2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
4 \section[Specialise]{Stamping out overloading, and (optionally) polymorphism}
7 module Specialise ( specProgram ) where
9 #include "HsVersions.h"
20 import CoreUtils ( exprIsTrivial, applyTypeToArgs, mkPiTypes )
21 import CoreFVs ( exprFreeVars, exprsFreeVars, idFreeVars )
22 import UniqSupply ( UniqSM, initUs_, MonadUnique(..) )
24 import MkId ( voidArgId, realWorldPrimId )
25 import Maybes ( catMaybes, isJust )
34 import qualified Data.Map as Map
35 import qualified FiniteMap as Map
38 %************************************************************************
40 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
42 %************************************************************************
44 These notes describe how we implement specialisation to eliminate
47 The specialisation pass works on Core
48 syntax, complete with all the explicit dictionary application,
49 abstraction and construction as added by the type checker. The
50 existing type checker remains largely as it is.
52 One important thought: the {\em types} passed to an overloaded
53 function, and the {\em dictionaries} passed are mutually redundant.
54 If the same function is applied to the same type(s) then it is sure to
55 be applied to the same dictionary(s)---or rather to the same {\em
56 values}. (The arguments might look different but they will evaluate
59 Second important thought: we know that we can make progress by
60 treating dictionary arguments as static and worth specialising on. So
61 we can do without binding-time analysis, and instead specialise on
62 dictionary arguments and no others.
71 and suppose f is overloaded.
73 STEP 1: CALL-INSTANCE COLLECTION
75 We traverse <body>, accumulating all applications of f to types and
78 (Might there be partial applications, to just some of its types and
79 dictionaries? In principle yes, but in practice the type checker only
80 builds applications of f to all its types and dictionaries, so partial
81 applications could only arise as a result of transformation, and even
82 then I think it's unlikely. In any case, we simply don't accumulate such
83 partial applications.)
88 So now we have a collection of calls to f:
92 Notice that f may take several type arguments. To avoid ambiguity, we
93 say that f is called at type t1/t2 and t3/t4.
95 We take equivalence classes using equality of the *types* (ignoring
96 the dictionary args, which as mentioned previously are redundant).
98 STEP 3: SPECIALISATION
100 For each equivalence class, choose a representative (f t1 t2 d1 d2),
101 and create a local instance of f, defined thus:
103 f@t1/t2 = <f_rhs> t1 t2 d1 d2
105 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
106 of simplification will now result. However we don't actually *do* that
107 simplification. Rather, we leave it for the simplifier to do. If we
108 *did* do it, though, we'd get more call instances from the specialised
109 RHS. We can work out what they are by instantiating the call-instance
110 set from f's RHS with the types t1, t2.
112 Add this new id to f's IdInfo, to record that f has a specialised version.
114 Before doing any of this, check that f's IdInfo doesn't already
115 tell us about an existing instance of f at the required type/s.
116 (This might happen if specialisation was applied more than once, or
117 it might arise from user SPECIALIZE pragmas.)
121 Wait a minute! What if f is recursive? Then we can't just plug in
122 its right-hand side, can we?
124 But it's ok. The type checker *always* creates non-recursive definitions
125 for overloaded recursive functions. For example:
127 f x = f (x+x) -- Yes I know its silly
131 f a (d::Num a) = let p = +.sel a d
133 letrec fl (y::a) = fl (p y y)
137 We still have recusion for non-overloaded functions which we
138 speciailise, but the recursive call should get specialised to the
139 same recursive version.
145 All this is crystal clear when the function is applied to *constant
146 types*; that is, types which have no type variables inside. But what if
147 it is applied to non-constant types? Suppose we find a call of f at type
148 t1/t2. There are two possibilities:
150 (a) The free type variables of t1, t2 are in scope at the definition point
151 of f. In this case there's no problem, we proceed just as before. A common
152 example is as follows. Here's the Haskell:
157 After typechecking we have
159 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
160 in +.sel a d (f a d y) (f a d y)
162 Notice that the call to f is at type type "a"; a non-constant type.
163 Both calls to f are at the same type, so we can specialise to give:
165 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
166 in +.sel a d (f@a y) (f@a y)
169 (b) The other case is when the type variables in the instance types
170 are *not* in scope at the definition point of f. The example we are
171 working with above is a good case. There are two instances of (+.sel a d),
172 but "a" is not in scope at the definition of +.sel. Can we do anything?
173 Yes, we can "common them up", a sort of limited common sub-expression deal.
176 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
177 f@a (x::a) = +.sel@a x x
178 in +.sel@a (f@a y) (f@a y)
180 This can save work, and can't be spotted by the type checker, because
181 the two instances of +.sel weren't originally at the same type.
185 * There are quite a few variations here. For example, the defn of
186 +.sel could be floated ouside the \y, to attempt to gain laziness.
187 It certainly mustn't be floated outside the \d because the d has to
190 * We don't want to inline f_rhs in this case, because
191 that will duplicate code. Just commoning up the call is the point.
193 * Nothing gets added to +.sel's IdInfo.
195 * Don't bother unless the equivalence class has more than one item!
197 Not clear whether this is all worth it. It is of course OK to
198 simply discard call-instances when passing a big lambda.
200 Polymorphism 2 -- Overloading
202 Consider a function whose most general type is
204 f :: forall a b. Ord a => [a] -> b -> b
206 There is really no point in making a version of g at Int/Int and another
207 at Int/Bool, because it's only instancing the type variable "a" which
208 buys us any efficiency. Since g is completely polymorphic in b there
209 ain't much point in making separate versions of g for the different
212 That suggests that we should identify which of g's type variables
213 are constrained (like "a") and which are unconstrained (like "b").
214 Then when taking equivalence classes in STEP 2, we ignore the type args
215 corresponding to unconstrained type variable. In STEP 3 we make
216 polymorphic versions. Thus:
218 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
227 f a (d::Num a) = let g = ...
229 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
231 Here, g is only called at one type, but the dictionary isn't in scope at the
232 definition point for g. Usually the type checker would build a
233 definition for d1 which enclosed g, but the transformation system
234 might have moved d1's defn inward. Solution: float dictionary bindings
235 outwards along with call instances.
239 f x = let g p q = p==q
245 Before specialisation, leaving out type abstractions we have
247 f df x = let g :: Eq a => a -> a -> Bool
249 h :: Num a => a -> a -> (a, Bool)
250 h dh r s = let deq = eqFromNum dh
251 in (+ dh r s, g deq r s)
255 After specialising h we get a specialised version of h, like this:
257 h' r s = let deq = eqFromNum df
258 in (+ df r s, g deq r s)
260 But we can't naively make an instance for g from this, because deq is not in scope
261 at the defn of g. Instead, we have to float out the (new) defn of deq
262 to widen its scope. Notice that this floating can't be done in advance -- it only
263 shows up when specialisation is done.
265 User SPECIALIZE pragmas
266 ~~~~~~~~~~~~~~~~~~~~~~~
267 Specialisation pragmas can be digested by the type checker, and implemented
268 by adding extra definitions along with that of f, in the same way as before
270 f@t1/t2 = <f_rhs> t1 t2 d1 d2
272 Indeed the pragmas *have* to be dealt with by the type checker, because
273 only it knows how to build the dictionaries d1 and d2! For example
275 g :: Ord a => [a] -> [a]
276 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
278 Here, the specialised version of g is an application of g's rhs to the
279 Ord dictionary for (Tree Int), which only the type checker can conjure
280 up. There might not even *be* one, if (Tree Int) is not an instance of
281 Ord! (All the other specialision has suitable dictionaries to hand
284 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
285 it is buried in a complex (as-yet-un-desugared) binding group.
288 f@t1/t2 = f* t1 t2 d1 d2
290 where f* is the Id f with an IdInfo which says "inline me regardless!".
291 Indeed all the specialisation could be done in this way.
292 That in turn means that the simplifier has to be prepared to inline absolutely
293 any in-scope let-bound thing.
296 Again, the pragma should permit polymorphism in unconstrained variables:
298 h :: Ord a => [a] -> b -> b
299 {-# SPECIALIZE h :: [Int] -> b -> b #-}
301 We *insist* that all overloaded type variables are specialised to ground types,
302 (and hence there can be no context inside a SPECIALIZE pragma).
