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
32 import CoreUtils ( applyTypeToArgs, mkPiTypes )
33 import CoreFVs ( exprFreeVars, exprsFreeVars, idFreeVars )
34 import CoreLint ( showPass, endPass )
35 import UniqSupply ( UniqSupply,
40 import MkId ( voidArgId, realWorldPrimId )
42 import Maybes ( catMaybes, maybeToBool )
43 import ErrUtils ( dumpIfSet_dyn )
51 %************************************************************************
53 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
55 %************************************************************************
57 These notes describe how we implement specialisation to eliminate
60 The specialisation pass works on Core
61 syntax, complete with all the explicit dictionary application,
62 abstraction and construction as added by the type checker. The
63 existing type checker remains largely as it is.
65 One important thought: the {\em types} passed to an overloaded
66 function, and the {\em dictionaries} passed are mutually redundant.
67 If the same function is applied to the same type(s) then it is sure to
68 be applied to the same dictionary(s)---or rather to the same {\em
69 values}. (The arguments might look different but they will evaluate
72 Second important thought: we know that we can make progress by
73 treating dictionary arguments as static and worth specialising on. So
74 we can do without binding-time analysis, and instead specialise on
75 dictionary arguments and no others.
84 and suppose f is overloaded.
86 STEP 1: CALL-INSTANCE COLLECTION
88 We traverse <body>, accumulating all applications of f to types and
91 (Might there be partial applications, to just some of its types and
92 dictionaries? In principle yes, but in practice the type checker only
93 builds applications of f to all its types and dictionaries, so partial
94 applications could only arise as a result of transformation, and even
95 then I think it's unlikely. In any case, we simply don't accumulate such
96 partial applications.)
101 So now we have a collection of calls to f:
105 Notice that f may take several type arguments. To avoid ambiguity, we
106 say that f is called at type t1/t2 and t3/t4.
108 We take equivalence classes using equality of the *types* (ignoring
109 the dictionary args, which as mentioned previously are redundant).
111 STEP 3: SPECIALISATION
113 For each equivalence class, choose a representative (f t1 t2 d1 d2),
114 and create a local instance of f, defined thus:
116 f@t1/t2 = <f_rhs> t1 t2 d1 d2
118 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
119 of simplification will now result. However we don't actually *do* that
120 simplification. Rather, we leave it for the simplifier to do. If we
121 *did* do it, though, we'd get more call instances from the specialised
122 RHS. We can work out what they are by instantiating the call-instance
123 set from f's RHS with the types t1, t2.
125 Add this new id to f's IdInfo, to record that f has a specialised version.
127 Before doing any of this, check that f's IdInfo doesn't already
128 tell us about an existing instance of f at the required type/s.
129 (This might happen if specialisation was applied more than once, or
130 it might arise from user SPECIALIZE pragmas.)
134 Wait a minute! What if f is recursive? Then we can't just plug in
135 its right-hand side, can we?
137 But it's ok. The type checker *always* creates non-recursive definitions
138 for overloaded recursive functions. For example:
140 f x = f (x+x) -- Yes I know its silly
144 f a (d::Num a) = let p = +.sel a d
146 letrec fl (y::a) = fl (p y y)
150 We still have recusion for non-overloaded functions which we
151 speciailise, but the recursive call should get specialised to the
152 same recursive version.
158 All this is crystal clear when the function is applied to *constant
159 types*; that is, types which have no type variables inside. But what if
160 it is applied to non-constant types? Suppose we find a call of f at type
161 t1/t2. There are two possibilities:
163 (a) The free type variables of t1, t2 are in scope at the definition point
164 of f. In this case there's no problem, we proceed just as before. A common
165 example is as follows. Here's the Haskell:
170 After typechecking we have
172 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
173 in +.sel a d (f a d y) (f a d y)
175 Notice that the call to f is at type type "a"; a non-constant type.
176 Both calls to f are at the same type, so we can specialise to give:
178 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
179 in +.sel a d (f@a y) (f@a y)
182 (b) The other case is when the type variables in the instance types
183 are *not* in scope at the definition point of f. The example we are
184 working with above is a good case. There are two instances of (+.sel a d),
185 but "a" is not in scope at the definition of +.sel. Can we do anything?
186 Yes, we can "common them up", a sort of limited common sub-expression deal.
189 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
190 f@a (x::a) = +.sel@a x x
191 in +.sel@a (f@a y) (f@a y)
193 This can save work, and can't be spotted by the type checker, because
194 the two instances of +.sel weren't originally at the same type.
198 * There are quite a few variations here. For example, the defn of
199 +.sel could be floated ouside the \y, to attempt to gain laziness.
200 It certainly mustn't be floated outside the \d because the d has to
203 * We don't want to inline f_rhs in this case, because
204 that will duplicate code. Just commoning up the call is the point.
206 * Nothing gets added to +.sel's IdInfo.
208 * Don't bother unless the equivalence class has more than one item!
210 Not clear whether this is all worth it. It is of course OK to
211 simply discard call-instances when passing a big lambda.
213 Polymorphism 2 -- Overloading
215 Consider a function whose most general type is
217 f :: forall a b. Ord a => [a] -> b -> b
219 There is really no point in making a version of g at Int/Int and another
220 at Int/Bool, because it's only instancing the type variable "a" which
221 buys us any efficiency. Since g is completely polymorphic in b there
222 ain't much point in making separate versions of g for the different
225 That suggests that we should identify which of g's type variables
226 are constrained (like "a") and which are unconstrained (like "b").
