"; a non-constant type.
178 Both calls to f are at the same type, so we can specialise to give:
180 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
181 in +.sel a d (f@a y) (f@a y)
184 (b) The other case is when the type variables in the instance types
185 are *not* in scope at the definition point of f. The example we are
186 working with above is a good case. There are two instances of (+.sel a d),
187 but "a" is not in scope at the definition of +.sel. Can we do anything?
188 Yes, we can "common them up", a sort of limited common sub-expression deal.
191 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
192 f@a (x::a) = +.sel@a x x
193 in +.sel@a (f@a y) (f@a y)
195 This can save work, and can't be spotted by the type checker, because
196 the two instances of +.sel weren't originally at the same type.
200 * There are quite a few variations here. For example, the defn of
201 +.sel could be floated ouside the \y, to attempt to gain laziness.
202 It certainly mustn't be floated outside the \d because the d has to
205 * We don't want to inline f_rhs in this case, because
206 that will duplicate code. Just commoning up the call is the point.
208 * Nothing gets added to +.sel's IdInfo.
210 * Don't bother unless the equivalence class has more than one item!
212 Not clear whether this is all worth it. It is of course OK to
213 simply discard call-instances when passing a big lambda.
215 Polymorphism 2 -- Overloading
217 Consider a function whose most general type is
219 f :: forall a b. Ord a => [a] -> b -> b
221 There is really no point in making a version of g at Int/Int and another
222 at Int/Bool, because it's only instancing the type variable "a" which
223 buys us any efficiency. Since g is completely polymorphic in b there
224 ain't much point in making separate versions of g for the different
227 That suggests that we should identify which of g's type variables
228 are constrained (like "a") and which are unconstrained (like "b").
229 Then when taking equivalence classes in STEP 2, we ignore the type args
230 corresponding to unconstrained type variable. In STEP 3 we make
231 polymorphic versions. Thus:
233 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
242 f a (d::Num a) = let g = ...
244 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
246 Here, g is only called at one type, but the dictionary isn't in scope at the
247 definition point for g. Usually the type checker would build a
248 definition for d1 which enclosed g, but the transformation system
249 might have moved d1's defn inward. Solution: float dictionary bindings
250 outwards along with call instances.
254 f x = let g p q = p==q
260 Before specialisation, leaving out type abstractions we have
262 f df x = let g :: Eq a => a -> a -> Bool
264 h :: Num a => a -> a -> (a, Bool)
265 h dh r s = let deq = eqFromNum dh
266 in (+ dh r s, g deq r s)
270 After specialising h we get a specialised version of h, like this:
272 h' r s = let deq = eqFromNum df
273 in (+ df r s, g deq r s)
275 But we can't naively make an instance for g from this, because deq is not in scope
276 at the defn of g. Instead, we have to float out the (new) defn of deq
277 to widen its scope. Notice that this floating can't be done in advance -- it only
278 shows up when specialisation is done.
280 User SPECIALIZE pragmas
281 ~~~~~~~~~~~~~~~~~~~~~~~
282 Specialisation pragmas can be digested by the type checker, and implemented
283 by adding extra definitions along with that of f, in the same way as before
285 f@t1/t2 = <f_rhs> t1 t2 d1 d2
287 Indeed the pragmas *have* to be dealt with by the type checker, because
288 only it knows how to build the dictionaries d1 and d2! For example
290 g :: Ord a => [a] -> [a]
291 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
293 Here, the specialised version of g is an application of g's rhs to the
294 Ord dictionary for (Tree Int), which only the type checker can conjure
295 up. There might not even *be* one, if (Tree Int) is not an instance of
296 Ord! (All the other specialision has suitable dictionaries to hand
299 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
300 it is buried in a complex (as-yet-un-desugared) binding group.
303 f@t1/t2 = f* t1 t2 d1 d2
305 where f* is the Id f with an IdInfo which says "inline me regardless!".
306 Indeed all the specialisation could be done in this way.
307 That in turn means that the simplifier has to be prepared to inline absolutely
308 any in-scope let-bound thing.
311 Again, the pragma should permit polymorphism in unconstrained variables:
313 h :: Ord a => [a] -> b -> b
314 {-# SPECIALIZE h :: [Int] -> b -> b #-}
316 We *insist* that all overloaded type variables are specialised to ground types,
317 (and hence there can be no context inside a SPECIALIZE pragma).
318 We *permit* unconstrained type variables to be specialised to
320 - or left as a polymorphic type variable
321 but nothing in between. So
323 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
325 is *illegal*. (It can be handled, but it adds complication, and gains the
329 SPECIALISING INSTANCE DECLARATIONS
330 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
333 instance Foo a => Foo [a] where
335 {-# SPECIALIZE instance Foo [Int] #-}
337 The original instance decl creates a dictionary-function
340 dfun.Foo.List :: forall a. Foo a -> Foo [a]
342 The SPECIALIZE pragma just makes a specialised copy, just as for
343 ordinary function definitions:
345 dfun.Foo.List@Int :: Foo [Int]
346 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
348 The information about what instance of the dfun exist gets added to
349 the dfun's IdInfo in the same way as a user-defined function too.