303 We *permit* unconstrained type variables to be specialised to
305 - or left as a polymorphic type variable
306 but nothing in between. So
308 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
310 is *illegal*. (It can be handled, but it adds complication, and gains the
314 SPECIALISING INSTANCE DECLARATIONS
315 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
318 instance Foo a => Foo [a] where
320 {-# SPECIALIZE instance Foo [Int] #-}
322 The original instance decl creates a dictionary-function
325 dfun.Foo.List :: forall a. Foo a -> Foo [a]
327 The SPECIALIZE pragma just makes a specialised copy, just as for
328 ordinary function definitions:
330 dfun.Foo.List@Int :: Foo [Int]
331 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
333 The information about what instance of the dfun exist gets added to
334 the dfun's IdInfo in the same way as a user-defined function too.
337 Automatic instance decl specialisation?
338 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
339 Can instance decls be specialised automatically? It's tricky.
340 We could collect call-instance information for each dfun, but
341 then when we specialised their bodies we'd get new call-instances
342 for ordinary functions; and when we specialised their bodies, we might get
343 new call-instances of the dfuns, and so on. This all arises because of
344 the unrestricted mutual recursion between instance decls and value decls.
346 Still, there's no actual problem; it just means that we may not do all
347 the specialisation we could theoretically do.
349 Furthermore, instance decls are usually exported and used non-locally,
350 so we'll want to compile enough to get those specialisations done.
352 Lastly, there's no such thing as a local instance decl, so we can
353 survive solely by spitting out *usage* information, and then reading that
354 back in as a pragma when next compiling the file. So for now,
355 we only specialise instance decls in response to pragmas.
358 SPITTING OUT USAGE INFORMATION
359 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
361 To spit out usage information we need to traverse the code collecting
362 call-instance information for all imported (non-prelude?) functions
363 and data types. Then we equivalence-class it and spit it out.
365 This is done at the top-level when all the call instances which escape
366 must be for imported functions and data types.
368 *** Not currently done ***
371 Partial specialisation by pragmas
372 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
373 What about partial specialisation:
375 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
376 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
380 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
382 Seems quite reasonable. Similar things could be done with instance decls:
384 instance (Foo a, Foo b) => Foo (a,b) where
386 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
387 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
389 Ho hum. Things are complex enough without this. I pass.
392 Requirements for the simplifer
393 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
394 The simplifier has to be able to take advantage of the specialisation.
396 * When the simplifier finds an application of a polymorphic f, it looks in
397 f's IdInfo in case there is a suitable instance to call instead. This converts
399 f t1 t2 d1 d2 ===> f_t1_t2
401 Note that the dictionaries get eaten up too!
403 * Dictionary selection operations on constant dictionaries must be
406 +.sel Int d ===> +Int
408 The obvious way to do this is in the same way as other specialised
409 calls: +.sel has inside it some IdInfo which tells that if it's applied
410 to the type Int then it should eat a dictionary and transform to +Int.
412 In short, dictionary selectors need IdInfo inside them for constant
415 * Exactly the same applies if a superclass dictionary is being
418 Eq.sel Int d ===> dEqInt
420 * Something similar applies to dictionary construction too. Suppose
421 dfun.Eq.List is the function taking a dictionary for (Eq a) to
422 one for (Eq [a]). Then we want
424 dfun.Eq.List Int d ===> dEq.List_Int
426 Where does the Eq [Int] dictionary come from? It is built in
427 response to a SPECIALIZE pragma on the Eq [a] instance decl.
429 In short, dfun Ids need IdInfo with a specialisation for each
430 constant instance of their instance declaration.
432 All this uses a single mechanism: the SpecEnv inside an Id
435 What does the specialisation IdInfo look like?
436 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
438 The SpecEnv of an Id maps a list of types (the template) to an expression
442 For example, if f has this SpecInfo:
444 [Int, a] -> \d:Ord Int. f' a
446 it means that we can replace the call
448 f Int t ===> (\d. f' t)
450 This chucks one dictionary away and proceeds with the
451 specialised version of f, namely f'.
454 What can't be done this way?
455 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
456 There is no way, post-typechecker, to get a dictionary for (say)
457 Eq a from a dictionary for Eq [a]. So if we find
461 we can't transform to
466 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
468 Of course, we currently have no way to automatically derive
469 eqList, nor to connect it to the Eq [a] instance decl, but you
470 can imagine that it might somehow be possible. Taking advantage
471 of this is permanently ruled out.
473 Still, this is no great hardship, because we intend to eliminate
474 overloading altogether anyway!
476 A note about non-tyvar dictionaries
477 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
478 Some Ids have types like
480 forall a,b,c. Eq a -> Ord [a] -> tau
482 This seems curious at first, because we usually only have dictionary
483 args whose types are of the form (C a) where a is a type variable.
484 But this doesn't hold for the functions arising from instance decls,
485 which sometimes get arguements with types of form (C (T a)) for some
488 Should we specialise wrt this compound-type dictionary? We used to say
490 "This is a heuristic judgement, as indeed is the fact that we
491 specialise wrt only dictionaries. We choose *not* to specialise
492 wrt compound dictionaries because at the moment the only place
493 they show up is in instance decls, where they are simply plugged
494 into a returned dictionary. So nothing is gained by specialising
497 But it is simpler and more uniform to specialise wrt these dicts too;
498 and in future GHC is likely to support full fledged type signatures
500 f :: Eq [(a,b)] => ...
503 %************************************************************************
505 \subsubsection{The new specialiser}
507 %************************************************************************
509 Our basic game plan is this. For let(rec) bound function
510 f :: (C a, D c) => (a,b,c,d) -> Bool
512 * Find any specialised calls of f, (f ts ds), where
513 ts are the type arguments t1 .. t4, and
514 ds are the dictionary arguments d1 .. d2.
516 * Add a new definition for f1 (say):
518 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
520 Note that we abstract over the unconstrained type arguments.
524 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
526 to the specialisations of f. This will be used by the
527 simplifier to replace calls
528 (f t1 t2 t3 t4) da db
530 (\d1 d1 -> f1 t2 t4) da db
532 All the stuff about how many dictionaries to discard, and what types
533 to apply the specialised function to, are handled by the fact that the
534 SpecEnv contains a template for the result of the specialisation.
536 We don't build *partial* specialisations for f. For example:
538 f :: Eq a => a -> a -> Bool
539 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
541 Here, little is gained by making a specialised copy of f.
542 There's a distinct danger that the specialised version would
543 first build a dictionary for (Eq b, Eq c), and then select the (==)
544 method from it! Even if it didn't, not a great deal is saved.
546 We do, however, generate polymorphic, but not overloaded, specialisations:
548 f :: Eq a => [a] -> b -> b -> b
549 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
551 Hence, the invariant is this:
553 *** no specialised version is overloaded ***
556 %************************************************************************
558 \subsubsection{The exported function}
560 %************************************************************************
563 specProgram :: ModGuts -> CoreM ModGuts
565 = do { hpt_rules <- getRuleBase
566 ; let local_rules = mg_rules guts
567 rule_base = extendRuleBaseList hpt_rules (mg_rules guts)
569 -- Specialise the bindings of this module
570 ; (binds', uds) <- runSpecM (go (mg_binds guts))
572 -- Specialise imported functions
573 ; (new_rules, spec_binds) <- specImports emptyVarSet rule_base uds
575 ; return (guts { mg_binds = spec_binds ++ binds'
576 , mg_rules = local_rules ++ new_rules }) }
578 -- We need to start with a Subst that knows all the things
579 -- that are in scope, so that the substitution engine doesn't
580 -- accidentally re-use a unique that's already in use
581 -- Easiest thing is to do it all at once, as if all the top-level
582 -- decls were mutually recursive
583 top_subst = mkEmptySubst $ mkInScopeSet $ mkVarSet $
584 bindersOfBinds $ mg_binds guts
586 go [] = return ([], emptyUDs)
587 go (bind:binds) = do (binds', uds) <- go binds
588 (bind', uds') <- specBind top_subst bind uds
589 return (bind' ++ binds', uds')
591 specImports :: VarSet -- Don't specialise these ones
592 -- See Note [Avoiding recursive specialisation]
593 -> RuleBase -- Rules from this module and the home package
594 -- (but not external packages, which can change)
595 -> UsageDetails -- Calls for imported things, and floating bindings
596 -> CoreM ( [CoreRule] -- New rules
597 , [CoreBind] ) -- Specialised bindings and floating bindings
598 specImports done rb uds
599 = do { let import_calls = varEnvElts (ud_calls uds)
600 ; (rules, spec_binds) <- go rb import_calls
601 ; return (rules, wrapDictBinds (ud_binds uds) spec_binds) }
603 go _ [] = return ([], [])
604 go rb (CIS fn calls_for_fn : other_calls)
605 = do { (rules1, spec_binds1) <- specImport done rb fn (Map.toList calls_for_fn)
606 ; (rules2, spec_binds2) <- go (extendRuleBaseList rb rules1) other_calls
607 ; return (rules1 ++ rules2, spec_binds1 ++ spec_binds2) }
609 specImport :: VarSet -- Don't specialise these
610 -- See Note [Avoiding recursive specialisation]
611 -> RuleBase -- Rules from this module
612 -> Id -> [CallInfo] -- Imported function and calls for it
613 -> CoreM ( [CoreRule] -- New rules
614 , [CoreBind] ) -- Specialised bindings
615 specImport done rb fn calls_for_fn
616 | not (fn `elemVarSet` done)
617 , isInlinablePragma (idInlinePragma fn)
618 , Just rhs <- maybeUnfoldingTemplate (realIdUnfolding fn)
619 = do { -- Get rules from the external package state
620 -- We keep doing this in case we "page-fault in"
621 -- more rules as we go along
622 ; hsc_env <- getHscEnv
623 ; eps <- liftIO $ hscEPS hsc_env
624 ; let full_rb = unionRuleBase rb (eps_rule_base eps)
625 rules_for_fn = getRules full_rb fn
627 ; (rules1, spec_pairs, uds) <- runSpecM $
628 specCalls emptySubst rules_for_fn calls_for_fn fn rhs
629 ; let spec_binds1 = [NonRec b r | (b,r) <- spec_pairs]
630 -- After the rules kick in we may get recursion, but
631 -- we rely on a global GlomBinds to sort that out later
633 -- Now specialise any cascaded calls
634 ; (rules2, spec_binds2) <- specImports (extendVarSet done fn)
635 (extendRuleBaseList rb rules1)
638 ; return (rules2 ++ rules1, spec_binds2 ++ spec_binds1) }
641 = WARN( True, ptext (sLit "specImport discard") <+> ppr fn <+> ppr calls_for_fn )
645 Avoiding recursive specialisation
646 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
647 When we specialise 'f' we may find new overloaded calls to 'g', 'h' in
648 'f's RHS. So we want to specialise g,h. But we don't want to
649 specialise f any more! It's possible that f's RHS might have a
650 recursive yet-more-specialised call, so we'd diverge in that case.
