2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
4 \section[Specialise]{Stamping out overloading, and (optionally) polymorphism}
8 -- The above warning supression flag is a temporary kludge.
9 -- While working on this module you are encouraged to remove it and fix
10 -- any warnings in the module. See
11 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
14 module Specialise ( specProgram ) where
16 #include "HsVersions.h"
18 import DynFlags ( DynFlags, DynFlag(..) )
19 import Id ( Id, idName, idType, mkUserLocal,
20 idInlinePragma, setInlinePragma )
21 import TcType ( Type, mkTyVarTy, tcSplitSigmaTy,
22 tyVarsOfTypes, tyVarsOfTheta, isClassPred,
23 tcCmpType, isUnLiftedType
25 import CoreSubst ( Subst, mkEmptySubst, extendTvSubstList, lookupIdSubst,
26 substBndr, substBndrs, substTy, substInScope,
27 cloneIdBndr, cloneIdBndrs, cloneRecIdBndrs
32 import CoreUtils ( applyTypeToArgs, mkPiTypes )
33 import CoreFVs ( exprFreeVars, exprsFreeVars, idFreeVars )
34 import CoreTidy ( tidyRules )
35 import CoreLint ( showPass, endPass )
36 import Rules ( addIdSpecialisations, mkLocalRule, lookupRule, emptyRuleBase, rulesOfBinds )
37 import PprCore ( pprRules )
38 import UniqSupply ( UniqSupply,
43 import MkId ( voidArgId, realWorldPrimId )
45 import Maybes ( catMaybes, maybeToBool )
46 import ErrUtils ( dumpIfSet_dyn )
47 import BasicTypes ( Activation( AlwaysActive ) )
49 import List ( partition )
50 import Util ( zipEqual, zipWithEqual, cmpList, lengthIs,
51 equalLength, lengthAtLeast, notNull )
57 %************************************************************************
59 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
61 %************************************************************************
63 These notes describe how we implement specialisation to eliminate
66 The specialisation pass works on Core
67 syntax, complete with all the explicit dictionary application,
68 abstraction and construction as added by the type checker. The
69 existing type checker remains largely as it is.
71 One important thought: the {\em types} passed to an overloaded
72 function, and the {\em dictionaries} passed are mutually redundant.
73 If the same function is applied to the same type(s) then it is sure to
74 be applied to the same dictionary(s)---or rather to the same {\em
75 values}. (The arguments might look different but they will evaluate
78 Second important thought: we know that we can make progress by
79 treating dictionary arguments as static and worth specialising on. So
80 we can do without binding-time analysis, and instead specialise on
81 dictionary arguments and no others.
90 and suppose f is overloaded.
92 STEP 1: CALL-INSTANCE COLLECTION
94 We traverse <body>, accumulating all applications of f to types and
97 (Might there be partial applications, to just some of its types and
98 dictionaries? In principle yes, but in practice the type checker only
99 builds applications of f to all its types and dictionaries, so partial
100 applications could only arise as a result of transformation, and even
101 then I think it's unlikely. In any case, we simply don't accumulate such
102 partial applications.)
107 So now we have a collection of calls to f:
111 Notice that f may take several type arguments. To avoid ambiguity, we
112 say that f is called at type t1/t2 and t3/t4.
114 We take equivalence classes using equality of the *types* (ignoring
115 the dictionary args, which as mentioned previously are redundant).
117 STEP 3: SPECIALISATION
119 For each equivalence class, choose a representative (f t1 t2 d1 d2),
120 and create a local instance of f, defined thus:
122 f@t1/t2 = <f_rhs> t1 t2 d1 d2
124 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
125 of simplification will now result. However we don't actually *do* that
126 simplification. Rather, we leave it for the simplifier to do. If we
127 *did* do it, though, we'd get more call instances from the specialised
128 RHS. We can work out what they are by instantiating the call-instance
129 set from f's RHS with the types t1, t2.
131 Add this new id to f's IdInfo, to record that f has a specialised version.
133 Before doing any of this, check that f's IdInfo doesn't already
134 tell us about an existing instance of f at the required type/s.
135 (This might happen if specialisation was applied more than once, or
136 it might arise from user SPECIALIZE pragmas.)
140 Wait a minute! What if f is recursive? Then we can't just plug in
141 its right-hand side, can we?
143 But it's ok. The type checker *always* creates non-recursive definitions
144 for overloaded recursive functions. For example:
146 f x = f (x+x) -- Yes I know its silly
150 f a (d::Num a) = let p = +.sel a d
152 letrec fl (y::a) = fl (p y y)
156 We still have recusion for non-overloaded functions which we
157 speciailise, but the recursive call should get specialised to the
158 same recursive version.