352 Automatic instance decl specialisation?
353 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
354 Can instance decls be specialised automatically? It's tricky.
355 We could collect call-instance information for each dfun, but
356 then when we specialised their bodies we'd get new call-instances
357 for ordinary functions; and when we specialised their bodies, we might get
358 new call-instances of the dfuns, and so on. This all arises because of
359 the unrestricted mutual recursion between instance decls and value decls.
361 Still, there's no actual problem; it just means that we may not do all
362 the specialisation we could theoretically do.
364 Furthermore, instance decls are usually exported and used non-locally,
365 so we'll want to compile enough to get those specialisations done.
367 Lastly, there's no such thing as a local instance decl, so we can
368 survive solely by spitting out *usage* information, and then reading that
369 back in as a pragma when next compiling the file. So for now,
370 we only specialise instance decls in response to pragmas.
373 SPITTING OUT USAGE INFORMATION
374 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
376 To spit out usage information we need to traverse the code collecting
377 call-instance information for all imported (non-prelude?) functions
378 and data types. Then we equivalence-class it and spit it out.
380 This is done at the top-level when all the call instances which escape
381 must be for imported functions and data types.
383 *** Not currently done ***
386 Partial specialisation by pragmas
387 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
388 What about partial specialisation:
390 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
391 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
395 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
397 Seems quite reasonable. Similar things could be done with instance decls:
399 instance (Foo a, Foo b) => Foo (a,b) where
401 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
402 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
404 Ho hum. Things are complex enough without this. I pass.
407 Requirements for the simplifer
408 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
409 The simplifier has to be able to take advantage of the specialisation.
411 * When the simplifier finds an application of a polymorphic f, it looks in
412 f's IdInfo in case there is a suitable instance to call instead. This converts
414 f t1 t2 d1 d2 ===> f_t1_t2
416 Note that the dictionaries get eaten up too!
418 * Dictionary selection operations on constant dictionaries must be
421 +.sel Int d ===> +Int
423 The obvious way to do this is in the same way as other specialised
424 calls: +.sel has inside it some IdInfo which tells that if it's applied
425 to the type Int then it should eat a dictionary and transform to +Int.
427 In short, dictionary selectors need IdInfo inside them for constant
430 * Exactly the same applies if a superclass dictionary is being
433 Eq.sel Int d ===> dEqInt
435 * Something similar applies to dictionary construction too. Suppose
436 dfun.Eq.List is the function taking a dictionary for (Eq a) to
437 one for (Eq [a]). Then we want
439 dfun.Eq.List Int d ===> dEq.List_Int
441 Where does the Eq [Int] dictionary come from? It is built in
442 response to a SPECIALIZE pragma on the Eq [a] instance decl.
444 In short, dfun Ids need IdInfo with a specialisation for each
445 constant instance of their instance declaration.
447 All this uses a single mechanism: the SpecEnv inside an Id
450 What does the specialisation IdInfo look like?
451 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
453 The SpecEnv of an Id maps a list of types (the template) to an expression
457 For example, if f has this SpecInfo:
459 [Int, a] -> \d:Ord Int. f' a
461 it means that we can replace the call
463 f Int t ===> (\d. f' t)
465 This chucks one dictionary away and proceeds with the
466 specialised version of f, namely f'.
469 What can't be done this way?
470 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
471 There is no way, post-typechecker, to get a dictionary for (say)
472 Eq a from a dictionary for Eq [a]. So if we find
476 we can't transform to
481 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
483 Of course, we currently have no way to automatically derive
484 eqList, nor to connect it to the Eq [a] instance decl, but you
485 can imagine that it might somehow be possible. Taking advantage
486 of this is permanently ruled out.
488 Still, this is no great hardship, because we intend to eliminate
489 overloading altogether anyway!
493 A note about non-tyvar dictionaries
494 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
495 Some Ids have types like
497 forall a,b,c. Eq a -> Ord [a] -> tau
499 This seems curious at first, because we usually only have dictionary
500 args whose types are of the form (C a) where a is a type variable.
501 But this doesn't hold for the functions arising from instance decls,
502 which sometimes get arguements with types of form (C (T a)) for some
505 Should we specialise wrt this compound-type dictionary? We used to say
507 "This is a heuristic judgement, as indeed is the fact that we
508 specialise wrt only dictionaries. We choose *not* to specialise
509 wrt compound dictionaries because at the moment the only place
510 they show up is in instance decls, where they are simply plugged
511 into a returned dictionary. So nothing is gained by specialising
514 But it is simpler and more uniform to specialise wrt these dicts too;
515 and in future GHC is likely to support full fledged type signatures
517 f ;: Eq [(a,b)] => ...
520 %************************************************************************
522 \subsubsection{The new specialiser}
524 %************************************************************************
526 Our basic game plan is this. For let(rec) bound function
527 f :: (C a, D c) => (a,b,c,d) -> Bool
529 * Find any specialised calls of f, (f ts ds), where
530 ts are the type arguments t1 .. t4, and
531 ds are the dictionary arguments d1 .. d2.