651 And if the call is to the same type, one specialisation is enough.
652 Avoiding this recursive specialisation loop is the reason for the
653 'done' VarSet passed to specImports and specImport.
655 %************************************************************************
657 \subsubsection{@specExpr@: the main function}
659 %************************************************************************
662 specVar :: Subst -> Id -> CoreExpr
663 specVar subst v = lookupIdSubst (text "specVar") subst v
665 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
666 -- We carry a substitution down:
667 -- a) we must clone any binding that might float outwards,
668 -- to avoid name clashes
669 -- b) we carry a type substitution to use when analysing
670 -- the RHS of specialised bindings (no type-let!)
672 ---------------- First the easy cases --------------------
673 specExpr subst (Type ty) = return (Type (CoreSubst.substTy subst ty), emptyUDs)
674 specExpr subst (Var v) = return (specVar subst v, emptyUDs)
675 specExpr _ (Lit lit) = return (Lit lit, emptyUDs)
676 specExpr subst (Cast e co) = do
677 (e', uds) <- specExpr subst e
678 return ((Cast e' (CoreSubst.substTy subst co)), uds)
679 specExpr subst (Note note body) = do
680 (body', uds) <- specExpr subst body
681 return (Note (specNote subst note) body', uds)
684 ---------------- Applications might generate a call instance --------------------
685 specExpr subst expr@(App {})
688 go (App fun arg) args = do (arg', uds_arg) <- specExpr subst arg
689 (fun', uds_app) <- go fun (arg':args)
690 return (App fun' arg', uds_arg `plusUDs` uds_app)
692 go (Var f) args = case specVar subst f of
693 Var f' -> return (Var f', mkCallUDs f' args)
694 e' -> return (e', emptyUDs) -- I don't expect this!
695 go other _ = specExpr subst other
697 ---------------- Lambda/case require dumping of usage details --------------------
698 specExpr subst e@(Lam _ _) = do
699 (body', uds) <- specExpr subst' body
700 let (free_uds, dumped_dbs) = dumpUDs bndrs' uds
701 return (mkLams bndrs' (wrapDictBindsE dumped_dbs body'), free_uds)
703 (bndrs, body) = collectBinders e
704 (subst', bndrs') = substBndrs subst bndrs
705 -- More efficient to collect a group of binders together all at once
706 -- and we don't want to split a lambda group with dumped bindings
708 specExpr subst (Case scrut case_bndr ty alts)
709 = do { (scrut', scrut_uds) <- specExpr subst scrut
710 ; (scrut'', case_bndr', alts', alts_uds)
711 <- specCase subst scrut' case_bndr alts
712 ; return (Case scrut'' case_bndr' (CoreSubst.substTy subst ty) alts'
713 , scrut_uds `plusUDs` alts_uds) }
715 ---------------- Finally, let is the interesting case --------------------
716 specExpr subst (Let bind body) = do
718 (rhs_subst, body_subst, bind') <- cloneBindSM subst bind
720 -- Deal with the body
721 (body', body_uds) <- specExpr body_subst body
723 -- Deal with the bindings
724 (binds', uds) <- specBind rhs_subst bind' body_uds
727 return (foldr Let body' binds', uds)
729 -- Must apply the type substitution to coerceions
730 specNote :: Subst -> Note -> Note
731 specNote _ note = note
735 -> CoreExpr -- Scrutinee, already done
737 -> SpecM ( CoreExpr -- New scrutinee
741 specCase subst scrut' case_bndr [(con, args, rhs)]
742 | isDictId case_bndr -- See Note [Floating dictionaries out of cases]
743 , interestingDict scrut'
744 , not (isDeadBinder case_bndr && null sc_args')
745 = do { (case_bndr_flt : sc_args_flt) <- mapM clone_me (case_bndr' : sc_args')
747 ; let sc_rhss = [ Case (Var case_bndr_flt) case_bndr' (idType sc_arg')
748 [(con, args', Var sc_arg')]
749 | sc_arg' <- sc_args' ]
751 -- Extend the substitution for RHS to map the *original* binders
752 -- to their floated verions. Attach an unfolding to these floated
753 -- binders so they look interesting to interestingDict
754 mb_sc_flts :: [Maybe DictId]
755 mb_sc_flts = map (lookupVarEnv clone_env) args'
756 clone_env = zipVarEnv sc_args' (zipWith add_unf sc_args_flt sc_rhss)
757 subst_prs = (case_bndr, Var (add_unf case_bndr_flt scrut'))
758 : [ (arg, Var sc_flt)
759 | (arg, Just sc_flt) <- args `zip` mb_sc_flts ]
760 subst_rhs' = extendIdSubstList subst_rhs subst_prs
762 ; (rhs', rhs_uds) <- specExpr subst_rhs' rhs
763 ; let scrut_bind = mkDB (NonRec case_bndr_flt scrut')
764 case_bndr_set = unitVarSet case_bndr_flt
765 sc_binds = [(NonRec sc_arg_flt sc_rhs, case_bndr_set)
766 | (sc_arg_flt, sc_rhs) <- sc_args_flt `zip` sc_rhss ]
767 flt_binds = scrut_bind : sc_binds
768 (free_uds, dumped_dbs) = dumpUDs (case_bndr':args') rhs_uds
769 all_uds = flt_binds `addDictBinds` free_uds
770 alt' = (con, args', wrapDictBindsE dumped_dbs rhs')
771 ; return (Var case_bndr_flt, case_bndr', [alt'], all_uds) }
773 (subst_rhs, (case_bndr':args')) = substBndrs subst (case_bndr:args)
774 sc_args' = filter is_flt_sc_arg args'
776 clone_me bndr = do { uniq <- getUniqueM
777 ; return (mkUserLocal occ uniq ty loc) }
781 occ = nameOccName name
782 loc = getSrcSpan name
784 add_unf sc_flt sc_rhs -- Sole purpose: make sc_flt respond True to interestingDictId
785 = setIdUnfolding sc_flt (mkSimpleUnfolding sc_rhs)
787 arg_set = mkVarSet args'
788 is_flt_sc_arg var = isId var
789 && not (isDeadBinder var)
791 && not (tyVarsOfType var_ty `intersectsVarSet` arg_set)
796 specCase subst scrut case_bndr alts
797 = do { (alts', uds_alts) <- mapAndCombineSM spec_alt alts
798 ; return (scrut, case_bndr', alts', uds_alts) }
800 (subst_alt, case_bndr') = substBndr subst case_bndr
801 spec_alt (con, args, rhs) = do
802 (rhs', uds) <- specExpr subst_rhs rhs
803 let (free_uds, dumped_dbs) = dumpUDs (case_bndr' : args') uds
804 return ((con, args', wrapDictBindsE dumped_dbs rhs'), free_uds)
806 (subst_rhs, args') = substBndrs subst_alt args
809 Note [Floating dictionaries out of cases]
810 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
812 g = \d. case d of { MkD sc ... -> ...(f sc)... }
813 Naively we can't float d2's binding out of the case expression,
814 because 'sc' is bound by the case, and that in turn means we can't
815 specialise f, which seems a pity.