164 All this is crystal clear when the function is applied to *constant
165 types*; that is, types which have no type variables inside. But what if
166 it is applied to non-constant types? Suppose we find a call of f at type
167 t1/t2. There are two possibilities:
169 (a) The free type variables of t1, t2 are in scope at the definition point
170 of f. In this case there's no problem, we proceed just as before. A common
171 example is as follows. Here's the Haskell:
176 After typechecking we have
178 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
179 in +.sel a d (f a d y) (f a d y)
181 Notice that the call to f is at type type "a"; a non-constant type.
182 Both calls to f are at the same type, so we can specialise to give:
184 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
185 in +.sel a d (f@a y) (f@a y)
188 (b) The other case is when the type variables in the instance types
189 are *not* in scope at the definition point of f. The example we are
190 working with above is a good case. There are two instances of (+.sel a d),
191 but "a" is not in scope at the definition of +.sel. Can we do anything?
192 Yes, we can "common them up", a sort of limited common sub-expression deal.
195 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
196 f@a (x::a) = +.sel@a x x
197 in +.sel@a (f@a y) (f@a y)
199 This can save work, and can't be spotted by the type checker, because
200 the two instances of +.sel weren't originally at the same type.
204 * There are quite a few variations here. For example, the defn of
205 +.sel could be floated ouside the \y, to attempt to gain laziness.
206 It certainly mustn't be floated outside the \d because the d has to
209 * We don't want to inline f_rhs in this case, because
210 that will duplicate code. Just commoning up the call is the point.
212 * Nothing gets added to +.sel's IdInfo.
214 * Don't bother unless the equivalence class has more than one item!
216 Not clear whether this is all worth it. It is of course OK to
217 simply discard call-instances when passing a big lambda.
219 Polymorphism 2 -- Overloading
221 Consider a function whose most general type is
223 f :: forall a b. Ord a => [a] -> b -> b
225 There is really no point in making a version of g at Int/Int and another
226 at Int/Bool, because it's only instancing the type variable "a" which
227 buys us any efficiency. Since g is completely polymorphic in b there
228 ain't much point in making separate versions of g for the different
231 That suggests that we should identify which of g's type variables
232 are constrained (like "a") and which are unconstrained (like "b").
233 Then when taking equivalence classes in STEP 2, we ignore the type args
234 corresponding to unconstrained type variable. In STEP 3 we make
235 polymorphic versions. Thus:
237 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
246 f a (d::Num a) = let g = ...
248 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
250 Here, g is only called at one type, but the dictionary isn't in scope at the
251 definition point for g. Usually the type checker would build a
252 definition for d1 which enclosed g, but the transformation system
253 might have moved d1's defn inward. Solution: float dictionary bindings
254 outwards along with call instances.
258 f x = let g p q = p==q
264 Before specialisation, leaving out type abstractions we have
266 f df x = let g :: Eq a => a -> a -> Bool
268 h :: Num a => a -> a -> (a, Bool)
269 h dh r s = let deq = eqFromNum dh
270 in (+ dh r s, g deq r s)
274 After specialising h we get a specialised version of h, like this:
276 h' r s = let deq = eqFromNum df
277 in (+ df r s, g deq r s)
279 But we can't naively make an instance for g from this, because deq is not in scope
280 at the defn of g. Instead, we have to float out the (new) defn of deq
281 to widen its scope. Notice that this floating can't be done in advance -- it only
282 shows up when specialisation is done.
284 User SPECIALIZE pragmas
285 ~~~~~~~~~~~~~~~~~~~~~~~
286 Specialisation pragmas can be digested by the type checker, and implemented
287 by adding extra definitions along with that of f, in the same way as before
289 f@t1/t2 = <f_rhs> t1 t2 d1 d2
291 Indeed the pragmas *have* to be dealt with by the type checker, because
292 only it knows how to build the dictionaries d1 and d2! For example
294 g :: Ord a => [a] -> [a]
295 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
297 Here, the specialised version of g is an application of g's rhs to the
298 Ord dictionary for (Tree Int), which only the type checker can conjure
299 up. There might not even *be* one, if (Tree Int) is not an instance of
300 Ord! (All the other specialision has suitable dictionaries to hand
303 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
304 it is buried in a complex (as-yet-un-desugared) binding group.
307 f@t1/t2 = f* t1 t2 d1 d2
309 where f* is the Id f with an IdInfo which says "inline me regardless!".
310 Indeed all the specialisation could be done in this way.
311 That in turn means that the simplifier has to be prepared to inline absolutely
312 any in-scope let-bound thing.