227 Then when taking equivalence classes in STEP 2, we ignore the type args
228 corresponding to unconstrained type variable. In STEP 3 we make
229 polymorphic versions. Thus:
231 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
240 f a (d::Num a) = let g = ...
242 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
244 Here, g is only called at one type, but the dictionary isn't in scope at the
245 definition point for g. Usually the type checker would build a
246 definition for d1 which enclosed g, but the transformation system
247 might have moved d1's defn inward. Solution: float dictionary bindings
248 outwards along with call instances.
252 f x = let g p q = p==q
258 Before specialisation, leaving out type abstractions we have
260 f df x = let g :: Eq a => a -> a -> Bool
262 h :: Num a => a -> a -> (a, Bool)
263 h dh r s = let deq = eqFromNum dh
264 in (+ dh r s, g deq r s)
268 After specialising h we get a specialised version of h, like this:
270 h' r s = let deq = eqFromNum df
271 in (+ df r s, g deq r s)
273 But we can't naively make an instance for g from this, because deq is not in scope
274 at the defn of g. Instead, we have to float out the (new) defn of deq
275 to widen its scope. Notice that this floating can't be done in advance -- it only
276 shows up when specialisation is done.
278 User SPECIALIZE pragmas
279 ~~~~~~~~~~~~~~~~~~~~~~~
280 Specialisation pragmas can be digested by the type checker, and implemented
281 by adding extra definitions along with that of f, in the same way as before
283 f@t1/t2 = <f_rhs> t1 t2 d1 d2
285 Indeed the pragmas *have* to be dealt with by the type checker, because
286 only it knows how to build the dictionaries d1 and d2! For example
288 g :: Ord a => [a] -> [a]
289 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
291 Here, the specialised version of g is an application of g's rhs to the
292 Ord dictionary for (Tree Int), which only the type checker can conjure
293 up. There might not even *be* one, if (Tree Int) is not an instance of
294 Ord! (All the other specialision has suitable dictionaries to hand
297 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
298 it is buried in a complex (as-yet-un-desugared) binding group.
301 f@t1/t2 = f* t1 t2 d1 d2
303 where f* is the Id f with an IdInfo which says "inline me regardless!".
304 Indeed all the specialisation could be done in this way.
305 That in turn means that the simplifier has to be prepared to inline absolutely
306 any in-scope let-bound thing.
309 Again, the pragma should permit polymorphism in unconstrained variables:
311 h :: Ord a => [a] -> b -> b
312 {-# SPECIALIZE h :: [Int] -> b -> b #-}
314 We *insist* that all overloaded type variables are specialised to ground types,
315 (and hence there can be no context inside a SPECIALIZE pragma).
316 We *permit* unconstrained type variables to be specialised to
318 - or left as a polymorphic type variable
319 but nothing in between. So
321 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
323 is *illegal*. (It can be handled, but it adds complication, and gains the
327 SPECIALISING INSTANCE DECLARATIONS
328 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
331 instance Foo a => Foo [a] where
333 {-# SPECIALIZE instance Foo [Int] #-}
335 The original instance decl creates a dictionary-function
338 dfun.Foo.List :: forall a. Foo a -> Foo [a]
340 The SPECIALIZE pragma just makes a specialised copy, just as for
341 ordinary function definitions:
343 dfun.Foo.List@Int :: Foo [Int]
344 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
346 The information about what instance of the dfun exist gets added to
347 the dfun's IdInfo in the same way as a user-defined function too.
350 Automatic instance decl specialisation?
351 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
352 Can instance decls be specialised automatically? It's tricky.
353 We could collect call-instance information for each dfun, but
354 then when we specialised their bodies we'd get new call-instances
355 for ordinary functions; and when we specialised their bodies, we might get
356 new call-instances of the dfuns, and so on. This all arises because of
357 the unrestricted mutual recursion between instance decls and value decls.
359 Still, there's no actual problem; it just means that we may not do all
360 the specialisation we could theoretically do.
362 Furthermore, instance decls are usually exported and used non-locally,
363 so we'll want to compile enough to get those specialisations done.
365 Lastly, there's no such thing as a local instance decl, so we can
366 survive solely by spitting out *usage* information, and then reading that
367 back in as a pragma when next compiling the file. So for now,
368 we only specialise instance decls in response to pragmas.
371 SPITTING OUT USAGE INFORMATION
372 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
374 To spit out usage information we need to traverse the code collecting
375 call-instance information for all imported (non-prelude?) functions
376 and data types. Then we equivalence-class it and spit it out.
378 This is done at the top-level when all the call instances which escape
379 must be for imported functions and data types.
381 *** Not currently done ***
384 Partial specialisation by pragmas
385 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
386 What about partial specialisation:
388 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
389 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
393 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
395 Seems quite reasonable. Similar things could be done with instance decls:
397 instance (Foo a, Foo b) => Foo (a,b) where
399 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
400 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
402 Ho hum. Things are complex enough without this. I pass.
405 Requirements for the simplifer
406 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
407 The simplifier has to be able to take advantage of the specialisation.
409 * When the simplifier finds an application of a polymorphic f, it looks in
410 f's IdInfo in case there is a suitable instance to call instead. This converts
412 f t1 t2 d1 d2 ===> f_t1_t2
414 Note that the dictionaries get eaten up too!
416 * Dictionary selection operations on constant dictionaries must be
419 +.sel Int d ===> +Int
421 The obvious way to do this is in the same way as other specialised
422 calls: +.sel has inside it some IdInfo which tells that if it's applied
423 to the type Int then it should eat a dictionary and transform to +Int.