817 So we invert the case, by floating out a binding
819 sc_flt = case d of { MkD sc ... -> sc }
820 Now we can float the call instance for 'f'. Indeed this is just
821 what'll happen if 'sc' was originally bound with a let binding,
822 but case is more efficient, and necessary with equalities. So it's
823 good to work with both.
825 You might think that this won't make any difference, because the
826 call instance will only get nuked by the \d. BUT if 'g' itself is
827 specialised, then transitively we should be able to specialise f.
830 case e of cb { MkD sc ... -> ...(f sc)... }
833 sc_flt = case cb_flt of { MkD sc ... -> sc }
835 case cb_flt of bg { MkD sc ... -> ....(f sc_flt)... }
837 The "_flt" things are the floated binds; we use the current substitution
838 to substitute sc -> sc_flt in the RHS
840 %************************************************************************
842 Dealing with a binding
844 %************************************************************************
847 specBind :: Subst -- Use this for RHSs
849 -> UsageDetails -- Info on how the scope of the binding
850 -> SpecM ([CoreBind], -- New bindings
851 UsageDetails) -- And info to pass upstream
853 -- Returned UsageDetails:
854 -- No calls for binders of this bind
855 specBind rhs_subst (NonRec fn rhs) body_uds
856 = do { (rhs', rhs_uds) <- specExpr rhs_subst rhs
857 ; (fn', spec_defns, body_uds1) <- specDefn rhs_subst body_uds fn rhs
859 ; let pairs = spec_defns ++ [(fn', rhs')]
860 -- fn' mentions the spec_defns in its rules,
861 -- so put the latter first
863 combined_uds = body_uds1 `plusUDs` rhs_uds
864 -- This way round a call in rhs_uds of a function f
865 -- at type T will override a call of f at T in body_uds1; and
866 -- that is good because it'll tend to keep "earlier" calls
867 -- See Note [Specialisation of dictionary functions]
869 (free_uds, dump_dbs, float_all) = dumpBindUDs [fn] combined_uds
870 -- See Note [From non-recursive to recursive]
872 final_binds | isEmptyBag dump_dbs = [NonRec b r | (b,r) <- pairs]
873 | otherwise = [Rec (flattenDictBinds dump_dbs pairs)]
876 -- Rather than discard the calls mentioning the bound variables
877 -- we float this binding along with the others
878 return ([], free_uds `snocDictBinds` final_binds)
880 -- No call in final_uds mentions bound variables,
881 -- so we can just leave the binding here
882 return (final_binds, free_uds) }
885 specBind rhs_subst (Rec pairs) body_uds
886 -- Note [Specialising a recursive group]
887 = do { let (bndrs,rhss) = unzip pairs
888 ; (rhss', rhs_uds) <- mapAndCombineSM (specExpr rhs_subst) rhss
889 ; let scope_uds = body_uds `plusUDs` rhs_uds
890 -- Includes binds and calls arising from rhss
892 ; (bndrs1, spec_defns1, uds1) <- specDefns rhs_subst scope_uds pairs
894 ; (bndrs3, spec_defns3, uds3)
895 <- if null spec_defns1 -- Common case: no specialisation
896 then return (bndrs1, [], uds1)
897 else do { -- Specialisation occurred; do it again
898 (bndrs2, spec_defns2, uds2)
899 <- specDefns rhs_subst uds1 (bndrs1 `zip` rhss)
900 ; return (bndrs2, spec_defns2 ++ spec_defns1, uds2) }
902 ; let (final_uds, dumped_dbs, float_all) = dumpBindUDs bndrs uds3
903 bind = Rec (flattenDictBinds dumped_dbs $
904 spec_defns3 ++ zip bndrs3 rhss')
907 return ([], final_uds `snocDictBind` bind)
909 return ([bind], final_uds) }
912 ---------------------------
914 -> UsageDetails -- Info on how it is used in its scope
915 -> [(Id,CoreExpr)] -- The things being bound and their un-processed RHS
916 -> SpecM ([Id], -- Original Ids with RULES added
917 [(Id,CoreExpr)], -- Extra, specialised bindings
918 UsageDetails) -- Stuff to fling upwards from the specialised versions
920 -- Specialise a list of bindings (the contents of a Rec), but flowing usages
921 -- upwards binding by binding. Example: { f = ...g ...; g = ...f .... }
922 -- Then if the input CallDetails has a specialised call for 'g', whose specialisation
923 -- in turn generates a specialised call for 'f', we catch that in this one sweep.
924 -- But not vice versa (it's a fixpoint problem).
926 specDefns _subst uds []
927 = return ([], [], uds)
928 specDefns subst uds ((bndr,rhs):pairs)
929 = do { (bndrs1, spec_defns1, uds1) <- specDefns subst uds pairs
930 ; (bndr1, spec_defns2, uds2) <- specDefn subst uds1 bndr rhs
931 ; return (bndr1 : bndrs1, spec_defns1 ++ spec_defns2, uds2) }
933 ---------------------------
935 -> UsageDetails -- Info on how it is used in its scope
936 -> Id -> CoreExpr -- The thing being bound and its un-processed RHS
937 -> SpecM (Id, -- Original Id with added RULES
938 [(Id,CoreExpr)], -- Extra, specialised bindings
939 UsageDetails) -- Stuff to fling upwards from the specialised versions
941 specDefn subst body_uds fn rhs
942 = do { let (body_uds_without_me, calls_for_me) = callsForMe fn body_uds
943 rules_for_me = idCoreRules fn
944 ; (rules, spec_defns, spec_uds) <- specCalls subst rules_for_me
946 ; return ( fn `addIdSpecialisations` rules
948 , body_uds_without_me `plusUDs` spec_uds) }
949 -- It's important that the `plusUDs` is this way
950 -- round, because body_uds_without_me may bind
951 -- dictionaries that are used in calls_for_me passed
952 -- to specDefn. So the dictionary bindings in
953 -- spec_uds may mention dictionaries bound in
954 -- body_uds_without_me
956 ---------------------------
958 -> [CoreRule] -- Existing RULES for the fn
961 -> SpecM ([CoreRule], -- New RULES for the fn
962 [(Id,CoreExpr)], -- Extra, specialised bindings
963 UsageDetails) -- New usage details from the specialised RHSs
965 -- This function checks existing rules, and does not create
966 -- duplicate ones. So the caller does not nneed to do this filtering.