315 Again, the pragma should permit polymorphism in unconstrained variables:
317 h :: Ord a => [a] -> b -> b
318 {-# SPECIALIZE h :: [Int] -> b -> b #-}
320 We *insist* that all overloaded type variables are specialised to ground types,
321 (and hence there can be no context inside a SPECIALIZE pragma).
322 We *permit* unconstrained type variables to be specialised to
324 - or left as a polymorphic type variable
325 but nothing in between. So
327 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
329 is *illegal*. (It can be handled, but it adds complication, and gains the
333 SPECIALISING INSTANCE DECLARATIONS
334 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
337 instance Foo a => Foo [a] where
339 {-# SPECIALIZE instance Foo [Int] #-}
341 The original instance decl creates a dictionary-function
344 dfun.Foo.List :: forall a. Foo a -> Foo [a]
346 The SPECIALIZE pragma just makes a specialised copy, just as for
347 ordinary function definitions:
349 dfun.Foo.List@Int :: Foo [Int]
350 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
352 The information about what instance of the dfun exist gets added to
353 the dfun's IdInfo in the same way as a user-defined function too.
356 Automatic instance decl specialisation?
357 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
358 Can instance decls be specialised automatically? It's tricky.
359 We could collect call-instance information for each dfun, but
360 then when we specialised their bodies we'd get new call-instances
361 for ordinary functions; and when we specialised their bodies, we might get
362 new call-instances of the dfuns, and so on. This all arises because of
363 the unrestricted mutual recursion between instance decls and value decls.
365 Still, there's no actual problem; it just means that we may not do all
366 the specialisation we could theoretically do.
368 Furthermore, instance decls are usually exported and used non-locally,
369 so we'll want to compile enough to get those specialisations done.
371 Lastly, there's no such thing as a local instance decl, so we can
372 survive solely by spitting out *usage* information, and then reading that
373 back in as a pragma when next compiling the file. So for now,
374 we only specialise instance decls in response to pragmas.
377 SPITTING OUT USAGE INFORMATION
378 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
380 To spit out usage information we need to traverse the code collecting
381 call-instance information for all imported (non-prelude?) functions
382 and data types. Then we equivalence-class it and spit it out.
384 This is done at the top-level when all the call instances which escape
385 must be for imported functions and data types.
387 *** Not currently done ***
390 Partial specialisation by pragmas
391 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
392 What about partial specialisation:
394 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
395 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
399 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
401 Seems quite reasonable. Similar things could be done with instance decls:
403 instance (Foo a, Foo b) => Foo (a,b) where
405 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
406 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
408 Ho hum. Things are complex enough without this. I pass.
411 Requirements for the simplifer
412 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
413 The simplifier has to be able to take advantage of the specialisation.
415 * When the simplifier finds an application of a polymorphic f, it looks in
416 f's IdInfo in case there is a suitable instance to call instead. This converts
418 f t1 t2 d1 d2 ===> f_t1_t2
420 Note that the dictionaries get eaten up too!
422 * Dictionary selection operations on constant dictionaries must be
425 +.sel Int d ===> +Int
427 The obvious way to do this is in the same way as other specialised
428 calls: +.sel has inside it some IdInfo which tells that if it's applied
429 to the type Int then it should eat a dictionary and transform to +Int.
431 In short, dictionary selectors need IdInfo inside them for constant
434 * Exactly the same applies if a superclass dictionary is being
437 Eq.sel Int d ===> dEqInt
439 * Something similar applies to dictionary construction too. Suppose
440 dfun.Eq.List is the function taking a dictionary for (Eq a) to
441 one for (Eq [a]). Then we want
443 dfun.Eq.List Int d ===> dEq.List_Int
445 Where does the Eq [Int] dictionary come from? It is built in
446 response to a SPECIALIZE pragma on the Eq [a] instance decl.
448 In short, dfun Ids need IdInfo with a specialisation for each
449 constant instance of their instance declaration.
451 All this uses a single mechanism: the SpecEnv inside an Id
454 What does the specialisation IdInfo look like?
455 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
457 The SpecEnv of an Id maps a list of types (the template) to an expression
461 For example, if f has this SpecInfo:
463 [Int, a] -> \d:Ord Int. f' a
465 it means that we can replace the call
467 f Int t ===> (\d. f' t)
469 This chucks one dictionary away and proceeds with the
470 specialised version of f, namely f'.
473 What can't be done this way?
474 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
475 There is no way, post-typechecker, to get a dictionary for (say)
476 Eq a from a dictionary for Eq [a]. So if we find
480 we can't transform to
485 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
487 Of course, we currently have no way to automatically derive
488 eqList, nor to connect it to the Eq [a] instance decl, but you
489 can imagine that it might somehow be possible. Taking advantage
490 of this is permanently ruled out.