425 In short, dictionary selectors need IdInfo inside them for constant
428 * Exactly the same applies if a superclass dictionary is being
431 Eq.sel Int d ===> dEqInt
433 * Something similar applies to dictionary construction too. Suppose
434 dfun.Eq.List is the function taking a dictionary for (Eq a) to
435 one for (Eq [a]). Then we want
437 dfun.Eq.List Int d ===> dEq.List_Int
439 Where does the Eq [Int] dictionary come from? It is built in
440 response to a SPECIALIZE pragma on the Eq [a] instance decl.
442 In short, dfun Ids need IdInfo with a specialisation for each
443 constant instance of their instance declaration.
445 All this uses a single mechanism: the SpecEnv inside an Id
448 What does the specialisation IdInfo look like?
449 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
451 The SpecEnv of an Id maps a list of types (the template) to an expression
455 For example, if f has this SpecInfo:
457 [Int, a] -> \d:Ord Int. f' a
459 it means that we can replace the call
461 f Int t ===> (\d. f' t)
463 This chucks one dictionary away and proceeds with the
464 specialised version of f, namely f'.
467 What can't be done this way?
468 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
469 There is no way, post-typechecker, to get a dictionary for (say)
470 Eq a from a dictionary for Eq [a]. So if we find
474 we can't transform to
479 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
481 Of course, we currently have no way to automatically derive
482 eqList, nor to connect it to the Eq [a] instance decl, but you
483 can imagine that it might somehow be possible. Taking advantage
484 of this is permanently ruled out.
486 Still, this is no great hardship, because we intend to eliminate
487 overloading altogether anyway!
491 A note about non-tyvar dictionaries
492 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
493 Some Ids have types like
495 forall a,b,c. Eq a -> Ord [a] -> tau
497 This seems curious at first, because we usually only have dictionary
498 args whose types are of the form (C a) where a is a type variable.
499 But this doesn't hold for the functions arising from instance decls,
500 which sometimes get arguements with types of form (C (T a)) for some
503 Should we specialise wrt this compound-type dictionary? We used to say
505 "This is a heuristic judgement, as indeed is the fact that we
506 specialise wrt only dictionaries. We choose *not* to specialise
507 wrt compound dictionaries because at the moment the only place
508 they show up is in instance decls, where they are simply plugged
509 into a returned dictionary. So nothing is gained by specialising
512 But it is simpler and more uniform to specialise wrt these dicts too;
513 and in future GHC is likely to support full fledged type signatures
515 f ;: Eq [(a,b)] => ...
518 %************************************************************************
520 \subsubsection{The new specialiser}
522 %************************************************************************
524 Our basic game plan is this. For let(rec) bound function
525 f :: (C a, D c) => (a,b,c,d) -> Bool
527 * Find any specialised calls of f, (f ts ds), where
528 ts are the type arguments t1 .. t4, and
529 ds are the dictionary arguments d1 .. d2.
531 * Add a new definition for f1 (say):
533 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
535 Note that we abstract over the unconstrained type arguments.
539 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
541 to the specialisations of f. This will be used by the
542 simplifier to replace calls
543 (f t1 t2 t3 t4) da db
545 (\d1 d1 -> f1 t2 t4) da db
547 All the stuff about how many dictionaries to discard, and what types
548 to apply the specialised function to, are handled by the fact that the
549 SpecEnv contains a template for the result of the specialisation.
551 We don't build *partial* specialisations for f. For example:
553 f :: Eq a => a -> a -> Bool
554 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
556 Here, little is gained by making a specialised copy of f.
557 There's a distinct danger that the specialised version would
558 first build a dictionary for (Eq b, Eq c), and then select the (==)
559 method from it! Even if it didn't, not a great deal is saved.
561 We do, however, generate polymorphic, but not overloaded, specialisations:
563 f :: Eq a => [a] -> b -> b -> b
564 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
566 Hence, the invariant is this:
568 *** no specialised version is overloaded ***
571 %************************************************************************
573 \subsubsection{The exported function}
575 %************************************************************************
578 specProgram :: DynFlags -> UniqSupply -> [CoreBind] -> IO [CoreBind]
579 specProgram dflags us binds = do
581 showPass dflags "Specialise"
583 let binds' = initSM us (do (binds', uds') <- go binds
584 return (dumpAllDictBinds uds' binds'))
586 endPass dflags "Specialise" Opt_D_dump_spec binds'
588 dumpIfSet_dyn dflags Opt_D_dump_rules "Top-level specialisations"
589 (pprRulesForUser (rulesOfBinds binds'))
593 -- We need to start with a Subst that knows all the things
594 -- that are in scope, so that the substitution engine doesn't
595 -- accidentally re-use a unique that's already in use
596 -- Easiest thing is to do it all at once, as if all the top-level
597 -- decls were mutually recursive
598 top_subst = mkEmptySubst (mkInScopeSet (mkVarSet (bindersOfBinds binds)))
600 go [] = return ([], emptyUDs)
601 go (bind:binds) = do (binds', uds) <- go binds
602 (bind', uds') <- specBind top_subst bind uds
603 return (bind' ++ binds', uds')
606 %************************************************************************
608 \subsubsection{@specExpr@: the main function}
610 %************************************************************************
613 specVar :: Subst -> Id -> CoreExpr
614 specVar subst v = lookupIdSubst subst v
616 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
617 -- We carry a substitution down:
618 -- a) we must clone any binding that might flaot outwards,
619 -- to avoid name clashes
620 -- b) we carry a type substitution to use when analysing
621 -- the RHS of specialised bindings (no type-let!)