967 -- See 'already_covered'
969 specCalls subst rules_for_me calls_for_me fn rhs
970 -- The first case is the interesting one
971 | rhs_tyvars `lengthIs` n_tyvars -- Rhs of fn's defn has right number of big lambdas
972 && rhs_ids `lengthAtLeast` n_dicts -- and enough dict args
973 && notNull calls_for_me -- And there are some calls to specialise
974 && not (isNeverActive (idInlineActivation fn))
975 -- Don't specialise NOINLINE things
976 -- See Note [Auto-specialisation and RULES]
978 -- && not (certainlyWillInline (idUnfolding fn)) -- And it's not small
979 -- See Note [Inline specialisation] for why we do not
980 -- switch off specialisation for inline functions
982 = -- pprTrace "specDefn: some" (ppr fn $$ ppr calls_for_me $$ ppr rules_for_me) $
983 do { stuff <- mapM spec_call calls_for_me
984 ; let (spec_defns, spec_uds, spec_rules) = unzip3 (catMaybes stuff)
985 ; return (spec_rules, spec_defns, plusUDList spec_uds) }
987 | otherwise -- No calls or RHS doesn't fit our preconceptions
988 = WARN( notNull calls_for_me, ptext (sLit "Missed specialisation opportunity for") <+> ppr fn )
989 -- Note [Specialisation shape]
990 -- pprTrace "specDefn: none" (ppr fn $$ ppr calls_for_me) $
991 return ([], [], emptyUDs)
995 fn_arity = idArity fn
996 fn_unf = realIdUnfolding fn -- Ignore loop-breaker-ness here
997 (tyvars, theta, _) = tcSplitSigmaTy fn_type
998 n_tyvars = length tyvars
999 n_dicts = length theta
1000 inl_prag = idInlinePragma fn
1001 inl_act = inlinePragmaActivation inl_prag
1002 is_local = isLocalId fn
1004 -- Figure out whether the function has an INLINE pragma
1005 -- See Note [Inline specialisations]
1007 spec_arity = unfoldingArity fn_unf - n_dicts -- Arity of the *specialised* inline rule
1009 (rhs_tyvars, rhs_ids, rhs_body) = collectTyAndValBinders rhs
1011 rhs_dict_ids = take n_dicts rhs_ids
1012 body = mkLams (drop n_dicts rhs_ids) rhs_body
1013 -- Glue back on the non-dict lambdas
1015 already_covered :: [CoreExpr] -> Bool
1016 already_covered args -- Note [Specialisations already covered]
1017 = isJust (lookupRule (const True) realIdUnfolding
1018 (substInScope subst)
1019 fn args rules_for_me)
1021 mk_ty_args :: [Maybe Type] -> [CoreExpr]
1022 mk_ty_args call_ts = zipWithEqual "spec_call" mk_ty_arg rhs_tyvars call_ts
1024 mk_ty_arg rhs_tyvar Nothing = Type (mkTyVarTy rhs_tyvar)
1025 mk_ty_arg _ (Just ty) = Type ty
1027 ----------------------------------------------------------
1028 -- Specialise to one particular call pattern
1029 spec_call :: CallInfo -- Call instance
1030 -> SpecM (Maybe ((Id,CoreExpr), -- Specialised definition
1031 UsageDetails, -- Usage details from specialised body
1032 CoreRule)) -- Info for the Id's SpecEnv
1033 spec_call (CallKey call_ts, (call_ds, _))
1034 = ASSERT( call_ts `lengthIs` n_tyvars && call_ds `lengthIs` n_dicts )
1036 -- Suppose f's defn is f = /\ a b c -> \ d1 d2 -> rhs
1037 -- Supppose the call is for f [Just t1, Nothing, Just t3] [dx1, dx2]
1039 -- Construct the new binding
1040 -- f1 = SUBST[a->t1,c->t3, d1->d1', d2->d2'] (/\ b -> rhs)
1041 -- PLUS the usage-details
1042 -- { d1' = dx1; d2' = dx2 }
1043 -- where d1', d2' are cloned versions of d1,d2, with the type substitution
1044 -- applied. These auxiliary bindings just avoid duplication of dx1, dx2
1046 -- Note that the substitution is applied to the whole thing.
1047 -- This is convenient, but just slightly fragile. Notably:
1048 -- * There had better be no name clashes in a/b/c
1050 -- poly_tyvars = [b] in the example above
1051 -- spec_tyvars = [a,c]
1052 -- ty_args = [t1,b,t3]
1053 poly_tyvars = [tv | (tv, Nothing) <- rhs_tyvars `zip` call_ts]
1054 spec_tv_binds = [(tv,ty) | (tv, Just ty) <- rhs_tyvars `zip` call_ts]
1055 spec_ty_args = map snd spec_tv_binds
1056 ty_args = mk_ty_args call_ts
1057 rhs_subst = CoreSubst.extendTvSubstList subst spec_tv_binds
1059 ; (rhs_subst1, inst_dict_ids) <- newDictBndrs rhs_subst rhs_dict_ids
1060 -- Clone rhs_dicts, including instantiating their types
1062 ; let (rhs_subst2, dx_binds) = bindAuxiliaryDicts rhs_subst1 $
1063 (my_zipEqual rhs_dict_ids inst_dict_ids call_ds)
1064 inst_args = ty_args ++ map Var inst_dict_ids
1066 ; if already_covered inst_args then
1069 { -- Figure out the type of the specialised function
1070 let body_ty = applyTypeToArgs rhs fn_type inst_args
1071 (lam_args, app_args) -- Add a dummy argument if body_ty is unlifted
1072 | isUnLiftedType body_ty -- C.f. WwLib.mkWorkerArgs
1073 = (poly_tyvars ++ [voidArgId], poly_tyvars ++ [realWorldPrimId])
1074 | otherwise = (poly_tyvars, poly_tyvars)
1075 spec_id_ty = mkPiTypes lam_args body_ty
1077 ; spec_f <- newSpecIdSM fn spec_id_ty
1078 ; (spec_rhs, rhs_uds) <- specExpr rhs_subst2 (mkLams lam_args body)
1080 -- The rule to put in the function's specialisation is:
1081 -- forall b, d1',d2'. f t1 b t3 d1' d2' = f1 b
1082 rule_name = mkFastString ("SPEC " ++ showSDoc (ppr fn <+> ppr spec_ty_args))
1083 spec_env_rule = mkRule True {- Auto generated -} is_local
1085 inl_act -- Note [Auto-specialisation and RULES]
1087 (poly_tyvars ++ inst_dict_ids)
1089 (mkVarApps (Var spec_f) app_args)
1091 -- Add the { d1' = dx1; d2' = dx2 } usage stuff
1092 final_uds = foldr consDictBind rhs_uds dx_binds
1094 -- Add an InlineRule if the parent has one
1095 -- See Note [Inline specialisations]
1097 = case inlinePragmaSpec inl_prag of
1098 Inline -> mkInlineUnfolding (Just spec_arity) spec_rhs
1099 Inlinable -> mkInlinableUnfolding spec_rhs
1102 -- Adding arity information just propagates it a bit faster
1103 -- See Note [Arity decrease] in Simplify
1104 -- Copy InlinePragma information from the parent Id.
1105 -- So if f has INLINE[1] so does spec_f
1106 spec_f_w_arity = spec_f `setIdArity` max 0 (fn_arity - n_dicts)
1107 `setInlinePragma` inl_prag
1108 `setIdUnfolding` spec_unf
1110 ; return (Just ((spec_f_w_arity, spec_rhs), final_uds, spec_env_rule)) } }
1112 my_zipEqual xs ys zs
1113 | debugIsOn && not (equalLength xs ys && equalLength ys zs)
1114 = pprPanic "my_zipEqual" (vcat [ ppr xs, ppr ys
1115 , ppr fn <+> ppr call_ts
1116 , ppr (idType fn), ppr theta
1117 , ppr n_dicts, ppr rhs_dict_ids
1119 | otherwise = zip3 xs ys zs
1123 -> [(DictId,DictId,CoreExpr)] -- (orig_dict, inst_dict, dx)
1124 -> (Subst, -- Substitute for all orig_dicts
1125 [CoreBind]) -- Auxiliary bindings
1126 -- Bind any dictionary arguments to fresh names, to preserve sharing
1127 -- Substitution already substitutes orig_dict -> inst_dict
1128 bindAuxiliaryDicts subst triples = go subst [] triples
1130 go subst binds [] = (subst, binds)
1131 go subst binds ((d, dx_id, dx) : pairs)
1132 | exprIsTrivial dx = go (extendIdSubst subst d dx) binds pairs
1133 -- No auxiliary binding necessary
1134 -- Note that we bind the *original* dict in the substitution,
1135 -- overriding any d->dx_id binding put there by substBndrs
1137 | otherwise = go subst_w_unf (NonRec dx_id dx : binds) pairs
1139 dx_id1 = dx_id `setIdUnfolding` mkSimpleUnfolding dx
1140 subst_w_unf = extendIdSubst subst d (Var dx_id1)
1141 -- Important! We're going to substitute dx_id1 for d
1142 -- and we want it to look "interesting", else we won't gather *any*
1143 -- consequential calls. E.g.