533 * Add a new definition for f1 (say):
535 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
537 Note that we abstract over the unconstrained type arguments.
541 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
543 to the specialisations of f. This will be used by the
544 simplifier to replace calls
545 (f t1 t2 t3 t4) da db
547 (\d1 d1 -> f1 t2 t4) da db
549 All the stuff about how many dictionaries to discard, and what types
550 to apply the specialised function to, are handled by the fact that the
551 SpecEnv contains a template for the result of the specialisation.
553 We don't build *partial* specialisations for f. For example:
555 f :: Eq a => a -> a -> Bool
556 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
558 Here, little is gained by making a specialised copy of f.
559 There's a distinct danger that the specialised version would
560 first build a dictionary for (Eq b, Eq c), and then select the (==)
561 method from it! Even if it didn't, not a great deal is saved.
563 We do, however, generate polymorphic, but not overloaded, specialisations:
565 f :: Eq a => [a] -> b -> b -> b
566 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
568 Hence, the invariant is this:
570 *** no specialised version is overloaded ***
573 %************************************************************************
575 \subsubsection{The exported function}
577 %************************************************************************
580 specProgram :: DynFlags -> UniqSupply -> [CoreBind] -> IO [CoreBind]
581 specProgram dflags us binds = do
583 showPass dflags "Specialise"
585 let binds' = initSM us (do (binds', uds') <- go binds
586 return (dumpAllDictBinds uds' binds'))
588 endPass dflags "Specialise" Opt_D_dump_spec binds'
590 dumpIfSet_dyn dflags Opt_D_dump_rules "Top-level specialisations"
591 (pprRules (tidyRules emptyTidyEnv (rulesOfBinds binds')))
595 -- We need to start with a Subst that knows all the things
596 -- that are in scope, so that the substitution engine doesn't
597 -- accidentally re-use a unique that's already in use
598 -- Easiest thing is to do it all at once, as if all the top-level
599 -- decls were mutually recursive
600 top_subst = mkEmptySubst (mkInScopeSet (mkVarSet (bindersOfBinds binds)))
602 go [] = return ([], emptyUDs)
603 go (bind:binds) = do (binds', uds) <- go binds
604 (bind', uds') <- specBind top_subst bind uds
605 return (bind' ++ binds', uds')
608 %************************************************************************
610 \subsubsection{@specExpr@: the main function}
612 %************************************************************************
615 specVar :: Subst -> Id -> CoreExpr
616 specVar subst v = lookupIdSubst subst v
618 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
619 -- We carry a substitution down:
620 -- a) we must clone any binding that might flaot outwards,
621 -- to avoid name clashes
622 -- b) we carry a type substitution to use when analysing
623 -- the RHS of specialised bindings (no type-let!)
625 ---------------- First the easy cases --------------------
626 specExpr subst (Type ty) = return (Type (substTy subst ty), emptyUDs)
627 specExpr subst (Var v) = return (specVar subst v, emptyUDs)
628 specExpr _ (Lit lit) = return (Lit lit, emptyUDs)
629 specExpr subst (Cast e co) = do
630 (e', uds) <- specExpr subst e
631 return ((Cast e' (substTy subst co)), uds)
632 specExpr subst (Note note body) = do
633 (body', uds) <- specExpr subst body
634 return (Note (specNote subst note) body', uds)
637 ---------------- Applications might generate a call instance --------------------
638 specExpr subst expr@(App {})
641 go (App fun arg) args = do (arg', uds_arg) <- specExpr subst arg
642 (fun', uds_app) <- go fun (arg':args)
643 return (App fun' arg', uds_arg `plusUDs` uds_app)
645 go (Var f) args = case specVar subst f of
646 Var f' -> return (Var f', mkCallUDs subst f' args)
647 e' -> return (e', emptyUDs) -- I don't expect this!