623 ---------------- First the easy cases --------------------
624 specExpr subst (Type ty) = return (Type (substTy subst ty), emptyUDs)
625 specExpr subst (Var v) = return (specVar subst v, emptyUDs)
626 specExpr _ (Lit lit) = return (Lit lit, emptyUDs)
627 specExpr subst (Cast e co) = do
628 (e', uds) <- specExpr subst e
629 return ((Cast e' (substTy subst co)), uds)
630 specExpr subst (Note note body) = do
631 (body', uds) <- specExpr subst body
632 return (Note (specNote subst note) body', uds)
635 ---------------- Applications might generate a call instance --------------------
636 specExpr subst expr@(App {})
639 go (App fun arg) args = do (arg', uds_arg) <- specExpr subst arg
640 (fun', uds_app) <- go fun (arg':args)
641 return (App fun' arg', uds_arg `plusUDs` uds_app)
643 go (Var f) args = case specVar subst f of
644 Var f' -> return (Var f', mkCallUDs subst f' args)
645 e' -> return (e', emptyUDs) -- I don't expect this!
646 go other _ = specExpr subst other
648 ---------------- Lambda/case require dumping of usage details --------------------
649 specExpr subst e@(Lam _ _) = do
650 (body', uds) <- specExpr subst' body
651 let (filtered_uds, body'') = dumpUDs bndrs' uds body'
652 return (mkLams bndrs' body'', filtered_uds)
654 (bndrs, body) = collectBinders e
655 (subst', bndrs') = substBndrs subst bndrs
656 -- More efficient to collect a group of binders together all at once
657 -- and we don't want to split a lambda group with dumped bindings
659 specExpr subst (Case scrut case_bndr ty alts) = do
660 (scrut', uds_scrut) <- specExpr subst scrut
661 (alts', uds_alts) <- mapAndCombineSM spec_alt alts
662 return (Case scrut' case_bndr' (substTy subst ty) alts', uds_scrut `plusUDs` uds_alts)
664 (subst_alt, case_bndr') = substBndr subst case_bndr
665 -- No need to clone case binder; it can't float like a let(rec)
667 spec_alt (con, args, rhs) = do
668 (rhs', uds) <- specExpr subst_rhs rhs
669 let (uds', rhs'') = dumpUDs args uds rhs'
670 return ((con, args', rhs''), uds')
672 (subst_rhs, args') = substBndrs subst_alt args
674 ---------------- Finally, let is the interesting case --------------------
675 specExpr subst (Let bind body) = do
677 (rhs_subst, body_subst, bind') <- cloneBindSM subst bind
679 -- Deal with the body
680 (body', body_uds) <- specExpr body_subst body
682 -- Deal with the bindings
683 (binds', uds) <- specBind rhs_subst bind' body_uds
686 return (foldr Let body' binds', uds)
688 -- Must apply the type substitution to coerceions
689 specNote :: Subst -> Note -> Note
690 specNote _ note = note
693 %************************************************************************
695 \subsubsection{Dealing with a binding}
697 %************************************************************************
700 specBind :: Subst -- Use this for RHSs
702 -> UsageDetails -- Info on how the scope of the binding
703 -> SpecM ([CoreBind], -- New bindings
704 UsageDetails) -- And info to pass upstream
706 specBind rhs_subst bind body_uds
707 = do { (bind', bind_uds) <- specBindItself rhs_subst bind (calls body_uds)
708 ; return (finishSpecBind bind' bind_uds body_uds) }
710 finishSpecBind :: CoreBind -> UsageDetails -> UsageDetails -> ([CoreBind], UsageDetails)
712 (MkUD { dict_binds = rhs_dbs, calls = rhs_calls, ud_fvs = rhs_fvs })
713 (MkUD { dict_binds = body_dbs, calls = body_calls, ud_fvs = body_fvs })
714 | not (mkVarSet bndrs `intersectsVarSet` all_fvs)
715 -- Common case 1: the bound variables are not
716 -- mentioned in the dictionary bindings
717 = ([bind], MkUD { dict_binds = body_dbs `unionBags` rhs_dbs
718 -- It's important that the `unionBags` is this way round,
719 -- because body_uds may bind dictionaries that are
720 -- used in the calls passed to specDefn. So the
721 -- dictionary bindings in rhs_uds may mention
722 -- dictionaries bound in body_uds.