1145 -- If we specialise f for a call (f (dfun dNumInt)), we'll get
1146 -- a consequent call (g d') with an auxiliary definition
1148 -- We want that consequent call to look interesting
1150 -- Again, note that we bind the *original* dict in the substitution,
1151 -- overriding any d->dx_id binding put there by substBndrs
1154 Note [From non-recursive to recursive]
1155 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1156 Even in the non-recursive case, if any dict-binds depend on 'fn' we might
1157 have built a recursive knot
1160 MkUD { ud_binds = d7 = MkD ..f..
1161 , ud_calls = ...(f T d7)... }
1165 Rec { fs x = <blah>[T/a, d7/d]
1170 Here the recursion is only through the RULE.
1173 Note [Specialisation of dictionary functions]
1174 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1175 Here is a nasty example that bit us badly: see Trac #3591
1178 instance Eq [a] => C [a]
1181 dfun :: Eq [a] -> C [a]
1182 dfun a d = MkD a d (meth d)
1184 d4 :: Eq [T] = <blah>
1185 d2 :: C [T] = dfun T d4
1186 d1 :: Eq [T] = $p1 d2
1187 d3 :: C [T] = dfun T d1
1189 None of these definitions is recursive. What happened was that we
1190 generated a specialisation:
1192 RULE forall d. dfun T d = dT :: C [T]
1193 dT = (MkD a d (meth d)) [T/a, d1/d]
1194 = MkD T d1 (meth d1)
1196 But now we use the RULE on the RHS of d2, to get
1198 d2 = dT = MkD d1 (meth d1)
1201 and now d1 is bottom! The problem is that when specialising 'dfun' we
1202 should first dump "below" the binding all floated dictionary bindings
1203 that mention 'dfun' itself. So d2 and d3 (and hence d1) must be
1204 placed below 'dfun', and thus unavailable to it when specialising
1205 'dfun'. That in turn means that the call (dfun T d1) must be
1206 discarded. On the other hand, the call (dfun T d4) is fine, assuming
1207 d4 doesn't mention dfun.
1211 class C a where { foo,bar :: [a] -> [a] }
1213 instance C Int where
1217 r_bar :: C a => [a] -> [a]
1218 r_bar xs = bar (xs ++ xs)
1222 r_bar a (c::C a) (xs::[a]) = bar a d (xs ++ xs)
1224 Rec { $fCInt :: C Int = MkC foo_help reverse
1225 foo_help (xs::[Int]) = r_bar Int $fCInt xs }
1227 The call (r_bar $fCInt) mentions $fCInt,
1228 which mentions foo_help,
1229 which mentions r_bar
1230 But we DO want to specialise r_bar at Int:
1232 Rec { $fCInt :: C Int = MkC foo_help reverse
1233 foo_help (xs::[Int]) = r_bar Int $fCInt xs
1235 r_bar a (c::C a) (xs::[a]) = bar a d (xs ++ xs)
1236 RULE r_bar Int _ = r_bar_Int
1238 r_bar_Int xs = bar Int $fCInt (xs ++ xs)
1241 Note that, because of its RULE, r_bar joins the recursive
1242 group. (In this case it'll unravel a short moment later.)
1245 Conclusion: we catch the nasty case using filter_dfuns in
1246 callsForMe. To be honest I'm not 100% certain that this is 100%
1247 right, but it works. Sigh.
1250 Note [Specialising a recursive group]
1251 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1253 let rec { f x = ...g x'...
1254 ; g y = ...f y'.... }
1256 Here we specialise 'f' at Char; but that is very likely to lead to
1257 a specialisation of 'g' at Char. We must do the latter, else the
1258 whole point of specialisation is lost.
1260 But we do not want to keep iterating to a fixpoint, because in the
1261 presence of polymorphic recursion we might generate an infinite number
1264 So we use the following heuristic:
1265 * Arrange the rec block in dependency order, so far as possible
1266 (the occurrence analyser already does this)
1268 * Specialise it much like a sequence of lets
1270 * Then go through the block a second time, feeding call-info from
1271 the RHSs back in the bottom, as it were
1273 In effect, the ordering maxmimises the effectiveness of each sweep,
1274 and we do just two sweeps. This should catch almost every case of
1275 monomorphic recursion -- the exception could be a very knotted-up
1276 recursion with multiple cycles tied up together.
1278 This plan is implemented in the Rec case of specBindItself.
1280 Note [Specialisations already covered]
1281 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1282 We obviously don't want to generate two specialisations for the same
1283 argument pattern. There are two wrinkles
1285 1. We do the already-covered test in specDefn, not when we generate
1286 the CallInfo in mkCallUDs. We used to test in the latter place, but
1287 we now iterate the specialiser somewhat, and the Id at the call site
1288 might therefore not have all the RULES that we can see in specDefn
1290 2. What about two specialisations where the second is an *instance*
1291 of the first? If the more specific one shows up first, we'll generate
1292 specialisations for both. If the *less* specific one shows up first,
1293 we *don't* currently generate a specialisation for the more specific
1294 one. (See the call to lookupRule in already_covered.) Reasons:
1295 (a) lookupRule doesn't say which matches are exact (bad reason)
1296 (b) if the earlier specialisation is user-provided, it's
1297 far from clear that we should auto-specialise further
1299 Note [Auto-specialisation and RULES]
1300 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1302 g :: Num a => a -> a
1305 f :: (Int -> Int) -> Int
1307 {-# RULE f g = 0 #-}
1309 Suppose that auto-specialisation makes a specialised version of
1310 g::Int->Int That version won't appear in the LHS of the RULE for f.
1311 So if the specialisation rule fires too early, the rule for f may
1314 It might be possible to add new rules, to "complete" the rewrite system.
1316 RULE forall d. g Int d = g_spec
1320 But that's a bit complicated. For now we ask the programmer's help,
1321 by *copying the INLINE activation pragma* to the auto-specialised
1322 rule. So if g says {-# NOINLINE[2] g #-}, then the auto-spec rule
1323 will also not be active until phase 2. And that's what programmers
1324 should jolly well do anyway, even aside from specialisation, to ensure
1325 that g doesn't inline too early.
1327 This in turn means that the RULE would never fire for a NOINLINE
1328 thing so not much point in generating a specialisation at all.
1330 Note [Specialisation shape]
1331 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
1332 We only specialise a function if it has visible top-level lambdas
1333 corresponding to its overloading. E.g. if
1334 f :: forall a. Eq a => ....
1335 then its body must look like
1338 Reason: when specialising the body for a call (f ty dexp), we want to
1339 substitute dexp for d, and pick up specialised calls in the body of f.
1341 This doesn't always work. One example I came across was this:
1342 newtype Gen a = MkGen{ unGen :: Int -> a }
1344 choose :: Eq a => a -> Gen a
1345 choose n = MkGen (\r -> n)
1347 oneof = choose (1::Int)
1349 It's a silly exapmle, but we get
1350 choose = /\a. g `cast` co
1351 where choose doesn't have any dict arguments. Thus far I have not
1352 tried to fix this (wait till there's a real example).
1354 Note [Inline specialisations]
1355 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1356 We transfer to the specialised function any INLINE stuff from the
1357 original. This means
1358 (a) the Activation for its inlining (from its InlinePragma)
1361 This is a change (Jun06). Previously the idea is that the point of
1362 inlining was precisely to specialise the function at its call site,
1363 and that's not so important for the specialised copies. But
1364 *pragma-directed* specialisation now takes place in the
1365 typechecker/desugarer, with manually specified INLINEs. The
1366 specialiation here is automatic. It'd be very odd if a function
1367 marked INLINE was specialised (because of some local use), and then
1368 forever after (including importing modules) the specialised version
1369 wasn't INLINEd. After all, the programmer said INLINE!
1371 You might wonder why we don't just not specialise INLINE functions.
1372 It's because even INLINE functions are sometimes not inlined, when
1373 they aren't applied to interesting arguments. But perhaps the type
1374 arguments alone are enough to specialise (even though the args are too
1375 boring to trigger inlining), and it's certainly better to call the
1376 specialised version.
1379 %************************************************************************
1381 \subsubsection{UsageDetails and suchlike}
1383 %************************************************************************
1388 ud_binds :: !(Bag DictBind),
1389 -- Floated dictionary bindings
1390 -- The order is important;
1391 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
1392 -- (Remember, Bags preserve order in GHC.)