648 go other _ = specExpr subst other
650 ---------------- Lambda/case require dumping of usage details --------------------
651 specExpr subst e@(Lam _ _) = do
652 (body', uds) <- specExpr subst' body
653 let (filtered_uds, body'') = dumpUDs bndrs' uds body'
654 return (mkLams bndrs' body'', filtered_uds)
656 (bndrs, body) = collectBinders e
657 (subst', bndrs') = substBndrs subst bndrs
658 -- More efficient to collect a group of binders together all at once
659 -- and we don't want to split a lambda group with dumped bindings
661 specExpr subst (Case scrut case_bndr ty alts) = do
662 (scrut', uds_scrut) <- specExpr subst scrut
663 (alts', uds_alts) <- mapAndCombineSM spec_alt alts
664 return (Case scrut' case_bndr' (substTy subst ty) alts', uds_scrut `plusUDs` uds_alts)
666 (subst_alt, case_bndr') = substBndr subst case_bndr
667 -- No need to clone case binder; it can't float like a let(rec)
669 spec_alt (con, args, rhs) = do
670 (rhs', uds) <- specExpr subst_rhs rhs
671 let (uds', rhs'') = dumpUDs args uds rhs'
672 return ((con, args', rhs''), uds')
674 (subst_rhs, args') = substBndrs subst_alt args
676 ---------------- Finally, let is the interesting case --------------------
677 specExpr subst (Let bind body) = do
679 (rhs_subst, body_subst, bind') <- cloneBindSM subst bind
681 -- Deal with the body
682 (body', body_uds) <- specExpr body_subst body
684 -- Deal with the bindings
685 (binds', uds) <- specBind rhs_subst bind' body_uds
688 return (foldr Let body' binds', uds)
690 -- Must apply the type substitution to coerceions
691 specNote :: Subst -> Note -> Note
692 specNote _ note = note
695 %************************************************************************
697 \subsubsection{Dealing with a binding}
699 %************************************************************************
702 specBind :: Subst -- Use this for RHSs
704 -> UsageDetails -- Info on how the scope of the binding
705 -> SpecM ([CoreBind], -- New bindings
706 UsageDetails) -- And info to pass upstream
708 specBind rhs_subst bind body_uds
709 = do { (bind', bind_uds) <- specBindItself rhs_subst bind (calls body_uds)
710 ; return (finishSpecBind bind' bind_uds body_uds) }
712 finishSpecBind :: CoreBind -> UsageDetails -> UsageDetails -> ([CoreBind], UsageDetails)
714 (MkUD { dict_binds = rhs_dbs, calls = rhs_calls, ud_fvs = rhs_fvs })
715 (MkUD { dict_binds = body_dbs, calls = body_calls, ud_fvs = body_fvs })
716 | not (mkVarSet bndrs `intersectsVarSet` all_fvs)
717 -- Common case 1: the bound variables are not
718 -- mentioned in the dictionary bindings
719 = ([bind], MkUD { dict_binds = body_dbs `unionBags` rhs_dbs
720 -- It's important that the `unionBags` is this way round,
721 -- because body_uds may bind dictionaries that are
722 -- used in the calls passed to specDefn. So the
723 -- dictionary bindings in rhs_uds may mention
724 -- dictionaries bound in body_uds.
726 , ud_fvs = all_fvs })
728 | case bind of { NonRec {} -> True; Rec {} -> False }
729 -- Common case 2: no specialisation happened, and binding
730 -- is non-recursive. But the binding may be
731 -- mentioned in body_dbs, so we should put it first
732 = ([], MkUD { dict_binds = rhs_dbs `unionBags` ((bind, b_fvs) `consBag` body_dbs)
734 , ud_fvs = all_fvs `unionVarSet` b_fvs })
736 | otherwise -- General case: make a huge Rec (sigh)
737 = ([], MkUD { dict_binds = unitBag (Rec all_db_prs, all_db_fvs)
739 , ud_fvs = all_fvs `unionVarSet` b_fvs })
741 all_fvs = rhs_fvs `unionVarSet` body_fvs
742 all_calls = zapCalls bndrs (rhs_calls `unionCalls` body_calls)
744 bndrs = bindersOf bind
745 b_fvs = bind_fvs bind
747 (all_db_prs, all_db_fvs) = add (bind, b_fvs) $
748 foldrBag add ([], emptyVarSet) $
749 rhs_dbs `unionBags` body_dbs
750 add (NonRec b r, b_fvs) (prs, fvs) = ((b,r) : prs, b_fvs `unionVarSet` fvs)
751 add (Rec b_prs, b_fvs) (prs, fvs) = (b_prs ++ prs, b_fvs `unionVarSet` fvs)
753 specBindItself :: Subst -> CoreBind -> CallDetails -> SpecM (CoreBind, UsageDetails)
755 -- specBindItself deals with the RHS, specialising it according
756 -- to the calls found in the body (if any)
757 specBindItself rhs_subst (NonRec bndr rhs) call_info = do
758 ((bndr',rhs'), spec_defns, spec_uds) <- specDefn rhs_subst call_info (bndr,rhs)
760 new_bind | null spec_defns = NonRec bndr' rhs'
761 | otherwise = Rec ((bndr',rhs'):spec_defns)
762 -- bndr' mentions the spec_defns in its SpecEnv
763 -- Not sure why we couln't just put the spec_defns first
764 return (new_bind, spec_uds)
766 specBindItself rhs_subst (Rec pairs) call_info = do
767 stuff <- mapM (specDefn rhs_subst call_info) pairs
769 (pairs', spec_defns_s, spec_uds_s) = unzip3 stuff
770 spec_defns = concat spec_defns_s
771 spec_uds = plusUDList spec_uds_s
772 new_bind = Rec (spec_defns ++ pairs')
773 return (new_bind, spec_uds)
776 specDefn :: Subst -- Subst to use for RHS
777 -> CallDetails -- Info on how it is used in its scope
778 -> (Id, CoreExpr) -- The thing being bound and its un-processed RHS
779 -> SpecM ((Id, CoreExpr), -- The thing and its processed RHS
780 -- the Id may now have specialisations attached
781 [(Id,CoreExpr)], -- Extra, specialised bindings