724 , ud_fvs = all_fvs })
726 | case bind of { NonRec {} -> True; Rec {} -> False }
727 -- Common case 2: no specialisation happened, and binding
728 -- is non-recursive. But the binding may be
729 -- mentioned in body_dbs, so we should put it first
730 = ([], MkUD { dict_binds = rhs_dbs `unionBags` ((bind, b_fvs) `consBag` body_dbs)
732 , ud_fvs = all_fvs `unionVarSet` b_fvs })
734 | otherwise -- General case: make a huge Rec (sigh)
735 = ([], MkUD { dict_binds = unitBag (Rec all_db_prs, all_db_fvs)
737 , ud_fvs = all_fvs `unionVarSet` b_fvs })
739 all_fvs = rhs_fvs `unionVarSet` body_fvs
740 all_calls = zapCalls bndrs (rhs_calls `unionCalls` body_calls)
742 bndrs = bindersOf bind
743 b_fvs = bind_fvs bind
745 (all_db_prs, all_db_fvs) = add (bind, b_fvs) $
746 foldrBag add ([], emptyVarSet) $
747 rhs_dbs `unionBags` body_dbs
748 add (NonRec b r, b_fvs) (prs, fvs) = ((b,r) : prs, b_fvs `unionVarSet` fvs)
749 add (Rec b_prs, b_fvs) (prs, fvs) = (b_prs ++ prs, b_fvs `unionVarSet` fvs)
751 specBindItself :: Subst -> CoreBind -> CallDetails -> SpecM (CoreBind, UsageDetails)
753 -- specBindItself deals with the RHS, specialising it according
754 -- to the calls found in the body (if any)
755 specBindItself rhs_subst (NonRec bndr rhs) call_info = do
756 ((bndr',rhs'), spec_defns, spec_uds) <- specDefn rhs_subst call_info (bndr,rhs)
758 new_bind | null spec_defns = NonRec bndr' rhs'
759 | otherwise = Rec ((bndr',rhs'):spec_defns)
760 -- bndr' mentions the spec_defns in its SpecEnv
761 -- Not sure why we couln't just put the spec_defns first
762 return (new_bind, spec_uds)
764 specBindItself rhs_subst (Rec pairs) call_info = do
765 stuff <- mapM (specDefn rhs_subst call_info) pairs
767 (pairs', spec_defns_s, spec_uds_s) = unzip3 stuff
768 spec_defns = concat spec_defns_s
769 spec_uds = plusUDList spec_uds_s
770 new_bind = Rec (spec_defns ++ pairs')
771 return (new_bind, spec_uds)
774 specDefn :: Subst -- Subst to use for RHS
775 -> CallDetails -- Info on how it is used in its scope
776 -> (Id, CoreExpr) -- The thing being bound and its un-processed RHS
777 -> SpecM ((Id, CoreExpr), -- The thing and its processed RHS
778 -- the Id may now have specialisations attached
779 [(Id,CoreExpr)], -- Extra, specialised bindings
780 UsageDetails -- Stuff to fling upwards from the RHS and its
781 ) -- specialised versions
783 specDefn subst calls (fn, rhs)
784 -- The first case is the interesting one
785 | rhs_tyvars `lengthIs` n_tyvars -- Rhs of fn's defn has right number of big lambdas
786 && rhs_ids `lengthAtLeast` n_dicts -- and enough dict args
787 && notNull calls_for_me -- And there are some calls to specialise
789 -- && not (certainlyWillInline (idUnfolding fn)) -- And it's not small
790 -- See Note [Inline specialisation] for why we do not
791 -- switch off specialisation for inline functions = do
793 -- Specialise the body of the function
794 (rhs', rhs_uds) <- specExpr subst rhs
796 -- Make a specialised version for each call in calls_for_me
797 stuff <- mapM spec_call calls_for_me
799 (spec_defns, spec_uds, spec_rules) = unzip3 stuff
801 fn' = addIdSpecialisations fn spec_rules
805 rhs_uds `plusUDs` plusUDList spec_uds)
807 | otherwise -- No calls or RHS doesn't fit our preconceptions
808 = WARN( notNull calls_for_me, ptext (sLit "Missed specialisation opportunity for") <+> ppr fn )
809 -- Note [Specialisation shape]
810 (do { (rhs', rhs_uds) <- specExpr subst rhs
811 ; return ((fn, rhs'), [], rhs_uds) })
815 (tyvars, theta, _) = tcSplitSigmaTy fn_type
816 n_tyvars = length tyvars
817 n_dicts = length theta
818 inline_prag = idInlinePragma fn
820 -- It's important that we "see past" any INLINE pragma
821 -- else we'll fail to specialise an INLINE thing
822 (inline_rhs, rhs_inside) = dropInline rhs
823 (rhs_tyvars, rhs_ids, rhs_body) = collectTyAndValBinders rhs_inside
825 rhs_dicts = take n_dicts rhs_ids
826 body = mkLams (drop n_dicts rhs_ids) rhs_body
827 -- Glue back on the non-dict lambdas
829 calls_for_me = case lookupFM calls fn of
831 Just cs -> fmToList cs
833 ----------------------------------------------------------
834 -- Specialise to one particular call pattern
835 spec_call :: (CallKey, ([DictExpr], VarSet)) -- Call instance
836 -> SpecM ((Id,CoreExpr), -- Specialised definition
837 UsageDetails, -- Usage details from specialised body
838 CoreRule) -- Info for the Id's SpecEnv
839 spec_call (CallKey call_ts, (call_ds, _))
840 = ASSERT( call_ts `lengthIs` n_tyvars && call_ds `lengthIs` n_dicts ) do
841 -- Calls are only recorded for properly-saturated applications
843 -- Suppose f's defn is f = /\ a b c d -> \ d1 d2 -> rhs
844 -- Supppose the call is for f [Just t1, Nothing, Just t3, Nothing] [dx1, dx2]
846 -- Construct the new binding
847 -- f1 = SUBST[a->t1,c->t3, d1->d1', d2->d2'] (/\ b d -> rhs)
848 -- PLUS the usage-details
849 -- { d1' = dx1; d2' = dx2 }
850 -- where d1', d2' are cloned versions of d1,d2, with the type substitution applied.
852 -- Note that the substitution is applied to the whole thing.