492 Still, this is no great hardship, because we intend to eliminate
493 overloading altogether anyway!
497 A note about non-tyvar dictionaries
498 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
499 Some Ids have types like
501 forall a,b,c. Eq a -> Ord [a] -> tau
503 This seems curious at first, because we usually only have dictionary
504 args whose types are of the form (C a) where a is a type variable.
505 But this doesn't hold for the functions arising from instance decls,
506 which sometimes get arguements with types of form (C (T a)) for some
509 Should we specialise wrt this compound-type dictionary? We used to say
511 "This is a heuristic judgement, as indeed is the fact that we
512 specialise wrt only dictionaries. We choose *not* to specialise
513 wrt compound dictionaries because at the moment the only place
514 they show up is in instance decls, where they are simply plugged
515 into a returned dictionary. So nothing is gained by specialising
518 But it is simpler and more uniform to specialise wrt these dicts too;
519 and in future GHC is likely to support full fledged type signatures
521 f ;: Eq [(a,b)] => ...
524 %************************************************************************
526 \subsubsection{The new specialiser}
528 %************************************************************************
530 Our basic game plan is this. For let(rec) bound function
531 f :: (C a, D c) => (a,b,c,d) -> Bool
533 * Find any specialised calls of f, (f ts ds), where
534 ts are the type arguments t1 .. t4, and
535 ds are the dictionary arguments d1 .. d2.
537 * Add a new definition for f1 (say):
539 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
541 Note that we abstract over the unconstrained type arguments.
545 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
547 to the specialisations of f. This will be used by the
548 simplifier to replace calls
549 (f t1 t2 t3 t4) da db
551 (\d1 d1 -> f1 t2 t4) da db
553 All the stuff about how many dictionaries to discard, and what types
554 to apply the specialised function to, are handled by the fact that the
555 SpecEnv contains a template for the result of the specialisation.
557 We don't build *partial* specialisations for f. For example:
559 f :: Eq a => a -> a -> Bool
560 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
562 Here, little is gained by making a specialised copy of f.
563 There's a distinct danger that the specialised version would
564 first build a dictionary for (Eq b, Eq c), and then select the (==)
565 method from it! Even if it didn't, not a great deal is saved.
567 We do, however, generate polymorphic, but not overloaded, specialisations:
569 f :: Eq a => [a] -> b -> b -> b
570 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
572 Hence, the invariant is this:
574 *** no specialised version is overloaded ***
577 %************************************************************************
579 \subsubsection{The exported function}
581 %************************************************************************
584 specProgram :: DynFlags -> UniqSupply -> [CoreBind] -> IO [CoreBind]
585 specProgram dflags us binds = do
587 showPass dflags "Specialise"
589 let binds' = initSM us (do (binds', uds') <- go binds
590 return (dumpAllDictBinds uds' binds'))
592 endPass dflags "Specialise" Opt_D_dump_spec binds'
594 dumpIfSet_dyn dflags Opt_D_dump_rules "Top-level specialisations"
595 (pprRules (tidyRules emptyTidyEnv (rulesOfBinds binds')))
599 -- We need to start with a Subst that knows all the things
600 -- that are in scope, so that the substitution engine doesn't
601 -- accidentally re-use a unique that's already in use
602 -- Easiest thing is to do it all at once, as if all the top-level
603 -- decls were mutually recursive
604 top_subst = mkEmptySubst (mkInScopeSet (mkVarSet (bindersOfBinds binds)))
606 go [] = return ([], emptyUDs)
607 go (bind:binds) = do (binds', uds) <- go binds
608 (bind', uds') <- specBind top_subst bind uds
609 return (bind' ++ binds', uds')
612 %************************************************************************
614 \subsubsection{@specExpr@: the main function}
616 %************************************************************************
619 specVar :: Subst -> Id -> CoreExpr
620 specVar subst v = lookupIdSubst subst v
622 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
623 -- We carry a substitution down:
624 -- a) we must clone any binding that might flaot outwards,
625 -- to avoid name clashes
626 -- b) we carry a type substitution to use when analysing
627 -- the RHS of specialised bindings (no type-let!)