782 UsageDetails -- Stuff to fling upwards from the RHS and its
783 ) -- specialised versions
785 specDefn subst calls (fn, rhs)
786 -- The first case is the interesting one
787 | rhs_tyvars `lengthIs` n_tyvars -- Rhs of fn's defn has right number of big lambdas
788 && rhs_ids `lengthAtLeast` n_dicts -- and enough dict args
789 && notNull calls_for_me -- And there are some calls to specialise
791 -- && not (certainlyWillInline (idUnfolding fn)) -- And it's not small
792 -- See Note [Inline specialisation] for why we do not
793 -- switch off specialisation for inline functions = do
795 -- Specialise the body of the function
796 (rhs', rhs_uds) <- specExpr subst rhs
798 -- Make a specialised version for each call in calls_for_me
799 stuff <- mapM spec_call calls_for_me
801 (spec_defns, spec_uds, spec_rules) = unzip3 stuff
803 fn' = addIdSpecialisations fn spec_rules
807 rhs_uds `plusUDs` plusUDList spec_uds)
809 | otherwise -- No calls or RHS doesn't fit our preconceptions
810 = WARN( notNull calls_for_me, ptext (sLit "Missed specialisation opportunity for") <+> ppr fn )
811 -- Note [Specialisation shape]
812 (do { (rhs', rhs_uds) <- specExpr subst rhs
813 ; return ((fn, rhs'), [], rhs_uds) })
817 (tyvars, theta, _) = tcSplitSigmaTy fn_type
818 n_tyvars = length tyvars
819 n_dicts = length theta
820 inline_prag = idInlinePragma fn
822 -- It's important that we "see past" any INLINE pragma
823 -- else we'll fail to specialise an INLINE thing
824 (inline_rhs, rhs_inside) = dropInline rhs
825 (rhs_tyvars, rhs_ids, rhs_body) = collectTyAndValBinders rhs_inside
827 rhs_dicts = take n_dicts rhs_ids
828 body = mkLams (drop n_dicts rhs_ids) rhs_body
829 -- Glue back on the non-dict lambdas
831 calls_for_me = case lookupFM calls fn of
833 Just cs -> fmToList cs
835 ----------------------------------------------------------
836 -- Specialise to one particular call pattern
837 spec_call :: (CallKey, ([DictExpr], VarSet)) -- Call instance
838 -> SpecM ((Id,CoreExpr), -- Specialised definition
839 UsageDetails, -- Usage details from specialised body
840 CoreRule) -- Info for the Id's SpecEnv
841 spec_call (CallKey call_ts, (call_ds, _))
842 = ASSERT( call_ts `lengthIs` n_tyvars && call_ds `lengthIs` n_dicts ) do
843 -- Calls are only recorded for properly-saturated applications
845 -- Suppose f's defn is f = /\ a b c d -> \ d1 d2 -> rhs
846 -- Supppose the call is for f [Just t1, Nothing, Just t3, Nothing] [dx1, dx2]
848 -- Construct the new binding
849 -- f1 = SUBST[a->t1,c->t3, d1->d1', d2->d2'] (/\ b d -> rhs)
850 -- PLUS the usage-details
851 -- { d1' = dx1; d2' = dx2 }
852 -- where d1', d2' are cloned versions of d1,d2, with the type substitution applied.
854 -- Note that the substitution is applied to the whole thing.
855 -- This is convenient, but just slightly fragile. Notably:
856 -- * There had better be no name clashes in a/b/c/d
859 -- poly_tyvars = [b,d] in the example above
860 -- spec_tyvars = [a,c]
861 -- ty_args = [t1,b,t3,d]
862 poly_tyvars = [tv | (tv, Nothing) <- rhs_tyvars `zip` call_ts]
863 spec_tyvars = [tv | (tv, Just _) <- rhs_tyvars `zip` call_ts]
864 ty_args = zipWithEqual "spec_call" mk_ty_arg rhs_tyvars call_ts
866 mk_ty_arg rhs_tyvar Nothing = Type (mkTyVarTy rhs_tyvar)
867 mk_ty_arg _ (Just ty) = Type ty
868 rhs_subst = extendTvSubstList subst (spec_tyvars `zip` [ty | Just ty <- call_ts])
870 (rhs_subst', rhs_dicts') <- cloneBinders rhs_subst rhs_dicts
872 inst_args = ty_args ++ map Var rhs_dicts'
874 -- Figure out the type of the specialised function
875 body_ty = applyTypeToArgs rhs fn_type inst_args
876 (lam_args, app_args) -- Add a dummy argument if body_ty is unlifted
877 | isUnLiftedType body_ty -- C.f. WwLib.mkWorkerArgs
878 = (poly_tyvars ++ [voidArgId], poly_tyvars ++ [realWorldPrimId])
879 | otherwise = (poly_tyvars, poly_tyvars)
880 spec_id_ty = mkPiTypes lam_args body_ty
882 spec_f <- newIdSM fn spec_id_ty
883 (spec_rhs, rhs_uds) <- specExpr rhs_subst' (mkLams lam_args body)
885 -- The rule to put in the function's specialisation is:
886 -- forall b,d, d1',d2'. f t1 b t3 d d1' d2' = f1 b d
887 spec_env_rule = mkLocalRule (mkFastString ("SPEC " ++ showSDoc (ppr fn)))
888 inline_prag -- Note [Auto-specialisation and RULES]
890 (poly_tyvars ++ rhs_dicts')
892 (mkVarApps (Var spec_f) app_args)
894 -- Add the { d1' = dx1; d2' = dx2 } usage stuff
895 final_uds = foldr addDictBind rhs_uds (my_zipEqual "spec_call" rhs_dicts' call_ds)
897 spec_pr | inline_rhs = (spec_f `setInlinePragma` inline_prag, Note InlineMe spec_rhs)
898 | otherwise = (spec_f, spec_rhs)
900 return (spec_pr, final_uds, spec_env_rule)
903 my_zipEqual doc xs ys
904 | debugIsOn && not (equalLength xs ys)
905 = pprPanic "my_zipEqual" (vcat
907 , ppr fn <+> ppr call_ts
908 , ppr (idType fn), ppr theta
909 , ppr n_dicts, ppr rhs_dicts
911 | otherwise = zipEqual doc xs ys
914 Note [Auto-specialisation and RULES]
915 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
920 f :: (Int -> Int) -> Int
924 Suppose that auto-specialisation makes a specialised version of
925 g::Int->Int That version won't appear in the LHS of the RULE for f.