853 -- This is convenient, but just slightly fragile. Notably:
854 -- * There had better be no name clashes in a/b/c/d
857 -- poly_tyvars = [b,d] in the example above
858 -- spec_tyvars = [a,c]
859 -- ty_args = [t1,b,t3,d]
860 poly_tyvars = [tv | (tv, Nothing) <- rhs_tyvars `zip` call_ts]
861 spec_tyvars = [tv | (tv, Just _) <- rhs_tyvars `zip` call_ts]
862 ty_args = zipWithEqual "spec_call" mk_ty_arg rhs_tyvars call_ts
864 mk_ty_arg rhs_tyvar Nothing = Type (mkTyVarTy rhs_tyvar)
865 mk_ty_arg _ (Just ty) = Type ty
867 spec_ty_args = [ty | Just ty <- call_ts]
868 rhs_subst = extendTvSubstList subst (spec_tyvars `zip` spec_ty_args)
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 rule_name = mkFastString ("SPEC " ++ showSDoc (ppr fn <+> ppr spec_ty_args))
888 spec_env_rule = mkLocalRule
890 inline_prag -- Note [Auto-specialisation and RULES]
892 (poly_tyvars ++ rhs_dicts')
894 (mkVarApps (Var spec_f) app_args)
896 -- Add the { d1' = dx1; d2' = dx2 } usage stuff
897 final_uds = foldr addDictBind rhs_uds (my_zipEqual "spec_call" rhs_dicts' call_ds)
899 spec_pr | inline_rhs = (spec_f `setInlinePragma` inline_prag, Note InlineMe spec_rhs)
900 | otherwise = (spec_f, spec_rhs)
902 return (spec_pr, final_uds, spec_env_rule)
905 my_zipEqual doc xs ys
906 | debugIsOn && not (equalLength xs ys)
907 = pprPanic "my_zipEqual" (vcat
909 , ppr fn <+> ppr call_ts
910 , ppr (idType fn), ppr theta
911 , ppr n_dicts, ppr rhs_dicts
913 | otherwise = zipEqual doc xs ys
916 Note [Auto-specialisation and RULES]
917 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
922 f :: (Int -> Int) -> Int
926 Suppose that auto-specialisation makes a specialised version of
927 g::Int->Int That version won't appear in the LHS of the RULE for f.
928 So if the specialisation rule fires too early, the rule for f may
931 It might be possible to add new rules, to "complete" the rewrite system.
933 RULE forall d. g Int d = g_spec
937 But that's a bit complicated. For now we ask the programmer's help,
938 by *copying the INLINE activation pragma* to the auto-specialised rule.
939 So if g says {-# NOINLINE[2] g #-}, then the auto-spec rule will also
940 not be active until phase 2.
943 Note [Specialisation shape]
944 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
945 We only specialise a function if it has visible top-level lambdas
946 corresponding to its overloading. E.g. if
947 f :: forall a. Eq a => ....
948 then its body must look like
951 Reason: when specialising the body for a call (f ty dexp), we want to
952 substitute dexp for d, and pick up specialised calls in the body of f.
954 This doesn't always work. One example I came across was htis:
955 newtype Gen a = MkGen{ unGen :: Int -> a }
957 choose :: Eq a => a -> Gen a
958 choose n = MkGen (\r -> n)
960 oneof = choose (1::Int)
962 It's a silly exapmle, but we get
963 choose = /\a. g `cast` co
964 where choose doesn't have any dict arguments. Thus far I have not
965 tried to fix this (wait till there's a real example).
968 Note [Inline specialisations]
969 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
970 We transfer to the specialised function any INLINE stuff from the
971 original. This means (a) the Activation in the IdInfo, and (b) any
974 This is a change (Jun06). Previously the idea is that the point of
975 inlining was precisely to specialise the function at its call site,
976 and that's not so important for the specialised copies. But
977 *pragma-directed* specialisation now takes place in the
978 typechecker/desugarer, with manually specified INLINEs. The
979 specialiation here is automatic. It'd be very odd if a function
980 marked INLINE was specialised (because of some local use), and then
981 forever after (including importing modules) the specialised version
982 wasn't INLINEd. After all, the programmer said INLINE!
984 You might wonder why we don't just not specialise INLINE functions.
985 It's because even INLINE functions are sometimes not inlined, when
986 they aren't applied to interesting arguments. But perhaps the type
987 arguments alone are enough to specialise (even though the args are too
988 boring to trigger inlining), and it's certainly better to call the
991 A case in point is dictionary functions, which are current marked
992 INLINE, but which are worth specialising.
995 dropInline :: CoreExpr -> (Bool, CoreExpr)
996 dropInline (Note InlineMe rhs) = (True, rhs)
997 dropInline rhs = (False, rhs)
1000 %************************************************************************
1002 \subsubsection{UsageDetails and suchlike}
1004 %************************************************************************
1009 dict_binds :: !(Bag DictBind),
1010 -- Floated dictionary bindings
1011 -- The order is important;
1012 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
1013 -- (Remember, Bags preserve order in GHC.)