629 ---------------- First the easy cases --------------------
630 specExpr subst (Type ty) = return (Type (substTy subst ty), emptyUDs)
631 specExpr subst (Var v) = return (specVar subst v, emptyUDs)
632 specExpr subst (Lit lit) = return (Lit lit, emptyUDs)
633 specExpr subst (Cast e co) = do
634 (e', uds) <- specExpr subst e
635 return ((Cast e' (substTy subst co)), uds)
636 specExpr subst (Note note body) = do
637 (body', uds) <- specExpr subst body
638 return (Note (specNote subst note) body', uds)
641 ---------------- Applications might generate a call instance --------------------
642 specExpr subst expr@(App fun arg)
645 go (App fun arg) args = do (arg', uds_arg) <- specExpr subst arg
646 (fun', uds_app) <- go fun (arg':args)
647 return (App fun' arg', uds_arg `plusUDs` uds_app)
649 go (Var f) args = case specVar subst f of
650 Var f' -> return (Var f', mkCallUDs subst f' args)
651 e' -> return (e', emptyUDs) -- I don't expect this!
652 go other args = specExpr subst other
654 ---------------- Lambda/case require dumping of usage details --------------------
655 specExpr subst e@(Lam _ _) = do
656 (body', uds) <- specExpr subst' body
657 let (filtered_uds, body'') = dumpUDs bndrs' uds body'
658 return (mkLams bndrs' body'', filtered_uds)
660 (bndrs, body) = collectBinders e
661 (subst', bndrs') = substBndrs subst bndrs
662 -- More efficient to collect a group of binders together all at once
663 -- and we don't want to split a lambda group with dumped bindings
665 specExpr subst (Case scrut case_bndr ty alts) = do
666 (scrut', uds_scrut) <- specExpr subst scrut
667 (alts', uds_alts) <- mapAndCombineSM spec_alt alts
668 return (Case scrut' case_bndr' (substTy subst ty) alts', uds_scrut `plusUDs` uds_alts)
670 (subst_alt, case_bndr') = substBndr subst case_bndr
671 -- No need to clone case binder; it can't float like a let(rec)
673 spec_alt (con, args, rhs) = do
674 (rhs', uds) <- specExpr subst_rhs rhs
675 let (uds', rhs'') = do dumpUDs args uds rhs'
676 return ((con, args', rhs''), uds')
678 (subst_rhs, args') = substBndrs subst_alt args
680 ---------------- Finally, let is the interesting case --------------------
681 specExpr subst (Let bind body) = do
683 (rhs_subst, body_subst, bind') <- cloneBindSM subst bind
685 -- Deal with the body
686 (body', body_uds) <- specExpr body_subst body
688 -- Deal with the bindings
689 (binds', uds) <- specBind rhs_subst bind' body_uds
692 return (foldr Let body' binds', uds)
694 -- Must apply the type substitution to coerceions
695 specNote subst note = note
698 %************************************************************************
700 \subsubsection{Dealing with a binding}
702 %************************************************************************
705 specBind :: Subst -- Use this for RHSs
707 -> UsageDetails -- Info on how the scope of the binding
708 -> SpecM ([CoreBind], -- New bindings
709 UsageDetails) -- And info to pass upstream
711 specBind rhs_subst bind body_uds = do
712 (bind', bind_uds) <- specBindItself rhs_subst bind (calls body_uds)
714 bndrs = bindersOf bind
715 all_uds = zapCalls bndrs (body_uds `plusUDs` bind_uds)
716 -- It's important that the `plusUDs` is this way round,
717 -- because body_uds may bind dictionaries that are
718 -- used in the calls passed to specDefn. So the
719 -- dictionary bindings in bind_uds may mention
720 -- dictionaries bound in body_uds.