926 So if the specialisation rule fires too early, the rule for f may
929 It might be possible to add new rules, to "complete" the rewrite system.
931 RULE forall d. g Int d = g_spec
935 But that's a bit complicated. For now we ask the programmer's help,
936 by *copying the INLINE activation pragma* to the auto-specialised rule.
937 So if g says {-# NOINLINE[2] g #-}, then the auto-spec rule will also
938 not be active until phase 2.
941 Note [Specialisation shape]
942 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
943 We only specialise a function if it has visible top-level lambdas
944 corresponding to its overloading. E.g. if
945 f :: forall a. Eq a => ....
946 then its body must look like
949 Reason: when specialising the body for a call (f ty dexp), we want to
950 substitute dexp for d, and pick up specialised calls in the body of f.
952 This doesn't always work. One example I came across was htis:
953 newtype Gen a = MkGen{ unGen :: Int -> a }
955 choose :: Eq a => a -> Gen a
956 choose n = MkGen (\r -> n)
958 oneof = choose (1::Int)
960 It's a silly exapmle, but we get
961 choose = /\a. g `cast` co
962 where choose doesn't have any dict arguments. Thus far I have not
963 tried to fix this (wait till there's a real example).
966 Note [Inline specialisations]
967 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
968 We transfer to the specialised function any INLINE stuff from the
969 original. This means (a) the Activation in the IdInfo, and (b) any
972 This is a change (Jun06). Previously the idea is that the point of
973 inlining was precisely to specialise the function at its call site,
974 and that's not so important for the specialised copies. But
975 *pragma-directed* specialisation now takes place in the
976 typechecker/desugarer, with manually specified INLINEs. The
977 specialiation here is automatic. It'd be very odd if a function
978 marked INLINE was specialised (because of some local use), and then
979 forever after (including importing modules) the specialised version
980 wasn't INLINEd. After all, the programmer said INLINE!
982 You might wonder why we don't just not specialise INLINE functions.
983 It's because even INLINE functions are sometimes not inlined, when
984 they aren't applied to interesting arguments. But perhaps the type
985 arguments alone are enough to specialise (even though the args are too
986 boring to trigger inlining), and it's certainly better to call the
989 A case in point is dictionary functions, which are current marked
990 INLINE, but which are worth specialising.
993 dropInline :: CoreExpr -> (Bool, CoreExpr)
994 dropInline (Note InlineMe rhs) = (True, rhs)
995 dropInline rhs = (False, rhs)
998 %************************************************************************
1000 \subsubsection{UsageDetails and suchlike}
1002 %************************************************************************
1007 dict_binds :: !(Bag DictBind),
1008 -- Floated dictionary bindings
1009 -- The order is important;
1010 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
1011 -- (Remember, Bags preserve order in GHC.)