1394 ud_calls :: !CallDetails
1396 -- INVARIANT: suppose bs = bindersOf ud_binds
1397 -- Then 'calls' may *mention* 'bs',
1398 -- but there should be no calls *for* bs
1401 instance Outputable UsageDetails where
1402 ppr (MkUD { ud_binds = dbs, ud_calls = calls })
1403 = ptext (sLit "MkUD") <+> braces (sep (punctuate comma
1404 [ptext (sLit "binds") <+> equals <+> ppr dbs,
1405 ptext (sLit "calls") <+> equals <+> ppr calls]))
1407 type DictBind = (CoreBind, VarSet)
1408 -- The set is the free vars of the binding
1409 -- both tyvars and dicts
1411 type DictExpr = CoreExpr
1413 emptyUDs :: UsageDetails
1414 emptyUDs = MkUD { ud_binds = emptyBag, ud_calls = emptyVarEnv }
1416 ------------------------------------------------------------
1417 type CallDetails = IdEnv CallInfoSet
1418 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
1420 -- CallInfo uses a Map, thereby ensuring that
1421 -- we record only one call instance for any key
1423 -- The list of types and dictionaries is guaranteed to
1424 -- match the type of f
1425 data CallInfoSet = CIS Id (Map CallKey ([DictExpr], VarSet))
1426 -- Range is dict args and the vars of the whole
1427 -- call (including tyvars)
1428 -- [*not* include the main id itself, of course]
1430 type CallInfo = (CallKey, ([DictExpr], VarSet))
1432 instance Outputable CallInfoSet where
1433 ppr (CIS fn map) = hang (ptext (sLit "CIS") <+> ppr fn)
1436 instance Outputable CallKey where
1437 ppr (CallKey ts) = ppr ts
1439 -- Type isn't an instance of Ord, so that we can control which
1440 -- instance we use. That's tiresome here. Oh well
1441 instance Eq CallKey where
1442 k1 == k2 = case k1 `compare` k2 of { EQ -> True; _ -> False }
1444 instance Ord CallKey where
1445 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
1447 cmp Nothing Nothing = EQ
1448 cmp Nothing (Just _) = LT
1449 cmp (Just _) Nothing = GT
1450 cmp (Just t1) (Just t2) = tcCmpType t1 t2
1452 unionCalls :: CallDetails -> CallDetails -> CallDetails
1453 unionCalls c1 c2 = plusVarEnv_C unionCallInfoSet c1 c2
1455 unionCallInfoSet :: CallInfoSet -> CallInfoSet -> CallInfoSet
1456 unionCallInfoSet (CIS f calls1) (CIS _ calls2) = CIS f (calls1 `Map.union` calls2)
1458 callDetailsFVs :: CallDetails -> VarSet
1459 callDetailsFVs calls = foldVarEnv (unionVarSet . callInfoFVs) emptyVarSet calls
1461 callInfoFVs :: CallInfoSet -> VarSet
1462 callInfoFVs (CIS _ call_info) = Map.foldRight (\(_,fv) vs -> unionVarSet fv vs) emptyVarSet call_info
1464 ------------------------------------------------------------
1465 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> UsageDetails
1466 singleCall id tys dicts
1467 = MkUD {ud_binds = emptyBag,
1468 ud_calls = unitVarEnv id $ CIS id $
1469 Map.singleton (CallKey tys) (dicts, call_fvs) }
1471 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
1472 tys_fvs = tyVarsOfTypes (catMaybes tys)
1473 -- The type args (tys) are guaranteed to be part of the dictionary
1474 -- types, because they are just the constrained types,
1475 -- and the dictionary is therefore sure to be bound
1476 -- inside the binding for any type variables free in the type;
1477 -- hence it's safe to neglect tyvars free in tys when making
1478 -- the free-var set for this call
1479 -- BUT I don't trust this reasoning; play safe and include tys_fvs
1481 -- We don't include the 'id' itself.
1483 mkCallUDs :: Id -> [CoreExpr] -> UsageDetails
1485 | not (want_calls_for f) -- Imported from elsewhere
1486 || null theta -- Not overloaded
1487 || not (all isClassPred theta)
1488 -- Only specialise if all overloading is on class params.
1489 -- In ptic, with implicit params, the type args
1490 -- *don't* say what the value of the implicit param is!
1491 || not (spec_tys `lengthIs` n_tyvars)
1492 || not ( dicts `lengthIs` n_dicts)
1493 || not (any interestingDict dicts) -- Note [Interesting dictionary arguments]
1494 -- See also Note [Specialisations already covered]
1495 = -- pprTrace "mkCallUDs: discarding" (vcat [ppr f, ppr args, ppr n_tyvars, ppr n_dicts, ppr (map interestingDict dicts)])
1496 emptyUDs -- Not overloaded, or no specialisation wanted
1499 = -- pprTrace "mkCallUDs: keeping" (vcat [ppr f, ppr args, ppr n_tyvars, ppr n_dicts, ppr (map interestingDict dicts)])
1500 singleCall f spec_tys dicts
1502 (tyvars, theta, _) = tcSplitSigmaTy (idType f)
1503 constrained_tyvars = tyVarsOfTheta theta
1504 n_tyvars = length tyvars
1505 n_dicts = length theta
1507 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
1508 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
1511 | tyvar `elemVarSet` constrained_tyvars = Just ty
1512 | otherwise = Nothing
1514 want_calls_for f = isLocalId f || isInlinablePragma (idInlinePragma f)
1517 Note [Interesting dictionary arguments]
1518 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1520 \a.\d:Eq a. let f = ... in ...(f d)...
1521 There really is not much point in specialising f wrt the dictionary d,
1522 because the code for the specialised f is not improved at all, because
1523 d is lambda-bound. We simply get junk specialisations.
1525 What is "interesting"? Just that it has *some* structure.
1528 interestingDict :: CoreExpr -> Bool
1529 -- A dictionary argument is interesting if it has *some* structure
1530 interestingDict (Var v) = hasSomeUnfolding (idUnfolding v)
1531 || isDataConWorkId v
1532 interestingDict (Type _) = False
1533 interestingDict (App fn (Type _)) = interestingDict fn
1534 interestingDict (Note _ a) = interestingDict a
1535 interestingDict (Cast e _) = interestingDict e
1536 interestingDict _ = True
1540 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
1541 plusUDs (MkUD {ud_binds = db1, ud_calls = calls1})
1542 (MkUD {ud_binds = db2, ud_calls = calls2})
1543 = MkUD { ud_binds = db1 `unionBags` db2
1544 , ud_calls = calls1 `unionCalls` calls2 }
1546 plusUDList :: [UsageDetails] -> UsageDetails
1547 plusUDList = foldr plusUDs emptyUDs
1549 -----------------------------
1550 _dictBindBndrs :: Bag DictBind -> [Id]
1551 _dictBindBndrs dbs = foldrBag ((++) . bindersOf . fst) [] dbs
1553 mkDB :: CoreBind -> DictBind
1554 mkDB bind = (bind, bind_fvs bind)
1556 bind_fvs :: CoreBind -> VarSet
1557 bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
1558 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1561 rhs_fvs = unionVarSets (map pair_fvs prs)
1563 pair_fvs :: (Id, CoreExpr) -> VarSet
1564 pair_fvs (bndr, rhs) = exprFreeVars rhs `unionVarSet` idFreeVars bndr
1565 -- Don't forget variables mentioned in the
1566 -- rules of the bndr. C.f. OccAnal.addRuleUsage
1567 -- Also tyvars mentioned in its type; they may not appear in the RHS
1571 flattenDictBinds :: Bag DictBind -> [(Id,CoreExpr)] -> [(Id,CoreExpr)]
1572 flattenDictBinds dbs pairs
1573 = foldrBag add pairs dbs
1575 add (NonRec b r,_) pairs = (b,r) : pairs
1576 add (Rec prs1, _) pairs = prs1 ++ pairs
1578 snocDictBinds :: UsageDetails -> [CoreBind] -> UsageDetails
1579 -- Add ud_binds to the tail end of the bindings in uds
1580 snocDictBinds uds dbs
1581 = uds { ud_binds = ud_binds uds `unionBags`
1582 foldr (consBag . mkDB) emptyBag dbs }
1584 consDictBind :: CoreBind -> UsageDetails -> UsageDetails
1585 consDictBind bind uds = uds { ud_binds = mkDB bind `consBag` ud_binds uds }
1587 addDictBinds :: [DictBind] -> UsageDetails -> UsageDetails
1588 addDictBinds binds uds = uds { ud_binds = listToBag binds `unionBags` ud_binds uds }
1590 snocDictBind :: UsageDetails -> CoreBind -> UsageDetails
1591 snocDictBind uds bind = uds { ud_binds = ud_binds uds `snocBag` mkDB bind }
1593 wrapDictBinds :: Bag DictBind -> [CoreBind] -> [CoreBind]
1594 wrapDictBinds dbs binds
1595 = foldrBag add binds dbs
1597 add (bind,_) binds = bind : binds
1599 wrapDictBindsE :: Bag DictBind -> CoreExpr -> CoreExpr