1015 calls :: !CallDetails,
1017 ud_fvs :: !VarSet -- A superset of the variables mentioned in
1018 -- either dict_binds or calls
1021 instance Outputable UsageDetails where
1022 ppr (MkUD { dict_binds = dbs, calls = calls, ud_fvs = fvs })
1023 = ptext (sLit "MkUD") <+> braces (sep (punctuate comma
1024 [ptext (sLit "binds") <+> equals <+> ppr dbs,
1025 ptext (sLit "calls") <+> equals <+> ppr calls,
1026 ptext (sLit "fvs") <+> equals <+> ppr fvs]))
1028 type DictBind = (CoreBind, VarSet)
1029 -- The set is the free vars of the binding
1030 -- both tyvars and dicts
1032 type DictExpr = CoreExpr
1034 emptyUDs :: UsageDetails
1035 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM, ud_fvs = emptyVarSet }
1037 ------------------------------------------------------------
1038 type CallDetails = FiniteMap Id CallInfo
1039 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
1040 type CallInfo = FiniteMap CallKey
1041 ([DictExpr], VarSet) -- Dict args and the vars of the whole
1042 -- call (including tyvars)
1043 -- [*not* include the main id itself, of course]
1044 -- The finite maps eliminate duplicates
1045 -- The list of types and dictionaries is guaranteed to
1046 -- match the type of f
1048 instance Outputable CallKey where
1049 ppr (CallKey ts) = ppr ts
1051 -- Type isn't an instance of Ord, so that we can control which
1052 -- instance we use. That's tiresome here. Oh well
1053 instance Eq CallKey where
1054 k1 == k2 = case k1 `compare` k2 of { EQ -> True; _ -> False }
1056 instance Ord CallKey where
1057 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
1059 cmp Nothing Nothing = EQ
1060 cmp Nothing (Just _) = LT
1061 cmp (Just _) Nothing = GT
1062 cmp (Just t1) (Just t2) = tcCmpType t1 t2
1064 unionCalls :: CallDetails -> CallDetails -> CallDetails
1065 unionCalls c1 c2 = plusFM_C plusFM c1 c2
1067 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> UsageDetails
1068 singleCall id tys dicts
1069 = MkUD {dict_binds = emptyBag,
1070 calls = unitFM id (unitFM (CallKey tys) (dicts, call_fvs)),
1073 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
1074 tys_fvs = tyVarsOfTypes (catMaybes tys)
1075 -- The type args (tys) are guaranteed to be part of the dictionary
1076 -- types, because they are just the constrained types,
1077 -- and the dictionary is therefore sure to be bound
1078 -- inside the binding for any type variables free in the type;
1079 -- hence it's safe to neglect tyvars free in tys when making
1080 -- the free-var set for this call
1081 -- BUT I don't trust this reasoning; play safe and include tys_fvs
1083 -- We don't include the 'id' itself.
1085 mkCallUDs :: Subst -> Id -> [CoreExpr] -> UsageDetails
1086 mkCallUDs subst f args
1088 || not (all isClassPred theta)
1089 -- Only specialise if all overloading is on class params.
1090 -- In ptic, with implicit params, the type args
1091 -- *don't* say what the value of the implicit param is!
1092 || not (spec_tys `lengthIs` n_tyvars)
1093 || not ( dicts `lengthIs` n_dicts)
1094 || maybeToBool (lookupRule (\_act -> True) (substInScope subst) emptyRuleBase f args)
1095 -- There's already a rule covering this call. A typical case
1096 -- is where there's an explicit user-provided rule. Then
1097 -- we don't want to create a specialised version
1098 -- of the function that overlaps.
1099 = emptyUDs -- Not overloaded, or no specialisation wanted
1102 = singleCall f spec_tys dicts
1104 (tyvars, theta, _) = tcSplitSigmaTy (idType f)
1105 constrained_tyvars = tyVarsOfTheta theta
1106 n_tyvars = length tyvars
1107 n_dicts = length theta
1109 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
1110 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
1113 | tyvar `elemVarSet` constrained_tyvars = Just ty
1114 | otherwise = Nothing
1116 ------------------------------------------------------------
1117 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
1118 plusUDs (MkUD {dict_binds = db1, calls = calls1, ud_fvs = fvs1})
1119 (MkUD {dict_binds = db2, calls = calls2, ud_fvs = fvs2})
1120 = MkUD {dict_binds = d, calls = c, ud_fvs = fvs1 `unionVarSet` fvs2}
1122 d = db1 `unionBags` db2
1123 c = calls1 `unionCalls` calls2
1125 plusUDList :: [UsageDetails] -> UsageDetails
1126 plusUDList = foldr plusUDs emptyUDs
1128 -- zapCalls deletes calls to ids from uds
1129 zapCalls :: [Id] -> CallDetails -> CallDetails
1130 zapCalls ids calls = delListFromFM calls ids
1132 mkDB :: CoreBind -> DictBind
1133 mkDB bind = (bind, bind_fvs bind)
1135 bind_fvs :: CoreBind -> VarSet
1136 bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
1137 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1140 rhs_fvs = unionVarSets (map pair_fvs prs)
1142 pair_fvs :: (Id, CoreExpr) -> VarSet
1143 pair_fvs (bndr, rhs) = exprFreeVars rhs `unionVarSet` idFreeVars bndr
1144 -- Don't forget variables mentioned in the
1145 -- rules of the bndr. C.f. OccAnal.addRuleUsage
1146 -- Also tyvars mentioned in its type; they may not appear in the RHS
1150 addDictBind :: (Id,CoreExpr) -> UsageDetails -> UsageDetails
1151 addDictBind (dict,rhs) uds
1152 = uds { dict_binds = db `consBag` dict_binds uds