1013 calls :: !CallDetails,
1015 ud_fvs :: !VarSet -- A superset of the variables mentioned in
1016 -- either dict_binds or calls
1019 instance Outputable UsageDetails where
1020 ppr (MkUD { dict_binds = dbs, calls = calls, ud_fvs = fvs })
1021 = ptext (sLit "MkUD") <+> braces (sep (punctuate comma
1022 [ptext (sLit "binds") <+> equals <+> ppr dbs,
1023 ptext (sLit "calls") <+> equals <+> ppr calls,
1024 ptext (sLit "fvs") <+> equals <+> ppr fvs]))
1026 type DictBind = (CoreBind, VarSet)
1027 -- The set is the free vars of the binding
1028 -- both tyvars and dicts
1030 type DictExpr = CoreExpr
1032 emptyUDs :: UsageDetails
1033 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM, ud_fvs = emptyVarSet }
1035 ------------------------------------------------------------
1036 type CallDetails = FiniteMap Id CallInfo
1037 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
1038 type CallInfo = FiniteMap CallKey
1039 ([DictExpr], VarSet) -- Dict args and the vars of the whole
1040 -- call (including tyvars)
1041 -- [*not* include the main id itself, of course]
1042 -- The finite maps eliminate duplicates
1043 -- The list of types and dictionaries is guaranteed to
1044 -- match the type of f
1046 instance Outputable CallKey where
1047 ppr (CallKey ts) = ppr ts
1049 -- Type isn't an instance of Ord, so that we can control which
1050 -- instance we use. That's tiresome here. Oh well
1051 instance Eq CallKey where
1052 k1 == k2 = case k1 `compare` k2 of { EQ -> True; _ -> False }
1054 instance Ord CallKey where
1055 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
1057 cmp Nothing Nothing = EQ
1058 cmp Nothing (Just _) = LT
1059 cmp (Just _) Nothing = GT
1060 cmp (Just t1) (Just t2) = tcCmpType t1 t2
1062 unionCalls :: CallDetails -> CallDetails -> CallDetails
1063 unionCalls c1 c2 = plusFM_C plusFM c1 c2
1065 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> UsageDetails
1066 singleCall id tys dicts
1067 = MkUD {dict_binds = emptyBag,
1068 calls = unitFM id (unitFM (CallKey tys) (dicts, call_fvs)),
1071 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
1072 tys_fvs = tyVarsOfTypes (catMaybes tys)
1073 -- The type args (tys) are guaranteed to be part of the dictionary
1074 -- types, because they are just the constrained types,
1075 -- and the dictionary is therefore sure to be bound
1076 -- inside the binding for any type variables free in the type;
1077 -- hence it's safe to neglect tyvars free in tys when making
1078 -- the free-var set for this call
1079 -- BUT I don't trust this reasoning; play safe and include tys_fvs
1081 -- We don't include the 'id' itself.
1083 mkCallUDs :: Subst -> Id -> [CoreExpr] -> UsageDetails
1084 mkCallUDs subst f args
1086 || not (all isClassPred theta)
1087 -- Only specialise if all overloading is on class params.
1088 -- In ptic, with implicit params, the type args
1089 -- *don't* say what the value of the implicit param is!
1090 || not (spec_tys `lengthIs` n_tyvars)
1091 || not ( dicts `lengthIs` n_dicts)
1092 || maybeToBool (lookupRule (\_act -> True) (substInScope subst) emptyRuleBase f args)
1093 -- There's already a rule covering this call. A typical case
1094 -- is where there's an explicit user-provided rule. Then
1095 -- we don't want to create a specialised version
1096 -- of the function that overlaps.
1097 = emptyUDs -- Not overloaded, or no specialisation wanted
1100 = singleCall f spec_tys dicts
1102 (tyvars, theta, _) = tcSplitSigmaTy (idType f)
1103 constrained_tyvars = tyVarsOfTheta theta
1104 n_tyvars = length tyvars
1105 n_dicts = length theta
1107 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
1108 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
1111 | tyvar `elemVarSet` constrained_tyvars = Just ty
1112 | otherwise = Nothing
1114 ------------------------------------------------------------
1115 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
1116 plusUDs (MkUD {dict_binds = db1, calls = calls1, ud_fvs = fvs1})
1117 (MkUD {dict_binds = db2, calls = calls2, ud_fvs = fvs2})
1118 = MkUD {dict_binds = d, calls = c, ud_fvs = fvs1 `unionVarSet` fvs2}
1120 d = db1 `unionBags` db2
1121 c = calls1 `unionCalls` calls2
1123 plusUDList :: [UsageDetails] -> UsageDetails
1124 plusUDList = foldr plusUDs emptyUDs
1126 -- zapCalls deletes calls to ids from uds
1127 zapCalls :: [Id] -> CallDetails -> CallDetails
1128 zapCalls ids calls = delListFromFM calls ids
1130 mkDB :: CoreBind -> DictBind
1131 mkDB bind = (bind, bind_fvs bind)
1133 bind_fvs :: CoreBind -> VarSet
1134 bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
1135 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1138 rhs_fvs = unionVarSets (map pair_fvs prs)
1140 pair_fvs :: (Id, CoreExpr) -> VarSet
1141 pair_fvs (bndr, rhs) = exprFreeVars rhs `unionVarSet` idFreeVars bndr
1142 -- Don't forget variables mentioned in the
1143 -- rules of the bndr. C.f. OccAnal.addRuleUsage
1144 -- Also tyvars mentioned in its type; they may not appear in the RHS
1148 addDictBind :: (Id,CoreExpr) -> UsageDetails -> UsageDetails
1149 addDictBind (dict,rhs) uds
1150 = uds { dict_binds = db `consBag` dict_binds uds