721 case splitUDs bndrs all_uds of
723 (_, ([],[])) -- This binding doesn't bind anything needed
724 -- in the UDs, so put the binding here
725 -- This is the case for most non-dict bindings, except
726 -- for the few that are mentioned in a dict binding
727 -- that is floating upwards in body_uds
728 -> return ([bind'], all_uds)
730 (float_uds, (dict_binds, calls)) -- This binding is needed in the UDs, so float it out
731 -> return ([], float_uds `plusUDs` mkBigUD bind' dict_binds calls)
734 -- A truly gruesome function
735 mkBigUD bind@(NonRec _ _) dbs calls
736 = -- Common case: non-recursive and no specialisations
737 -- (if there were any specialistions it would have been made recursive)
738 MkUD { dict_binds = listToBag (mkDB bind : dbs),
739 calls = listToCallDetails calls }
741 mkBigUD bind dbs calls
743 MkUD { dict_binds = unitBag (mkDB (Rec (bind_prs bind ++ dbsToPairs dbs))),
745 calls = listToCallDetails calls }
747 bind_prs (NonRec b r) = [(b,r)]
748 bind_prs (Rec prs) = prs
751 dbsToPairs ((bind,_):dbs) = bind_prs bind ++ dbsToPairs dbs
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 rhs_bndrs = rhs_tyvars ++ rhs_dicts
827 body = mkLams (drop n_dicts rhs_ids) rhs_body
828 -- Glue back on the non-dict lambdas
830 calls_for_me = case lookupFM calls fn of
832 Just cs -> fmToList cs
834 ----------------------------------------------------------
835 -- Specialise to one particular call pattern
836 spec_call :: (CallKey, ([DictExpr], VarSet)) -- Call instance
837 -> SpecM ((Id,CoreExpr), -- Specialised definition
838 UsageDetails, -- Usage details from specialised body
839 CoreRule) -- Info for the Id's SpecEnv
840 spec_call (CallKey call_ts, (call_ds, call_fvs))
841 = ASSERT( call_ts `lengthIs` n_tyvars && call_ds `lengthIs` n_dicts ) do
842 -- Calls are only recorded for properly-saturated applications
844 -- Suppose f's defn is f = /\ a b c d -> \ d1 d2 -> rhs
845 -- Supppose the call is for f [Just t1, Nothing, Just t3, Nothing] [dx1, dx2]
847 -- Construct the new binding
848 -- f1 = SUBST[a->t1,c->t3, d1->d1', d2->d2'] (/\ b d -> rhs)
849 -- PLUS the usage-details
850 -- { d1' = dx1; d2' = dx2 }
851 -- where d1', d2' are cloned versions of d1,d2, with the type substitution applied.
853 -- Note that the substitution is applied to the whole thing.
854 -- This is convenient, but just slightly fragile. Notably:
855 -- * There had better be no name clashes in a/b/c/d
858 -- poly_tyvars = [b,d] in the example above
859 -- spec_tyvars = [a,c]
860 -- ty_args = [t1,b,t3,d]
861 poly_tyvars = [tv | (tv, Nothing) <- rhs_tyvars `zip` call_ts]
862 spec_tyvars = [tv | (tv, Just _) <- rhs_tyvars `zip` call_ts]
863 ty_args = zipWithEqual "spec_call" mk_ty_arg rhs_tyvars call_ts
865 mk_ty_arg rhs_tyvar Nothing = Type (mkTyVarTy rhs_tyvar)
866 mk_ty_arg rhs_tyvar (Just ty) = Type ty
867 rhs_subst = extendTvSubstList subst (spec_tyvars `zip` [ty | Just ty <- call_ts])
869 (rhs_subst', rhs_dicts') <- cloneBinders rhs_subst rhs_dicts
871 inst_args = ty_args ++ map Var rhs_dicts'
873 -- Figure out the type of the specialised function
874 body_ty = applyTypeToArgs rhs fn_type inst_args
875 (lam_args, app_args) -- Add a dummy argument if body_ty is unlifted
876 | isUnLiftedType body_ty -- C.f. WwLib.mkWorkerArgs
877 = (poly_tyvars ++ [voidArgId], poly_tyvars ++ [realWorldPrimId])
878 | otherwise = (poly_tyvars, poly_tyvars)
879 spec_id_ty = mkPiTypes lam_args body_ty
881 spec_f <- newIdSM fn spec_id_ty
882 (spec_rhs, rhs_uds) <- specExpr rhs_subst' (mkLams lam_args body)
884 -- The rule to put in the function's specialisation is:
885 -- forall b,d, d1',d2'. f t1 b t3 d d1' d2' = f1 b d
886 spec_env_rule = mkLocalRule (mkFastString ("SPEC " ++ showSDoc (ppr fn)))
887 inline_prag -- Note [Auto-specialisation and RULES]
889 (poly_tyvars ++ rhs_dicts')
891 (mkVarApps (Var spec_f) app_args)
893 -- Add the { d1' = dx1; d2' = dx2 } usage stuff
894 final_uds = foldr addDictBind rhs_uds (my_zipEqual "spec_call" rhs_dicts' call_ds)
896 spec_pr | inline_rhs = (spec_f `setInlinePragma` inline_prag, Note InlineMe spec_rhs)
897 | otherwise = (spec_f, spec_rhs)
899 return (spec_pr, final_uds, spec_env_rule)
902 my_zipEqual doc xs ys
904 | not (equalLength xs ys) = pprPanic "my_zipEqual" (vcat
906 , ppr fn <+> ppr call_ts
907 , ppr (idType fn), ppr theta
908 , 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
1016 type DictBind = (CoreBind, VarSet)