1600 wrapDictBindsE dbs expr
1601 = foldrBag add expr dbs
1603 add (bind,_) expr = Let bind expr
1605 ----------------------
1606 dumpUDs :: [CoreBndr] -> UsageDetails -> (UsageDetails, Bag DictBind)
1607 -- Used at a lambda or case binder; just dump anything mentioning the binder
1608 dumpUDs bndrs uds@(MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
1609 | null bndrs = (uds, emptyBag) -- Common in case alternatives
1610 | otherwise = -- pprTrace "dumpUDs" (ppr bndrs $$ ppr free_uds $$ ppr dump_dbs) $
1611 (free_uds, dump_dbs)
1613 free_uds = MkUD { ud_binds = free_dbs, ud_calls = free_calls }
1614 bndr_set = mkVarSet bndrs
1615 (free_dbs, dump_dbs, dump_set) = splitDictBinds orig_dbs bndr_set
1616 free_calls = deleteCallsMentioning dump_set $ -- Drop calls mentioning bndr_set on the floor
1617 deleteCallsFor bndrs orig_calls -- Discard calls for bndr_set; there should be
1618 -- no calls for any of the dicts in dump_dbs
1620 dumpBindUDs :: [CoreBndr] -> UsageDetails -> (UsageDetails, Bag DictBind, Bool)
1621 -- Used at a lambda or case binder; just dump anything mentioning the binder
1622 dumpBindUDs bndrs (MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
1623 = -- pprTrace "dumpBindUDs" (ppr bndrs $$ ppr free_uds $$ ppr dump_dbs) $
1624 (free_uds, dump_dbs, float_all)
1626 free_uds = MkUD { ud_binds = free_dbs, ud_calls = free_calls }
1627 bndr_set = mkVarSet bndrs
1628 (free_dbs, dump_dbs, dump_set) = splitDictBinds orig_dbs bndr_set
1629 free_calls = deleteCallsFor bndrs orig_calls
1630 float_all = dump_set `intersectsVarSet` callDetailsFVs free_calls
1632 callsForMe :: Id -> UsageDetails -> (UsageDetails, [CallInfo])
1633 callsForMe fn (MkUD { ud_binds = orig_dbs, ud_calls = orig_calls })
1634 = -- pprTrace ("callsForMe")
1636 -- text "Orig dbs =" <+> ppr (_dictBindBndrs orig_dbs),
1637 -- text "Orig calls =" <+> ppr orig_calls,
1638 -- text "Dep set =" <+> ppr dep_set,
1639 -- text "Calls for me =" <+> ppr calls_for_me]) $
1640 (uds_without_me, calls_for_me)
1642 uds_without_me = MkUD { ud_binds = orig_dbs, ud_calls = delVarEnv orig_calls fn }
1643 calls_for_me = case lookupVarEnv orig_calls fn of
1645 Just (CIS _ calls) -> filter_dfuns (Map.toList calls)
1647 dep_set = foldlBag go (unitVarSet fn) orig_dbs
1648 go dep_set (db,fvs) | fvs `intersectsVarSet` dep_set
1649 = extendVarSetList dep_set (bindersOf db)
1650 | otherwise = dep_set
1652 -- Note [Specialisation of dictionary functions]
1653 filter_dfuns | isDFunId fn = filter ok_call
1654 | otherwise = \cs -> cs
1656 ok_call (_, (_,fvs)) = not (fvs `intersectsVarSet` dep_set)
1658 ----------------------
1659 splitDictBinds :: Bag DictBind -> IdSet -> (Bag DictBind, Bag DictBind, IdSet)
1660 -- Returns (free_dbs, dump_dbs, dump_set)
1661 splitDictBinds dbs bndr_set
1662 = foldlBag split_db (emptyBag, emptyBag, bndr_set) dbs
1663 -- Important that it's foldl not foldr;
1664 -- we're accumulating the set of dumped ids in dump_set
1666 split_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1667 | dump_idset `intersectsVarSet` fvs -- Dump it
1668 = (free_dbs, dump_dbs `snocBag` db,
1669 extendVarSetList dump_idset (bindersOf bind))
1671 | otherwise -- Don't dump it
1672 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1675 ----------------------
1676 deleteCallsMentioning :: VarSet -> CallDetails -> CallDetails
1677 -- Remove calls *mentioning* bs
1678 deleteCallsMentioning bs calls
1679 = mapVarEnv filter_calls calls
1681 filter_calls :: CallInfoSet -> CallInfoSet
1682 filter_calls (CIS f calls) = CIS f (Map.filter keep_call calls)
1683 keep_call (_, fvs) = not (fvs `intersectsVarSet` bs)
1685 deleteCallsFor :: [Id] -> CallDetails -> CallDetails
1686 -- Remove calls *for* bs
1687 deleteCallsFor bs calls = delVarEnvList calls bs
1691 %************************************************************************
1693 \subsubsection{Boring helper functions}
1695 %************************************************************************
1698 type SpecM a = UniqSM a
1700 runSpecM:: SpecM a -> CoreM a
1701 runSpecM spec = do { us <- getUniqueSupplyM
1702 ; return (initUs_ us spec) }
1704 mapAndCombineSM :: (a -> SpecM (b, UsageDetails)) -> [a] -> SpecM ([b], UsageDetails)
1705 mapAndCombineSM _ [] = return ([], emptyUDs)
1706 mapAndCombineSM f (x:xs) = do (y, uds1) <- f x
1707 (ys, uds2) <- mapAndCombineSM f xs
1708 return (y:ys, uds1 `plusUDs` uds2)
1710 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1711 -- Clone the binders of the bind; return new bind with the cloned binders
1712 -- Return the substitution to use for RHSs, and the one to use for the body
1713 cloneBindSM subst (NonRec bndr rhs) = do
1714 us <- getUniqueSupplyM
1715 let (subst', bndr') = cloneIdBndr subst us bndr
1716 return (subst, subst', NonRec bndr' rhs)
1718 cloneBindSM subst (Rec pairs) = do
1719 us <- getUniqueSupplyM
1720 let (subst', bndrs') = cloneRecIdBndrs subst us (map fst pairs)
1721 return (subst', subst', Rec (bndrs' `zip` map snd pairs))
1723 newDictBndrs :: Subst -> [CoreBndr] -> SpecM (Subst, [CoreBndr])
1724 -- Make up completely fresh binders for the dictionaries
1725 -- Their bindings are going to float outwards
1726 newDictBndrs subst bndrs
1727 = do { bndrs' <- mapM new bndrs
1728 ; let subst' = extendIdSubstList subst
1729 [(d, Var d') | (d,d') <- bndrs `zip` bndrs']
1730 ; return (subst', bndrs' ) }
1732 new b = do { uniq <- getUniqueM
1734 ty' = CoreSubst.substTy subst (idType b)
1735 ; return (mkUserLocal (nameOccName n) uniq ty' (getSrcSpan n)) }
1737 newSpecIdSM :: Id -> Type -> SpecM Id
1738 -- Give the new Id a similar occurrence name to the old one
1739 newSpecIdSM old_id new_ty
1740 = do { uniq <- getUniqueM
1741 ; let name = idName old_id
1742 new_occ = mkSpecOcc (nameOccName name)
1743 new_id = mkUserLocal new_occ uniq new_ty (getSrcSpan name)
1748 Old (but interesting) stuff about unboxed bindings
1749 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1751 What should we do when a value is specialised to a *strict* unboxed value?
1753 map_*_* f (x:xs) = let h = f x
1757 Could convert let to case:
1759 map_*_Int# f (x:xs) = case f x of h# ->
1763 This may be undesirable since it forces evaluation here, but the value
1764 may not be used in all branches of the body. In the general case this
1765 transformation is impossible since the mutual recursion in a letrec
1766 cannot be expressed as a case.
1768 There is also a problem with top-level unboxed values, since our
1769 implementation cannot handle unboxed values at the top level.
1771 Solution: Lift the binding of the unboxed value and extract it when it
1774 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1779 Now give it to the simplifier and the _Lifting will be optimised away.
1781 The benfit is that we have given the specialised "unboxed" values a
1782 very simplep lifted semantics and then leave it up to the simplifier to
1783 optimise it --- knowing that the overheads will be removed in nearly
1786 In particular, the value will only be evaluted in the branches of the
1787 program which use it, rather than being forced at the point where the
1788 value is bound. For example:
1790 filtermap_*_* p f (x:xs)
1797 filtermap_*_Int# p f (x:xs)
1798 = let h = case (f x) of h# -> _Lift h#
1801 True -> case h of _Lift h#
1805 The binding for h can still be inlined in the one branch and the
1806 _Lifting eliminated.
1809 Question: When won't the _Lifting be eliminated?
1811 Answer: When they at the top-level (where it is necessary) or when
1812 inlining would duplicate work (or possibly code depending on
1813 options). However, the _Lifting will still be eliminated if the
1814 strictness analyser deems the lifted binding strict.