1151 , ud_fvs = ud_fvs uds `unionVarSet` fvs }
1153 db@(_, fvs) = mkDB (NonRec dict rhs)
1155 dumpAllDictBinds :: UsageDetails -> [CoreBind] -> [CoreBind]
1156 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
1157 = foldrBag add binds dbs
1159 add (bind,_) binds = bind : binds
1161 dumpUDs :: [CoreBndr]
1162 -> UsageDetails -> CoreExpr
1163 -> (UsageDetails, CoreExpr)
1164 dumpUDs bndrs (MkUD { dict_binds = orig_dbs
1165 , calls = orig_calls
1166 , ud_fvs = fvs}) body
1167 = (MkUD { dict_binds = free_dbs
1168 , calls = free_calls
1169 , ud_fvs = fvs `minusVarSet` bndr_set}, -- This may delete fewer variables
1170 foldrBag add_let body dump_dbs) -- than in priciple possible
1172 bndr_set = mkVarSet bndrs
1173 add_let (bind,_) body = Let bind body
1175 (free_dbs, dump_dbs, dump_set)
1176 = foldlBag dump_db (emptyBag, emptyBag, bndr_set) orig_dbs
1177 -- Important that it's foldl not foldr;
1178 -- we're accumulating the set of dumped ids in dump_set
1180 free_calls = filterCalls dump_set orig_calls
1182 dump_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1183 | dump_idset `intersectsVarSet` fvs -- Dump it
1184 = (free_dbs, dump_dbs `snocBag` db,
1185 extendVarSetList dump_idset (bindersOf bind))
1187 | otherwise -- Don't dump it
1188 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1190 filterCalls :: VarSet -> CallDetails -> CallDetails
1191 -- Remove any calls that mention the variables
1192 filterCalls bs calls
1193 = mapFM (\_ cs -> filter_calls cs) $
1194 filterFM (\k _ -> k `elemVarSet` bs) calls
1196 filter_calls :: CallInfo -> CallInfo
1197 filter_calls = filterFM (\_ (_, fvs) -> fvs `intersectsVarSet` bs)
1201 %************************************************************************
1203 \subsubsection{Boring helper functions}
1205 %************************************************************************
1208 type SpecM a = UniqSM a
1210 initSM :: UniqSupply -> SpecM a -> a
1213 mapAndCombineSM :: (a -> SpecM (b, UsageDetails)) -> [a] -> SpecM ([b], UsageDetails)
1214 mapAndCombineSM _ [] = return ([], emptyUDs)
1215 mapAndCombineSM f (x:xs) = do (y, uds1) <- f x
1216 (ys, uds2) <- mapAndCombineSM f xs
1217 return (y:ys, uds1 `plusUDs` uds2)
1219 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1220 -- Clone the binders of the bind; return new bind with the cloned binders
1221 -- Return the substitution to use for RHSs, and the one to use for the body
1222 cloneBindSM subst (NonRec bndr rhs) = do
1223 us <- getUniqueSupplyM
1224 let (subst', bndr') = cloneIdBndr subst us bndr
1225 return (subst, subst', NonRec bndr' rhs)
1227 cloneBindSM subst (Rec pairs) = do
1228 us <- getUniqueSupplyM
1229 let (subst', bndrs') = cloneRecIdBndrs subst us (map fst pairs)
1230 return (subst', subst', Rec (bndrs' `zip` map snd pairs))
1232 cloneBinders :: Subst -> [CoreBndr] -> SpecM (Subst, [CoreBndr])
1233 cloneBinders subst bndrs = do
1234 us <- getUniqueSupplyM
1235 return (cloneIdBndrs subst us bndrs)
1237 newIdSM :: Id -> Type -> SpecM Id
1238 newIdSM old_id new_ty = do
1241 -- Give the new Id a similar occurrence name to the old one
1242 name = idName old_id
1243 new_id = mkUserLocal (mkSpecOcc (nameOccName name)) uniq new_ty (getSrcSpan name)
1248 Old (but interesting) stuff about unboxed bindings
1249 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1251 What should we do when a value is specialised to a *strict* unboxed value?
1253 map_*_* f (x:xs) = let h = f x
1257 Could convert let to case:
1259 map_*_Int# f (x:xs) = case f x of h# ->
1263 This may be undesirable since it forces evaluation here, but the value
1264 may not be used in all branches of the body. In the general case this
1265 transformation is impossible since the mutual recursion in a letrec
1266 cannot be expressed as a case.
1268 There is also a problem with top-level unboxed values, since our
1269 implementation cannot handle unboxed values at the top level.
1271 Solution: Lift the binding of the unboxed value and extract it when it
1274 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1279 Now give it to the simplifier and the _Lifting will be optimised away.
1281 The benfit is that we have given the specialised "unboxed" values a
1282 very simplep lifted semantics and then leave it up to the simplifier to
1283 optimise it --- knowing that the overheads will be removed in nearly
1286 In particular, the value will only be evaluted in the branches of the
1287 program which use it, rather than being forced at the point where the
1288 value is bound. For example:
1290 filtermap_*_* p f (x:xs)
1297 filtermap_*_Int# p f (x:xs)
1298 = let h = case (f x) of h# -> _Lift h#
1301 True -> case h of _Lift h#
1305 The binding for h can still be inlined in the one branch and the
1306 _Lifting eliminated.
1309 Question: When won't the _Lifting be eliminated?
1311 Answer: When they at the top-level (where it is necessary) or when
1312 inlining would duplicate work (or possibly code depending on
1313 options). However, the _Lifting will still be eliminated if the
1314 strictness analyser deems the lifted binding strict.