1017 -- The set is the free vars of the binding
1018 -- both tyvars and dicts
1020 type DictExpr = CoreExpr
1022 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM }
1024 type ProtoUsageDetails = ([DictBind],
1025 [(Id, CallKey, ([DictExpr], VarSet))]
1028 ------------------------------------------------------------
1029 type CallDetails = FiniteMap Id CallInfo
1030 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
1031 type CallInfo = FiniteMap CallKey
1032 ([DictExpr], VarSet) -- Dict args and the vars of the whole
1033 -- call (including tyvars)
1034 -- [*not* include the main id itself, of course]
1035 -- The finite maps eliminate duplicates
1036 -- The list of types and dictionaries is guaranteed to
1037 -- match the type of f
1039 -- Type isn't an instance of Ord, so that we can control which
1040 -- instance we use. That's tiresome here. Oh well
1041 instance Eq CallKey where
1042 k1 == k2 = case k1 `compare` k2 of { EQ -> True; other -> False }
1044 instance Ord CallKey where
1045 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
1047 cmp Nothing Nothing = EQ
1048 cmp Nothing (Just t2) = LT
1049 cmp (Just t1) Nothing = GT
1050 cmp (Just t1) (Just t2) = tcCmpType t1 t2
1052 unionCalls :: CallDetails -> CallDetails -> CallDetails
1053 unionCalls c1 c2 = plusFM_C plusFM c1 c2
1055 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> CallDetails
1056 singleCall id tys dicts
1057 = unitFM id (unitFM (CallKey tys) (dicts, call_fvs))
1059 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
1060 tys_fvs = tyVarsOfTypes (catMaybes tys)
1061 -- The type args (tys) are guaranteed to be part of the dictionary
1062 -- types, because they are just the constrained types,
1063 -- and the dictionary is therefore sure to be bound
1064 -- inside the binding for any type variables free in the type;
1065 -- hence it's safe to neglect tyvars free in tys when making
1066 -- the free-var set for this call
1067 -- BUT I don't trust this reasoning; play safe and include tys_fvs
1069 -- We don't include the 'id' itself.
1071 listToCallDetails calls
1072 = foldr (unionCalls . mk_call) emptyFM calls
1074 mk_call (id, tys, dicts_w_fvs) = unitFM id (unitFM tys dicts_w_fvs)
1075 -- NB: the free vars of the call are provided
1077 callDetailsToList calls = [ (id,tys,dicts)
1078 | (id,fm) <- fmToList calls,
1079 (tys, dicts) <- fmToList fm
1082 mkCallUDs subst f args
1084 || not (all isClassPred theta)
1085 -- Only specialise if all overloading is on class params.
1086 -- In ptic, with implicit params, the type args
1087 -- *don't* say what the value of the implicit param is!
1088 || not (spec_tys `lengthIs` n_tyvars)
1089 || not ( dicts `lengthIs` n_dicts)
1090 || maybeToBool (lookupRule (\act -> True) (substInScope subst) emptyRuleBase f args)
1091 -- There's already a rule covering this call. A typical case
1092 -- is where there's an explicit user-provided rule. Then
1093 -- we don't want to create a specialised version
1094 -- of the function that overlaps.
1095 = emptyUDs -- Not overloaded, or no specialisation wanted
1098 = MkUD {dict_binds = emptyBag,
1099 calls = 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})
1117 (MkUD {dict_binds = db2, calls = calls2})
1118 = MkUD {dict_binds = d, calls = c}
1120 d = db1 `unionBags` db2
1121 c = calls1 `unionCalls` calls2
1123 plusUDList = foldr plusUDs emptyUDs
1125 -- zapCalls deletes calls to ids from uds
1126 zapCalls ids uds = uds {calls = delListFromFM (calls uds) ids}
1128 mkDB bind = (bind, bind_fvs bind)
1130 bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
1131 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1134 rhs_fvs = unionVarSets (map pair_fvs prs)
1136 pair_fvs (bndr, rhs) = exprFreeVars rhs `unionVarSet` idFreeVars bndr
1137 -- Don't forget variables mentioned in the
1138 -- rules of the bndr. C.f. OccAnal.addRuleUsage
1139 -- Also tyvars mentioned in its type; they may not appear in the RHS
1143 addDictBind (dict,rhs) uds = uds { dict_binds = mkDB (NonRec dict rhs) `consBag` dict_binds uds }
1145 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
1146 = foldrBag add binds dbs
1148 add (bind,_) binds = bind : binds
1150 dumpUDs :: [CoreBndr]
1151 -> UsageDetails -> CoreExpr
1152 -> (UsageDetails, CoreExpr)
1153 dumpUDs bndrs uds body
1154 = (free_uds, foldr add_let body dict_binds)
1156 (free_uds, (dict_binds, _)) = splitUDs bndrs uds