1153 , ud_fvs = ud_fvs uds `unionVarSet` fvs }
1155 db@(_, fvs) = mkDB (NonRec dict rhs)
1157 dumpAllDictBinds :: UsageDetails -> [CoreBind] -> [CoreBind]
1158 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
1159 = foldrBag add binds dbs
1161 add (bind,_) binds = bind : binds
1163 dumpUDs :: [CoreBndr]
1164 -> UsageDetails -> CoreExpr
1165 -> (UsageDetails, CoreExpr)
1166 dumpUDs bndrs (MkUD { dict_binds = orig_dbs
1167 , calls = orig_calls
1168 , ud_fvs = fvs}) body
1169 = (new_uds, foldrBag add_let body dump_dbs)
1170 -- This may delete fewer variables
1171 -- than in priciple possible
1174 MkUD { dict_binds = free_dbs
1175 , calls = free_calls
1176 , ud_fvs = fvs `minusVarSet` bndr_set}
1178 bndr_set = mkVarSet bndrs
1179 add_let (bind,_) body = Let bind body
1181 (free_dbs, dump_dbs, dump_set)
1182 = foldlBag dump_db (emptyBag, emptyBag, bndr_set) orig_dbs
1183 -- Important that it's foldl not foldr;
1184 -- we're accumulating the set of dumped ids in dump_set
1186 free_calls = filterCalls dump_set orig_calls
1188 dump_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1189 | dump_idset `intersectsVarSet` fvs -- Dump it
1190 = (free_dbs, dump_dbs `snocBag` db,
1191 extendVarSetList dump_idset (bindersOf bind))
1193 | otherwise -- Don't dump it
1194 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1196 filterCalls :: VarSet -> CallDetails -> CallDetails
1197 -- Remove any calls that mention the variables
1198 filterCalls bs calls
1199 = mapFM (\_ cs -> filter_calls cs) $
1200 filterFM (\k _ -> not (k `elemVarSet` bs)) calls
1202 filter_calls :: CallInfo -> CallInfo
1203 filter_calls = filterFM (\_ (_, fvs) -> not (fvs `intersectsVarSet` bs))
1207 %************************************************************************
1209 \subsubsection{Boring helper functions}
1211 %************************************************************************
1214 type SpecM a = UniqSM a
1216 initSM :: UniqSupply -> SpecM a -> a
1219 mapAndCombineSM :: (a -> SpecM (b, UsageDetails)) -> [a] -> SpecM ([b], UsageDetails)
1220 mapAndCombineSM _ [] = return ([], emptyUDs)
1221 mapAndCombineSM f (x:xs) = do (y, uds1) <- f x
1222 (ys, uds2) <- mapAndCombineSM f xs
1223 return (y:ys, uds1 `plusUDs` uds2)
1225 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1226 -- Clone the binders of the bind; return new bind with the cloned binders
1227 -- Return the substitution to use for RHSs, and the one to use for the body
1228 cloneBindSM subst (NonRec bndr rhs) = do
1229 us <- getUniqueSupplyM
1230 let (subst', bndr') = cloneIdBndr subst us bndr
1231 return (subst, subst', NonRec bndr' rhs)
1233 cloneBindSM subst (Rec pairs) = do
1234 us <- getUniqueSupplyM
1235 let (subst', bndrs') = cloneRecIdBndrs subst us (map fst pairs)
1236 return (subst', subst', Rec (bndrs' `zip` map snd pairs))
1238 cloneBinders :: Subst -> [CoreBndr] -> SpecM (Subst, [CoreBndr])
1239 cloneBinders subst bndrs = do
1240 us <- getUniqueSupplyM
1241 return (cloneIdBndrs subst us bndrs)
1243 newIdSM :: Id -> Type -> SpecM Id
1244 newIdSM old_id new_ty = do
1247 -- Give the new Id a similar occurrence name to the old one
1248 name = idName old_id
1249 new_id = mkUserLocal (mkSpecOcc (nameOccName name)) uniq new_ty (getSrcSpan name)
1254 Old (but interesting) stuff about unboxed bindings
1255 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1257 What should we do when a value is specialised to a *strict* unboxed value?
1259 map_*_* f (x:xs) = let h = f x
1263 Could convert let to case:
1265 map_*_Int# f (x:xs) = case f x of h# ->
1269 This may be undesirable since it forces evaluation here, but the value
1270 may not be used in all branches of the body. In the general case this
1271 transformation is impossible since the mutual recursion in a letrec
1272 cannot be expressed as a case.
1274 There is also a problem with top-level unboxed values, since our
1275 implementation cannot handle unboxed values at the top level.
1277 Solution: Lift the binding of the unboxed value and extract it when it
1280 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1285 Now give it to the simplifier and the _Lifting will be optimised away.
1287 The benfit is that we have given the specialised "unboxed" values a
1288 very simplep lifted semantics and then leave it up to the simplifier to
1289 optimise it --- knowing that the overheads will be removed in nearly
1292 In particular, the value will only be evaluted in the branches of the
1293 program which use it, rather than being forced at the point where the
1294 value is bound. For example:
1296 filtermap_*_* p f (x:xs)
1303 filtermap_*_Int# p f (x:xs)
1304 = let h = case (f x) of h# -> _Lift h#
1307 True -> case h of _Lift h#
1311 The binding for h can still be inlined in the one branch and the
1312 _Lifting eliminated.
1315 Question: When won't the _Lifting be eliminated?
1317 Answer: When they at the top-level (where it is necessary) or when
1318 inlining would duplicate work (or possibly code depending on
1319 options). However, the _Lifting will still be eliminated if the
1320 strictness analyser deems the lifted binding strict.