1157 add_let (bind,_) body = Let bind body
1159 splitUDs :: [CoreBndr]
1161 -> (UsageDetails, -- These don't mention the binders
1162 ProtoUsageDetails) -- These do
1164 splitUDs bndrs uds@(MkUD {dict_binds = orig_dbs,
1165 calls = orig_calls})
1167 = if isEmptyBag dump_dbs && null dump_calls then
1168 -- Common case: binder doesn't affect floats
1172 -- Binders bind some of the fvs of the floats
1173 (MkUD {dict_binds = free_dbs,
1174 calls = listToCallDetails free_calls},
1175 (bagToList dump_dbs, dump_calls)
1179 bndr_set = mkVarSet bndrs
1181 (free_dbs, dump_dbs, dump_idset)
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 -- Filter out any calls that mention things that are being dumped
1187 orig_call_list = callDetailsToList orig_calls
1188 (dump_calls, free_calls) = partition captured orig_call_list
1189 captured (id,tys,(dicts, fvs)) = fvs `intersectsVarSet` dump_idset
1190 || id `elemVarSet` dump_idset
1192 dump_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1193 | dump_idset `intersectsVarSet` fvs -- Dump it
1194 = (free_dbs, dump_dbs `snocBag` db,
1195 extendVarSetList dump_idset (bindersOf bind))
1197 | otherwise -- Don't dump it
1198 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1202 %************************************************************************
1204 \subsubsection{Boring helper functions}
1206 %************************************************************************
1209 type SpecM a = UniqSM a
1213 mapAndCombineSM f [] = return ([], emptyUDs)
1214 mapAndCombineSM f (x:xs) = do (y, uds1) <- f x
1215 (ys, uds2) <- mapAndCombineSM f xs
1216 return (y:ys, uds1 `plusUDs` uds2)
1218 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1219 -- Clone the binders of the bind; return new bind with the cloned binders
1220 -- Return the substitution to use for RHSs, and the one to use for the body
1221 cloneBindSM subst (NonRec bndr rhs) = do
1222 us <- getUniqueSupplyM
1223 let (subst', bndr') = do cloneIdBndr subst us bndr
1224 return (subst, subst', NonRec bndr' rhs)
1226 cloneBindSM subst (Rec pairs) = do
1227 us <- getUniqueSupplyM
1228 let (subst', bndrs') = cloneRecIdBndrs subst us (map fst pairs)
1229 return (subst', subst', Rec (bndrs' `zip` map snd pairs))
1231 cloneBinders subst bndrs = do
1232 us <- getUniqueSupplyM
1233 return (cloneIdBndrs subst us bndrs)
1235 newIdSM old_id new_ty = do
1238 -- Give the new Id a similar occurrence name to the old one
1239 name = idName old_id
1240 new_id = mkUserLocal (mkSpecOcc (nameOccName name)) uniq new_ty (getSrcSpan name)
1245 Old (but interesting) stuff about unboxed bindings
1246 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1248 What should we do when a value is specialised to a *strict* unboxed value?
1250 map_*_* f (x:xs) = let h = f x
1254 Could convert let to case:
1256 map_*_Int# f (x:xs) = case f x of h# ->
1260 This may be undesirable since it forces evaluation here, but the value
1261 may not be used in all branches of the body. In the general case this
1262 transformation is impossible since the mutual recursion in a letrec
1263 cannot be expressed as a case.
1265 There is also a problem with top-level unboxed values, since our
1266 implementation cannot handle unboxed values at the top level.
1268 Solution: Lift the binding of the unboxed value and extract it when it
1271 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1276 Now give it to the simplifier and the _Lifting will be optimised away.
1278 The benfit is that we have given the specialised "unboxed" values a
1279 very simplep lifted semantics and then leave it up to the simplifier to
1280 optimise it --- knowing that the overheads will be removed in nearly
1283 In particular, the value will only be evaluted in the branches of the
1284 program which use it, rather than being forced at the point where the
1285 value is bound. For example:
1287 filtermap_*_* p f (x:xs)
1294 filtermap_*_Int# p f (x:xs)
1295 = let h = case (f x) of h# -> _Lift h#
1298 True -> case h of _Lift h#
1302 The binding for h can still be inlined in the one branch and the
1303 _Lifting eliminated.
1306 Question: When won't the _Lifting be eliminated?
1308 Answer: When they at the top-level (where it is necessary) or when
1309 inlining would duplicate work (or possibly code depending on
1310 options). However, the _Lifting will still be eliminated if the
1311 strictness analyser deems the lifted binding strict.